research paper on cancer detection

Machine Learning Approaches in Cancer Detection and Diagnosis: Mini Review

• November 2017

• Deenbandhu Chhotu Ram University of Science and Technology, Murthal

• UNSW Sydney

Abstract and Figures

Discover the world's research

• 25+ million members
• 160+ million publication pages
• 2.3+ billion citations

• Gunji Srilikhita

• Mohsen Guizani

• Pengwei Yang

• Wen-Jia Sun
• Han-Fei Guo
• Sachin Urabinahatti
• Devappa Jayadevappa

• Vijay Gaikwad
• Samvedh Shetty
• Shrinivas Saraf

• Henri J. Boersma
• Shyam Vasudeva Rao

• Jitendra Singh
• Solveig Badillo

• Fabian Birzele

• Radhakrishnan Nagarajan

• Hamed Habibi Aghdam
• Elnaz Jahani Heravi
• Younus Ahmad Malla
• EXPERT SYST APPL

• Mehrbakhsh Nilashi

• Recruit researchers
• Join for free
• Login Email Tip: Most researchers use their institutional email address as their ResearchGate login Password Forgot password? Keep me logged in Log in or Continue with Google Welcome back! Please log in. Email · Hint Tip: Most researchers use their institutional email address as their ResearchGate login Password Forgot password? Keep me logged in Log in or Continue with Google No account? Sign up

• Open access
• Published: 27 September 2021

Deep learning in cancer diagnosis, prognosis and treatment selection

• Khoa A. Tran 1 , 2 ,
• Olga Kondrashova 1 ,
• Andrew Bradley 4 ,
• Elizabeth D. Williams 2 , 3 ,
• John V. Pearson 1 &
• Nicola Waddell   ORCID: orcid.org/0000-0002-3950-2476 1  

Genome Medicine volume  13 , Article number:  152 ( 2021 ) Cite this article

49k Accesses

298 Citations

38 Altmetric

Metrics details

Deep learning is a subdiscipline of artificial intelligence that uses a machine learning technique called artificial neural networks to extract patterns and make predictions from large data sets. The increasing adoption of deep learning across healthcare domains together with the availability of highly characterised cancer datasets has accelerated research into the utility of deep learning in the analysis of the complex biology of cancer. While early results are promising, this is a rapidly evolving field with new knowledge emerging in both cancer biology and deep learning. In this review, we provide an overview of emerging deep learning techniques and how they are being applied to oncology. We focus on the deep learning applications for omics data types, including genomic, methylation and transcriptomic data, as well as histopathology-based genomic inference, and provide perspectives on how the different data types can be integrated to develop decision support tools. We provide specific examples of how deep learning may be applied in cancer diagnosis, prognosis and treatment management. We also assess the current limitations and challenges for the application of deep learning in precision oncology, including the lack of phenotypically rich data and the need for more explainable deep learning models. Finally, we conclude with a discussion of how current obstacles can be overcome to enable future clinical utilisation of deep learning.

Artificial intelligence (AI) encompasses multiple technologies with the common aim to computationally simulate human intelligence. Machine learning (ML) is a subgroup of AI that focuses on making predictions by identifying patterns in data using mathematical algorithms. Deep learning (DL) is a subgroup of ML that focuses on making predictions using multi-layered neural network algorithms inspired by the neurological architecture of the brain. Compared to other ML methods such as logistic regression, the neural network architecture of DL enables the models to scale exponentially with the growing quantity and dimensionality of data [ 1 ]. This makes DL particularly useful for solving complex computational problems such as large-scale image classification, natural language processing and speech recognition and translation [ 1 ].

Cancer care is undergoing a shift towards precision healthcare enabled by the increasing availability and integration of multiple data types including genomic, transcriptomic and histopathologic data (Fig. 1 ). The use and interpretation of diverse and high-dimensionality data types for translational research or clinical tasks require significant time and expertise. Moreover, the integration of multiple data types is more resource-intensive than the interpretation of individual data types and needs modelling algorithms that can learn from tremendous numbers of intricate features. The use of ML algorithms to automate these tasks and aid cancer detection (identifying the presence of cancer) and diagnosis (characterising the cancer) has become increasingly prevalent [ 2 , 3 ]. Excitingly, DL models have the potential to harness this complexity to provide meaningful insights and identify relevant granular features from multiple data types [ 4 , 5 ]. In this review, we describe the latest applications of deep learning in cancer diagnosis, prognosis and treatment selection. We focus on DL applications for omics and histopathological data, as well as the integration of multiple data types. We provide a brief introduction to emerging DL methods relevant to applications covered in this review. Next, we discuss specific applications of DL in oncology, including cancer origin detection, molecular subtypes identification, prognosis and survivability prediction, histological inference of genomic traits, tumour microenvironment profiling and future applications in spatial transcriptomics, metagenomics and pharmacogenomics. We conclude with an examination of current challenges and potential strategies that would enable DL to be routinely applied in clinical settings.

Deep learning may impact clinical oncology during diagnosis, prognosis and treatment. Specific areas of clinical oncology where deep learning is showing promise include cancer of unknown primary, molecular subtyping of cancers, prognosis and survivability and precision oncology. Examples of deep learning applications within each of these areas are listed. The data modalities utilised by deep learning models are numerous and include genomic, transcriptomic and histopathology data categories covered in this review

Emerging deep learning methods

Covering all DL methods in detail is outside the scope of this review; rather, we provide a high-level summary of emerging DL methods in oncology. DL utilises artificial neural networks to extract non-linear, entangled and representative features from massive and high-dimensional data [ 1 ]. A deep neural network is typically constructed of millions of densely interconnected computing neurons organised into consecutive layers. Within each layer, a neuron is connected to other neurons in the layer before it, from which it receives data, and other neurons in the layer after it, to which it sends data. When presented with data, a neural network feeds each training sample, with known ground truth, to its input layer before passing the information down to all succeeding layers (usually called hidden layers). This information is then multiplied, divided, added and subtracted millions of times before it reaches the output layer, which becomes the prediction. For supervised deep learning, each pair of training sample and label is fed through a neural network while its weights and thresholds are being adjusted to get the prediction closer to the provided label. When faced with unseen (test) data, these trained weights and thresholds are frozen and used to make predictions.

Fundamental neural network methods

There are multiple neural network-based methods, all with different advantages and applications. Multilayer perceptron (MLP), recurrent neural network (RNN) and convolutional neural network (CNN) are the most fundamental and are frequently used as building blocks for more advanced techniques. MLPs are the simplest type of neural networks, where neurons are organised in consecutive layers so that signals travel through the network in one direction (from input to output) [ 1 ]. Although MLPs can perform well for generic predictions, they are also prone to overfitting [ 6 ]. RNNs process an input sequence one element at a time, while maintaining history of all past elements in hidden ‘state vector(s)’. Output predictions are made at every element using information from the current element and also previous elements [ 1 , 7 ]. RNNs are typically used for analysing sequential data such as text, speech or DNA sequences. By contrast, CNNs are designed to draw spatial relationships from image data. CNNs traverse an image and apply small feature-filter matrices, i.e. convolution filters, to extract granular features [ 1 ]. Features extracted by the last convolution layer are then used for making predictions. CNNs have also been adapted for analysis of non-image data, e.g. genomic data represented in a vector, matrix or tensor format [ 8 ]. A review by Dias and Torkamani [ 7 ] described in detail how MLPs, RNNs and CNNs operate on biomedical and genomics data. Moreover, the use of MLPs, RNNs and CNNs to assist clinicians and researchers has been proposed across multiple oncology areas, including radiotherapy [ 9 ], digital histopathology [ 10 , 11 ] and clinical and genomic diagnostics [ 7 ]. While routine clinical use is still limited, some of the models have already been FDA-approved and adopted into a clinical setting, for example CNNs for the prediction of malignancy in pulmonary nodules detected by CT [ 12 ], and prostate and breast cancer diagnosis prediction using digital histopathology [ 13 , 14 ].

Advanced neural-network methods

Graph convolutional neural networks (GCNNs) generalise CNNs beyond regular structures (Euclidean domains) to non-Euclidean domains such as graphs which have arbitrary structure. GCNNs are specifically designed to analyse graph data, e.g. using prior biological knowledge of an interconnected network of proteins with nodes representing proteins and pairwise connections representing protein–protein interactions (PPI) [ 15 ], using resources such as the STRING PPI database [ 16 ] (Fig. 2 a). This enables GCNNs to incorporate known biological associations between genetic features and perceive their cooperative patterns, which have been shown to be useful in cancer diagnostics [ 17 ].

An overview of Deep Learning techniques and concepts in oncology. a Graph convolutional neural networks (GCNN) are designed to operate on graph-structured data. In this particular example inspired by [ 17 , 18 , 19 ], gene expression values (upper left panel) are represented as graph signals structured by a protein–protein interactions graph (lower left panel) that serve as inputs to GCNN. For a single sample (highlighted with red outline), each node represents one gene with its expression value assigned to the corresponding protein node, and inter-node connections represent known protein–protein interactions. GCNN methods covered in this review require a graph to be undirected. Graph convolution filters are applied on each gene to extract meaningful gene expression patterns from the gene’s neighbourhood (nodes connected by orange edges). Pooling, i.e. combining clusters of nodes, can be applied following graph convolution to obtain a coarser representation of the graph. Output of the final graph convolution/pooling layer would then be passed through fully connected layers producing GCNN’s decision. b Semantic segmentation is applied to image data where it assigns a class label to each pixel within an image. A semantic segmentation model usually consists of an encoder, a decoder and a softmax function. The encoder consists of feature extraction layers to ‘learn’ meaningful and granular features from the input, while the decoder learns features to generate a coloured map of major object classes in the input (through the use of the softmax function). The example shows a H&E tumour section with infiltrating lymphocyte map generated by Saltz et al. [ 20 ] DL model c multimodal learning allows multiple datasets representing the same underlying phenotype to be combined to increase predictive power. Multimodal learning usually starts with encoding each input modality into a representation vector of lower dimension, followed by a feature combination step to aggregate these vectors together. d Explainability methods take a trained neural network and mathematically quantify how each input feature influences the model’s prediction. The outputs are usually feature contribution scores, capable of explaining the most salient features that dictate the model’s predictions. In this example, each input gene is assigned a contribution score by the explainability model (colour scale indicates the influence on the model prediction). An example of gene interaction network is shown coloured by contribution scores (links between red dots represent biological connections between genes)

Semantic segmentation is an important CNN-based visual learning method specifically for image data (Fig. 2 b). The purpose of semantic segmentation is to produce a class label for every single pixel in an image and cluster parts of an image together into each class, where the class represents an object or component of the image. Semantic segmentation models are generally supervised, i.e. they are given class labels for each pixel and are trained to detect the major ‘semantics’ for each class.

To enhance the predictive power of DL models, different data types (modalities) can be combined using multimodal learning (Fig. 2 c). In clinical oncology, data modalities can include image, numeric and descriptive data. Cancer is a complex and multi-faceted disease with layers of microscopic, macroscopic and molecular features that can separately or together influence treatment responses and patient prognosis. Therefore, combining clinical data (e.g. diagnostic test results and pathology reports), medical images (e.g. histopathology and computed tomography) and different types of omics data, such as genomic, transcriptomic and proteomic profiles, may be useful. The two most important requirements for a multimodal network are the ability to create representations that contain dense meaningful features of the original input, and a mathematical method to combine representations from all modalities. There are several methods capable of performing the representative learning task, e.g. CNNs, RNNs, deep belief networks and autoencoders (AE) [ 21 ]; score-level fusion [ 22 ]; or multimodal data fusion [ 23 ]. The multimodal learning applications discussed in this review are based on AE models. In simplistic terms, AE architecture comprises of an encoder and a decoder working in tandem. The encoder is responsible for creating a representation vector of lower dimension than the input, while the decoder is responsible for reconstructing the original input using this low-dimensional vector [ 24 ]. This forces the encoder to ‘learn’ to encapsulate meaningful features from the input and has been shown to have good generalisability [ 24 ]. Moreover, it provides DL models the unique ability to readily integrate different data modalities, e.g. medical images, genomic data and clinical information, into a single ‘end-to-end optimised’ model [ 8 ].

A major challenge with implementing DL into clinical practice is the ‘black box’ nature of the models [ 25 ]. High-stake medical decisions, such as diagnosis, prognosis and treatment selection, require trustworthy and explainable decision processes. Most DL models have limited interpretability, i.e. it is very difficult to dissect a neural network and understand how millions of parameters work simultaneously. Some even argue that more interpretable models such as Decision Trees should be ultimately preferred for making medical decisions [ 26 ]. An alternative approach is explainability—mathematical quantification of how influential, or ‘salient’, the features are towards a certain prediction (Fig. 2 d). This information can be used to ‘explain’ the decision-making process of a neural network model and identify features that contribute to a prediction. This knowledge can enable resolution of potential disagreements between DL models and clinicians and thus increase trust in DL systems [ 27 ]. Moreover, DL models do not always have perfect performance due to either imperfect training data (e.g. assay noise or errors in recording) or systematic errors caused by bias within DL models themselves, which can result from the training data not being representative of the population where DL is later applied [ 27 ]. In these circumstances, explainability can assist clinicians in evaluating predictions [ 27 ]. While some explainability methods were developed specifically for neural networks [ 28 , 29 ], others offer a more model- and data-agnostic solution [ 30 , 31 , 32 , 33 ]. Excitingly, explainability methods can be used in conjunction with multi-modal learning for data integration and discovery of cross-modality insights, e.g. how cancer traits across different omics types correlate and influence each other.

Another challenge in applying DL in oncology is the requirement for large amounts of robust, well-phenotyped training data to achieve good model generalisability. Large curated ‘ground-truth’ datasets of matched genomic, histopathological and clinical outcome data are scarce beyond the publicly available datasets, such as The Cancer Genome Atlas (TCGA) [ 34 ], International Cancer Genome Consortium (ICGC) [ 35 ], Gene Expression Omnibus (GEO) [ 36 ], European Genome-Phenome Archive (EGA) [ 37 ] and Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) [ 38 ]. Pre-training on abundant datasets from other domains may help overcome the challenges of limited data (a process known as transfer learning). The pre-trained neural network would then be reconfigured and trained again on data from the domain of interest. This approach usually results in a considerable reduction in computational and time resources for models training, and a significant increase in predictive performance, compared to training on small domain-specific datasets [ 39 ].

Deep learning in oncology

A variety of DL approaches that utilise a combination of genomic, transcriptomic or histopathology data have been applied in clinical and translational oncology with the aim of enhancing patient diagnosis, prognosis and treatment selection (Fig. 1 , Table 1 ). However, even with the emerging DL approaches, human intervention remains essential in oncology. Therefore, the goal of DL is not to outperform or replace humans, but to provide decision support tools that assist cancer researchers to study the disease and health professionals in the clinical management of people with cancer [ 79 ].

Deep learning for microscopy-based assessment of cancer

Cancers are traditionally diagnosed by histopathology or cytopathology to confirm the presence of tumour cells within a patient sample, assess markers relevant to cancer and to characterise features such as tumour type, stage and grade. This microscopy-based assessment is crucial; however, the process is relatively labour-intensive and somewhat subjective [ 80 , 81 ]. A histology image viewed at high magnification (typically 20x or 40x) can reveal millions of subtle cellular features, and deep CNN models are exceptionally good at extracting features from high-resolution image data [ 82 ]. Automating cancer grading with histology-based deep CNNs has proven successful, with studies showing that performance of deep CNNs can be comparable with pathologists in grading prostate [ 40 , 41 , 42 ], breast [ 43 ], colon cancer [ 44 ] and lymphoma [ 45 ]. Explainability methods can enable and improve histology-based classification models by allowing pathologists to validate DL-generated predictions. For example, Hägele et al. applied the Layer-wise Relevance Propagation (LRP) [ 29 ] method on DL models classifying healthy versus cancerous tissues using whole-slide images of lung cancer [ 46 ]. The LRP algorithm assigned a relevance score for each pixel, and pixel-wise relevance scores were aggregated into cell-level scores and compared against pathologists’ annotations. These scores were then used to evaluate DL model performance and identify how multiple data biases affected the performance at cellular levels [ 46 ]. These insights allow clinician and software developers to gain insights into DL models during development and deployment phases.

In addition to classification and explainability, semantic segmentation approaches can also be applied on histopathology images to localise specific regions. One notable approach to perform semantic segmentation is to use generative adversarial networks (GANs) [ 47 ]. GAN is a versatile generative DL method comprising a pair of two neural networks: a generator and a discriminator [ 83 ]. In the context of semantic segmentation, the generator learns to label each pixel of an image to a class object (Fig. 2 b), while the discriminator learns to distinguish the predicted class labels from the ground truth [ 84 ]. This ‘adversarial’ mechanism forces the generator to be as accurate as possible in localising objects so that the discriminator cannot recognise the difference between predicted and ground-truth class labels [ 84 ]. Using this approach, Poojitha and Lal Sharma trained a CNN-based generator to segment cancer tissue to ‘help’ a CNN-based classifier predict prostate cancer grading [ 47 ]. The GAN-annotated tissue maps helped the CNN classifier achieve comparable accuracy to the grading produced by anatomical pathologists, indicating DL models can detect relevant cell regions in pathology images for decision making.

Molecular subtyping of cancers

Transcriptomic profiling can be used to assign cancers into clinically meaningful molecular subtypes that have diagnostic, prognostic or treatment selection relevance. Molecular subtypes were first described for breast cancer [ 85 , 86 ], then later for other cancers including colorectal [ 87 ], ovarian cancer [ 88 ] and sarcomas [ 89 ]. Standard computational methods, such as support vector machines (SVMs) or k-nearest neighbours, used to subtype cancers can be prone to errors due to batch effects [ 90 ] and may rely only on a handful of signature genes, omitting important biological information [ 91 , 92 , 93 ]. Deep learning algorithms can overcome these limitations by learning patterns from the whole transcriptome. A neural network model DeepCC trained on TCGA RNA-seq colon and breast cancer data, then tested on independent gene expression microarray data showed superior accuracy, sensitivity and specificity when compared to traditional ML approaches including random forest, logistic regression, SVM and gradient boosting machine [ 48 ]. Neural networks have also been successfully applied to transcriptomic data for molecular subtyping of lung [ 94 ], gastric and ovarian cancers [ 95 ]. DL methods have the potential to be highly generalisable in profiling cancer molecular subtypes due to their ability to train on a large number of features that are generated by transcriptomic profiling. Furthermore, due to their flexibility, DL methods can incorporate prior biological knowledge to achieve improved performance. For example, Rhee et al. trained a hybrid GCNN model on expression profiles of a cancer hallmark gene set, connected in a graph using the STRING PPI network [ 16 ] to predict breast cancer molecular subtypes, PAM50 [ 18 ]. This approach outperformed other ML methods in subtype classification. Furthermore, the granular features extracted by the GCNN model naturally clustered tumours into PAM50 subtypes without relying on a classification model demonstrating that the method successfully learned the latent properties in the gene expression profiles [ 18 ].

The use of multimodal learning to integrate transcriptomic with other omics data may enable enhanced subtype predictions. A novel multimodal method using two CNN models trained separately on copy number alterations (CNAs) and gene expression before concatenating their representations for predictions was able to predict PAM50 breast cancer subtypes better than CNNs trained on individual data types [ 54 ]. As multi-omics analysis becomes increasingly popular, multimodal learning methods are expected to become more prevalent in cancer diagnostics. However, the challenges of generating multi-omic data from patient samples in the clinical setting, as opposed to samples bio-banked for research, may hinder the clinical implementation of these approaches.

Digital histopathology images are an integral part of the oncology workflow [ 11 ] and can be an alternative to transcriptomic-based methods for molecular subtyping. CNN models have been applied on haematoxylin and eosin (H&E) sections to predict molecular subtypes of lung [ 49 ], colorectal [ 50 ], breast [ 51 , 52 ] and bladder cancer [ 53 ], with greater accuracy when compared to traditional ML methods.

Diagnosing cancers of unknown primary

Determining the primary cancer site can be important during the diagnostic process, as it can be a significant indicator of how the cancer will behave clinically, and the treatment strategies are sometimes decided by the tumour origin [ 96 , 97 ]. However, 3–5% of cancer cases are metastatic cancers of unknown origin, termed cancers of unknown primary (CUPs) [ 98 , 99 ]. Genomic, methylation and transcriptomic profiles of metastatic tumours have unique patterns that can reveal their tissues of origin [ 100 , 101 , 102 ].

Traditional ML methods, such as regression and SVMs, applied to these omics data can predict tumour origin; however, they usually rely on a small subset of genes, which can be limiting in predicting a broad range of cancer types and subtypes. In contrast, DL algorithms can utilise large number of genomic and transcriptomic features. The Pan-Cancer Analysis of Whole Genomes (PCAWG) Consortium [ 103 ] used a DL model to predict the origins of 24 cancer types individually and collectively using thousands of somatic mutation features across 2 different classes (mutational distribution and driver gene and pathway features) [ 55 ]. Remarkably, the study found that driver genes and pathways are not among the most salient features, highlighting why previous efforts in panel and exome sequencing for CUP produced mixed results [ 104 , 105 , 106 , 107 ]. Deep learning approaches utilising transcriptome data have also shown utility in predicting tumour site of origin [ 56 , 57 ]. A neural network called SCOPE, trained on whole transcriptome TCGA data, was able to predict the origins of treatment-resistant metastatic cancers, even for rare cancers such as metastatic adenoid cystic carcinoma [ 56 ]. The CUP-AI-Dx algorithm, built upon a widely used CNN model called Inception [ 108 ], achieved similar results on 32 cancer types from TCGA and ICGC [ 57 ]. As whole genome sequencing becomes increasingly available, these models show great potential for future DL methods to incorporate multiple omics features to accurately categorise tumours into clinically meaningful subtypes by their molecular features.

In addition to genomic and transcriptomic data, a new model call TOAD trained on whole slide images (WSIs) was able to simultaneously predict metastasis status and origin of 18 tumour types [ 58 ]. Moreover, the model employed an explainability method called attention [ 109 , 110 ] to assign diagnostic relevance scores to image regions and revealed that regions with cancer cells contributed most to both metastasis and origin decision making [ 58 ]. These results suggested TOAD can ‘focus’ on biologically relevant image patterns and is a good candidate for clinical deployment.

Cancer prognosis and survival

Prognosis prediction is an essential part of clinical oncology, as the expected disease path and likelihood of survival can inform treatment decisions [ 111 ]. DL applied to genomic, transcriptomic and other data types has the potential to predict prognosis and patient survival [ 59 , 60 , 61 , 62 , 112 ]. The most common survival prediction method is the Cox proportional hazard regression model (Cox-PH) [ 113 , 114 , 115 ], which is a multivariate linear regression model finding correlations between survival time and predictor variables. A challenge of applying Cox-PH on genomic and transcriptomic data is its linear nature, which can potentially neglect complex and possibly nonlinear relationships between features [ 116 ]. By contrast, deep neural networks are naturally nonlinear, and in theory could excel at this task. Interestingly, many studies have incorporated Cox regression used for survival analysis into DL and trained these models on transcriptomic data for enhanced prognosis predictions [ 59 , 60 , 61 , 62 , 112 ]. Among them, Cox-nnet was a pioneering approach that made Cox regression the output layer of neural networks, effectively using millions of deep features extracted by hidden layers as input for the Cox regression model [ 59 ]. Cox-nnet was trained on RNA-seq data from 10 TCGA cancer types and benchmarked against two variations of Cox-PH (Cox-PH and CoxBoost). Cox-nnet showed superior accuracy and was the only model able to uniquely identify important pathways including p53 signalling, endocytosis and adherens junctions [ 59 ], demonstrating that the combination of Cox-PH and neural networks has the potential to capture biological information relating to prognosis. The potential of DL was confirmed by Huang et al. [ 62 ] who found that 3 different DL versions of Cox Regression (Cox-nnet, DeepSurv [ 60 ] and AECOX [ 62 ]) outperformed Cox-PH and traditional ML models. These results suggest that DL models can provide better accuracy than traditional models in predicting prognosis by learning from complex molecular interactions using their flexible architecture.

The incorporation of biological pathways in DL has enabled the elucidation of key survival drivers among thousands of features. PASNET [ 63 ] and its Cox-regression version Cox-PASNet [ 64 ] are among the most advanced DL models in this area. Both models incorporate a pathway layer between the input and the hidden layers of the neural network, where each node of the pathway layer represents a pathway (based on pathway databases such as Reactome [ 117 ] and KEGG [ 118 ]), and the connections between the two layers represent the gene-pathway relationships. These trained pathway nodes have different weights. By analysing the weight differences across different survival groups and identifying genes connected to each node, PASNet and Cox-PASNet were able to identify clinically actionable genetic traits of glioblastoma multiforme (GBM) and ovarian cancer [ 63 , 64 ]. In GBM, Cox-PASNet correctly identified PI3K cascade, a pathway highly involved in tumour proliferation, invasion and migration in GBM [ 119 ]. Cox-PASNet also correctly detected MAPK9, a gene strongly associated with GBM carcinogenesis and a novel potential therapeutic, as one the most influential genes [ 120 ]. The GCNN-explainability model from Chereda et al. is the latest example of incorporating molecular networks in cancer prognosis [ 19 ]. The study used gene expression profiles, structured by a PPI from Human Protein Reference Database (HPRD) [ 121 ], to predict metastasis of breast cancer samples. The explainability method, LRP [ 29 ], was then used to identify and analyse the biological relevance of the most relevant genes for predictions [ 19 ]. Pathway analysis of these genes showed that they include oncogenes, molecular-subtype-specific and therapeutically targetable genes, such as EGFR and ESR1 [ 19 ].

In addition to prognosis predictions from transcriptomic data, CNN models trained on histopathology images have been used to infer survival in several cancers including brain [ 122 ], colorectal [ 123 ], renal cell [ 124 ], liver cancers [ 125 ] and mesothelioma [ 65 ]. Among them, MesoNet [ 65 ] stands out for incorporating a feature contribution explainability algorithm called CHOWDER [ 126 ] on H&E tissue sections of mesothelioma to identify that the features contributing the most to survival predictions were primarily stromal cells associated with inflammation, cellular diversity and vacuolisation [ 65 ]. The CHOWDER algorithm enabled MesoNet to utilise large H&E images as well as segment and detect important regions for survival predictions without any local annotations by pathologists [ 65 ]. These findings suggest that ‘white-box’ DL models like MesoNet could be useful companion diagnostic tools in clinical setting by assisting clinicians in identifying known and novel histological features associated with a survival outcome.

Multi-modal DL analysis integrating histopathology images and, if available, omics data has the potential to better stratify patients into prognostic groups, as well as suggest more personalised and targeted treatments. Most multi-modal prognostic studies have focussed on three aspects: individual feature extraction from a single modality, multi-modal data integration and cross-modal analysis of prognostic features. The model PAGE-Net performed these tasks by using a CNN to create representations of WSIs and Cox-PASNet [ 64 ] to extract genetic pathway information from gene expression [ 66 ]. This architecture allowed PAGE-NET to not only integrate histopathological and transcriptomic data, but also identify patterns across both modalities that cause different survival rates [ 66 ]. More interestingly, the combination of multi-modal and explainability methods is particularly promising. PathME [ 67 ] is a pioneer of this approach by bringing together representation-extraction AEs and an explainability algorithm called SHAP [ 31 , 32 , 33 , 127 ]. The AEs captured important features from gene expression, miRNA expression, DNA methylation and CNAs for survival prediction, while SHAP scores each feature from each omic based on how relevant it is to the prediction [ 67 ]. Together, the two algorithms detected clinically relevant cross-omics features that affect survival across GBM, colorectal, breast and lung cancer [ 67 ]. The PathME methodology is cancer-agnostic, which makes it a great candidate for clinical implementations to explore actionable biomarkers in large-scale multi-omics data. Additionally, other studies [ 128 , 129 , 130 ] have employed Principal Component Analysis (PCA) [ 131 ] to compress gene expression, mutational signatures and methylation status into eigengene vectors [ 132 ], which were then combined with CNN-extracted histopathology features for survival predictions. While these methods could integrate histopathology data with multi-omics, they are not as explainable as PAGE-Net [ 66 ] or PathME [ 67 ] and thus less clinically suitable, as the conversion of genes into eigengenes makes exploration of cross-modality interactions challenging.

• Precision oncology

The promise of precision medicine is to use high-resolution omics data to enable optimised management and treatment of patients to improve survival. An important part of precision oncology involves understanding cancer genomics and the tumour microenvironment (TME). DL offers the potential to infer important genomic features from readily available histopathology data, as well as disentangle the complex heterogeneity of TME to enable precision oncology.

Genomic traits such as tumour mutation burden (TMB) and microsatellite instability (MSI) have been shown to be important biomarkers of immunotherapy response across cancer types [ 133 , 134 , 135 , 136 ]. Assessment of these traits requires sequencing (comprehensive panel, exome or whole genome), which is still expensive and is not readily available in the clinic.

Routinely used histopathological images are a potential window to genomic features and may in future prove useful for predictions of specific clinically meaningful molecular features without the need for tumour sequencing. Several CNN methods have been developed to infer TMB, MSI and other clinically relevant genomic features from H&E sections [ 68 , 69 , 70 , 137 ]. A model called Image2TMB used ensemble learning to predict TMB in lung cancer using H&E images. Image2TMB was able to achieve the same average accuracy as large panel sequencing with significantly less variance. It also attempted to estimate TMB for each region of an image [ 69 ], which could enable studies of histological features associated with molecular heterogeneity.

Another DL model called HE2RNA used weakly supervised learning to infer gene expression from histopathology images, which were then used to infer MSI status in colorectal cancer [ 68 ]. When compared with another DL method to predict MSI directly from H&E slides [ 137 ], HE2RNA showed superior performance on both formalin-fixed paraffin-embedded (FFPE) and frozen sections, indicating a high level of robustness across tissue processing approaches.

Kather et al. [ 70 ] has also showed that CNN models trained and evaluated on TCGA H&E slides can accurately predict a range of actionable genetic alterations across multiple cancer types, including mutational status of key genes, molecular subtypes and gene expression of standard biomarkers such as hormone receptor status. While these molecular inference methods demonstrate an intriguing application of DL in histopathology, their current clinical utility is likely to be limited as features such as MSI and hormone receptor status are already part of the routine diagnostic workflows (immunohistochemistry staining for mismatch-repair proteins in colorectal and endometrial cancer or ER, PR in breast cancer). However, these studies serve as proof-of-concept, and the developed models could in future be adapted to predict clinically important molecular features that are not routinely assessed. Thus, future investigations into histopathology-based genomic inference are warranted, with the understanding that the accuracy of such DL models needs to be exceptional for them to replace current assays.

The tumour microenvironment

The TME plays a key role in cancer progression, metastasis and response to therapy [ 138 ]. However, there remain many unknowns in the complex molecular and cellular interactions within the TME. The rise of DL in cancer research, coupled with large publicly available catalogues of genomic, transcriptomic and histopathology data, have created a strong technical framework for the use of neural networks in profiling the heterogeneity of TME.

Infiltrating immune cell populations, such as CD4+ and CD8+ T cells, are potential important biomarkers of immunotherapy response [ 139 , 140 ]. Traditional ML methods can accurately estimate TME cell compositions using transcriptomic [ 141 , 142 ] or methylation data [ 143 ]. However, most of these methods rely on the generation of signature Gene Expression Profiles (GEPs) or the selection of a limited number of CpG sites, biassed to previously known biomarkers. This can lead to models susceptible to noise and bias and unable to discover novel genetic biomarkers. DL methods can be trained on the whole dataset (i.e. the whole transcriptome) to identify the optimal features without relying on GEPs. Recently developed DL TME methods include Scaden [ 71 ], a transcriptomic-based neural network model, and MethylNet, a methylation-based model [ 72 ]. MethylNet also incorporated the SHAP explainability method [ 31 , 32 , 33 , 127 ] to quantify how relevant each CpG site is for deconvolution. While these methods currently focus on showing DL models are more robust against noise, bias and batch effects compared to traditional ML models, future follow-up studies are likely to reveal additional cellular heterogeneity traits of the TME and possibly inform treatment decisions. For example, a CNN trained on H&E slides of 13 cancer types [ 20 ] showed a strong correlation between spatial tumour infiltrating lymphocytes (TIL) patterns and cellular compositions derived by CIBERSORT (a Support Vector Regression model) [ 141 ]. These models have significant clinical implications, as rapid and automated identification of the composition, amount and spatial organisation of TIL can support the clinical decision making for prognosis predictions (for example, for breast cancer) and infer treatment options, specifically immunotherapy. We expect future DL methods will further explore the integrations of histopathology and omics in profiling tumour immune landscape [ 144 ]. We also expect future DL methods to incorporate single-cell transcriptomics (scRNA-Seq) data to improve TME predictions and even infer transcriptomic profiles of individual cell types. Several DL methods have already been developed to address batch correction, normalisation, imputation, dimensionality reduction and cell annotations for scRNA-Seq cancer data [ 145 , 146 , 147 ]. However, these studies are still experimental and require further effort and validation to be clinically applicable [ 148 ].

The new frontiers

An exciting new approach for studying the TME is spatial transcriptomics which allows quantification of gene expression in individual cells or regions while maintaining their positional representation, thus capturing spatial heterogeneity of gene expression at high resolution [ 149 , 150 ]. Given the complexity of this data, DL approaches are well suited for its analysis and interpretation. For example, by integrating histopathology images and spatial transcriptomics, DL can predict localised gene expression from tissue slides, as demonstrated by ST-Net, a neural network capable of predicting expressions of clinically relevant genes in breast cancer using tissue spots from H&E slides [ 73 ]. As the cost of spatial transcriptomics decreases in the future, it is expected more translational applications of DL will arise, for example utilising spatial transcriptomics information for improved prognosis predictions, subtype classification and refining our understanding of tumour heterogeneity [ 151 ].

In addition, gut microbiome, i.e. metagenome, has been an emerging field and shown to play an important role in cancer treatment efficacy and outcomes [ 152 , 153 ]. As more multi-omics datasets (genomics, transcriptomics, proteomics, microbiotics) are being generated, annotated and made available, we speculate that integrative analysis between these data types will help mapping omics profiles of each individual patient to the metagenome, which will unlock effective new exciting options.

Lastly, pharmacogenomics, to predict drug responses and the mechanisms of action using genomic characteristics, is an important and exciting area in precision oncology where DL methods have significant potential [ 154 ]. The increasing availability of public omics data has facilitated recent growth of DL applications in cancer pharmacogenomics [ 155 , 156 , 157 ]. Most common applications include therapy response and resistance (e.g. Dr.VAE [ 158 ] or CDRscan [ 74 ]), drug combination synergy (e.g. DeepSynergy [ 75 ] and Jiang et al. [ 76 ]), drug repositioning (e.g. deepDR [ 77 ]) and drug-target interactions (e.g. DeepDTI [ 78 ]). As pharmacogenomics is a highly translational field, we expect many such DL models will be applied in clinical setting in the future.

Challenges and limitations: the road to clinical implementation

This review provides an overview of exciting potential DL applications in oncology. However, there are several challenges to the widespread implementation of DL in clinical practice. Here, we discuss challenges and limitations of DL in clinical oncology and provide our perspective for future improvements.

Data variability

Data variability is a major challenge for applying DL to oncology. For example, in immunohistochemistry each lab may have different intensity of staining or have different qualities of staining. It is currently unclear how DL systems would deal with this inter- and intra-laboratory variability. For transcriptomic data, one of the principal difficulties is establishing the exact processing applied to generate a sequence library and processed dataset. Even properties as basic as ‘the list of human genes’ are not settled and multiple authorities publish and regularly update lists of genes, observed spliceforms, so any analysis should specify both the source and version of the gene model used. Additionally, there are a large range of data transformations (log, linear, etc.) and data normalisations (FPKM, TMM, TPM), with implementations in multiple programming languages resulting in a combinatorially large number of possible processing paths that should theoretically return the same results but without any formal process to ensure that that assumption is true.

Paucity of public phenotypically characterised datasets

One challenge of implementing DL into clinical practice is the need for large phenotypically characterised datasets that enable development and training of DL models with good generalisation performance. High-quality cancer datasets that have undergone omics profiling are difficult to acquire in the clinical setting due to cost, sample availability and quality. In addition, clinical tumour samples can be small and are typically stored as FFPE blocks, resulting in degraded RNA and crosslinked DNA not suitable for comprehensive molecular profiling. To overcome this, explanability methods, such as SHAP, could be applied on the current DL models, that are developed in research setting, to identify the most salient features and design targeted profiling workflows suitable for clinical samples. This way, the DL models could still capture the complexity and possible non-linear gene relationships, but be retrained to make clinical predictions using only the select salient features. Multi-modal based DL models coupled with explainability could also be explored due to their potential of using features in one modality to complement missing data in another. Transfer learning can also overcome challenges of requiring large datasets by pre-training DL models from other domains. In practice, however, large data sets with thousands of samples per class are still needed for accurate predictions in the clinic, as patient outcomes are complex and there is clinical heterogeneity between patients including responses, treatment courses, comorbidities and other lifestyle factors that may impact prognosis and survival. As more data is being routinely generated and clinical information centrally collected in digital health databases, we expect to see more DL models developed for treatment response predictions as well as the general prognosis predictions. More interestingly, DL’s ability to continue learning from and become more accurate with new training samples, i.e. active learning, can significantly help pathologists reduce time spent on training histopathology data annotation. For example, a histopathology-based DL model by Saltz et al. only required pathologists to annotate a few training images at a time, and stopping the manual annotation process when the model’s performance is satisfactory [ 20 ].

Lastly, clinical data about a sample or piece of data usually do not capture all the complexities of the samples and phenotype and can be prone to incompleteness, inconsistencies and errors. A potential strategy to address this issue is to design DL models less reliant on or independent from clinical annotations, for example the MesoNet model was able to detect prognostically meaningful regions from H&E images without any pathologist-derived annotations [ 65 ].

AI explainability and uncertainty

Finally, for DL to be implemented and accepted in the clinic, the models need to be designed to complement and enhance clinical workflows. For human experts to effectively utilise these models, they need to be not only explainable, but also capable of estimating the uncertainty in their predictions.

Over the last 5 years, research into explainable AI has accelerated. For DL to obtain regulatory approval and be used as a diagnostic tool, comprehensive studies of the biological relevance of explainability are imperative. In medical imaging, this entails validating DL-identified clinically relevant regions against pathology review, and in some cases, cross-validation with genomic features [ 46 ]. In genomics, this entails validating DL-identified relevant genetic features against those identified by conventional bioinformatics methods, for example confirming that the most discriminatory genes in predicting tissue types, as identified by SHAP, were also identified by pairwise differential expression analysis using edgeR [ 159 ] or showing that patient-specific molecular interaction networks produced in predicting metastasis status of breast cancer were not only linked to benign/malignant phenotype, but also indicative of tumour progression and therapeutic targets [ 19 ].

Furthermore, DL model’s ability to produce the ‘I don’t know’ output, when uncertain about predictions, is critical. Most DL applications covered in this review are point-estimate methods, i.e. the predictions are simply the best guess with the highest probability. In critical circumstances, overconfident predictions, e.g. predicting cancer primary site with only 40% certainty, can result in inaccurate diagnosis or cancer management decisions. Furthermore, when uncertainty estimates are too high, companion diagnostic tools should be able to abstain from making predictions and ask for medical experts’ opinion [ 160 ]. Probabilistic DL methods capable of quantifying prediction uncertainty, such as Bayesian DL [ 161 ], are great candidates to address these issues and have recently started to be applied in cancer diagnosis tasks [ 162 , 163 , 164 ]. We expect probabilistic models to become mainstream in oncology in the near future.

Conclusions

In summary, DL has the potential to dramatically transformed cancer care and bring it a step closer to the promise of precision oncology. In an era where genomics is being implemented into health delivery and health data is becoming increasingly digitised, it is anticipated that artificial intelligence and DL will be used in the development, validation and implementation of decision support tools to facilitate precision oncology. In this review, we showcased a number of promising applications of DL in various areas of oncology, including digital histopathology, molecular subtyping, cancer diagnosis, prognostication, histological inference of genomic characteristics, tumour microenvironment and emerging frontiers such as spatial transcriptomics and pharmacogenomics. As the research matures, the future of applied DL in oncology will likely focus on integration of medical images and omics data using multimodal learning that can identify biologically meaningful biomarkers. Excitingly, the combination of multimodal learning and explainability can reveal novel insights. Important prerequisites of widespread adoption of DL in clinical setting are phenotypically rich data for training models and clinical validation of the biological relevance of DL-generated insights. We expect as new technologies such as single-cell sequencing, spatial transcriptomics and multiplexed imaging become more accessible, more efforts will be dedicated to improving both the quantity and quality of labelling/annotation of medical data. Finally, for DL to be accepted in routine patient care, clinical validation of explainable DL methods will play a vital role.

Availability of data and materials

Not applicable

Abbreviations

Autoencoder

• Artificial intelligence
• Cancer of unknown primary

Copy number aberrations

Convolutional neural network

Cox proportional hazard regression model

• Deep learning

European Genome Atlas

Formalin-fixed, paraffin-embedded

Glioblastoma multiforme

Graph convolutional neural network

Gene Expression Omnibus

Graphical Processing Units

Human Protein Reference Database

Haematoxylin and Eosin

International Cancer Genome Consortium

Layer-wise Relevance Propagation

Microsatellite instability

Machine learning

Multilayer perceptron

Pan-Cancer Analysis of Whole Genomes

Protein-protein interactions

RNA sequencing

Recurrent neural network

Support vector machine

The Cancer Genome Atlas

Tumour infiltrating lymphocytes

• Tumour microenvironment

Tumour mutation burden

Weighted correlation network analysis

LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436–44.

Article   CAS   PubMed   Google Scholar  

Libbrecht MW, Noble WS. Machine learning applications in genetics and genomics. Nat Rev Genet. 2015;16:321–32.

Article   CAS   PubMed   PubMed Central   Google Scholar  

Jones W, Alasoo K, Fishman D, Parts L. Computational biology: deep learning. Skolnick J, editor. Emerg Top Life Sci. 2017;1:257–74.

Article   PubMed   Google Scholar  

Wainberg M, Merico D, Delong A, Frey BJ. Deep learning in biomedicine. Nat Biotechnol. 2018;36:829–38.

Zou J, Huss M, Abid A, Mohammadi P, Torkamani A, Telenti A. A primer on deep learning in genomics. Nat Genet. 2019;51:12–8.

Montesinos-López OA, Montesinos-López A, Pérez-Rodríguez P, Barrón-López JA, Martini JWR, Fajardo-Flores SB, et al. A review of deep learning applications for genomic selection. BMC Genomics. 2021;22:19.

Article   PubMed   PubMed Central   Google Scholar  

Dias R, Torkamani A. Artificial intelligence in clinical and genomic diagnostics. Genome Med. 2019;11(1):70. https://doi.org/10.1186/s13073-019-0689-8 .

Eraslan G, Avsec Ž, Gagneur J, Theis FJ. Deep learning: new computational modelling techniques for genomics. Nat Rev Genet. 2019;20(7):389–403. https://doi.org/10.1038/s41576-019-0122-6 .

Huynh E, Hosny A, Guthier C, Bitterman DS, Petit SF, Haas-Kogan DA, et al. Artificial intelligence in radiation oncology. Nat Rev Clin Oncol. 2020;17:771–81.

Bera K, Schalper KA, Rimm DL, Velcheti V, Madabhushi A. Artificial intelligence in digital pathology—new tools for diagnosis and precision oncology. Nat Rev Clin Oncol. 2019;16:703–15.

Huss R, Coupland SE. Software-assisted decision support in digital histopathology. J Pathol. 2020;250:685–92.

Massion PP, Antic S, Ather S, Arteta C, Brabec J, Chen H, et al. Assessing the accuracy of a deep learning method to risk stratify indeterminate pulmonary nodules. Am J Respir Crit Care Med. 2020;202:241–9.

Kanan C, Sue J, Grady L, Fuchs TJ, Chandarlapaty S, Reis-Filho JS, et al. Independent validation of paige prostate: assessing clinical benefit of an artificial intelligence tool within a digital diagnostic pathology laboratory workflow. J Clin Oncol. 2020;38(15_suppl):e14076. https://doi.org/10.1200/JCO.2020.38.15_suppl.e14076 .

Article   Google Scholar  

Silva LM, Pereira EM, Salles PG, Godrich R, Ceballos R, Kunz JD, et al. Independent real-world application of a clinical-grade automated prostate cancer detection system. J Pathol. 2021;path:5662.

Google Scholar  

Schulte-Sasse R, Budach S, Hnisz D, Marsico A. Graph convolutional networks improve the prediction of cancer driver genes. Artif Neural Netw Mach Learn – ICANN 2019 [Internet]. Munich: Springer; 2019. p. 658–68. Available from: https://link.springer.com/chapter/10.1007%2F978-3-030-30493-5_60

Szklarczyk D, Franceschini A, Wyder S, Forslund K, Heller D, Huerta-Cepas J, et al. STRING v10: protein–protein interaction networks, integrated over the tree of life. Nucleic Acids Res. 2015;43:D447–52.

Ramirez R, Chiu Y-C, Hererra A, Mostavi M, Ramirez J, Chen Y, et al. Classification of cancer types using graph convolutional neural networks. Front Phys. 2020;8:203. https://doi.org/10.3389/fphy.2020.00203 .

Rhee S, Seo S, Kim S. Hybrid approach of relation network and localized graph convolutional filtering for breast cancer subtype classification. Proc Twenty-Seventh Int Jt Conf Artif Intell [Internet]. Stockholm: International Joint Conferences on Artificial Intelligence Organization; 2018. p. 3527–34. [cited 2021 Apr 30]. Available from: https://www.ijcai.org/proceedings/2018/490

Chereda H, Bleckmann A, Menck K, Perera-Bel J, Stegmaier P, Auer F, et al. Explaining decisions of graph convolutional neural networks: patient-specific molecular subnetworks responsible for metastasis prediction in breast cancer. Genome Med. 2021;13:42.

Saltz J, Gupta R, Hou L, Kurc T, Singh P, Nguyen V, et al. Spatial organization and molecular correlation of tumor-infiltrating lymphocytes using deep learning on pathology images. Cell Rep. 2018;23:181–193.e7.

Gao J, Li P, Chen Z, Zhang J. A survey on deep learning for multimodal data fusion. Neural Comput. 2020;32:829–64.

Sun D, Wang M, Li A. A multimodal deep neural network for human breast cancer prognosis prediction by integrating multi-dimensional data. IEEE/ACM Trans Comput Biol Bioinform. 2019;16:841–50.

Cheerla A, Gevaert O. Deep learning with multimodal representation for pancancer prognosis prediction. Bioinformatics. 2019;35(14):i446–54. https://doi.org/10.1093/bioinformatics/btz342 .

Tschannen M, Bachem O, Lucic M. Recent advances in autoencoder-based representation learning. ArXiv181205069 Cs Stat [Internet]. 2018; [cited 2020 Apr 21]; Available from: http://arxiv.org/abs/1812.05069 .

Kelly CJ, Karthikesalingam A, Suleyman M, Corrado G, King D. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. 2019;17:195.

Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. 2019;1:206–15.

The Precise4Q consortium, Amann J, Blasimme A, Vayena E, Frey D, Madai VI. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC Med Inform Decis Mak. 2020;20:310.

Article   PubMed Central   Google Scholar  

Shrikumar A, Greenside P, Kundaje A. Learning important features through propagating activation differences. ArXiv170402685 Cs [Internet]. 2019; [cited 2020 Apr 20]; Available from: http://arxiv.org/abs/1704.02685 .

Bach S, Binder A, Montavon G, Klauschen F, Müller K-R, Samek W. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. Suarez OD, editor. PLoS One. 2015;10:e0130140.

Ribeiro MT, Singh S, Guestrin C. “Why Should I Trust You?”: explaining the predictions of any classifier. Proc 22nd ACM SIGKDD Int Conf Knowl Discov Data Min [Internet]. San Francisco: ACM; 2016. p. 1135–44. [cited 2020 Dec 8]. Available from: https://dl.acm.org/doi/10.1145/2939672.2939778

Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. NIPS17 Proc 31st. Int Conf Neural Inf Process Syst Curran Associates Inc. 2017;30:4768–77.

Erion G, Janizek JD, Sturmfels P, Lundberg S, Lee S-I. Learning explainable models using attribution priors. ArXiv190610670 Cs Stat [Internet]. 2019; [cited 2020 Jun 22]; Available from: http://arxiv.org/abs/1906.10670 .

Lundberg SM, Erion G, Chen H, DeGrave A, Prutkin JM, Nair B, et al. From local explanations to global understanding with explainable AI for trees. Nat Mach Intell. 2020;2(1):56–67. https://doi.org/10.1038/s42256-019-0138-9 .

The Cancer Genome Atlas Research Network, Weinstein JN, Collisson EA, Mills GB, Shaw KRM, Ozenberger BA, et al. The Cancer Genome Atlas Pan-Cancer analysis project. Nat Genet. 2013;45:1113–20.

Article   CAS   Google Scholar  

The International Cancer Genome Consortium. International network of cancer genome projects. Nature. 2010;464:993–8.

Article   CAS   PubMed Central   Google Scholar  

Edgar R. Gene Expression Omnibus: NCBI gene expression and hybridization array data repository. Nucleic Acids Res. 2002;30:207–10.

Lappalainen I, Almeida-King J, Kumanduri V, Senf A, Spalding JD, ur-Rehman S, et al. The European Genome-phenome Archive of human data consented for biomedical research. Nat Genet. 2015;47(7):692–5. https://doi.org/10.1038/ng.3312 .

METABRIC Group, Curtis C, Shah SP, Chin S-F, Turashvili G, Rueda OM, et al. The genomic and transcriptomic architecture of 2,000 breast tumours reveals novel subgroups. Nature. 2012;486:346–52.

Zhuang F, Qi Z, Duan K, Xi D, Zhu Y, Zhu H, et al. A comprehensive survey on transfer learning. ArXiv191102685 Cs Stat [Internet]. 2020; [cited 2020 Dec 6]; Available from: http://arxiv.org/abs/1911.02685 .

Ryu HS, Jin M-S, Park JH, Lee S, Cho J, Oh S, et al. Automated gleason scoring and tumor quantification in prostate core needle biopsy images using deep neural networks and its comparison with pathologist-based assessment. Cancers. 2019;11:1860.

Nir G, Karimi D, Goldenberg SL, Fazli L, Skinnider BF, Tavassoli P, et al. Comparison of artificial intelligence techniques to evaluate performance of a classifier for automatic grading of prostate cancer from digitized histopathologic images. JAMA Netw Open. 2019;2:e190442.

Ström P, Kartasalo K, Olsson H, Solorzano L, Delahunt B, Berney DM, et al. Artificial intelligence for diagnosis and grading of prostate cancer in biopsies: a population-based, diagnostic study. Lancet Oncol. 2020;21:222–32.

Ehteshami Bejnordi B, Mullooly M, Pfeiffer RM, Fan S, Vacek PM, Weaver DL, et al. Using deep convolutional neural networks to identify and classify tumor-associated stroma in diagnostic breast biopsies. Mod Pathol. 2018;31:1502–12.

Vuong TLT, Lee D, Kwak JT, Kim K. Multi-task deep learning for colon cancer grading. 2020 Int Conf Electron Inf Commun ICEIC [Internet]. Barcelona: IEEE; 2020. p. 1–2. [cited 2020 Nov 9]. Available from: https://ieeexplore.ieee.org/document/9051305/

El Achi HE, Khoury JD. Artificial intelligence and digital microscopy applications in diagnostic hematopathology. Cancers. 2020;12(4):797. https://doi.org/10.3390/cancers12040797 .

Hägele M, Seegerer P, Lapuschkin S, Bockmayr M, Samek W, Klauschen F, et al. Resolving challenges in deep learning-based analyses of histopathological images using explanation methods. Sci Rep. 2020;10:6423.

Poojitha UP, Lal SS. Hybrid unified deep learning network for highly precise gleason grading of prostate cancer. 2019 41st Annu Int Conf IEEE Eng Med Biol Soc EMBC [Internet]. Berlin: IEEE; 2019. p. 899–903. [cited 2020 Apr 3]Available from: https://ieeexplore.ieee.org/document/8856912/

Gao F, Wang W, Tan M, Zhu L, Zhang Y, Fessler E, et al. DeepCC: a novel deep learning-based framework for cancer molecular subtype classification. Oncogenesis. 2019;8:44.

Yu K-H, Wang F, Berry GJ, Ré C, Altman RB, Snyder M, et al. Classifying non-small cell lung cancer types and transcriptomic subtypes using convolutional neural networks. J Am Med Inform Assoc. 2020;27:757–69.

Sirinukunwattana K, Domingo E, Richman SD, Redmond KL, Blake A, Verrill C, et al. Image-based consensus molecular subtype (imCMS) classification of colorectal cancer using deep learning. Gut. 2020;gutjnl-2019:319866.

Stålhammar G, Fuentes Martinez N, Lippert M, Tobin NP, Mølholm I, Kis L, et al. Digital image analysis outperforms manual biomarker assessment in breast cancer. Mod Pathol. 2016;29(4):318–29. https://doi.org/10.1038/modpathol.2016.34 .

Couture HD, Williams LA, Geradts J, Nyante SJ, Butler EN, Marron JS, et al. Image analysis with deep learning to predict breast cancer grade, ER status, histologic subtype, and intrinsic subtype. NPJ Breast Cancer. 2018;4:30.

Woerl A-C, Eckstein M, Geiger J, Wagner DC, Daher T, Stenzel P, et al. Deep Learning Predicts Molecular Subtype of Muscle-invasive bladder cancer from conventional histopathological slides. Eur Urol. 2020;78:256–64.

Md MI, Huang S, Ajwad R, Chi C, Wang Y, Hu P. An integrative deep learning framework for classifying molecular subtypes of breast cancer. Comput Struct Biotechnol J. 2020;18:2185–99.

PCAWG Tumor Subtypes and Clinical Translation Working Group, PCAWG Consortium, Jiao W, Atwal G, Polak P, Karlic R, et al. A deep learning system accurately classifies primary and metastatic cancers using passenger mutation patterns. Nat Commun. 2020;11:728.

Grewal JK, Tessier-Cloutier B, Jones M, Gakkhar S, Ma Y, Moore R, et al. Application of a neural network whole transcriptome–based pan-cancer method for diagnosis of primary and metastatic cancers. JAMA Netw Open. 2019;2(4):e192597. https://doi.org/10.1001/jamanetworkopen.2019.2597 .

Zhao Y, Pan Z, Namburi S, Pattison A, Posner A, Balachander S, et al. CUP-AI-Dx: A tool for inferring cancer tissue of origin and molecular subtype using RNA gene-expression data and artificial intelligence. EBioMedicine. 2020;61:103030.

Lu MY, Chen TY, Williamson DFK, Zhao M, Shady M, Lipkova J, et al. AI-based pathology predicts origins for cancers of unknown primary. Nature. 2021;594:106–10.

Ching T, Zhu X, Garmire LX. Cox-nnet: an artificial neural network method for prognosis prediction of high-throughput omics data. Markowetz F, editor. PLoS Comput Biol. 2018;14:e1006076.

Katzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Med Res Methodol. 2018;18:24.

Jing B, Zhang T, Wang Z, Jin Y, Liu K, Qiu W, et al. A deep survival analysis method based on ranking. Artif Intell Med. 2019;98:1–9.

Huang Z, Johnson TS, Han Z, Helm B, Cao S, Zhang C, et al. Deep learning-based cancer survival prognosis from RNA-seq data: approaches and evaluations. BMC Med Genet. 2020;13:41.

Hao J, Kim Y, Kim T-K, Kang M. PASNet: pathway-associated sparse deep neural network for prognosis prediction from high-throughput data. BMC Bioinformatics. 2018;19:510.

Hao J, Kim Y, Mallavarapu T, Oh JH, Kang M. Interpretable deep neural network for cancer survival analysis by integrating genomic and clinical data. BMC Med Genet. 2019;12:189.

Courtiol P, Maussion C, Moarii M, Pronier E, Pilcer S, Sefta M, et al. Deep learning-based classification of mesothelioma improves prediction of patient outcome. Nat Med. 2019;25:1519–25.

Hao J, Kosaraju SC, Tsaku NZ, Song DH, Kang M. PAGE-Net: interpretable and integrative deep learning for survival analysis using histopathological images and genomic data. Biocomput 2020 [Internet]. Kohala Coast: WORLD SCIENTIFIC; 2019. p. 355–66. [cited 2020 Apr 6]. Available from: https://www.worldscientific.com/doi/abs/10.1142/9789811215636_0032

Lemsara A, Ouadfel S, Fröhlich H. PathME: pathway based multi-modal sparse autoencoders for clustering of patient-level multi-omics data. BMC Bioinformatics. 2020;21:146.

Schmauch B, Romagnoni A, Pronier E, Saillard C, Maillé P, Calderaro J, et al. A deep learning model to predict RNA-Seq expression of tumours from whole slide images. Nat Commun. 2020;11:3877.

Jain MS, Massoud TF. Predicting tumour mutational burden from histopathological images using multiscale deep learning. Nat Mach Intell. 2020;2:356–62.

Kather JN, Heij LR, Grabsch HI, Loeffler C, Echle A, Muti HS, et al. Pan-cancer image-based detection of clinically actionable genetic alterations. Nat Can. 2020;1(8):789–99. https://doi.org/10.1038/s43018-020-0087-6 .

Menden K, Marouf M, Oller S, Dalmia A, Magruder DS, Kloiber K, et al. Deep learning–based cell composition analysis from tissue expression profiles. Sci Adv [Internet]. 2020;6 Available from: https://advances.sciencemag.org/content/6/30/eaba2619 .

Levy JJ, Titus AJ, Petersen CL, Chen Y, Salas LA, Christensen BC. MethylNet: an automated and modular deep learning approach for DNA methylation analysis. BMC Bioinformatics. 2020;21:108.

He B, Bergenstråhle L, Stenbeck L, Abid A, Andersson A, Borg Å, et al. Integrating spatial gene expression and breast tumour morphology via deep learning. Nat Biomed Eng. 2020;4:827–34.

Chang Y, Park H, Yang H-J, Lee S, Lee K-Y, Kim TS, et al. Cancer Drug Response Profile scan (CDRscan): a deep learning model that predicts drug effectiveness from cancer genomic signature. Sci Rep. 2018;8:8857.

Preuer K, Lewis RPI, Hochreiter S, Bender A, Bulusu KC, Klambauer G. DeepSynergy: predicting anti-cancer drug synergy with Deep Learning. Wren J, editor. Bioinformatics. 2018;34(9):1538–46. https://doi.org/10.1093/bioinformatics/btx806 .

Jiang P, Huang S, Fu Z, Sun Z, Lakowski TM, Hu P. Deep graph embedding for prioritizing synergistic anticancer drug combinations. Comput Struct Biotechnol J. 2020;18:427–38.

Zeng X, Zhu S, Liu X, Zhou Y, Nussinov R, Cheng F. deepDR: a network-based deep learning approach to in silico drug repositioning. Cowen L, editor. Bioinformatics. 2019;35:5191–5198.

Wen M, Zhang Z, Niu S, Sha H, Yang R, Yun Y, et al. Deep-learning-based drug−target interaction prediction. J Proteome Res. 2017;16(4):1401–9.

Walsh S, de Jong EEC, van Timmeren JE, Ibrahim A, Compter I, Peerlings J, et al. Decision support systems in oncology. JCO Clin Cancer Inform. 2019;3(3):1–9. https://doi.org/10.1200/CCI.18.00001 .

Gurcan M, Lozanski G, Pennell M, Shana′Ah A, Zhao W, Gewirtz A, et al. Inter-reader variability in follicular lymphoma grading: conventional and digital reading. J Pathol Inform. 2013;4:30.

Rabe K, Snir OL, Bossuyt V, Harigopal M, Celli R, Reisenbichler ES. Interobserver variability in breast carcinoma grading results in prognostic stage differences. Hum Pathol. 2019;94:51–7.

Maggiori E, Tarabalka Y, Charpiat G, Alliez P. High-resolution image classification with convolutional networks. 2017 IEEE Int Geosci Remote Sens Symp IGARSS [Internet]. Fort Worth: IEEE; 2017. p. 5157–60. [cited 2020 Dec 8]. Available from: http://ieeexplore.ieee.org/document/8128163/

Goodfellow IJ, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, et al. Generative adversarial networks. ArXiv14062661 Cs Stat [Internet]. 2014; [cited 2021 Apr 27]; Available from: http://arxiv.org/abs/1406.2661 .

Luc P, Couprie C, Chintala S, Verbeek J. Semantic segmentation using adversarial networks. ArXiv161108408 Cs [Internet]. 2016; [cited 2021 Aug 12]; Available from: http://arxiv.org/abs/1611.08408 .

Sorlie T, Perou CM, Tibshirani R, Aas T, Geisler S, Johnsen H, et al. Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc Natl Acad Sci. 2001;98:10869–74.

Yersal O. Biological subtypes of breast cancer: prognostic and therapeutic implications. World J Clin Oncol. 2014;5(3):412–24. https://doi.org/10.5306/wjco.v5.i3.412 .

Komor MA, Bosch LJ, Bounova G, Bolijn AS, Delis-van Diemen PM, Rausch C, et al. Consensus molecular subtype classification of colorectal adenomas: CMS classification of colorectal adenomas. J Pathol. 2018;246:266–76.

Tothill RW, Tinker AV, George J, Brown R, Fox SB, Lade S, et al. Novel molecular subtypes of serous and endometrioid ovarian cancer linked to clinical outcome. Clin Cancer Res. 2008;14(16):5198–208. https://doi.org/10.1158/1078-0432.CCR-08-0196 .

Jain S, Xu R, Prieto VG, Lee P. Molecular classification of soft tissue sarcomas and its clinical applications. Int J Clin Exp. 2010;3:416–29.

CAS   Google Scholar  

Leek JT, Scharpf RB, Bravo HC, Simcha D, Langmead B, Johnson WE, et al. Tackling the widespread and critical impact of batch effects in high-throughput data. Nat Rev Genet. 2010;11:733–9.

Haury A-C, Gestraud P, Vert J-P. The influence of feature selection methods on accuracy, stability and interpretability of molecular signatures. Teh M-T, editor. PLoS One. 2011;6:e28210.

Kela I, Ein-Dor L, Getz G, Givol D, Domany E. Outcome signature genes in breast cancer: is there a unique set? Breast Cancer Res. 2005;7:P4.38, bcr1168.

Drier Y, Domany E. Do two machine-learning based prognostic signatures for breast cancer capture the same biological processes? El-Rifai W, editor. PLoS One. 2011;6:e17795.

Hu F, Zhou Y, Wang Q, Yang Z, Shi Y, Chi Q. Gene expression classification of lung adenocarcinoma into molecular subtypes. IEEE/ACM Trans Comput Biol Bioinform. 2020;17:1187–97.

Wang K, Duan X, Gao F, Wang W, Liu L, Wang X. Dissecting cancer heterogeneity based on dimension reduction of transcriptomic profiles using extreme learning machines. Wong K-K, editor. PLoS One. 2018;13:e0203824.

Varadhachary GR, Abbruzzese JL, Lenzi R. Diagnostic strategies for unknown primary cancer. Cancer. 2004;100:1776–85.

Greco FA. Molecular diagnosis of the tissue of origin in cancer of unknown primary site: useful in patient management. Curr Treat Options in Oncol. 2013;14:634–42.

Pavlidis N, Pentheroudakis G. Cancer of unknown primary site. Lancet. 2012;379:1428–35.

Varadhachary GR, Raber MN. Cancer of unknown primary site. N Engl J Med. 2014;371(8):757–65. https://doi.org/10.1056/NEJMra1303917 .

Kandoth C, McLellan MD, Vandin F, Ye K, Niu B, Lu C, et al. Mutational landscape and significance across 12 major cancer types. Nature. 2013;502:333–9.

Lawrence MS, Stojanov P, Polak P, Kryukov GV, Cibulskis K, Sivachenko A, et al. Mutational heterogeneity in cancer and the search for new cancer-associated genes. Nature. 2013;499:214–8.

Ciriello G, Miller ML, Aksoy BA, Senbabaoglu Y, Schultz N, Sander C. Emerging landscape of oncogenic signatures across human cancers. Nat Genet. 2013;45:1127–33.

The ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium. Pan-cancer analysis of whole genomes. Nature. 2020;578:82–93.

Chen Y, Sun J, Huang L-C, Xu H, Zhao Z. Classification of cancer primary sites using machine learning and somatic mutations. Biomed Res Int. 2015;2015:1–9.

Tothill RW, Li J, Mileshkin L, Doig K, Siganakis T, Cowin P, et al. Massively-parallel sequencing assists the diagnosis and guided treatment of cancers of unknown primary: NGS in cancers of unknown primary. J Pathol. 2013;231:413–23.

Soh KP, Szczurek E, Sakoparnig T, Beerenwinkel N. Predicting cancer type from tumour DNA signatures. Genome Med. 2017;9:104.

Marquard AM, Birkbak NJ, Thomas CE, Favero F, Krzystanek M, Lefebvre C, et al. TumorTracer: a method to identify the tissue of origin from the somatic mutations of a tumor specimen. BMC Med Genet. 2015;8:58.

Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, et al. Going deeper with convolutions. 2015 IEEE Conf Comput Vis Pattern Recognit CVPR [Internet]. Boston: IEEE; 2015. p. 1–9. [cited 2020 Dec 8]. Available from: http://ieeexplore.ieee.org/document/7298594/

Ilse M, Tomczak JM, Welling M. Attention-based deep multiple instance learning. arXiv:1802.04712 Cs [Internet]. 2018. [cited 2021 Sep 17]. Available from https://arxiv.org/abs/1802.04712 .

Lu MY, Williamson DFK, Chen TY, Chen RJ, Barbieri M, Mahmood F. Data-efficient and weakly supervised computational pathology on whole-slide images. Nat Biomed Eng [Internet]. 2021; [cited 2021 May 10]; Available from: http://www.nature.com/articles/s41551-020-00682-w .

Nair M, Sandhu S, Sharma A. Prognostic and predictive biomarkers in cancer. Curr Cancer Drug Targets. 2014;14:477–504.

Lai Y-H, Chen W-N, Hsu T-C, Lin C, Tsao Y, Wu S. Overall survival prediction of non-small cell lung cancer by integrating microarray and clinical data with deep learning. Sci Rep. 2020;10:4679.

Cox DR. Regression Models and Life-Tables, vol. 35; 2020.

Ahmed FE, Vos PW, Holbert D. Modeling survival in colon cancer: a methodological review. Mol Cancer. 2007;6(1):15. https://doi.org/10.1186/1476-4598-6-15 .

de O Ferraz R, Moreira-Filho D de C. Survival analysis of women with breast cancer: competing risk models. Ciênc Saúde Coletiva. 2017;22:3743–54.

Solvang HK, Lingjærde OC, Frigessi A, Børresen-Dale A-L, Kristensen VN. Linear and non-linear dependencies between copy number aberrations and mRNA expression reveal distinct molecular pathways in breast cancer. BMC Bioinformatics. 2011;12:197.

Fabregat A, Jupe S, Matthews L, Sidiropoulos K, Gillespie M, Garapati P, et al. The Reactome Pathway Knowledgebase. Nucleic Acids Res. 2018;46(D1):D649–55. https://doi.org/10.1093/nar/gkx1132 .

Kanehisa M, Furumichi M, Tanabe M, Sato Y, Morishima K. KEGG: new perspectives on genomes, pathways, diseases and drugs. Nucleic Acids Res. 2017;45:D353–61.

Weber GL, Parat M-O, Binder ZA, Gallia GL, Riggins GJ. Abrogation of PIK3CA or PIK3R1 reduces proliferation, migration, and invasion in glioblastoma multiforme cells. Oncotarget. 2011;2:833–49.

Brahm CG, Walenkamp AME, Linde MEV, Verheul HMW, Stephan R, Fehrmann N. Identification of novel therapeutic targets in glioblastoma with functional genomic mRNA profiling. J Clin Oncol [Internet]. 2017;35 Available from: https://ascopubs.org/doi/10.1200/JCO.2017.35.15_suppl.2018 .

Keshava Prasad TS, Goel R, Kandasamy K, Keerthikumar S, Kumar S, Mathivanan S, et al. Human protein reference database--2009 update. Nucleic Acids Res. 2009;37:D767–72.

Zadeh Shirazi A, Fornaciari E, Bagherian NS, Ebert LM, Koszyca B, Gomez GA. DeepSurvNet: deep survival convolutional network for brain cancer survival rate classification based on histopathological images. Med Biol Eng Comput [Internet]. 2020; [cited 2020 Apr 6]; Available from: http://link.springer.com/10.1007/s11517-020-02147-3 .

Bychkov D, Linder N, Turkki R, Nordling S, Kovanen PE, Verrill C, et al. Deep learning based tissue analysis predicts outcome in colorectal cancer. Sci Rep. 2018;8:3395.

Tabibu S, Vinod PK, Jawahar CV. Pan-renal cell carcinoma classification and survival prediction from histopathology images using deep learning. Sci Rep. 2019;9:10509.

Saillard C, Schmauch B, Laifa O, Moarii M, Toldo S, Zaslavskiy M, et al. Predicting survival after hepatocellular carcinoma resection using deep-learning on histological slides. Hepatology. 2020;72(6):2000–13.

Courtiol P, Tramel EW, Sanselme M, Wainrib G. Classification and disease localization in histopathology using only global labels: a weakly-supervised approach. ArXiv180202212 Cs Stat [Internet]. 2020; [cited 2020 Apr 9]; Available from: http://arxiv.org/abs/1802.02212 .

Lundberg SM, Nair B, Vavilala MS, Horibe M, Eisses MJ, Adams T, et al. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nat Biomed Eng. 2018;2:749–60.

Shao W, Cheng J, Sun L, Han Z, Feng Q, Zhang D, et al. Ordinal multi-modal feature selection for survival analysis of early-stage renal cancer. In: Frangi AF, Schnabel JA, Davatzikos C, Alberola-López C, Fichtinger G, editors. Med Image Comput Comput Assist Interv – MICCAI 2018 [Internet]. Cham: Springer International Publishing; 2018. p. 648–56. [cited 2020 Apr 21]. Available from: http://link.springer.com/10.1007/978-3-030-00934-2_72 .

Ning Z, Pan W, Chen Y, Xiao Q, Zhang X, Luo J, et al. Integrative analysis of cross-modal features for the prognosis prediction of clear cell renal cell carcinoma. Schwartz R, editor. Bioinformatics. 2020;36(9):2888–95.

Shao W, Huang K, Han Z, Cheng J, Cheng L, Wang T, et al. Integrative analysis of pathological images and multi-dimensional genomic data for early-stage cancer prognosis. IEEE Trans Med Imaging. 2020;39(1):99–110. https://doi.org/10.1109/TMI.2019.2920608 .

Makiewicz A, Ratajczak W. Principal Components Analysis (PCA). Computers & Geosciences. 1993;19:303–42.

Langfelder P, Horvath S. WGCNA: an R package for weighted correlation network analysis. BMC Bioinformatics. 2008;9:559.

Article   PubMed   PubMed Central   CAS   Google Scholar  

Samstein RM, Lee C-H, Shoushtari AN, Hellmann MD, Shen R, Janjigian YY, et al. Tumor mutational load predicts survival after immunotherapy across multiple cancer types. Nat Genet. 2019;51:202–6.

Riviere P, Goodman AM, Okamura R, Barkauskas DA, Whitchurch TJ, Lee S, et al. High tumor mutational burden correlates with longer survival in immunotherapy-naïve patients with diverse cancers. Mol Cancer Ther. 2020;19(10):2139–45. https://doi.org/10.1158/1535-7163.MCT-20-0161 .

Bao X, Zhang H, Wu W, Cheng S, Dai X, Zhu X, et al. Analysis of the molecular nature associated with microsatellite status in colon cancer identifies clinical implications for immunotherapy. J Immunother Cancer. 2020;8:e001437.

Cortes-Ciriano I, Lee S, Park W-Y, Kim T-M, Park PJ. A molecular portrait of microsatellite instability across multiple cancers. Nat Commun. 2017;8:15180.

Kather JN, Pearson AT, Halama N, Jäger D, Krause J, Loosen SH, et al. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nat Med. 2019;25:1054–6.

Runa F, Hamalian S, Meade K, Shisgal P, Gray PC, Kelber JA. Tumor microenvironment heterogeneity: challenges and opportunities. Curr Mol Biol Rep. 2017;3:218–29.

Borst J, Ahrends T, Bąbała N, Melief CJM, Kastenmüller W. CD4+ T cell help in cancer immunology and immunotherapy. Nat Rev Immunol. 2018;18(10):635–47. https://doi.org/10.1038/s41577-018-0044-0 .

Tumeh PC, Harview CL, Yearley JH, Shintaku IP, Taylor EJM, Robert L, et al. PD-1 blockade induces responses by inhibiting adaptive immune resistance. Nature. 2014;515:568–71.

Newman AM, Liu CL, Green MR, Gentles AJ, Feng W, Xu Y, et al. Robust enumeration of cell subsets from tissue expression profiles. Nat Methods. 2015;12:453–7.

Newman AM, Steen CB, Liu CL, Gentles AJ, Chaudhuri AA, Scherer F, et al. Determining cell type abundance and expression from bulk tissues with digital cytometry. Nat Biotechnol. 2019;37:773–82.

Chakravarthy A, Furness A, Joshi K, Ghorani E, Ford K, Ward MJ, et al. Pan-cancer deconvolution of tumour composition using DNA methylation. Nat Commun. 2018;9:3220.

Klauschen F, Müller K-R, Binder A, Bockmayr M, Hägele M, Seegerer P, et al. Scoring of tumor-infiltrating lymphocytes: from visual estimation to machine learning. Semin Cancer Biol. 2018;52(Pt 2):151–7. https://doi.org/10.1016/j.semcancer.2018.07.001 .

Lopez R, Regier J, Cole MB, Jordan MI, Yosef N. Deep generative modeling for single-cell transcriptomics. Nat Methods. 2018;15:1053–8.

Amodio M, van Dijk D, Srinivasan K, Chen WS, Mohsen H, Moon KR, et al. Exploring single-cell data with deep multitasking neural networks. Nat Methods. 2019;16:1139–45.

Deng Y, Bao F, Dai Q, Wu LF, Altschuler SJ. Scalable analysis of cell-type composition from single-cell transcriptomics using deep recurrent learning. Nat Methods. 2019;16:311–4.

Fan J, Slowikowski K, Zhang F. Single-cell transcriptomics in cancer: computational challenges and opportunities. Exp Mol Med. 2020;52:1452–65.

Ståhl PL, Salmén F, Vickovic S, Lundmark A, Navarro JF, Magnusson J, et al. Visualization and analysis of gene expression in tissue sections by spatial transcriptomics. Science. 2016;353:78–82.

Gerlinger M, Rowan AJ, Horswell S, Larkin J, Endesfelder D, Gronroos E, et al. Intratumor heterogeneity and branched evolution revealed by multiregion sequencing. N Engl J Med. 2012;366:883–92.

Yoosuf N, Navarro JF, Salmén F, Ståhl PL, Daub CO. Identification and transfer of spatial transcriptomics signatures for cancer diagnosis. Breast Cancer Res. 2020;22:6.

Vivarelli S, Salemi R, Candido S, Falzone L, Santagati M, Stefani S, et al. Gut microbiota and cancer: from pathogenesis to therapy. Cancers. 2019;11:38.

Cammarota G, Ianiro G, Ahern A, Carbone C, Temko A, Claesson MJ, et al. Gut microbiome, big data and machine learning to promote precision medicine for cancer. Nat Rev Gastroenterol Hepatol. 2020;17:635–48.

Relling MV, Evans WE. Pharmacogenomics in the clinic. Nature. 2015;526(7573):343–50. https://doi.org/10.1038/nature15817 .

Adam G, Rampášek L, Safikhani Z, Smirnov P, Haibe-Kains B, Goldenberg A. Machine learning approaches to drug response prediction: challenges and recent progress. NPJ Precis Oncol. 2020;4(1):19. https://doi.org/10.1038/s41698-020-0122-1 .

Kalinin AA, Higgins GA, Reamaroon N, Soroushmehr S, Allyn-Feuer A, Dinov ID, et al. Deep learning in pharmacogenomics: from gene regulation to patient stratification. Pharmacogenomics. 2018;19:629–50.

Chiu Y-C, Chen H-IH, Gorthi A, Mostavi M, Zheng S, Huang Y, et al. Deep learning of pharmacogenomics resources: moving towards precision oncology. Brief Bioinform. 2020;21(6):2066–83.

Rampášek L, Hidru D, Smirnov P, Haibe-Kains B, Goldenberg A. Dr.VAE: improving drug response prediction via modeling of drug perturbation effects. Schwartz R, editor. Bioinformatics. 2019;35:3743–3751.

Robinson MD, McCarthy DJ, Smyth GK. edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics. 2010;26:139–40.

Kompa B, Snoek J, Beam AL. Second opinion needed: communicating uncertainty in medical machine learning. NPJ Digit Med. 2021;4:4.

Wang H, Yeung D-Y. Towards Bayesian deep learning: a framework and some existing methods. IEEE Trans Knowl Data Eng. 2016;28:3395–408.

Danaee P, Ghaeini R, Hendrix DA. A deep learning approach for cancer detection and relevant gene identification. Biocomput 2017 [Internet]. Kohala Coast: World Scientific; 2017. p. 219–29. [cited 2021 May 10]. Available from: http://www.worldscientific.com/doi/abs/10.1142/9789813207813_0022

Khairnar P, Thiagarajan P, Ghosh S. A modified Bayesian convolutional neural network for breast histopathology image classification and uncertainty quantification. ArXiv201012575 Cs Eess [Internet]. 2020; [cited 2021 May 10]; Available from: http://arxiv.org/abs/2010.12575 .

Abdar M, Samami M, Mahmoodabad SD, Doan T, Mazoure B, Hashemifesharaki R, et al. Uncertainty quantification in skin cancer classification using three-way decision-based Bayesian deep learning. Comput Biol Med. 2021;135:104418.

Download references

Acknowledgements

Khoa Tran was the recipient of the Maureen and Barry Stevenson PhD Scholarship, we are grateful to Maureen Stevenson for her support.

We would also like to thank Rebecca Johnston for her scientific advice and intellectual discussions.

Nicola Waddell is supported by a National Health and Medical Research Council of Australia (NHMRC) Senior Research Fellowship (APP1139071).

Author information

Authors and affiliations.

Department of Genetics and Computational Biology, QIMR Berghofer Medical Research Institute, Brisbane, 4006, Australia

Khoa A. Tran, Olga Kondrashova, John V. Pearson & Nicola Waddell

School of Biomedical Sciences, Faculty of Health, Queensland University of Technology (QUT), Brisbane, 4059, Australia

Khoa A. Tran & Elizabeth D. Williams

Australian Prostate Cancer Research Centre - Queensland (APCRC-Q) and Queensland Bladder Cancer Initiative (QBCI), Brisbane, 4102, Australia

Elizabeth D. Williams

Faculty of Engineering, Queensland University of Technology (QUT), Brisbane, 4000, Australia

Andrew Bradley

You can also search for this author in PubMed   Google Scholar

Contributions

Khoa Tran, Olga Kondrashova and Nicola Waddell co-wrote the paper. Andrew Bradley, Elizabeth Williams and John Pearson reviewed and edited the paper. The authors read and approved the final manuscript.

Corresponding author

Correspondence to Nicola Waddell .

Ethics declarations

Ethics approval and consent to participate, consent for publication, competing interests.

John V Pearson and Nicola Waddell are co-founders and Board members of genomiQa. The remaining authors declare that they have no competing interests.

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . The Creative Commons Public Domain Dedication waiver ( http://creativecommons.org/publicdomain/zero/1.0/ ) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Cite this article.

Tran, K.A., Kondrashova, O., Bradley, A. et al. Deep learning in cancer diagnosis, prognosis and treatment selection. Genome Med 13 , 152 (2021). https://doi.org/10.1186/s13073-021-00968-x

Download citation

Received : 13 December 2020

Accepted : 12 September 2021

Published : 27 September 2021

DOI : https://doi.org/10.1186/s13073-021-00968-x

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Multi-modal learning
• Explainability
• Cancer genomics
• Molecular subtypes
• Pharmacogenomics

Genome Medicine

ISSN: 1756-994X

• Submission enquiries: [email protected]
• General enquiries: [email protected]

• Open access
• Published: 24 July 2023

Deep learning applications to breast cancer detection by magnetic resonance imaging: a literature review

• Richard Adam 1 ,
• Kevin Dell’Aquila 1 ,
• Laura Hodges 1 ,
• Takouhie Maldjian 1 &
• Tim Q. Duong 1  

Breast Cancer Research volume  25 , Article number:  87 ( 2023 ) Cite this article

11k Accesses

14 Citations

1 Altmetric

Metrics details

Deep learning analysis of radiological images has the potential to improve diagnostic accuracy of breast cancer, ultimately leading to better patient outcomes. This paper systematically reviewed the current literature on deep learning detection of breast cancer based on magnetic resonance imaging (MRI). The literature search was performed from 2015 to Dec 31, 2022, using Pubmed. Other database included Semantic Scholar, ACM Digital Library, Google search, Google Scholar, and pre-print depositories (such as Research Square). Articles that were not deep learning (such as texture analysis) were excluded. PRISMA guidelines for reporting were used. We analyzed different deep learning algorithms, methods of analysis, experimental design, MRI image types, types of ground truths, sample sizes, numbers of benign and malignant lesions, and performance in the literature. We discussed lessons learned, challenges to broad deployment in clinical practice and suggested future research directions.

Breast cancer is the most common cancer and the second leading cause of cancer death in women. One in eight American women (13%) will be diagnosed with breast cancer in their lifetime, and one in 39 women (3%) will die from breast cancer (American Cancer Society Statistics, 2023). The American Cancer Society recommends yearly screening mammography for early detection of breast cancer for women, which may begin at age 40 [ 1 ]. About 2%–5% of women in the general population in the US have a lifetime risk of breast cancer of 20% or higher [ 1 ], although it can vary depending on the population being studied and the risk assessment method used. The ACS recommends yearly breast magnetic resonance imaging (MRI) in addition to mammography for women with 20–25% or greater lifetime risk [ 1 ]. Early detection and treatment are likely to result in better patient outcomes.

MRI is generally more sensitive and offers more detailed pathophysiological information but is less cost effective compared to mammography for population-based screening [ 2 , 3 ]. Breast MRI utilizes high-powered magnets and radio waves to generate 3D images. Cancer yield from MRI alone averages 22 cancers for every 1000 women screened, a rate of cancer detection roughly 10 times that achieved with screening mammography in average-risk women, and roughly twice the yield achieved with screening mammography in high-risk women [ 4 ]. Many recent studies have established contrast-enhanced breast MRI as a screening modality for women with a hereditary or familial increased risk for the development of breast cancer [ 5 ].

Interpretation of breast cancer on MRI relies on the expertise of radiologists. The growing demand for breast MRI and the shortage of radiologists has resulted in increased workload for radiologists [ 6 , 7 ], leading to long wait times and delays in diagnosis [ 8 , 9 ]. Machine learning methods show promise in assisting radiologists, in improving accuracy with the interpretation of breast MRI images and supporting clinical decision-making and improving patient outcomes [ 10 , 11 ]. By analyzing large datasets of MRIs, machine learning algorithms can learn to identify and classify suspicious areas, potentially reducing the number of false positives and false negatives [ 11 , 12 ] and thus improving diagnostic accuracy. A few studies have shown that machine learning can outperform radiologists in detecting breast cancer on MRIs [ 13 ]. Machine learning could also help to prioritize worklists in a radiology department.

In recent years, deep learning (DL) methods have revolutionized the field of computer vision with wide range of applications, from image classification and object detection to semantic segmentation and medical image analysis [ 14 ]. Deep learning is superior to traditional machine learning because of its ability to learn from unstructured or unlabeled data [ 14 ]. Unlike traditional machine algorithms which require time-consuming data labeling, deep learning algorithms are more flexible and adaptable as they can learn from data that are not labeled or structured [ 15 ]. There have been a few reviews on deep learning breast cancer detection. Oza et al. reviewed detection and classification on mammography [ 16 ]. Saba et al. [ 17 ] presented a compendium of state-of-the-art techniques for diagnosing breast cancers and other cancers. Hu et al. [ 18 ] provided a broad overview on the research and development of artificial intelligence systems for clinical breast cancer image analysis, discussing the clinical role of artificial intelligence in risk assessment, detection, diagnosis, prognosis, and treatment response assessment. Mahoro et al. [ 10 ] reviewed the applications of deep learning to breast cancer diagnosis across multiple imaging modalities. Sechopoulos et al. [ 19 ] discussed the advances of AI in the realm of mammography and digital tomosynthesis. AI-based workflows integrating multiple datastreams, including breast imaging, can support clinical decision-making and help facilitate personalized medicine [ 20 ]. To our knowledge, there is currently no review that systematically compares different deep learning studies of breast cancer detection using MRI. Such a review would be important because it could help to delineate the path forward.

Figure  1 shows a graphic representation of a deep learning workflow. The input layer represents the breast cancer image that serves as input to the CNN. The multiple convolutional layers are stacked on top of the input layer. Each convolutional layer applies filters or kernels to extract specific features from the input image. These filters learn to detect patterns such as edges, textures, or other relevant features related to breast cancer. After each convolutional layer, activation functions like rectified linear unit (ReLU) are typically applied to introduce nonlinearity into the network. Following some of the convolutional layers, pooling layers are used to downsample the spatial dimensions of the feature maps. Common pooling techniques include max-pooling or average pooling. Pooling helps reduce the computational complexity and extract the most salient features. After the convolutional and pooling layers, fully connected layers are employed. These layers connect all the neurons from the previous layers to the subsequent layers. Fully connected layers enable the network to learn complex relationships between features. The final layer is the output layer, which provides the classification or prediction. In the case of breast cancer detection, it might output the probability or prediction of malignancy or benignity.

The input layer represents the breast cancer image that serves as input to the CNN. The multiple convolutional layers are stacked on top of the input layer. Pooling layers are used to downsample the spatial dimensions of the feature maps. Fully connected layers are then employed to connect all the neurons from the previous layers to the subsequent layers. The final layer is the output layer, which provides the classification

The goal of this study was to review the current literature on deep learning detection of breast cancer using breast MRI. We included literature in which DL was used for both primary screening setting and as a supplemental detection tool. We compared different deep learning algorithms, methods of analysis, types of ground truths, sample size, numbers of benign and malignant lesions, MRI image types, and performance indices, among others. We also discussed lessons learned, challenges of deployment in clinical practice and suggested future research directions.

Materials and methods

No ethics committee approval was required for this review.

Search strategy and eligibility criteria

PRISMA guidelines for reporting were adopted in our systematic review. The literature search was performed from 2017 to Dec 31, 2022, using the following key words: “breast MRI,” “breast magnetic resonance imaging,” “deep learning,” “breast cancer detection,” and “breast cancer screening.” The database included Pubmed, Semantic Scholar, ACM Digital Library, Google search, Google Scholar, and pre-print depositories (such as Research Square). We noted that many of the computing or machine learning journals were found on sites other than Pubmed. Some were full-length peer-reviewed conference papers, in contrast with small conference abstracts. Articles that were not deep learning (such as texture analysis) were excluded. Only original articles written in English were selected. Figure  2 shows the flowchart demonstrating how articles were included and excluded for our review. The search and initial screening for eligibility were performed by RA and independently verified by KD and/or TD. This study did not review DL prediction of neoadjuvant chemotherapy which has recently been reviewed [ 21 ].

PRISMA selection flowchart

Pubmed search yielded 59 articles, of which 22 were review articles, 30 were not related to breast cancer detection on MRI, and two had unclear/unconventional methodologies. Five articles were found in Pubmed search after exclusion (Fig.  2 ). In addition, 13 articles were found on different databases outside of Pubmed, because many computing and machine learning journals were not indexed by Pubmed. A total of 18 articles were included in our study (Table 1 ). Two of the studies stated that the patient populations were moderate/high risk [ 22 , 23 ] or high risk [ 23 ], while the remaining papers do not state whether the dataset was from screening or supplemental MRI.

In this review, we first summarized individual papers and followed by generalization of lessons learned. We then discussed challenges of deployment in the clinics and suggested future research directions.

Summary of individual papers

Adachi et al. [ 13 ] performed a retrospective study using RetinaNet as a CNN architecture to analyze and detect breast cancer in MIPs of DCE fat-suppressed MRI images. Images of breast lesions were annotated with a rectangular region-of-interest (ROI) and labeled as “benign” or “malignant” by an experienced breast radiologist. The AUCs, sensitivities, and specificities of four readers were also evaluated as well as those of readers combined with CNN. RetinaNet alone had a higher area under the curve (AUC) and sensitivity (0.925 and 0.926, respectively) than any of the readers. In two cases, the AI system misdiagnosed normal breast as malignancy, which may be the result of variations in normal breast tissue. Invasive ductal carcinoma near the axilla was missed by AI, possibly due to confusion for normal axillary lymph node. Wider variety of data and larger datasets for training could alleviate these problems.

Antropova et al. [ 24 ] compared MIP derived from the second post-contrast subtraction T1-weighted image to the central slice of the second post-contrast image with and without subtraction. The ground truth was ROIs based on radiology assessment with biopsy-proven malignancy. MIP images showed the highest AUC. Feature extraction and classifier training for each slice for DCE-MRI sequences, with slices in the hundreds, would have been computationally expensive at the time. MIP images, in widespread use clinically, contain enhancement information throughout the tumor volume. MIP images, which represent a volume data, avoid using a plethora of slices, and are, therefore, faster and computationally less intensive and less expensive. MIP (AUC = 0.88) outperformed one-slice DCE image, and subtracted DCE image (AUC = 0.83) outperformed single-slice DCE image (AUC = 0.80). The subtracted DCE image is derived from 2 timepoints, the pre-contrast image subtracted from the post-contrast image, which produces a higher AUC. Using multiple slices and/or multiple timepoints could further increase the AUC with DCE images, possibly even exceeding that of the MIP image (0.88). This would be an area for further exploration.

Ayatollahi et al. [ 22 ] performed a retrospective study using 3D RetinaNet as a CNN architecture to analyze and detect breast cancer in ultrafast TWIST DCE-MRI images. They used 572 images (365 malignant and 207 benign) taken from 462 patients. Bounding boxes drawn around the lesion in the images were used as ground truth. They found a detection rate of 0.90 and a sensitivity of 0.95 with tenfold cross validation.

Feng et al. [ 23 ] implemented the Knowledge-Driven Feature Learning and Integration model (KFLI) using DWI and DCE-MRI data from 100 high-risk female patients with 32 benign and 68 malignant lesions, segmented by two experienced radiologists. They reported 0.85 accuracy. The model formulated a sequence division module and adaptive weighting module. The sequence division module based on lesion characteristics is proposed for feature learning, and the adaptive weighting module proposed is used for automatic feature integration while improving the performance of cooperative diagnosis. This model provides the contribution of sub-sequences and guides the radiologists to focus on characteristic-related sequences with high contribution to lesion diagnosis. This can save time for the radiologists and helps them to better understand the output results of the deep networks. As such, it can extract sufficient and effective features from each sub-sequence for a comprehensive diagnosis of breast cancer. This model is a deep network and domain knowledge ensemble that achieved high sensitivity, specificity, and accuracy.

Fujioka et al. [ 25 ] used 3D MIP projection from early phase (1–2 min) of dynamic contrast-enhanced axial fat-suppressed DCE mages, with performance of CNN models compared to two human readers (Reader 1 = breast surgeon with 5 years of experience and Reader 2 = radiologist with 20 years of experience) in distinguishing benign from malignant lesions. The highest AUC achieved with deep learning was with InceptionResNetV2 CNN model, at 0.895. Mean AUC across the different CNN models was 0.830, and range was 0.750–0.895, performing comparably to human readers. False-positive masses tended to be relatively large with fast pattern of strong enhancement, and false-negative masses tended to be relatively small with medium to slow pattern of enhancement. One false positive and one false negative for non-mass enhancing lesion that was observed were also incorrectly diagnosed by the human readers. The main limitation of their study was small sample size.

Haarburger et al. [ 26 ] performed an analysis of 3D whole-volume images on a larger cohort ( N  = 408 patients), yielding an AUC of up to 0.89 and accuracy of 0.81, further establishing the feasibility of using 3D DCE whole images. Their method involved feeding DCE images from 5 timepoints (before contrast and 4 times post-contrast) and T2-weighted images to the algorithms. The multicurriculum ensemble consisted of a 3D CNN that generates feature maps and a classification component that performs classification based on the aggregated feature maps made by the previous components. AUC range of 0.50–0.89 was produced depending on the CNN models used. Multiscale curriculum training improved simple 3D ResNet18 from an AUC of 0.50 to an AUC of 0.89 (ResNet18 curriculum). A radiologist with 2 years of experience demonstrated AUC of 0.93 and accuracy of 0.93. An advantage of the multicurriculum ensemble is the elimination of the need for pixelwise segmentation for individual lesions, as only coarse localization coordinates for Stage 1 training (performed in 3D in this case) and one global label per breast for Stage 2 training is needed, where Stage 2 involved predictions of whole images in 3D in this study. The high performance of this model can be attributed to the high amount of context and global information provided. Their 3D data use whole breast volumes without time-consuming and cost prohibitive lesion segmentation. A major drawback of 3D images is the requirement of more RAM and many patients required to train the model.

Herent et al. [ 27 ] used T1-weighted fat-suppressed post-contrast MRI in a CNN model that detected and then characterized lesions ( N  = 335). Lesion characterization consisted of diagnosing malignancy and lesion classification. Their model, therefore, performed three tasks and thereby was a multitask technique, which limits overfitting. ResNET50 neural network performed feature extraction from images, and images were processed by the algorithm’s attention block which learned to detect abnormalities. Images were fed into a second branch where features were averaged over the selected regions, then fitted to a logistic regression to produce the output. On an independent test set of 168 images, a weighted mean AUC of 0.816 was achieved. The training dataset consisted of 17 different histopathologies, of which most were represented as very small percentages of the whole dataset of 335. Several of the listed lesion types represented less than 1% of the training dataset. This leads to the problem of overfitting. The authors mention that validation of the algorithm by applying it to 3D images in an independent dataset, rather than using the single 2D images as they did, would show if the model is generalizable. The authors state that training on larger databases and with multiparametric MRI would likely increase accuracy. This study shows good performance of a supervised attention model with deep learning for breast MRI.

Hu et al. [ 28 ] used multiparametric MR images (DCE-MRI sequence and a T2-weighted MRI sequence) in a CNN model including 616 patients with 927 unique breast lesions, 728 of which were malignant. A pre-trained CNN extracted features from both DCE and T2w sequences depicting lesions that were classified as benign or malignant by support vector machine classifiers. Sequences were integrated at different levels using image fusion, feature fusion, and classifier fusion. Feature fusion from multiparametric sequences outperformed DCE-MRI alone. The feature fusion model had an AUC of 0.87, sensitivity of 0.78, and specificity of 0.79. CNN models that used separate T2w and DCE images into combined RBG images or aggregates of the probability of malignancy output from DCE and T2w classifiers both did not perform significantly better than the CNN model using DCE-alone. Although other studies have demonstrated that single-sequence MRI is sufficient for high CNN performance, this study demonstrates that multiparametric MRI (as fusion of features from DCE-MRI and T2-weighted MRI) offers enough information to outperform single-sequence MRI.

Li et al. [ 29 ] used 3D CNNs in DCE-MR images to differentiate between benign and malignant tumors from 143 patients. In 2D and 3D DCE-MRI, a region-of-interest (ROI) and volume-of-interest (VOI) were segmented, and enhancement ratios for each MR series were calculated. The AUC value of 0.801 for the 3D CNN was higher than the value of 0.739 for 2D CNN. Furthermore, the 3D CNN achieved higher accuracy, sensitivity, and specificity values of 0.781, 0.744, and 0.823, respectively. The DCE-MRI enhancement maps had higher accuracy by using more information to diagnose breast cancer. The high values demonstrate that 3D CNN in breast cancer MR imaging can be used for the detection of breast cancer and reduce manual feature extraction.

Liu et al. [ 30 ] used CNN to analyze and detect breast cancer on T1 DCE-MRI images from 438 patients, 131 from I-SPY clinical trials and 307 from Columbia University. Segmentation was performed through an automated process involving fuzzy C-method after seed points were manually indicated. This study included analysis of commonly excluded image features such as background parenchymal enhancement, slice images of breast MRI, and axilla/axillary lymph node involvement. The methods also minimized annotations done at pixel level, to maximize automation of visual interpretation. These objectives increased efficiency, decreased subjective bias, and allowed for complete evaluation of the whole image. Obtaining images with multiple timepoints from multiple institutions made the algorithm more generalizable. The CNN model achieved AUC of 0.92, accuracy of 0.94, sensitivity of 0.74, and specificity of 0.95.

Marrone et al. [ 31 ] used CNN to evaluate 42 malignant and 25 benign lesions in 42 women. ROIs were obtained by an experienced radiologist, and manual segmentation was performed. Accuracy of up to 0.76 was achieved. AUC as high as 0.76 was seen on pre-trained AlexNet versus 0.73 on fine-tuning of pre-trained AlexNet where the last trained layers were replaced by untrained layers. The latter method could allow reduced number of training images needed. The training from scratch AlexNet model is accomplished when AlexNet pre-trained on the ImageNet database is used to extract a feature vector from the last internal CNN layer, and a new supervised training is employed, which yielded the lowest AUC of 0.68 and accuracy of 0.55.

Rasti et al. [ 32 ] analyzed DCE-MRI subtraction images from MRI studies ( N  = 112) using a multi-ensemble CNN (ME-CNN) functioning as a CAD system to distinguish benign from malignant masses, producing 0.96 accuracy with their method. The ME-CNN is a modular and image-based ensemble, which can stochastically partition the high-dimensional image space through simultaneous and competitive learning of its modules. It also has the advantages of fast execution time in both training and testing and a compact structure with a small number of free parameters. Among several promising directions, one could extend the ME-CNN approach to the pre-processing stage, by combining ME-CNN with recent advances in fully autonomous CNNs for semantic segmentation.

Truhn et al. [ 33 ] used T2-weighted images with one pre-contrast and four post-contrast DCE images in 447 patients with 1294 enhancing lesions (787 malignant and 507 benign) manually segmented by a breast radiologist. Deep learning with CNN demonstrated an AUC of 0.88 which was inferior to prospective interpretation by one of the three breast radiologists (7–25 years of experience) reading cases in equal proportion (0.98). When only half of the dataset was used for training ( n  = 647), the AUC was 0.83. The authors speculate that with increased training on a greater number of cases that their model could improve its performance.

Wu et al. [ 34 ] trained a CNN model to analyze and detect lesions from DCE T1-weighted images from 130 patients, 71 of which had malignant lesions and 59 had benign tumors. Fuzzy C-means clustering-based algorithm automatically segmented 3D tumor volumes from DCE images after rectangular region-of-interest were placed by an expert radiologist. An objective of the study was to demonstrate that single-sequence MRI at multiple timepoints provides sufficient information for CNN models to accurately classify lesions.

Yurtakkal et al. [ 35 ] utilized DCE images of 98 benign and 102 malignant lesions, producing 0.98 accuracy, 1.00 sensitivity, and 0.96 specificity. The multi-layer CNN architecture utilized consisted of six groups of convolutional, batch normalization, rectified linear activation function layers, and five max-pooling followed by one dropout layer, one fully connected layer, and one softmax layer.

Zheng et al. [ 36 ] used a dense convolutional long short-term memory (DC-LSTM) on a dataset of lesions obtained through a university hospital ( N  = 72). The method was inspired by DenseNet and built on convolutional LSTM. It first uses a three-layer convolutional LSTM to encode DCE-MRI as sequential data and extract time-intensity information then adds a simplified dense block to reduce the amount of information being processed and improve feature reuse. This lowered the variance and improved accuracy in the results. Compared to a ResNet-50 model trained only on the main task, the combined model of DC-LSTM + ResNet improved the accuracy from 0.625 to 0.847 on the same dataset. Additionally, the authors proposed a latent attributes method to efficiently use the information in diagnostic reports and accelerate the convergence of the network.

Jiejie Zhou et al. [ 37 ] evaluated 133 lesions (91 malignant and 62 benign) using ResNET50, which is similar to ResNET18 used by Truhn et al. [ 33 ] and Haarburger et al . [ 26 ]. Their investigation demonstrated that deep learning produced higher accuracy compared to ROI-based and radiomics-based models in distinguishing between benign and malignant lesions. They compared the metrics resulting from using five different bounding boxes. They found that using the tumor alone and smaller bounding boxes yielded the highest AUC of 0.97–0.99. They also found that the inclusion of a small amount of peritumoral tissue improved accuracy compared to smaller boxes that did not include peritumoral tissue (tumor alone boxes) or larger input boxes (that include tissue more remote from peritumoral tissue), with accuracy of 0.91 in the testing dataset. The tumor microenvironment influences tumor growth, and the tumor itself can alter its microenvironment to become more supportive of tumor growth. Therefore, the immediate peritumoral tissue, which would include the tumor microenvironment, was important in guiding the CNN to accurately differentiate between benign and malignant tumors. This dynamic peritumoral ecosystem can be influenced by the tumor directing heterogeneous cells to aggregate and promote angiogenesis, chronic inflammation, tumor growth, and invasion. Recognizing features displayed by biomarkers of the tumor microenvironment may help to identify and grade the aggressiveness of a lesion. This complex interaction between the tumor and its microenvironment may potentially be a predictor of outcomes as well and should be included in DL models that require segmentation. In DL models using whole images without segmentation of any sort, the peritumoral tissue would already be included, which would preclude the need for precise bounding boxes.

Juan Zhou et al. [ 38 ] used 3D deep learning models to classify and localize malignancy from cases ( N  = 1537) of MRIs. The deep 3D densely connected networks were utilized under image-level supervision (weakly supervised). Since 3D weakly supervised approach was not well studied compared to 2D methods, the purpose of this study was to develop a 3D deep learning model that could identify malignant cancer from benign lesions and could localize the cancer. The model configurations of global average pooling (GAP) and global max-pooling (GMP) that were used both achieved over 0.80 accuracy with AUC of 0.856 (GMP) and 0.858 (GAP) which demonstrated the effectiveness of the 3D DenseNet deep learning method in MRI scans to diagnose breast cancer. The model ensemble achieved AUC of 0.859.

Summary of lessons learned

Most studies were single-center studies, but they came from around the world, with the majority coming from the US, Asia, and Europe. All studies except one [ 33 ] were retrospective studies. The sample size of each study ranged from 42 to 690 patients, generally small for DL analysis. Sample sizes for patients with benign and malignant lesions were comparable and were not skewed toward either normal or malignant lesions, suggesting that these datasets were not from high-risk screening patients because high-risk screening dataset would have consisted of very low (i.e., typically < 5%) positive cases.

Image types

Most studies used private datasets as their image source. ISPY-1 data were the only public dataset noted ( https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=20643859 ). Most studies involved DCE data acquisition, but most analysis include only a single post-contrast MRI. For those that used multiple post-contrast MRI dynamics, most fed each dynamic into each separate independent channel, which does not optimally make use of the relationships between imaging dynamics. Some studies used subtraction of post- and pre-contrast or signal enhancement ratio (SER) [ 24 , 32 , 35 ]. Three studies utilized MIP DCE images to minimize computation cost [ 13 , 24 , 25 ]. However, collapsing images by MIP has drawbacks (i.e., collapse enhanced vascular structures into a single plane may be mistaken as cancer). There were only five studies [ 23 , 26 , 28 , 33 , 36 ] that utilized multiparametric data types (i.e., DCE, T2-weighted, and DWI). Although combining multiple types of MRIs should improve performance, this has not been conclusively demonstrated in practice.

Types of DL architectures

RetinaNet and KFLi are optimized for object detection, while VGGNet, InceptionResNet, and AlexNet are designed for image classification (see review [ 16 , 17 , 39 ]). LSTM is used for time-series modeling. DenseNet, on the other hand, can be used for a wide range of tasks, including image classification, object detection, and semantic segmentation. Ensemble methods, which combine multiple models, are useful for boosting the overall performance of a system. U-Net and R-Net are specialized deep learning models for semantic segmentation tasks in medical image analysis. U-Net uses an encoder–decoder architecture to segment images into multiple classes, while R-Net is a residual network that improves the accuracy and efficiency of the segmentation task.

The most used algorithm is CNN or CNN-based. There is no consensus that certain algorithms are better than others. Given the fact that different algorithms were tested on different datasets, it is not possible to conclude that a particular DL architecture performs better than others. Careful comparison of multiple algorithms on the same datasets is needed. Thus, we only discussed potential advantages and disadvantages of each DL architecture. Performance indices could be misleading.

Although each model has its own unique architecture and design principles, most of the above-mentioned methods utilized convolutional layers, pooling layers, activation functions, and regularization techniques (such as dropout and batch normalization) for model optimization. Additionally, the use of pre-trained models and transfer learning has become increasingly popular, allowing leverage of knowledge learned from large datasets such as ImageNet to improve the performance of their models on smaller, specialized datasets. However, the literature on transfer learning in breast cancer MRI detection is limited. A relatively new deep learning method known as transformer has found exciting applications in medical imaging [ 40 , 41 ].

Ground truths

Ground truths were either based on pathology (i.e., benign versus malignant cancer), radiology reports, radiologist annotation (ROI contoured on images), or a bounding box, with reference to pathology or clinical follow-up (i.e., absence of a positive clinical diagnosis). While the gold standard is pathology, imaging or clinical follow-up without adverse change over a prescribed period has been used as empiric evidence of non-malignancy. This is an acceptable form of ground truth.

Only four out of 18 studies provided heatmaps of the regions that the DL algorithms consider important. Heatmaps enable data to be presented visually in color showing whether the area of activity makes sense anatomically or if it is artifactual (i.e., biopsy clip, motion artifact, or outside of the breast). Heatmaps are important for interpretability and explainability of DL outputs.

Performance

All studies include some performance indices, and most include AUC, accuracy, sensitivity, and specificity. AUC ranged from 0.5 to 1.0, with the majority around 0.8–0.9. Other metrics also varied over a wide range. DL training methods varied, and they included leave-one-out method, hold-out method, and splitting the dataset (such as 80%/20% training/testing) with cross validation. Most studies utilized five- or tenfold cross validation for performance evaluation but some used a single hold-out method, and some did not include cross validation. Cross validation is important to avoid unintentional skewing of data due to partition for training and testing. Different training methods could affect performance. Interpretation of these metrics needs to be made with caution as there could be study reporting bias, small sample size, and overfitting, among others. High-performance indices of the DL algorithm performance are necessary for adoption in clinical use. However, good performance indices alone are not sufficient. Other measures such as heatmaps and experience to gain trust are needed for widespread clinical adoption of DL algorithms.

DL detection of axillary lymph node involvement

Accurate assessment of the axillary lymph node involvement in breast cancer patients is also essential for prognosis and treatment planning [ 42 , 43 ]. Current radiological staging of nodal metastasis has poor accuracy. DL detection of lymph node involvement is challenging because of their small sizes and difficulty in getting ground truths. Only a few studies have reported the use of DL to detect lymph node involvement [ 44 , 45 , 46 ].

Challenges for DL to achieve routine clinical applications

Although deep learning is a promising tool in the diagnosis of breast cancer, there are several challenges that need to be addressed before routine clinical applications can be broadly realized.

Data availability: One of the major challenges in medical image diagnosis (and breast cancer MRI in particular) is the availability of large, diverse, and well-annotated datasets. Deep learning models require a large amount of high-quality data to learn from, but, in many cases, medical datasets are small and imbalanced. In medical image diagnosis, it is important to have high-quality annotations of images, which can be time-consuming and costly to obtain. Annotating medical images requires specialized expertise, and there may be inconsistencies between different experts. This can lead to challenges in building accurate and generalizable models. Medical image datasets can lack diversity, which can lead to biased models. For example, a model trained on images with inadequate representation of racial or ethnicity subgroups may not be broadly generalizable. Private medical datasets obtained from one institution could be non-representative of certain racial or ethnic subgroups and, therefore, may not be generalizable. Publicly available data are unfortunately limited, one of which can be found on cancerimagingarchive.net. Collaborative learning facilitating training of DL models by sharing data without breaching privacy can be accomplished with federated learning [ 47 ].

Interpretability , explainability, and generalizability [ 48 ]: Deep learning models are often seen as “black boxes” that can be difficult to interpret. This is especially problematic in medical image diagnosis, where it is important to understand why a particular diagnosis is made. Recent research has focused on developing methods to explain the decision-making process of deep learning models, such as using attention mechanisms or generating heatmaps to highlight relevant regions in the MRI image. While efforts have been made to develop methods to explain the decision-making process of deep learning models, the explainability of these models is still limited [ 49 ]. This can make it difficult for clinicians to understand the model's decision and to trust the model. Deep learning models may perform well on the datasets on which they were trained but may not generalize well to new datasets or to patients with different characteristics. This can lead to challenges in deploying the model in a real-world setting.

Ethical concerns: Deep learning models can be used to make life-or-death decisions, such as the diagnosis of cancer. This raises ethical concerns about the safety, responsibility, privacy, fairness, and transparency of these models [ 50 ]. There are also social implications (including but not limited to equity) of using artificial intelligence in health care. This needs to be addressed as we develop more and more powerful DL algorithms.

Perspectives and conclusions

Artificial intelligence has the potential to revolutionize breast cancer screening and diagnosis, helping radiologists to be more efficient and more accurate, ultimately leading to better patient outcomes. It can also help to reduce the need for biopsy or unnecessary testing and treatment. However, some challenges exist that preclude broad deployment in clinical practice to date. There need to be large, diverse, and well-annotated images that are readily available for research. Deep learning results need to be more accurate, interpretable, explainable, and generalizable. A future research direction includes incorporation of other clinical data and risk factors into the model, such as age, family history, or genetic mutations, to improve diagnostic accuracy and enable personalized medicine. Another direction is to assess the impact of deep learning on health outcomes to enable more investment in hospital administrators and other stakeholders. Finally, it is important to address the ethical, legal, and social implications of using artificial intelligence.

Availability of data and materials

Not applicable.

Saslow D, Boetes C, Burke W, Harms S, Leach MO, Lehman CD, Morris E, Pisano E, Schnall M, Sener S, et al. American Cancer Society guidelines for breast screening with MRI as an adjunct to mammography. CA Cancer J Clin. 2007;57(2):75–89.

Article   PubMed   Google Scholar  

Feig S. Comparison of costs and benefits of breast cancer screening with mammography, ultrasonography, and MRI. Obstet Gynecol Clin North Am. 2011;38(1):179–96.

Kumar NA, Schnall MD. MR imaging: its current and potential utility in the diagnosis and management of breast cancer. Magn Reson Imaging Clin N Am. 2000;8(4):715–28.

Article   CAS   PubMed   Google Scholar  

Lehman CD, Smith RA. The role of MRI in breast cancer screening. J Natl Compr Canc Netw. 2009;7(10):1109–15.

Mann RM, Kuhl CK, Moy L. Contrast-enhanced MRI for breast cancer screening. J Magn Reson Imaging. 2019;50(2):377–90.

Article   PubMed   PubMed Central   Google Scholar  

Batchu S, Liu F, Amireh A, Waller J, Umair M. A review of applications of machine learning in mammography and future challenges. Oncology. 2021;99(8):483–90.

Wuni AR, Botwe BO, Akudjedu TN. Impact of artificial intelligence on clinical radiography practice: futuristic prospects in a low resource setting. Radiography (Lond). 2021;27(Suppl 1):S69–73.

Skegg D, Paul C, Benson-Cooper D, Chetwynd J, Clarke A, Fitzgerald N, Gray A, St George I, Simpson A. Mammographic screening for breast cancer: prospects for New Zealand. N Z Med J. 1988;101(852):531–3.

CAS   PubMed   Google Scholar  

Wood DA, Kafiabadi S, Busaidi AA, Guilhem E, Montvila A, Lynch J, Townend M, Agarwal S, Mazumder A, Barker GJ, et al. Deep learning models for triaging hospital head MRI examinations. Med Image Anal. 2022;78: 102391.

Mahoro E, Akhloufi MA. Applying deep learning for breast cancer detection in radiology. Curr Oncol. 2022;29(11):8767–93.

Deo RC, Nallamothu BK. Learning about machine learning: the promise and pitfalls of big data and the electronic health record. Circ Cardiovasc Qual Outcomes. 2016;9(6):618–20.

Meyer-Base A, Morra L, Tahmassebi A, Lobbes M, Meyer-Base U, Pinker K. AI-enhanced diagnosis of challenging lesions in breast mri: a methodology and application primer. J Magn Reson Imaging. 2021;54(3):686–702.

Adachi M, Fujioka T, Mori M, Kubota K, Kikuchi Y, Xiaotong W, Oyama J, Kimura K, Oda G, Nakagawa T, et al. Detection and diagnosis of breast cancer using artificial intelligence based assessment of maximum intensity projection dynamic contrast-enhanced magnetic resonance images. Diagnostics (Basel). 2020;10(5):330.

LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521(7553):436–44.

Koh DM, Papanikolaou N, Bick U, Illing R, Kahn CE Jr, Kalpathi-Cramer J, Matos C, Marti-Bonmati L, Miles A, Mun SK, et al. Artificial intelligence and machine learning in cancer imaging. Commun Med (Lond). 2022;2:133.

Oza P, Sharma P, Patel S, Bruno A. A bottom-up review of image analysis methods for suspicious region detection in mammograms. J Imaging. 2021;7(9):190.

Saba T. Recent advancement in cancer detection using machine learning: systematic survey of decades, comparisons and challenges. J Infect Public Health. 2020;13(9):1274–89.

Hu Q, Giger ML. Clinical artificial intelligence applications: breast imaging. Radiol Clin North Am. 2021;59(6):1027–43.

Sechopoulos I, Teuwen J, Mann R. Artificial intelligence for breast cancer detection in mammography and digital breast tomosynthesis: state of the art. Semin Cancer Biol. 2021;72:214–25.

Sheth D, Giger ML. Artificial intelligence in the interpretation of breast cancer on MRI. J Magn Reson Imaging. 2020;51(5):1310–24.

Khan N, Adam R, Huang P, Maldjian T, Duong TQ. Deep learning prediction of pathologic complete response in breast cancer using MRI and other clinical data: a systematic review. Tomography. 2022;8(6):2784–95.

Ayatollahi F, Shokouhi SB, Mann RM, Teuwen J. Automatic breast lesion detection in ultrafast DCE-MRI using deep learning. Med Phys. 2021;48(10):5897–907.

Feng H, Cao J, Wang H, Xie Y, Yang D, Feng J, Chen B. A knowledge-driven feature learning and integration method for breast cancer diagnosis on multi-sequence MRI. Magn Reson Imaging. 2020;69:40–8.

Antropova N, Abe H, Giger ML. Use of clinical MRI maximum intensity projections for improved breast lesion classification with deep convolutional neural networks. J Med Imaging (Bellingham). 2018;5(1): 014503.

PubMed   Google Scholar  

Fujioka T, Yashima Y, Oyama J, Mori M, Kubota K, Katsuta L, Kimura K, Yamaga E, Oda G, Nakagawa T, et al. Deep-learning approach with convolutional neural network for classification of maximum intensity projections of dynamic contrast-enhanced breast magnetic resonance imaging. Magn Reson Imaging. 2021;75:1–8.

Haarburger C, Baumgartner M, Truhn D, Broeckmann M, Schneider H, Schrading S, Kuhl C, Merhof D: Multi scale curriculum CNN for context-aware breast mri malignancy classification. In: Medical Image Computing and Computer Assisted Intervention—MICCAI ; 2019: 495–503.

Herent P, Schmauch B, Jehanno P, Dehaene O, Saillard C, Balleyguier C, Arfi-Rouche J, Jegou S. Detection and characterization of MRI breast lesions using deep learning. Diagn Interv Imaging. 2019;100(4):219–25.

Hu Q, Whitney HM, Giger ML. A deep learning methodology for improved breast cancer diagnosis using multiparametric MRI. Sci Rep. 2020;10(1):10536.

Article   CAS   PubMed   PubMed Central   Google Scholar  

Li J, Fan M, Zhang J, Li L: Discriminating between benign and malignant breast tumors using 3D convolutional neural network in dynamic contrast enhanced-MR images. In: SPIE Medical Imaging : SPIE; 2017: 10138.

Liu MZ, Swintelski C, Sun S, Siddique M, Desperito E, Jambawalikar S, Ha R. Weakly supervised deep learning approach to breast mri assessment. Acad Radiol. 2022;29(Suppl 1):S166–72.

Marrone S, Piantadosi G, Fusco R, Petrillo A, Sansone M, Sansone C: An investigation of deep learning for lesions malignancy classification in breast DCE-MRI. In: International Conference on Image Analysis and Processing (ICIAP) 2017: 479–489.

Rasti R, Teshnehlab M, Phung SL. Breast cancer diagnosis in DCE-MRI using mixture ensemble of convolutional neural networks. Pattern Recogn. 2017;72:381–90.

Article   Google Scholar  

Truhn D, Schrading S, Haarburger C, Schneider H, Merhof D, Kuhl C. Radiomic versus convolutional neural networks analysis for classification of contrast-enhancing lesions at multiparametric breast MRI. Radiology. 2019;290(2):290–7.

Wu W, Wu J, Dou Y, Rubert N, Wang Y. A deep learning fusion model with evidence-based confidence level analysis for differentiation of malignant and benign breast tumors using dynamic contrast enhanced MRI. Biomed Signal Process Control. 2022;72: 103319.

Yurttakal AH, Erbay H, İkizceli T, Karaçavuş S. Detection of breast cancer via deep convolution neural networks using MRI images. Multimed Tools Appl. 2020;79:15555–73.

Zheng H, Gu Y, Qin Y, Huang X, Yang J, Yang G-Z: Small lesion classification in dynamic contrast enhancement MRI for breast cancer early detection. In: International Conference on Medical Image Computing and Computer-Assisted Intervention; 2018.

Zhou J, Zhang Y, Chang KT, Lee KE, Wang O, Li J, Lin Y, Pan Z, Chang P, Chow D, et al. Diagnosis of benign and malignant breast lesions on DCE-MRI by using radiomics and deep learning with consideration of peritumor tissue. J Magn Reson Imaging. 2020;51(3):798–809.

Zhou J, Luo LY, Dou Q, Chen H, Chen C, Li GJ, Jiang ZF, Heng PA. Weakly supervised 3D deep learning for breast cancer classification and localization of the lesions in MR images. J Magn Reson Imaging. 2019;50(4):1144–51.

Alzubaidi L, Zhang J, Humaidi AJ, Al-Dujaili A, Duan Y, Al-Shamma O, Santamaria J, Fadhel MA, Al-Amidie M, Farhan L. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. J Big Data. 2021;8(1):53.

Li J, Chen J, Tang Y, Wang C, Landman BA, Zhou SK. Transforming medical imaging with transformers? A comparative review of key properties, current progresses, and future perspectives. Med Image Anal. 2023;85: 102762.

Moutik O, Sekkat H, Tigani S, Chehri A, Saadane R, Tchakoucht TA, Paul A. Convolutional neural networks or vision transformers: Who will win the race for action recognitions in visual data? Sensors (Basel). 2023;23(2):734.

Chang JM, Leung JWT, Moy L, Ha SM, Moon WK. Axillary nodal evaluation in breast cancer: state of the art. Radiology. 2020;295(3):500–15.

Zhou P, Wei Y, Chen G, Guo L, Yan D, Wang Y. Axillary lymph node metastasis detection by magnetic resonance imaging in patients with breast cancer: a meta-analysis. Thorac Cancer. 2018;9(8):989–96.

Ren T, Cattell R, Duanmu H, Huang P, Li H, Vanguri R, Liu MZ, Jambawalikar S, Ha R, Wang F, et al. Convolutional neural network detection of axillary lymph node metastasis using standard clinical breast MRI. Clin Breast Cancer. 2020;20(3):e301–8.

Ren T, Lin S, Huang P, Duong TQ. Convolutional neural network of multiparametric MRI accurately detects axillary lymph node metastasis in breast cancer patients with pre neoadjuvant chemotherapy. Clin Breast Cancer. 2022;22(2):170–7.

Golden JA. Deep learning algorithms for detection of lymph node metastases from breast cancer: helping artificial intelligence be seen. JAMA. 2017;318(22):2184–6.

Gupta S, Kumar S, Chang K, Lu C, Singh P, Kalpathy-Cramer J. Collaborative privacy-preserving approaches for distributed deep learning using multi-institutional data. Radiographics. 2023;43(4): e220107.

Miotto R, Wang F, Wang S, Jiang X, Dudley JT. Deep learning for healthcare: review, opportunities and challenges. Brief Bioinform. 2018;19(6):1236–46.

Holzinger A, Langs G, Denk H, Zatloukal K, Muller H. Causability and explainability of artificial intelligence in medicine. Wiley Interdiscip Rev Data Min Knowl Discov. 2019;9(4): e1312.

Smallman M. Multi scale ethics-why we need to consider the ethics of AI in healthcare at different scales. Sci Eng Ethics. 2022;28(6):63.

Download references

Author information

Authors and affiliations.

Department of Radiology, Albert Einstein College of Medicine and the Montefiore Medical Center, 1300 Morris Park Avenue, Bronx, NY, 10461, USA

Richard Adam, Kevin Dell’Aquila, Laura Hodges, Takouhie Maldjian & Tim Q. Duong

You can also search for this author in PubMed   Google Scholar

Contributions

RA performed literature search, analyzed data, and wrote paper. KD performed literature search, analyzed data, and edited paper. LH analyzed literature and edited paper. TM analyzed literature and edited paper. TQD wrote and edited paper, and supervised.

Corresponding author

Correspondence to Tim Q. Duong .

Ethics declarations

Ethics approval and consent to participate, consent for publication, competing interests.

The authors declare that they have no competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . The Creative Commons Public Domain Dedication waiver ( http://creativecommons.org/publicdomain/zero/1.0/ ) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Cite this article.

Adam, R., Dell’Aquila, K., Hodges, L. et al. Deep learning applications to breast cancer detection by magnetic resonance imaging: a literature review. Breast Cancer Res 25 , 87 (2023). https://doi.org/10.1186/s13058-023-01687-4

Download citation

Received : 09 March 2023

Accepted : 11 July 2023

Published : 24 July 2023

DOI : https://doi.org/10.1186/s13058-023-01687-4

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Machine learning
• Artificial intelligence
• Texture feature analysis
• Convolutional neural network
• Dynamic contrast enhancement

Breast Cancer Research

ISSN: 1465-542X

• Submission enquiries: [email protected]

• Open access
• Published: 23 March 2023

A review and comparative study of cancer detection using machine learning: SBERT and SimCSE application

• Mpho Mokoatle 1 ,
• Vukosi Marivate 1 ,
• Darlington Mapiye 2 ,
• Riana Bornman 4 &
• Vanessa. M. Hayes 3 , 4  

BMC Bioinformatics volume  24 , Article number:  112 ( 2023 ) Cite this article

10k Accesses

18 Citations

2 Altmetric

Metrics details

Using visual, biological, and electronic health records data as the sole input source, pretrained convolutional neural networks and conventional machine learning methods have been heavily employed for the identification of various malignancies. Initially, a series of preprocessing steps and image segmentation steps are performed to extract region of interest features from noisy features. Then, the extracted features are applied to several machine learning and deep learning methods for the detection of cancer.

In this work, a review of all the methods that have been applied to develop machine learning algorithms that detect cancer is provided. With more than 100 types of cancer, this study only examines research on the four most common and prevalent cancers worldwide: lung, breast, prostate, and colorectal cancer. Next, by using state-of-the-art sentence transformers namely: SBERT (2019) and the unsupervised SimCSE (2021), this study proposes a new methodology for detecting cancer. This method requires raw DNA sequences of matched tumor/normal pair as the only input. The learnt DNA representations retrieved from SBERT and SimCSE will then be sent to machine learning algorithms (XGBoost, Random Forest, LightGBM, and CNNs) for classification. As far as we are aware, SBERT and SimCSE transformers have not been applied to represent DNA sequences in cancer detection settings.

The XGBoost model, which had the highest overall accuracy of 73 ± 0.13 % using SBERT embeddings and 75 ± 0.12 % using SimCSE embeddings, was the best performing classifier. In light of these findings, it can be concluded that incorporating sentence representations from SimCSE’s sentence transformer only marginally improved the performance of machine learning models.

Peer Review reports

Introduction

Cancer is a disease where some cells in the body grow destructively and may spread to other body organs [ 1 ]. Typically, cells grow and expand through a cell division process to create new cells that can be used to repair old and damaged ones. However, this phenomenon can be interrupted resulting in abnormal cells growing uncontrollably to form tumors that can be malignant (harmful) or benign (harmless) [ 2 , 3 , 4 ].

With the introduction of genomic data that allows physicians and healthcare decision-makers to learn more about their patients and their response to the therapy they provide to them, this has facilitated the use of machine learning and deep learning to solve challenging cancer problems. These kinds of problems involve various tasks such as designing cancer risk-prediction models that try to identify patients that are at a higher risk of developing cancer than the general population, studying the progression of the disease to improve survival rates, and building methods that trace the effectiveness of treatment to improve treatment options [ 5 , 6 , 7 ].

Generally, the first step in analyzing genomic data to address cancer-related problems is selecting a data representation algorithm that will be used to estimate contiguous representations of the data. Examples of such algorithms include Word2vec [ 8 ], GloVe [ 9 ], and fastText [ 10 ]. The more recent and advanced versions of these algorithms are sentence transformers which are used to compute dense vector representations for sentences, paragraphs, and images. Similar texts are found close together in a vector space and dissimilar texts are far apart [ 11 ]. In this work, two such sentence transformers (SBERT and SimCSE) are proposed for detecting cancer in tumor/normal pairs of colorectal cancer patients. In this new approach, the classification algorithm relies on raw DNA sequences as the only input source. Moreover, this work provides a review of the most recent developments in cancers of the human body using machine learning and deep learning methods. While these kinds of similar reviews already exist in the literature, this study solely focuses on work that investigates four cancer types that have high prevalence rates worldwide [ 12 ] (lung, breast, prostate, and colorectal cancer) that have been published in the last five years (2018–2022).

Detection of cancer using machine learning

Lung cancer.

Lung cancer is the type of cancer that begins in the lungs and may spread to other organs in the body. This kind of cancer occurs when malignant cells develop in the tissue of the lung. There are two types of lung cancer: non-small-cell lung cancer (NSCLC) and small-cell lung cancer (SCLC). These cancers develop differently and thus their treatment therapies are different. Smoking (tobacco) is the leading cause of lung cancer. However, non-smokers can also develop lung cancer [ 13 , 14 ].

When it comes to the detection of lung cancer using machine learning (Fig.  1 ), a considerable amount of work has been done, a summary is provided (Table 1 ). Typically, a series of pre-processing steps using statistical methods and pretrained CNNs for feature extraction are carried out from several input sources (mostly images) to delineate the cancer region. Then, the extracted features are fed as input to several machine learning algorithms for classification of various lung cancer tasks such as the detection of malignant lung nodules from benign ones [ 15 , 16 , 17 ], the separation of a set of normalized biological data points into cancerous and non cancerous groups [ 18 ], and a basic comparative analysis of powerful machine learning algorithms for lung cancer detection [ 19 ].

Generalized machine learning framework for lung cancer prediction [ 33 ]

The lowest classification accuracy reported in Table 1 was 74.4% by work in [ 20 ]. In this work, a pretrained CNN model (DenseNet) was used to develop a lung cancer detection model. First, the model was fine-tuned to identify lung nodules from chest X-rays using the ChestX-ray14 dataset [ 21 ]. Second, the model was fine-tuned to identify lung cancer from images in the JSRT (Japanese Society of Radiological Technology) dataset [ 22 ].

The highest classification accuracy of 99.7% for lung cancer classification was reported by work in [ 18 ]. This study developed the Discrete AdaBoost Optimized Ensemble Learning Generalized Neural Network (DAELGNN) framework that uses a set of normalized biological data points to create a neural network that separates normal lung features from non-normal (cancerous) features.

Popular datasets used in lung cancer research using machine learning include the Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI) (LIDC-IDRI) database [ 23 ] initiated by the National Cancer Institute (NCI), and the histopathological images of lung and colon cancer (LC2500) database [ 24 ].

Breast cancer

Breast Cancer is a malignant tumor or growth that develops in the cells of the breast [ 34 ]. Similar to lung cancer, breast cancer also has the ability to metastasize to near by lymph nodes or to other body organs. Towards the end of 2020, there were approximately 7.8 million women who have been diagnosed with breast cancer, making this type of cancer the most prevalent cancer in the world. Risk factors of breast cancer include age, obesity, abuse of alcohol, and family history [ 35 , 36 , 37 ].

Currently, there is no identified prevention procedure for breast cancer. However, maintaining a healthy living habit such as physical exercise and less alcohol intake can reduce the risk of developing breast cancer [ 38 ]. It has also been said that early detection methods that rely on machine learning can improve the prognosis. As such, this type of cancer has been extensively studied using machine learning and deep learning [ 39 , 40 ].

As with lung cancer (Sect.  2.1 ), a great deal of work has been executed in developing breast cancer detection models, a generalized approach that illustrates the process using machine learning is provided (Fig.  2 ).

Generalized machine learning framework for breast cancer prediction [ 45 ]

Several classification problems have been studied that mainly focuses on the detection of breast cancer from thermogram images [ 41 ], handrafted features [ 42 ], mammograms [ 43 ], and whole slide images [ 44 ]. To develop a breast cancer detection model, initially, a pre-processing step is implemented that aims to extract features of interest. Then, the extracted features are provided as input to machine learning models for classification. This framework is implemented by several works such as [ 45 , 46 , 47 , 48 ].

One of the most popular datasets used for breast cancer detection using machine learning is the Wisconsin breast cancer dataset [ 42 ]. This dataset consists of features that describe the characteristics of the cell nuclei that is present in the image such as the diagnosis features (malignant or benign), radius, symmetry, and texture. Studies that used this dataset are [ 49 , 50 ]. In [ 49 ], the authors scaled the Wisconsin breast cancer features to be in the range between 0 and 1, then used a CNN for classification into benign or malignant. As opposed to using a CNN for classification, the authors [ 50 ] used traditional machine learning classifiers (Linear Regression, Multilayer Perceptron (MLP), Nearest Neighbor search, Softmax Regression, Gated recurrent Unit (GRU)-SVM, and SVM). For data pre-processing, the study used the Standard Scaler technique that standardizes data points by removing the mean and scaling the data to unit variance. The MLP model outperformed the other models by producing the highest accuracy of 99.04% which is almost similar to the accuracy of 99.6% that was reported by [ 49 ].

Different form binary classification of benign or malignant classes, a study [ 46 ] proposed a two-step approach to design a breast cancer multi-class classification model that predicts eight categories of breast cancer. In the first approach, the study used handcrafted features that are generated from histopathology images. These features were then fed as input to classical machine learning algorithms (RF, SVM, Linear Discriminant Analysis (LDA)). In the second approach, the study applied a transfer learning method to develop the multi-classification deep learning framework where pretained CNNs (ResNet50, VGG16 and VGG19) were used as feature extractors and baseline models. It was then found that the VGG16 pretrained CNN with the linear SVM provided the best accuracy in the range of 91.23% \(-\) 93.97%. This study also found that using pretrained CNNs as feature extractors improved the classification performance of the models.

The Table 2 provides a summary of the work that has been done to detect breast cancer using machine learning.

Prostate cancer

Prostate cancer is a type of cancer that develops when cells in the prostate gland start to grow uncontrollably (malignant). Prostate cancer often presents with no symptoms and grows at a slow rate. As a result, some men may die of other diseases before the cancer starts to cause notable problems. Comparably, prostate cancer can also be aggressive and metastasize to other body organs that are outside the confines of the prostate gland. Risk factors that are associated with this type of cancer include age, specifically, men that are above the age of 50. Other risk factors include ethnicity, family history of prostate cancer, breast or ovarian cancer, and obesity [ 61 , 62 , 63 ].

Transfer learning, which is defined as the reuse of a pretrained model on a new problem, was frequently applied to develop prostate cancer detection models using machine learning (Fig.  3 ). For example, a study [ 64 ] applied a transfer learning approach to detect prostate cancer on magnetic resonance images (MRI) by using a pretrained GoogleNet. A series of features such as texture, entropy, morphological, scale invariant feature transform (SIFT), and Elliptic Fourier Descriptors (EFDs) were extracted from the images as described by [ 65 , 66 ]. Other traditional machine learning classifiers were also evaluated such as Decision trees, and SVM Gaussian however, the GoogleNet model outperformed the other models.

Generalized machine learning framework for prostate cancer prediction using 3-d CNNs, pooling layers, and a fully connected layer for classification [ 69 ]

Also using transfer learning, a study [ 67 ] developed a prostate cancer detection model by using MRI images and ultrasound (US) images. The model was developed in two stages: first, pretrained CNNs were used for classification of the US and MRI images into benign or malignant. While the pretrained CNNs performed well on the US images (accuracy 97%), the performance on the MRI images was not adequate. As a result, the best-performing pretrained CNN(VGG16) was selected and used as a feature extractor. The extracted features were then provided as input to traditional machine learning classifiers.

Another study [ 68 ] also used the same dataset as in [ 64 ] to create a prostate cancer detection model. However, instead of using GoogleNet as seen previously by [ 64 ], this study used a ResNet-101 and an autoencoder for feature reduction. Other machine learning models were also evaluated but, the study concluded that the pretrained ResNet-101 outperformed the other models with an accuracy of 100%. These results are similar to a previous study [ 64 ] that showed how pretrained CNNs outperform traditional machine learning models for cancer detection.

Table 3 , gives a summary of recent work that has been executed to create prostate cancer detection models.

Colorectal cancer

Colorectal cancer is a type of cancer that starts in the colon or rectum. The colon and rectum are parts of the human body that make up the large intestine that is part of the digestive system. A large part of the large intestine is made up of the colon which is divided into a few parts namely: ascending colon, transverse colon, descending colon, and sigmoid colon. The main function of the colon is to absorb water and salt from the remaining food waste after it has passed through the small intestine. Then, the waste that is left after passing through the colon goes into the rectum and is stored there until it is passed through the anus. Some colorectal cancers called polyps first develop as growth that can be found in the inner lining of the colon or rectum. Overtime, these polyps can develop into cancer, however, not all of them can be cancerous. Some of the risk factors of colorectal cancer include obesity, lack of exercise, diets that are rich in red meat, smoking, and alcohol [ 82 , 83 , 84 ].

In relation to the advancements made in colorectal cancer research using machine learning (Fig.  4 ), various tasks have been investigated such as predicting high-risk colorectal cancer from images, predicting five-year disease-specific survival, colorectal cancer tissue multi-class classification, and identifying the risk factors for lymph node metastasis (LNM) in colorectal cancer patients [ 85 , 86 , 87 , 88 ]. As with prostate cancer, transfer learning was mostly applied to extract features from various input sources such as colonoscopic images, tissue microarrays (TMA), and H &E slide images. Then, the extracted features were fed as input to machine learning algorithms for classification.

Using a deep CNN network to predict colorectal cancer outcome using images [ 86 ]

One common observation with regards to colorectal cancer models, is that the predictions made from the models were compared to those of experts. For example, a study [ 85 ] developed a deep learning model that detects high risk colorectal cancer from whole slide images that were collected from colon biopsies. The deep learning model was created in two stages: first, a segmentation procedure was executed to extract high risk regions from whole slide images. This segmentation procedure applied Faster-Region Based Convolutional Neural Network (Faster-RCNN) that uses a ResNet-101 model as a backbone for feature extraction. The second stage of implementing the model applied a gradient-boosted decision tree on the output of the Faster-RCNN deep learning model to classify the slides into either high or low risk colorectal cancer, and achieved an AUC of 91.7%. The study then found that the predictions made from the validation set were in agreement with annotations made by expert pathologists.

Work in [ 89 ] also compared predictions made by the Microsatellite instability (MSI)-predictor model with those of expert pathologists and found that experts achieved a mean AUROC of 61% while the model achieved an AUROC of 93% on a hold-out set and 87% on a reader experiment.

A previous study [ 90 ] developed a model named CRCNet, based a pretrained dense CNN, that automatically detects colorecal cancer from colonoscopic images and found that the model exceeded the avarage performance of expert endoscopists on a recall rate of 91.3% versus 83.8%.

In Table 4 , a summary is provided that describes the work that has been executed in colorectal cancer research using machine learning.

In summary of the literature survey (Sect.  2 ), a series of machine learning approaches for the detection of cancer were analysed. Imaging datasets, biological and clinical data, and EHRs were primarily employed as the initial input source when developing cancer detection algorithms. This procedure involved a few preprocessing steps. First, the input source was typically preprocessed at the beginning stages of the experiment to extract regions or features of interest. Next, the retrieved set of features were then applied to downstream machine learning classifiers for cancer prediction. In this work, as opposed to using imaging datasets, clinical and biological data or, EHRs as the starting input source, this work proposes to use raw DNA sequences as the only input source. Moreover, contrary to using statistical methods or advanced CNNs for data extraction and representation, this work proposes to use state-of-the-art sentence transformers namely: SBERT and SimCSE. As far as we are aware, these two sentence transformer models have not been applied for learning representations in cancer research. The learned representations will then be fed as input to machine learning algorithms for cancer prediction.

Data description

In this study, 95 samples from colorectal cancer patients and matched-normal samples from previous work [ 104 ] were analysed. Exon sequences from two key genes: APC and ATM were used. The full details of the exons that were used in this study is shown Tables 5 and 6 . Table 7 shows the data distribution among the normal/tumor DNA sequences. Ethics approval was granted by the University of Pretoria EBIT Research Ethics Committee (EBIT/139/2020).

Data encoding

To encode the DNA sequences, state-of-the-art sentence transformers: Sentence-BERT [ 105 ] and SimCSE [ 105 ] were used. These transformers are explained in the next subsection.

Sentence-BERT

Sentence-BERT (SBERT) (Fig.  5 ) adapts the pretrained BERT [ 106 ] and RoBERTa [ 107 ] transformer network and modifies it to use a siamese and triplet network architectures to compute fixed-sized vectors for more than 100 languages. The sentence embeddings can then be contrasted using the cosine-similarity. SBERT was trained on the combination of SNLI data [ 108 ] and the Multi-Genre NLI dataset [ 109 ].

SBERT architecture with classification objective function (left) and the regression objective function (right) [ 105 ]

In its architecture, SBERT adds a default mean-pooling procedure on the output of the BERT or RoBERTa network to compute sentence embeddings. SBERT implements the following objective functions: classification objective function, regression objective function, and the triplet objective function. In the classification objective function, the sentence embeddings of two sentence pairs u and v are concatenated using the element-wise difference \(\mid u-v \mid\) and multiplied with the trainable weight \(W_{t} \epsilon {\mathbb {R}}^{3n *k}\) :

where n is the length or dimension of the sentence embeddings and k is the value of the target labels.

The regression objective function makes use of mean-squared-error loss as the objective function to compute the cosine-similarity between two sentence embeddings u and v .

The triplet objective function fine-tunes the network such that the distance between an anchor sentence a and a positive sentence p is smaller than the distance between sentence a and the negative sentence n .

Using the pretrained SBERT model: all-MiniLM-L6-v2 , each DNA sequence was represented by a 384-dimensional vector.

As with SBERT, Simple Contrastive Sentence Embedding (SimCSE) [ 110 ] (Fig.  6 is a transformer based model that modifies the BERT/RoberTa encoder to generate sentence embeddings. It uses a contrastive learning approach that aims to learn sentence representations by pulling close neighbours together and propelling non-neighbours. SimCSE comes in two learning forms: unsupervised and supervised SimCSE. In unsupervised SimCSE, the network is fine-tuned to predict the input sentence itself using dropout as noise then, the other sentences that are in the mini-batch are taken as negatives. In this case, dropout acts as a data augmentation method while previous [ 111 , 112 ] methods have used word deletion, reordering, and substitution as a way of generating positive instances. In unsupervised SimCSE, an input sentence is fed twice to the encoder then, two embeddings with different dropout masks z , \(z'\) are generated as output. The training objective for SimCSE is:

where z is the standard dropout mask that are found in Transformers and no additional dropout mask is added [ 110 ].

Unsupervised SimCSE ( a ) and supervised SimCSE ( b ) [ 110 ]

In supervised SimCSE, positive pairs are taken from the natural language inference (NLI) datasets and used to optimise the following equation:

where \(\tau\) is a temperature hyperparamter and \(sim(h_{1},h_{2})\) is the cosine similarity.

Using the unsupervised pretrained SimCSE model: unsup-simcse-bert-base-uncased , each DNA sequence was represented by a 768-dimensional vector.

K -means clustering

The k -means clustering algorithm was used to visualize the sentence representations generated from SBERT and SimCSE in an unsupervised approach. The k -means algorithm divides the data points into k clusters where each data point is said to belong to the cluster centroid closest to it. Since the data consists of two types of documents (tumor vs. normal), the k -means algorithm was asked to find 2 clusters n and assign each DNA sequence to its closest centroid [ 113 ].

Machine learning experiments

A total of three machine learning algorithms were used for classification: Light Gradient Boosting (LightGBM), eXtreme Gradient Boosting (XGBoost), and Random Forest (RF).

eXtreme gradient boosting (XGBoost)

eXtreme Gradient Boosting (XGBoost), is an efficient implementation of the gradient boosting algorithm. Gradient boosting belongs to a group of ensemble machine learning algorithms that be used to solve classification or regression problems. The ensembles are created from decision trees that are added one at a time to the ensemble, and fit to correct the classification error that were made by prior trees [ 114 ].

Light gradient boosting (LightGBM)

Light Gradient Boosting (LightGBM) machine is also a gradient boosting model that is used for ranking, classification, and regression. In contrast to XGBoost, LightGBM splits the tree vertically as opposed to horizontally. This method of growing the tree leaf vertically results in more loss reduction and provides higher accuracy while also being faster. LightGBM uses the Gradient-based One-Side Sampling (GOSS) method to filter out data instances for obtaining the best split value while XGBoost uses a pre-sorted and Histogram-based algorithm for calculating the best split value [ 115 ].

Random forest (RF)

Random forest (RF) is a supervised machine learning that is used in classification and regression tasks. It creates decision tress based on different samples and takes the majority vote for classification or average for regression. While XGBoost and LightGBM use a gradient boosting method, Random Forest uses a bagging method. The bagging method builds a different training subset from the training data with replacement. Each model is trained separately and the final result is based on a majority voting after consolidating the results of all the models [ 116 ].

Convolutional neural network (CNN)

Convolutional neural networks (CNNs) are a subset of neural networks that are frequently used to process speech, audio, and visual input signals. Convolutional, pooling, and fully connected (FC) layers are the three types of layers that are generally present in CNNs. The convolutional layer is the fundamental component of a CNN and is in charge of performing convolutional operations on the input before passing the outcome to the following layer. Then, the input is subjected to dimensionality reduction using pooling layers that reduces the number of parameters in the input. The FC layer uses a variety of activation functions, including the softmax activation function and the sigmoid activation function, to carry out the classification task using the features retrieved from the network’s prior layers [ 117 , 118 ]. In this work, a three-layer CNN model with a sigmoid activation function will be supplied with the embedding features that were retrieved by SBERT and SimCSE sentence transformers. Due to computational limitations, the network will be trained over 10 epochs using the RMSprop optimizer and cross-validated over five folds.

Performance evaluation metrics

To measure the performance of the machine learning models, the average performance of the models were reported using 5-fold cross validation and the following metrics were used: accuracy, precision, recall and F1 score. In Table 8 , the definition of these metrics is provided.

This section described the datasets used in the study as well as data representation methods and machine learning algorithms that were applied in this work. In the next section, the results of the applied methods are described.

Visualizations

In this subsection, unlabeled data from SBERT and SimCSE representations were explored and visualized with the k -means clustering algorithm. The representations of the SBERT algorithm (Fig.  7 ) revealed more overlap between the data points in comparison to the representations of the SimCSE algorithm (Fig.  8 ). In the next subsection, machine learning models are evaluated to reveal if there is sufficient signal in the representations of the two sentence transformers that can discriminate between tumor and normal DNA sequences.

Visualisation of the SBERT documents with k -means clustering

Visualisation of the SimCSE documents with k -means clustering

Comparative performance of the machine learning results

Sbert before smote.

Table 9 presents the performance of the machine learning models on the dev set in terms of the average accuracy, averaged over the five folds using the SBERT representations. More performance metrics such as F1 score, recall, and precision are reported in the Additional file 1 (Appendix A ).

Considering that the tumor DNA sequences belonging to the APC gene comprised of \(\approx\) 64% of the data before SMOTE sampling, the machine learning models classified most sequences as positive (tumor); with the CNN achieving the best overall with the highest accuracy of 67.3 ± 0.04%.

In contrast to the data distribution of the APC gene before SMOTE sampling, the original data distribution of sequences from the ATM gene were relatively balanced as the tumor sequences comprised of 53% of the total data, and normal DNA sequences made up 47%. Moreover, as opposed to predicting nearly all sequences as positive, the machine learning models demonstrated an unbiased above-average performance as the highest performing model (XGBoost) achieved an accuracy of 73. ± 0.13 %.

SBERT after SMOTE

The performance of the majority of the machine learning classifiers after applying SMOTE remained consistent in that very little improvement or decline was observed. Moreover, while the CNN model previously obtained the highest overall accuracy before SMOTE oversampling, it performed the worst after applying SMOTE with a reported accuracy of 47. ± 17.4 %. Although biased, the LightGBM classifier reached the highest accuracy of 64.9 ± 0.29 %. Its confusion matrix is shown (Fig.  9 ).

Confusion matrix of the LightGBM model using SBERT representations after SMOTE (dev set)

The same trend as seen in the previous Sect.  4.2.2 was also observed in this section with sequences from the ATM gene. Here, the performance of the machine learning models after SMOTE sampling was relatively similar to the performance of the machine learning models before SMOTE sampling as the XGBoost still maintained the best overall accuracy of 73. ± 0.13 % (Fig.  10 ).

Confusion matrix of the XGBoost model using SBERT representations after SMOTE (dev set)

SimCSE before SMOTE

Table 9 also presents the performance of the machine learning models in terms of the average accuracy, averaged over the five folds using the SimCSE representations. Supplementary performance metrics are reported (Additional file 1 : Appendix A).

In this experimental setting, the performance of the machine learning models with SBERT representations before SMOTE sampling was similar to the performance of the models with SimCSE representations before SMOTE sampling. Here, the CNN achieved the best accuracy of 67. ± 0.0 %.

A similar pattern as in the previous Sect. ( APC , SimCSE before SMOTE) was also detected in this setting when using sequences from the ATM gene in that the performance of the SimCSE models were almost similar to the performance of the SBERT models (before SMOTE) with slight improvement. The LightGBM model achieved the highest accuracy of 74. ± 0.18 % which was an improvement in accuracy of approximately 4 %.

SimCSE after SMOTE

The LightGBM model achieved the highest accuracy of 64.7 ± 0.29 (Fig.  11 ), which was indistinguishable to the performance reported before SMOTE oversampling.

Confusion matrix of the LightGBM model using SimCSE representations after SMOTE (dev set)

ATM In this final experimental setting, the results demonstrated a consistent performance before SMOTE sampling and after SMOTE sampling. The highest performing model was the Random forest model as it achieved an average accuracy of 71.6 ± 1.47 % (Fig.  12 ).

Confusion matrix of the Random forest model using SimCSE representations after SMOTE (dev set)

In Table 10 , the experiments were repeated on an additional unseen test set. Overall, the machine learning models demonstrated a slight increase in the accuracy as the highest performing model, XGBoost, achieved an average accuracy of 75. ± 0.12 % using SimCSE representations from the ATM gene.

This paper provided a literature review of how cancer has been detected using various machine learning methods. Additionally, this work developed machine learning models that detect cancer using raw DNA sequences as the only input source. The DNA sequences were retrieved from matched tumor/normal pairs of colorectal cancer patients as described by previous work [ 104 ]. For data representation, two state-of-the-art sentence transformers were proposed: SBERT and SimCSE. To the best of our knowledge, these two methods have not been used to represent DNA sequences in cancer detection problems using machine learning. In summary of the results, we note that using SimCSE representations only marginally improved the performance of the machine learning models.

The ability to detect cancer by relying on human DNA as the only input source to a learning algorithm was one of the significant contributions of this work. We acknowledge that similar research investigating the role that the DNA plays in various cancer types has been conducted in the past. In contrary, the way the DNA was represented for the learning algorithms in our work is different from that in earlier research. An example would be work performed by [ 120 ] that used cell-free DNA (cfDNA) data from shallow whole-genome sequencing to uncover patterns associated with a number of different cancers including Hodgkin lymphoma, diffuse large B-cell lymphoma, and multiple myeloma. This study used PCA transformed genome-wide coverage features and applied them as input to a support vector algorithm to predict cancer status rather than employing sentence transforms for data representation as was done in our study. Another study [ 121 ] also used cfDNA sequences to predict cancer tissue sequences from healthy ones. In this work, reads from hepatocellular carcinoma (HCC) patients and healthy individuals were integrated with methylation information and then, a deep learning model was created to predict the reads that originated from a cancer tissue. The deep learning model consisted of a 1-d CNN followed by a maxpooling layer, a bi-directional LSTM, a 1-d CNN, and three dense layers. To represent the cfDNA sequences and methylation information, the variables were encoded into a one-hot encoded matrix that was then provided as input to the deep learning model for classification. Different from relying on raw DNA or cfDNA data to develop cancer detection frameworks, a study [ 122 ] consolidated methods from variant calling and machine learning to develop a model that detects cancers of unknown primary (CUP) origin which account for approximately 3% of all cancer diagnoses. This work employed whole-genome-sequencing-based mutation features derived from structural variants that were generated through variant calling and fed them as input to an ensemble of random forest binary classifiers for the detection of 35 different cancers.

Limitations of the study

The machine learning experiments were only performed on two key genes: APC and APC , therefore it would have been interesting to see how the models generalize across various genes. The common disadvantage of conducting the experiments on multiple genes or whole genome sequencing data is that they require more computational resources which have a direct impact on cost. Another limitation of this work is that only two pretrained models were used for generating the sentence representations. Since there are several other pretrained models that are publicly available to choose from, some pretrained models were slower to execute than others hence a decision was made to focus on pretrained models that provided fast execution.

This article reviewed the literature and demonstrated how various machine learning techniques have been used to identify cancer. Given that they are the most common malignancies worldwide, this work placed a special emphasis on four cancer types: lung, breast, prostate, and colorectal cancer. Then, a new method for the identification of colorectal cancer employing SBERT and SimCSE sentence representations was presented. Raw DNA sequences from matched tumor/normal pairs of colorectal cancer served as the sole input for this approach. The learned representations were then provided as input to machine learning classifiers for classification. In light of the performance of the machine learning classifiers, XGBoost was found to be the best performing classifier overall. Moreover, using SimCSE representations only marginally improved the classification performance of the machine learning models.

Availability of data and materials

The data can be accessed at the host database (The European Genome-phenome Archive at the European Bioinformatics Institute, accession number: EGAD00001004582 Data access ).

Jones PA, Baylin SB. The epigenomics of cancer. Cell. 2007;128(4):683–92.

Article   CAS   PubMed   PubMed Central   Google Scholar  

What Is Cancer? National Cancer Institute. https://www.cancer.gov/about-cancer/understanding/what-is-cancer

Zheng R, Sun K, Zhang S, Zeng H, Zou X, Chen R, Gu X, Wei W, He J. Report of cancer epidemiology in china, 2015. Zhonghua zhong liu za zhi. 2019;41(1):19–28.

CAS   PubMed   Google Scholar  

Hegde PS, Chen DS. Top 10 challenges in cancer immunotherapy. Immunity. 2020;52(1):17–35.

Article   CAS   PubMed   Google Scholar  

Kourou K, Exarchos TP, Exarchos KP, Karamouzis MV, Fotiadis DI. Machine learning applications in cancer prognosis and prediction. Comput Struct Biotechnol J. 2015;13:8–17.

Iqbal MJ, Javed Z, Sadia H, Qureshi IA, Irshad A, Ahmed R, Malik K, Raza S, Abbas A, Pezzani R, et al. Clinical applications of artificial intelligence and machine learning in cancer diagnosis: looking into the future. Cancer Cell Int. 2021;21(1):1–11.

Article   Google Scholar  

Loud JT, Murphy J. Cancer screening and early detection in the 21st century. Semin Oncol Nurs. 2017;33:121–8.

Article   PubMed   PubMed Central   Google Scholar  

Goldberg Y, Levy O. word2vec explained: deriving mikolov et al.’s negative-sampling word-embedding method. 2014; arXiv preprint arXiv:1402.3722

Pennington J, Socher R, Manning CD. Glove: global vectors for word representation. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). 2014. p. 1532–43.

Bojanowski P, Grave E, Joulin A, Mikolov T. Enriching word vectors with subword information. Trans Assoc Comput Linguist. 2017;5:135–46.

Church KW. Word2vec. Natl Lang Eng. 2017;23(1):155–62.

Cancer. World Health Organization. https://www.who.int/news-room/fact-sheets/detail/cancer

Bade BC, Cruz CSD. Lung cancer 2020: epidemiology, etiology, and prevention. Clin Chest Med. 2020;41(1):1–24.

Article   PubMed   Google Scholar  

Barta JA, Powell CA, Wisnivesky JP. Global epidemiology of lung cancer. Ann Global Health. 2019;85:1.

de Carvalho Filho AO, Silva AC, de Paiva AC, Nunes RA, Gattass M. Classification of patterns of benignity and malignancy based on ct using topology-based phylogenetic diversity index and convolutional neural network. Pattern Recogn. 2018;81:200–12.

Rodrigues MB, Da Nobrega RVM, Alves SSA, Reboucas Filho PP, Duarte JBF, Sangaiah AK, De Albuquerque VHC. Health of things algorithms for malignancy level classification of lung nodules. IEEE Access. 2018;6:18592–601.

Asuntha A, Srinivasan A. Deep learning for lung cancer detection and classification. Multim Tools Appl. 2020;79(11):7731–62.

Shakeel PM, Tolba A, Al-Makhadmeh Z, Jaber MM. Automatic detection of lung cancer from biomedical data set using discrete adaboost optimized ensemble learning generalized neural networks. Neural Comput Appl. 2020;32(3):777–90.

Abdullah DM, Abdulazeez AM, Sallow AB. Lung cancer prediction and classification based on correlation selection method using machine learning techniques. Qubahan Acad J. 2021;1(2):141–9.

Ausawalaithong W, Thirach A, Marukatat S, Wilaiprasitporn T. Automatic lung cancer prediction from chest x-ray images using the deep learning approach. In: 2018 11th biomedical engineering international conference (BMEiCON). 2018; pp. 1–5. IEEE

Wang X, Peng Y, Lu L, Lu Z, Bagheri M, Summers RM. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2017; pp. 2097–106

Shiraishi J, Katsuragawa S, Ikezoe J, Matsumoto T, Kobayashi T, Komatsu K-I, Matsui M, Fujita H, Kodera Y, Doi K. Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists’ detection of pulmonary nodules. Am J Roentgenol. 2000;174(1):71–4.

Article   CAS   Google Scholar  

Armato SG III, McLennan G, Bidaut L, McNitt-Gray MF, Meyer CR, Reeves AP, Zhao B, Aberle DR, Henschke CI, Hoffman EA, et al. The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans. Med Phys. 2011;38(2):915–31.

Kaggle: Lung and Colon Cancer Histopathological Images. https://www.kaggle.com/andrewmvd/lung-and-colon-cancer-histopathological-images Accessed 16 July 2020.

Radhika P, Nair RA, Veena G. A comparative study of lung cancer detection using machine learning algorithms. In: 2019 IEEE international conference on electrical, computer and communication technologies (ICECCT). 2019; pp. 1–4. IEEE

Salaken SM, Khosravi A, Khatami A, Nahavandi S, Hosen MA. Lung cancer classification using deep learned features on low population dataset. In: 2017 IEEE 30th Canadian conference on electrical and computer engineering (CCECE). 2017; pp. 1–5. IEEE.

Bhattacharjee A, Richards WG, Staunton J, Li C, Monti S, Vasa P, Ladd C, Beheshti J, Bueno R, Gillette M, et al. Classification of human lung carcinomas by mrna expression profiling reveals distinct adenocarcinoma subclasses. Proc Natl Acad Sci. 2001;98(24):13790–5.

Bhatia S, Sinha Y, Goel L. Lung cancer detection: a deep learning approach. In: Soft computing for problem solving. 2019; p. 699–705. Springer.

Shin H, Oh S, Hong S, Kang M, Kang D, Ji Y-G, Choi BH, Kang K-W, Jeong H, Park Y, et al. Early-stage lung cancer diagnosis by deep learning-based spectroscopic analysis of circulating exosomes. ACS Nano. 2020;14(5):5435–44.

Masud M, Sikder N, Nahid A-A, Bairagi AK, AlZain MA. A machine learning approach to diagnosing lung and colon cancer using a deep learning-based classification framework. Sensors. 2021;21(3):748.

Naseer I, Akram S, Masood T, Jaffar A, Khan MA, Mosavi A. Performance analysis of state-of-the-art cnn architectures for luna16. Sensors. 2022;22(12):4426.

Setio AAA, Traverso A, De Bel T, Berens MS, Van Den Bogaard C, Cerello P, Chen H, Dou Q, Fantacci ME, Geurts B, et al. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Med Image Anal. 2017;42:1–13.

Saba T. Recent advancement in cancer detection using machine learning: systematic survey of decades, comparisons and challenges. J Infect Pub Health. 2020;13(9):1274–89.

Sun Y-S, Zhao Z, Yang Z-N, Xu F, Lu H-J, Zhu Z-Y, Shi W, Jiang J, Yao P-P, Zhu H-P. Risk factors and preventions of breast cancer. Int J Biol Sci. 2017;13(11):1387.

Breast cancer. World Health Organization. https://www.who.int/news-room/fact-sheets/detail/breast-cancer

Kelsey JL, Gammon MD. The epidemiology of breast cancer. CA Cancer J Clin. 1991;41(3):146–65.

Harbeck N, Penault-Llorca F, Cortes J, Gnant M, Houssami N, Poortmans P, Ruddy K, Tsang J, Cardoso F. Breast cancer. Nat Rev Dis Prim. 2019;5(1):1–31.

Google Scholar  

Waks AG, Winer EP. Breast cancer treatment: a review. JAMA. 2019;321(3):288–300.

Tahmooresi M, Afshar A, Rad BB, Nowshath K, Bamiah M. Early detection of breast cancer using machine learning techniques. J Telecommun Electr Comput Eng. 2018;10(3):21–7.

Sharma S, Aggarwal A, Choudhury T. Breast cancer detection using machine learning algorithms. In: 2018 international conference on computational techniques, electronics and mechanical systems (CTEMS). 2018; p. 114–8 . IEEE.

VisualLab: A Methodology for Breast Disease Computer-Aided Diagnosis Using Dynamic Thermography. http://visual.ic.uff.br/en/proeng/thiagoelias/

Wolberg WH, Street WN, Mangasarian OL. Breast cancer wisconsin (diagnostic) data set. UCI machine learning repository. http://archive.ics.uci.edu/ml/ ; 1992.

Suckling JP. The mammographic image analysis society digital mammogram database. Digital Mammo. 1994; pp. 375–86.

Roy A. Deep convolutional neural networks for breast cancer detection. In: 2019 IEEE 10th annual ubiquitous computing, electronics & mobile communication conference (UEMCON). 2019; pp. 0169–71 . IEEE.

Mambou SJ, Maresova P, Krejcar O, Selamat A, Kuca K. Breast cancer detection using infrared thermal imaging and a deep learning model. Sensors. 2018;18(9):2799.

Sharma S, Mehra R. Conventional machine learning and deep learning approach for multi-classification of breast cancer histopathology images-a comparative insight. J Digit Imag. 2020;33(3):632–54.

Remya R, Rajini NH. Transfer learning based breast cancer detection and classification using mammogram images. In: 2022 international conference on electronics and renewable systems (ICEARS). 2022; pp. 1060–5 . IEEE.

Vaka AR, Soni B, Reddy S. Breast cancer detection by leveraging machine learning. ICT Express. 2020;6(4):320–4.

Khuriwal N, Mishra N. Breast cancer detection from histopathological images using deep learning. In: 2018 3rd international conference and workshops on recent advances and innovations in engineering (ICRAIE). 2018; pp. 1–4 . IEEE.

Agarap AFM. On breast cancer detection: an application of machine learning algorithms on the wisconsin diagnostic dataset. In: proceedings of the 2nd international conference on machine learning and soft computing. 2018; pp. 5–9.

Shen L, Margolies LR, Rothstein JH, Fluder E, McBride R, Sieh W. Deep learning to improve breast cancer detection on screening mammography. Sci Rep. 2019;9(1):1–12.

Sawyer Lee R, Gimenez F, Hoogi A, Rubin D. Curated Breast Imaging Subset of DDSM. The cancer imaging archive, 2016.

Moreira IC, Amaral I, Domingues I, Cardoso A, Cardoso MJ, Cardoso JS. Inbreast: toward a full-field digital mammographic database. Acad Radiol. 2012;19(2):236–48.

VRI: Breast Cancer Histopathological Database (BreakHis). https://web.inf.ufpr.br/vri/databases/breast-cancer-histopathological-database-breakhis/

Alanazi SA, Kamruzzaman M, Islam Sarker MN, Alruwaili M, Alhwaiti Y, Alshammari N, Siddiqi MH. Boosting breast cancer detection using convolutional neural network. J Healthc Eng 2021;2021.

Janowczyk, A.: Use case 6: invasive ductal carcinoma (IDC) segmentation. http://www.andrewjanowczyk.com/use-case-6-invasive-ductal-carcinoma-idc-segmentation/

Arooj S, et al.: Breast cancer detection and classification empowered with transfer learning. Front Pub Health. 2022;10.

Nasir MU, Ghazal TM, Khan MA, Zubair M, Rahman A-u, Ahmed R, Hamadi HA, Yeun CY. Breast cancer prediction empowered with fine-tuning. Comput Intell Neurosci. 2022;2022.

Breast cancer patients mris. Kaggle. https://www.kaggle.com/uzairkhan45/breast-cancer-patients-mris

Khan MBS, Nawaz MS, Ahmed R, Khan MA, Mosavi A, et al. Intelligent breast cancer diagnostic system empowered by deep extreme gradient descent optimization. Mathem Biosci Eng. 2022;19(8):7978–8002.

What is Prostate Cancer. UCLA Health. https://www.uclahealth.org/urology/prostate-cancer/what-is-prostate-cancer

Desai MM, Cacciamani GE, Gill K, Zhang J, Liu L, Abreu A, Gill IS. Trends in incidence of metastatic prostate cancer in the us. JAMA Netw Open. 2022;5(3):222246.

Cackowski FC, Heath EI. Prostate cancer dormancy and recurrence. Cancer Lett. 2022;524:103–8.

Abbasi AA, Hussain L, Awan IA, Abbasi I, Majid A, Nadeem MSA, Chaudhary Q-A. Detecting prostate cancer using deep learning convolution neural network with transfer learning approach. Cogn Neurodyn. 2020;14(4):523–33.

Hussain L, Ahmed A, Saeed S, Rathore S, Awan IA, Shah SA, Majid A, Idris A, Awan AA. Prostate cancer detection using machine learning techniques by employing combination of features extracting strategies. Cancer Biomark. 2018;21(2):393–413.

Hussain L, et al. Detecting brain tumor using machines learning techniques based on different features extracting strategies. Curr Med Imag. 2019;15(6):595–606.

Hassan MR, Islam MF, Uddin MZ, Ghoshal G, Hassan MM, Huda S, Fortino G. Prostate cancer classification from ultrasound and mri images using deep learning based explainable artificial intelligence. Fut Gener Comput Syst. 2022;127:462–72.

Iqbal S, Siddiqui GF, Rehman A, Hussain L, Saba T, Tariq U, Abbasi AA. Prostate cancer detection using deep learning and traditional techniques. IEEE Access. 2021;9:27085–100.

Feng Y, Yang F, Zhou X, Guo Y, Tang F, Ren F, Guo J, Ji S. A deep learning approach for targeted contrast-enhanced ultrasound based prostate cancer detection. IEEE/ACM transactions on computational biology and bioinformatics. 2018;16(6):1794–801.

Reda I, Khalil A, Elmogy M, Abou El-Fetouh A, Shalaby A, Abou El-Ghar M, Elmaghraby A, Ghazal M, El-Baz A. Deep learning role in early diagnosis of prostate cancer. Technol Cancer Res Treat. 2018;17:1533034618775530.

Barlow H, Mao S, Khushi M. Predicting high-risk prostate cancer using machine learning methods. Data. 2019;4(3):129.

Yoo S, Gujrathi I, Haider MA, Khalvati F. Prostate cancer detection using deep convolutional neural networks. Sci Rep. 2019;9(1):1–10.

Tolkach Y, Dohmgörgen T, Toma M, Kristiansen G. High-accuracy prostate cancer pathology using deep learning. Nat Mach Intell. 2020;2(7):411–8.

Genomic Data Commons Data Portal. National Cancer Institute (NIH) GDC Data Portal. http://portal.gdc.cancer.gov

Zenodo. Zenodo. https://zenodo.org/deposit/3825933

Hosseinzadeh M, Saha A, Brand P, Slootweg I, de Rooij M, Huisman H. Deep learning–assisted prostate cancer detection on bi-parametric mri: minimum training data size requirements and effect of prior knowledge. Eur Radiol. 2021; 1–11.

Natarajan S, Priester A, Margolis D, Huang J, Marks L. Prostate mri and ultrasound with pathology and coordinates of tracked biopsy (prostate-mri-us-biopsy). 2020.

Clark K, Vendt B, Smith K, Freymann J, Kirby J, Koppel P, Moore S, Phillips S, Maffitt D, Pringle M, et al. The cancer imaging archive (tcia): maintaining and operating a public information repository. J Dig Imaging. 2013;26(6):1045–57.

Sonn GA, Natarajan S, Margolis DJ, MacAiran M, Lieu P, Huang J, Dorey FJ, Marks LS. Targeted biopsy in the detection of prostate cancer using an office based magnetic resonance ultrasound fusion device. J Urol. 2013;189(1):86–92.

Tsuneki M, Abe M, Kanavati F. A deep learning model for prostate adenocarcinoma classification in needle biopsy whole-slide images using transfer learning. Diagnostics. 2022;12(3):768.

Otsu N. A threshold selection method from gray-level histograms. IEEE Trans Syst Man Cybern. 1979;9(1):62–6.

What Is Colorectal Cancer? American Cancer Society. https://www.cancer.org/cancer/colon-rectal-cancer/about/what-is-colorectal-cancer.html

Center MM, Jemal A, Smith RA, Ward E. Worldwide variations in colorectal cancer. CA Cancer J Clin. 2009;59(6):366–78.

Weitz J, Koch M, Debus J, Höhler T, Galle PR, Büchler MW. Colorectal cancer. Lancet. 2005;365(9454):153–65.

Ho C, Zhao Z, Chen XF, Sauer J, Saraf SA, Jialdasani R, Taghipour K, Sathe A, Khor L-Y, Lim K-H, et al. A promising deep learning-assistive algorithm for histopathological screening of colorectal cancer. Sci Rep. 2022;12(1):1–9.

Bychkov D, Linder N, Turkki R, Nordling S, Kovanen PE, Verrill C, Walliander M, Lundin M, Haglund C, Lundin J. Deep learning based tissue analysis predicts outcome in colorectal cancer. Sci Rep. 2018;8(1):1–11.

Damkliang K, Wongsirichot T, Thongsuksai P. Tissue classification for colorectal cancer utilizing techniques of deep learning and machine learning. Biomed Eng Appl Basis Commun. 2021;33(03):2150022.

Brockmoeller S, Echle A, Ghaffari Laleh N, Eiholm S, Malmstrøm ML, Plato Kuhlmann T, Levic K, Grabsch HI, West NP, Saldanha OL, et al. Deep learning identifies inflamed fat as a risk factor for lymph node metastasis in early colorectal cancer. J Pathol. 2022;256(3):269–81.

Yamashita R, Long J, Longacre T, Peng L, Berry G, Martin B, Higgins J, Rubin DL, Shen J. Deep learning model for the prediction of microsatellite instability in colorectal cancer: a diagnostic study. Lancet Oncol. 2021;22(1):132–41.

Zhou D, Tian F, Tian X, Sun L, Huang X, Zhao F, Zhou N, Chen Z, Zhang Q, Yang M, et al. Diagnostic evaluation of a deep learning model for optical diagnosis of colorectal cancer. Nat Commun. 2020;11(1):1–9.

CAS   Google Scholar  

Wang Y-H, Nguyen PA, Islam MM, Li Y-C, Yang H-C, et al. Development of deep learning algorithm for detection of colorectal cancer in ehr data. In: MedInfo. 2019; pp. 438–41

Echle A, Grabsch HI, Quirke P, van den Brandt PA, West NP, Hutchins GG, Heij LR, Tan X, Richman SD, Krause J, et al. Clinical-grade detection of microsatellite instability in colorectal tumors by deep learning. Gastroenterology. 2020;159(4):1406–16.

Macenko M, et al. A method for normalizing histology slides for quantitative analysis. In: 2009 IEEE international symposium on biomedical imaging: from Nano to Macro, pp. 1107–10 (2009). IEEE.

Amitay EL, Carr PR, Jansen L, Walter V, Roth W, Herpel E, Kloor M, Bläker H, Chang-Claude J, Brenner H, et al. Association of aspirin and nonsteroidal anti-inflammatory drugs with colorectal cancer risk by molecular subtypes. JNCI J Natl Cancer Inst. 2019;111(5):475–83.

Group QC, et al. Adjuvant chemotherapy versus observation in patients with colorectal cancer: a randomised study. Lancet. 2007;370(9604):2020–9.

van den Brandt PA, Goldbohm RA, Veer PV, Volovics A, Hermus RJ, Sturmans F. A large-scale prospective cohort study on diet and cancer in the netherlands. Journal of clinical epidemiology. 1990;43(3):285–95.

Taylor J, Wright P, Rossington H, Mara J, Glover A, West N, Morris E, Quirke P. Regional multidisciplinary team intervention programme to improve colorectal cancer outcomes: study protocol for the yorkshire cancer research bowel cancer improvement programme (ycr bcip). BMJ Open. 2019;9(11): 030618.

Histological images for MSI vs. MSS classification in gastrointestinal cancer, FFPE samples. Zenodo. https://zenodo.org/record/2530835#.Ypib9C8RpQI

Sarwinda D, Paradisa RH, Bustamam A, Anggia P. Deep learning in image classification using residual network (resnet) variants for detection of colorectal cancer. Proc Comput Sci. 2021;179:423–31.

Tissue Image Analytics (TIA) Centre. warwick. https://warwick.ac.uk/fac/cross_fac/tia/data/glascontest/download

Lorenzovici N, Dulf E-H, Mocan T, Mocan L. Artificial intelligence in colorectal cancer diagnosis using clinical data: non-invasive approach. Diagnostics. 2021;11(3):514.

Kather JN, Weis C-A, Bianconi F, Melchers SM, Schad LR, Gaiser T, Marx A, Zöllner FG. Multi-class texture analysis in colorectal cancer histology. Sci Rep. 2016;6(1):1–11.

Muti H, Loeffler C, Echle A, Heij L, Buelow R, Krause J, et al. The aachen protocol for deep learning histopathology: a hands-on guide for data preprocessing. Zenodo Aachen. 2020;10

Poulos RC, Perera D, Packham D, Shah A, Janitz C, Pimanda JE, Hawkins N, Ward RL, Hesson LB, Wong JW. Scarcity of recurrent regulatory driver mutations in colorectal cancer revealed by targeted deep sequencing. JNCI Cancer spectr. 2019;3(2):012.

Reimers N, Gurevych I. Sentence-bert: Sentence embeddings using siamese bert-networks. 2019; arXiv preprint arXiv:1908.10084

Devlin J, Chang M-W, Lee K, Toutanova K. Bert: pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

Liu Y, Ott M, Goyal N, Du J, Joshi M, Chen D, Levy O, Lewis M, Zettlemoyer L, Stoyanov V. Roberta: a robustly optimized bert pretraining approach. 2019; arXiv preprint arXiv:1907.11692

Bowman, S.R., Angeli, G., Potts, C., Manning, C.D.: A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326 (2015)

Williams A, Nangia N, Bowman SR. A broad-coverage challenge corpus for sentence understanding through inference. arXiv preprint arXiv:1704.05426 (2017)

Gao T, Yao X, Chen D. Simcse: simple contrastive learning of sentence embeddings. 2021;arXiv preprint arXiv:2104.08821

Wu, Z., Wang, S., Gu, J., Khabsa, M., Sun, F., Ma, H.: Clear: Contrastive learning for sentence representation. arXiv preprint arXiv:2012.15466 (2020)

Meng Y, Xiong C, Bajaj P, Bennett P, Han J, Song X, et al. Coco-lm: correcting and contrasting text sequences for language model pretraining. Advances in Neural Information Processing Systems. 2021;34

Hartigan JA, Wong MA. Algorithm as 136: a k-means clustering algorithm. J R Stat Soc. 1979;28(1):100–8.

Chen T, Guestrin C. Xgboost: a scalable tree boosting system. In: Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 2016; pp. 785–94

Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, Ye Q, Liu T-Y. Lightgbm: A highly efficient gradient boosting decision tree. Adv Neural Inf Process Syst. 2017;30

Breiman L. Random forests. Mach Learn. 2001;45(1):5–32.

Albawi S, Mohammed TA, Al-Zawi S. Understanding of a convolutional neural network. In: 2017 International conference on engineering and technology (ICET). 2017; pp. 1–6. IEEE

O’Shea K, Nash R. An introduction to convolutional neural networks. 2015; arXiv preprint arXiv:1511.08458

Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A, Cournapeau D, Brucher M, Perrot M, Duchesnay E. Scikit-learn: machine learning in Python. J Mach Learn Res. 2011;12:2825–30.

Che H, Jatsenko T, Lenaerts L, Dehaspe L, Vancoillie L, Brison N, Parijs I, Van Den Bogaert K, Fischerova D, Heremans R, et al. Pan-cancer detection and typing by mining patterns in large genome-wide cell-free dna sequencing datasets. Clin Chem. 2022;68(9):1164–76.

Li J, Wei L, Zhang X, Zhang W, Wang H, Zhong B, Xie Z, Lv H, Wang X. Dismir: D eep learning-based noninvasive cancer detection by i ntegrating dna s equence and methylation information of i ndividual cell-free dna r eads. Brief Bioinf. 2021;22(6):250.

Nguyen L, Van Hoeck A, Cuppen E. Machine learning-based tissue of origin classification for cancer of unknown primary diagnostics using genome-wide mutation features. Nat Commun. 2022;13(1):4013.

Download references

Acknowledgements

The authors would like to thank the DAC for MCO colorectal cancer genomics at The University of New South Wales, for providing the data used in the study. The authors would also like to thank Prof. Jason Wong, for facilitating the data access requests and approvals.

The work reported herein was made possible through funding by the South African Medical Research Council (SAMRC) through its Division of Research Capacity Development under the Internship Scholarship Program from funding received from the South African National Treasury. The content hereof is the sole responsibility of the authors and does not necessarily represent the official views of the SAMRC or the funders.

Author information

Authors and affiliations.

Department of Computer Science, University of Pretoria, Pretoria, South Africa

Mpho Mokoatle & Vukosi Marivate

CapeBio TM Technologies, Centurion, South Africa

Darlington Mapiye

School of Medical Sciences, The University of Sydney, Sydney, Australia

Vanessa. M. Hayes

School of Health Systems and Public Health, University of Pretoria, Pretoria, South Africa

Riana Bornman & Vanessa. M. Hayes

You can also search for this author in PubMed   Google Scholar

Contributions

MM conceptualised the work and wrote the main manuscript. VM and DM co-supervised and validated the results of experiments reported on the paper. RB and VMH provided expert advice on the topic and also reviewed the manuscript. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Mpho Mokoatle .

Ethics declarations

Ethics approval and consent to participate.

Ethics approval was granted by the University of Pretoria EBIT Research Ethics Committee (EBIT/139/2020). Data approval was granted by the DAC for MCO colorectal cancer genomics at UNSW.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1..

Appendix A.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . The Creative Commons Public Domain Dedication waiver ( http://creativecommons.org/publicdomain/zero/1.0/ ) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Cite this article.

Mokoatle, M., Marivate, V., Mapiye, D. et al. A review and comparative study of cancer detection using machine learning: SBERT and SimCSE application. BMC Bioinformatics 24 , 112 (2023). https://doi.org/10.1186/s12859-023-05235-x

Download citation

Received : 28 November 2022

Accepted : 17 March 2023

Published : 23 March 2023

DOI : https://doi.org/10.1186/s12859-023-05235-x

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Cancer detection
• Machine learning
• SentenceBert

BMC Bioinformatics

ISSN: 1471-2105

• General enquiries: [email protected]

IEEE Account

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

Advancement in Lung Cancer Diagnosis: A Comprehensive Review of Deep Learning Approaches

• First Online: 08 August 2024

Cite this chapter

• Djamel Bouchaffra 1 , 2 ,
• Faycal Ykhlef 1 &
• Samir Benbelkacem 1  

Part of the book series: Interdisciplinary Cancer Research

Lung cancer continues to pose a significant global health challenge. To overcome this challenge, continuous advancements are being achieved in diagnostic methodologies to enhance early detection and improve patient outcomes. This chapter provides a thorough examination of recent progress in lung cancer diagnosis through an extensive survey of deep learning approaches. Focusing on the integration of artificial intelligence (AI) techniques with medical imaging, the chapter encompasses an analysis of convolutional neural networks (CNNs), recurrent neural networks (RNNs), including long short-term memory (LSTMs) networks, and generative-pretrained transformers (GPTs) or large language models (LLMs). The chapter delves into the evolution of deep learning models for lung cancer detection, emphasizing their performance in image classification, lesion segmentation, and overall diagnostic accuracy. Additionally, we also showcase the literature that explores the integration of diverse imaging modalities, such as computed tomography (CT), positron emission tomography (PET), and magnetic resonance imaging (MRI), within deep learning frameworks to enhance the robustness and reliability of diagnostic systems. Furthermore, the review addresses the challenges inherent in the exploration of deep learning in lung cancer diagnosis, including issues related to data quality, model interpretability, and generalizability. Strategies to address these challenges, such as transfer learning, data augmentation (based on generative adversarial networks), and transformers, are thoroughly discussed. The comprehensive analysis presented in this chapter aims to provide a consolidated understanding of the current landscape of deep learning approaches in lung cancer diagnosis. By highlighting recent advancements, challenges, and potential solutions, this chapter contributes to the ongoing dialogue within the scientific community, fostering the development of more effective and reliable tools for early detection and management of lung cancer.

This is a preview of subscription content, log in via an institution to check access.

Access this chapter

Institutional subscriptions

Ardila D, Kiraly AP, Bharadwaj S, Choi B, Reicher JJ, Peng L, Tse D, Etemadi M, Ye W, Corrado G, Naidich DP, Shetty S (2019) End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nat Med 25(6):954–961. https://doi.org/10.1038/s41591-019-0447-x

Article   Google Scholar  

Bandi P, Geessink O, Manson Q, Van Dijk M, Balkenhol M, Hermsen M, Ehteshami Bejnordi B, Lee B, Paeng K, Zhong A, Li Q, Zanjani FG, Zinger S, Fukuta K, Komura D, Ovtcharov V, Cheng S, Zeng S, Thagaard J, Dahl AB (2019) From detection of individual metastases to classification of lymph node status at the patient level: the CAMELYON17 challenge. IEEE Trans Med Imaging 38(2):550–560. https://doi.org/10.1109/tmi.2018.2867350

Chen CJ, Ding A, Li Z, Luo C, Wallach HS (2021) Weakly supervised lesion localization and classification in chest x-rays: attributes and categories matter. arXiv preprint arXiv:2103.10826

Google Scholar  

Coudray N, Ocampo PS, Sakellaropoulos T, Narula N, Snuderl M, Fenyö D, Moreira AL, Razavian N, Tsirigos A (2018) Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nat Med 24(10):1559–1567. https://doi.org/10.1038/s41591-018-0177-5

Cruz-Roa A, Gilmore H, Basavanhally A et al (2017) Accurate and reproducible invasive breast cancer detection in whole-slide images: a deep learning approach for quantifying tumor extent. Sci Rep 7:46450. https://doi.org/10.1038/srep46450

Dosovitskiy A, Brown T, Houlsby N, Kolesnikov A, Beyer L, Zhai X, Kavukcuoglu K (2020) An image is worth 16x16 words: transformers for image recognition at scale. arXiv preprint arXiv:2103.10826

Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S (2020) Dermatologist-level classification of skin cancer with deep neural networks. Nature 542(7639):115–118. https://doi.org/10.1038/nature21056

Huynh E, Hosny A, Guthier C (2016) A two-stage transfer learning algorithm in medical imaging. Proceedings of the IEEE conference on computer vision and pattern recognition workshops (CVPRW), pp 1–6. https://doi.org/10.1109/CVPRW.2016.16

Janowczyk A, Zuo R, Gilmore H (2017) CNN-based segmentation of histology images for prediction of cancer grade. J Pathol Inf 8:27. https://doi.org/10.4103/jpi.jpi_34_17

Nam JG, Park S, Hwang EJ, Lee JH, Jin KN, Lim KY, Park HJ (2022) Development and validation of deep learning-based automatic detection algorithm for malignant pulmonary nodules on chest radiographs. Radiology:211108. https://doi.org/10.1148/radiol.2018180237

Rajpurkar P, Irvin J, Zhu K, Yang B, Mehta H, Duan T, Ng AY (2017) Chexnet: radiologist-level pneumonia detection on chest x-rays with deep learning. arXiv preprint arXiv:1711.05225

Setio AAA, Traverso A, de Bel T, Berens MSN, van den Bogaard C, Cerello P, Chen H, Dou Q, Fantacci ME, Geurts B, van der Gugten R, Heng PA, Jansen B, de Kaste MMJ, Kotov V, Lin JY-H, Manders JTMC, Sóñora-Mengana A, García-Naranjo JC, Papavasileiou E (2017) Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the LUNA16 challenge. Med Image Anal 42:1–13. https://doi.org/10.1016/j.media.2017.06.015

Shin H, Roth HR, Gao M, Lu L, Xu Z, Nogues I, Summers RM (2016) Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning. IEEE Trans Med Imaging 35(5):1285–1298. https://doi.org/10.1109/TMI.2016.2528162

Sirinukunwattana K, Raza SE, Tsang YW, Snead DR, Cree IA, Raj-poot NM (2016) Locality sensitive deep learning for detection and classification of nuclei in routine colon cancer histology images. IEEE Trans Med Imaging 35(5):1196–1206. https://doi.org/10.1109/TMI.2016.2525803

Wang X, Kang J, Zhang X (2020) A novel lung nodule detection system for CT images based on region growing and support vector machine. IEEE Access 8:182475–182483. https://doi.org/10.1109/ACCESS.2020.3029034

Yang B, Chen J, Liu W, Han Z, Guo Z (2021) A novel deep learning model for the identification and classification of lung nodules using global and local receptive fields. Comput Med Imaging Graph 89:101824

Yuan Y, Bar-Yoseph H, Yu S, Jiang H, Dewan M, Lubin N (2019) Deep learning for automated contouring of primary tumor volumes by MRI for nasopharyngeal carcinoma. Radiat Oncol 13(1):1–10. https://doi.org/10.1186/s13014-018-1127-y

Zhang W, Xie Y, Li L, Liu S, Zhang L, Tian J (2019) A transfer learning strategy for deep learning-based classification of 18F-FDG-PET images. Med Phys 46(7):3084–3093. https://doi.org/10.1002/mp.13547

Zhang L, Lu L, Nogues I, Summers RM, Liu S, Yao J, Li Q (2020) DeepPulmonary: a deep learning-based detection system for pulmonary nodules using chest CTs. IEEE Trans Med Imaging 39(3):1169–1179. https://doi.org/10.1109/TMI.2019.2945131

Download references

Author information

Authors and affiliations.

Centre de D́eveloppement des Technologies Avanćees, Algiers, Algeria

Djamel Bouchaffra, Faycal Ykhlef & Samir Benbelkacem

Laboratoire LIPN, UMR CNRS 7030 Institut Galilée—Université Sorbonne Paris Nord, Paris, France

Djamel Bouchaffra

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Djamel Bouchaffra .

Rights and permissions

Reprints and permissions

Copyright information

© 2024 The Author(s), under exclusive license to Springer Nature Switzerland AG

About this chapter

Bouchaffra, D., Ykhlef, F., Benbelkacem, S. (2024). Advancement in Lung Cancer Diagnosis: A Comprehensive Review of Deep Learning Approaches. In: Interdisciplinary Cancer Research. Springer, Cham. https://doi.org/10.1007/16833_2024_302

Download citation

DOI : https://doi.org/10.1007/16833_2024_302

Published : 08 August 2024

Publisher Name : Springer, Cham

Share this chapter

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Publish with us

Policies and ethics

• Find a journal
• Track your research

Information

• Author Services

Initiatives

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess .

Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for future research directions and describes possible research applications.

Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive positive feedback from the reviewers.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Original Submission Date Received: .

• Active Journals
• Find a Journal
• Proceedings Series
• For Authors
• For Reviewers
• For Editors
• For Librarians
• For Publishers
• For Societies
• For Conference Organizers
• Open Access Policy
• Institutional Open Access Program
• Special Issues Guidelines
• Editorial Process
• Research and Publication Ethics
• Article Processing Charges
• Testimonials
• Preprints.org
• SciProfiles
• Encyclopedia

Article Menu

• Subscribe SciFeed
• Recommended Articles
• Author Biographies
• Google Scholar
• on Google Scholar
• Table of Contents

Find support for a specific problem in the support section of our website.

Please let us know what you think of our products and services.

Visit our dedicated information section to learn more about MDPI.

JSmol Viewer

Brain tumor detection and classification using an optimized convolutional neural network.

1. Introduction

Key contributions of our work.

• An optimized CNN hyperparameter model: The paper presents an advanced CNN hyperparameter model that has been carefully developed to optimize critical parameters in diagnosing brain tumors. The activation function, learning rate, batch, padding, filter size and numbers, and pooling layers are just a few of the carefully selected parameters that enhance the model performance and ability to generalize the model. The objective is to increase the model’s overall diagnostic accuracy and dependability by fine-tuning these hyperparameters.
• Datasets used: In this study, three publicly available brain MRI datasets sourced from Kaggle were utilized to test and validate the proposed model.
• Outstanding predictions: The proposed approach demonstrates exceptional results in average precision, recall, and f1-score values of 97% and an accuracy of 97.18% for dataset 1. These outcomes indicate the effectiveness of the optimized CNN model in accurately diagnosing brain tumors.
• Comparative analysis: The study extensively compares our optimized model with established techniques, affirming the strength and reliability of the findings. The proposed method consistently surpasses these approaches, showcasing its superiority in accuracy and reliability when it comes to diagnosing brain tumors.
• Practical implications: This model offers medical professionals a more accurate and effective tool to aid their decision-making in diagnosing brain tumors. By enhancing diagnostic accuracy and reliability, the model has the potential to advance medical imaging and improve patient care.

2. Related Work

3. materials and methods, 3.1. mri dataset, 3.2. pre-processing, 3.3. hyperparameters of ccn for training, 3.4. hyperparametric fine-tuning of cnn, 3.5. working of hyperparameteric cnn, 4.1. evaluation criteria, 4.2. applied model results.

Click here to enlarge figure

5. Discussion

6. limitations of the model and future work, 7. conclusions, author contributions, institutional review board statement, informed consent statement, data availability statement, conflicts of interest.

• Sadoon, T.A.; Ali, M.H. Deep learning model for glioma, meningioma and pituitary classification. Int. J. Adv. Appl. Sci. 2021 , 2252 , 8814. [ Google Scholar ] [ CrossRef ]
• Cheng, J.; Huang, W.; Cao, S.; Yang, R.; Yang, W.; Yun, Z.; Wang, Z.; Feng, Q. Enhanced performance of brain tumor classification via tumor region augmentation and partition. PLoS ONE 2015 , 10 , e0140381. [ Google Scholar ] [ CrossRef ]
• Wen, P.Y.; Kesari, S. Malignant gliomas in adults. N. Engl. J. Med. 2008 , 359 , 492–507. [ Google Scholar ] [ CrossRef ]
• Goodenberger, M.L.; Jenkins, R.B. Genetics of adult glioma. Cancer Genet. 2012 , 205 , 613–621. [ Google Scholar ] [ CrossRef ]
• Laws, E.R. Pituitary Disorders ; Wiley Online Library: Hoboken, NJ, USA, 2013. [ Google Scholar ]
• Devi, O.R.; Bindu, C.S.; Kumar, E.S. Identification and Classification of Brain Tumor from MRI Using Transfer Learning Approach. In Proceedings of the 2022 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS), Greater Noida, India, 4–5 November 2022. [ Google Scholar ]
• Nagaraj, P.; Muneeswaran, V.; Reddy, L.V.; Upendra, P.; Reddy, M.V. Programmed multi-classification of brain tumor images using deep neural network. In Proceedings of the 2020 4th international conference on intelligent computing and control systems (ICICCS), Madurai, India, 13–15 May 2020. [ Google Scholar ]
• Jans, L.B.; Chen, M.; Elewaut, D.; Bosch, F.V.D.; Carron, P.; Jacques, P.; Wittoek, R.; Jaremko, J.L.; Herregods, N. MRI-based synthetic CT in the detection of structural lesions in patients with suspected sacroiliitis: Comparison with MRI. Radiology 2021 , 298 , 343–349. [ Google Scholar ] [ CrossRef ]
• Ammirati, M.; Nahed, B.V.; Andrews, D.; Chen, C.C.; Olson, J.J. Congress of neurological surgeons systematic review and evidence-based guidelines on treatment options for adults with multiple metastatic brain tumors. Neurosurgery 2019 , 84 , E180–E182. [ Google Scholar ] [ CrossRef ]
• Wahlang, I.; Maji, A.K.; Saha, G.; Chakrabarti, P.; Jasinski, M.; Leonowicz, Z.; Jasinska, E. Brain magnetic resonance imaging classification using deep learning architectures with gender and age. Sensors 2022 , 22 , 1766. [ Google Scholar ] [ CrossRef ]
• Alanazi, M.F.; Ali, M.U.; Hussain, S.J.; Zafar, A.; Mohatram, M.; Irfan, M.; AlRuwaili, R.; Alruwaili, M.; Ali, N.H.; Albarrak, A.M. Brain tumor/mass classification framework using magnetic-resonance-imaging-based isolated and developed transfer deep-learning model. Sensors 2022 , 22 , 372. [ Google Scholar ] [ CrossRef ]
• Wohlfahrt, P.; Möhler, C.; Troost, E.G.; Greilich, S.; Richter, C. Dual-energy computed tomography to assess intra-and inter-patient tissue variability for proton treatment planning of patients with brain tumor. Int. J. Radiat. Oncol. Biol. Phys. 2019 , 105 , 504–513. [ Google Scholar ] [ CrossRef ]
• Lohmann, P.; Werner, J.-M.; Shah, N.J.; Fink, G.R.; Langen, K.-J.; Galldiks, N. Combined amino acid positron emission tomography and advanced magnetic resonance imaging in glioma patients. Cancers 2019 , 11 , 153. [ Google Scholar ] [ CrossRef ]
• Hu, R.; Hoch, M.J. Application of diffusion weighted imaging and diffusion tensor imaging in the pretreatment and post-treatment of brain tumor. Radiol. Clin. 2021 , 59 , 335–347. [ Google Scholar ] [ CrossRef ]
• Branzoli, F.; Pontoizeau, C.; Tchara, L.; Di Stefano, A.L.; Kamoun, A.; Deelchand, D.K.; Valabrègue, R.; Lehéricy, S.; Sanson, M.; Ottolenghi, C.; et al. Cystathionine as a marker for 1p/19q codeleted gliomas by in vivo magnetic resonance spectroscopy. Neuro-Oncology 2019 , 21 , 765–774. [ Google Scholar ] [ CrossRef ]
• Islam, S.; Udhayakumar, R.; Kar, S.; Pavitha, N.; Aswal, U.S.; Saini, D.K. Enhancing Brain Tumor Detection Classification by Computational Intelligence. In Proceedings of the 2023 Second International Conference on Electronics and Renewable Systems (ICEARS), Tuticorin, India, 2–4 March 2023. [ Google Scholar ]
• Chakraborty, P.; Das, S.S.; Dey, A.; Chakraborty, A.; Bhattacharyya, C.; Kandimalla, R.; Mukherjee, B.; Gopalakrishnan, A.V.; Singh, S.K.; Kant, S.; et al. Quantum dots: The cutting-edge nanotheranostics in brain cancer management. J. Control. Release 2022 , 350 , 698–715. [ Google Scholar ] [ CrossRef ]
• Biratu, E.S.; Schwenker, F.; Ayano, Y.M.; Debelee, T.G. A survey of brain tumor segmentation and classification algorithms. J. Imaging 2021 , 7 , 179. [ Google Scholar ] [ CrossRef ]
• Asiri, A.A.; Shaf, A.; Ali, T.; Aamir, M.; Irfan, M.; Alqahtani, S. Enhancing brain tumor diagnosis: An optimized CNN hyperparameter model for improved accuracy and reliability. PeerJ Comput. Sci. 2024 , 10 , e1878. [ Google Scholar ] [ CrossRef ]
• Images, B.M. Available online: https://github.com/muhammadaamir1234/Dataset-Link (accessed on 7 May 2024).
• ZainEldin, H.; Gamel, S.A.; El-Kenawy, E.-S.M.; Alharbi, A.H.; Khafaga, D.S.; Ibrahim, A.; Talaat, F.M. Brain tumor detection and classification using deep learning and sine-cosine fitness grey wolf optimization. Bioengineering 2022 , 10 , 18. [ Google Scholar ] [ CrossRef ]
• Umirzakova, S.; Mardieva, S.; Muksimova, S.; Ahmad, S.; Whangbo, T. Enhancing the Super-Resolution of Medical Images: Introducing the Deep Residual Feature Distillation Channel Attention Network for Optimized Performance and Efficiency. Bioengineering 2023 , 10 , 1332. [ Google Scholar ] [ CrossRef ]
• Ait Amou, M.; Xia, K.; Kamhi, S.; Mouhafid, M. A novel MRI diagnosis method for brain tumor classification based on CNN and Bayesian Optimization. Healthcare 2022 , 10 , 494. [ Google Scholar ] [ CrossRef ]
• Rasool, M.; Ismail, N.A.; Al-Dhaqm, A.; Yafooz, W.M.S.; Alsaeedi, A. A novel approach for classifying brain tumours combining a squeezenet model with svm and fine-tuning. Electronics 2022 , 12 , 149. [ Google Scholar ] [ CrossRef ]
• Balakrishnan, R.; Hernández, M.D.C.V.; Farrall, A.J. Automatic segmentation of white matter hyperintensities from brain magnetic resonance images in the era of deep learning and big data—A systematic review. Comput. Med. Imaging Graph. 2021 , 88 , 101867. [ Google Scholar ] [ CrossRef ]
• Piórkowski, A.; Lasek, J. Evaluation of local thresholding algorithms for segmentation of white matter hyperintensities in magnetic resonance images of the brain. In Proceedings of the Applied Informatics: Fourth International Conference, ICAI 2021, Buenos Aires, Argentina, 28–30 October 2021; Proceedings 4. Springer International Publishing: Cham, Swizterland, 2021; pp. 331–345. [ Google Scholar ]
• Reyes, D.; Sánchez, J. Performance of convolutional neural networks for the classification of brain tumors using magnetic resonance imaging. Heliyon 2024 , 10 , e25468. [ Google Scholar ] [ CrossRef ]
• Kurdi, S.Z.; Ali, M.H.; Jaber, M.M.; Saba, T.; Rehman, A.; Damaševičius, R. Brain tumor classification using meta-heuristic optimized convolutional neural networks. J. Pers. Med. 2023 , 13 , 181. [ Google Scholar ] [ CrossRef ]
• Takahashi, S.; Takahashi, M.; Kinoshita, M.; Miyake, M.; Kawaguchi, R.; Shinojima, N.; Mukasa, A.; Saito, K.; Nagane, M.; Otani, R.; et al. Fine-tuning approach for segmentation of gliomas in brain magnetic resonance images with a machine learning method to normalize image differences among facilities. Cancers 2021 , 13 , 1415. [ Google Scholar ] [ CrossRef ]
• Saba, T.; Mohamed, A.S.; El-Affendi, M.; Amin, J.; Sharif, M. Brain tumor detection using fusion of hand crafted and deep learning features. Cogn. Syst. Res. 2020 , 59 , 221–230. [ Google Scholar ] [ CrossRef ]
• Amin, J.; Sharif, M.; Raza, M.; Saba, T.; Anjum, M.A. Brain tumor detection using statistical and machine learning method. Comput. Methods Programs Biomed. 2019 , 177 , 69–79. [ Google Scholar ] [ CrossRef ]
• Amin, J.; Sharif, M.; Yasmin, M.; Fernandes, S.L. Big data analysis for brain tumor detection: Deep convolutional neural networks. Future Gener. Comput. Syst. 2018 , 87 , 290–297. [ Google Scholar ] [ CrossRef ]
• Amin, J.; Sharif, M.; Yasmin, M.; Fernandes, S.L. A distinctive approach in brain tumor detection and classification using MRI. Pattern Recognit. Lett. 2020 , 139 , 118–127. [ Google Scholar ] [ CrossRef ]
• Toğaçar, M.; Ergen, B.; Cömert, Z. BrainMRNet: Brain tumor detection using magnetic resonance images with a novel convolutional neural network model. Med. Hypotheses 2020 , 134 , 109531. [ Google Scholar ] [ CrossRef ]
• Shakeel, P.M.; Tobely, T.E.E.; Al-Feel, H.; Manogaran, G.; Baskar, S. Neural network based brain tumor detection using wireless infrared imaging sensor. IEEE Access 2019 , 7 , 5577–5588. [ Google Scholar ] [ CrossRef ]
• Hossain, T.; Shishir, F.S.; Ashraf, M.; Al Nasim, M.A.; Shah, F.M. Brain tumor detection using convolutional neural network. In Proceedings of the 2019 1st international conference on advances in science, engineering and robotics technology (ICASERT), Dhaka, Bangladesh, 3–5 May 2019. [ Google Scholar ]
• Özyurt, F.; Sert, E.; Avci, E.; Dogantekin, E. Brain tumor detection based on Convolutional Neural Network with neutrosophic expert maximum fuzzy sure entropy. Measurement 2019 , 147 , 106830. [ Google Scholar ] [ CrossRef ]
• Sajid, S.; Hussain, S.; Sarwar, A. Brain tumor detection and segmentation in MR images using deep learning. Arab. J. Sci. Eng. 2019 , 44 , 9249–9261. [ Google Scholar ] [ CrossRef ]
• Nazir, M.; Shakil, S.; Khurshid, K. Role of deep learning in brain tumor detection and classification (2015 to 2020): A review. Comput. Med. Imaging Graph. 2021 , 91 , 101940. [ Google Scholar ] [ CrossRef ]
• Woźniak, M.; Siłka, J.; Wieczorek, M. Deep neural network correlation learning mechanism for CT brain tumor detection. Neural Comput. Appl. 2023 , 35 , 14611–14626. [ Google Scholar ] [ CrossRef ]
• Majib, M.S.; Rahman, M.; Sazzad, T.M.S.; Khan, N.I.; Dey, S.K. Vgg-scnet: A vgg net-based deep learning framework for brain tumor detection on mri images. IEEE Access 2021 , 9 , 116942–116952. [ Google Scholar ] [ CrossRef ]
• Alnowami, M.; Taha, E.; Alsebaeai, S.; Anwar, S.M.; Alhawsawi, A. MR image normalization dilemma and the accuracy of brain tumor classification model. J. Radiat. Res. Appl. Sci. 2022 , 15 , 33–39. [ Google Scholar ] [ CrossRef ]
• Shehab, L.H.; Fahmy, O.M.; Gasser, S.M.; El-Mahallawy, M.S. An efficient brain tumor image segmentation based on deep residual networks (ResNets). J. King Saud Univ.-Eng. Sci. 2021 , 33 , 404–412. [ Google Scholar ] [ CrossRef ]
• Sankareswaran, S.P.; Krishnan, M. Unsupervised end-to-end brain tumor magnetic resonance image registration using RBCNN: Rigid transformation, B-spline transformation and convolutional neural network. Curr. Med. Imaging 2022 , 18 , 387–397. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Shukla, M.; Sharma, K.K. A comparative study to detect tumor in brain MRI images using clustering algorithms. In Proceedings of the 2020 2nd International Conference on Innovative Mechanisms for Industry Applications (ICIMIA), Bangalore, India, 5–7 March 2020. [ Google Scholar ]
• Minarno, A.E.; Mandiri, M.H.C.; Munarko, Y.; Hariyady, H. Convolutional neural network with hyperparameter tuning for brain tumor classification. Kinetik Game Technol. Inf. Syst. Comput. Netw. Comput. Electron. Control 2021 , 6 . [ Google Scholar ] [ CrossRef ]
• Latif, G.; Iskandar, D.A.; Alghazo, J.; Butt, M.M. Brain MR Image Classification for Glioma tumor detection using deep convolutional neural network features. Curr. Med. Imaging 2021 , 17 , 56–63. [ Google Scholar ]
• Grampurohit, S.; Shalavadi, V.; Dhotargavi, V.R.; Kudari, M.; Jolad, S. Brain tumor detection using deep learning models. In Proceedings of the 2020 IEEE India Council International Subsections Conference (INDISCON), Visakhapatnam, India, 3–4 October 2020. [ Google Scholar ]
• Noreen, N.; Palaniappan, S.; Qayyum, A.; Ahmad, I.; Alassafi, M.O. Brain Tumor Classification Based on Fine-Tuned Models and the Ensemble Method. Comput. Mater. Contin. 2021 , 67 , 3967–3982. [ Google Scholar ] [ CrossRef ]
• Asiri, A.A.; Aamir, M.; Ali, T.; Shaf, A.; Irfan, M.; Mehdar, K.M.; Alqhtani, S.M.; Alghamdi, A.H.; Alshamrani, A.F.A.; Alshehri, O.M. Next-Gen brain tumor classification: Pioneering with deep learning and fine-tuned conditional generative adversarial networks. PeerJ Comput. Sci. 2023 , 9 , e1667. [ Google Scholar ] [ CrossRef ]
• Khan, M.A.; Mostafa, R.R.; Zhang, Y.-D.; Baili, J.; Alhaisoni, M.; Tariq, U.; Khan, J.A.; Kim, Y.J.; Cha, J. Deep-Net: Fine-Tuned Deep Neural Network Multi-Features Fusion for Brain Tumor Recognition. Comput. Mater. Contin. 2023 , 76 , 3029–3047. [ Google Scholar ]

Share and Cite

Aamir, M.; Namoun, A.; Munir, S.; Aljohani, N.; Alanazi, M.H.; Alsahafi, Y.; Alotibi, F. Brain Tumor Detection and Classification Using an Optimized Convolutional Neural Network. Diagnostics 2024 , 14 , 1714. https://doi.org/10.3390/diagnostics14161714

Aamir M, Namoun A, Munir S, Aljohani N, Alanazi MH, Alsahafi Y, Alotibi F. Brain Tumor Detection and Classification Using an Optimized Convolutional Neural Network. Diagnostics . 2024; 14(16):1714. https://doi.org/10.3390/diagnostics14161714

Aamir, Muhammad, Abdallah Namoun, Sehrish Munir, Nasser Aljohani, Meshari Huwaytim Alanazi, Yaser Alsahafi, and Faris Alotibi. 2024. "Brain Tumor Detection and Classification Using an Optimized Convolutional Neural Network" Diagnostics 14, no. 16: 1714. https://doi.org/10.3390/diagnostics14161714

Article Metrics

Article access statistics, further information, mdpi initiatives, follow mdpi.

Subscribe to receive issue release notifications and newsletters from MDPI journals

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Published: 09 August 2024

A thousand and one tumors: the promise of AI for cancer biology

• Joy Linyue Fan 1 , 2   na1 ,
• Achille Nazaret 2 , 3   na1 &
• Elham Azizi   ORCID: orcid.org/0000-0001-5059-6971 1 , 2 , 3 , 4 , 5  

Nature Methods volume  21 ,  pages 1403–1406 ( 2024 ) Cite this article

748 Accesses

1 Citations

4 Altmetric

Metrics details

• Cancer genomics
• Machine learning

Breakthroughs in AI and multimodal genomics are unlocking the ability to study the tumor microenvironment. We explore promising machine learning techniques to integrate and interpret high-dimensional data, examine cellular dynamics and unravel gene regulatory mechanisms, ultimately enhancing our understanding of tumor progression and resistance.

This is a preview of subscription content, access via your institution

Access options

Access Nature and 54 other Nature Portfolio journals

Get Nature+, our best-value online-access subscription

24,99 € / 30 days

cancel any time

Subscribe to this journal

Receive 12 print issues and online access

251,40 € per year

only 20,95 € per issue

Buy this article

• Purchase on Springer Link
• Instant access to full article PDF

Prices may be subject to local taxes which are calculated during checkout

Hao, Y. et al. Cell 184 , 3573–3587.e29 (2021).

Article   CAS   PubMed   PubMed Central   Google Scholar  

Gayoso, A. et al. Nat. Biotechnol. 40 , 163–166 (2022).

Article   CAS   PubMed   Google Scholar  

Lange, M. et al. Nat. Methods 19 , 159–170 (2022).

Chari, T. & Pachter, L. PLOS Comput. Biol. 19 , e1011288 (2023).

Nazaret, A. et al. Preprint at bioRxiv https://doi.org/10.1101/2023.11.11.566719 (2023).

Lopez, R. et al. Preprint at arXiv https://doi.org/10.48550/arXiv.1905.02269 (2019).

Boyeau, P. et al. Preprint at bioRxiv https://doi.org/10.1101/2022.10.04.510898 (2024).

Rosen, Y. et al. Preprint at bioRxiv https://doi.org/10.1101/2023.11.28.568918 (2023).

Welch, J. D. et al. Cell 177 , 1873–1887.e17 (2019).

He, S. et al. Nat. Biotechnol. https://doi.org/10.1038/s41587-024-02173-8 (2024).

Liu, Y., Jin, Y., Azizi, E. & Blumberg, A. J. BMC Bioinformatics 24 , 480 (2023).

Article   PubMed   PubMed Central   Google Scholar  

Nazaret, A., Hong, J., Azizi, E. & Blei, D. Preprint at https://doi.org/10.48550/arXiv.2311.10263 (2023).

Hao, M. et al. Nat. Methods https://doi.org/10.1038/s41592-024-02305-7 (2024).

Wang, Y. et al. Nat. Genet. 55 , 19–25 (2023).

Maurer, K. et al. Preprint at bioRxiv https://doi.org/10.1101/2024.02.09.579677 (2024).

Download references

Acknowledgements

We thank J. McFaline-Figueroa for helpful discussions. J.L.F. acknowledges support from the Columbia University Van C. Mow fellowship and the Avanessians doctoral fellowship. A.N. acknowledges support from the Eric & Wendy Schmidt Center Ph.D. Fellowship and the Africk Family Fund. E.A. was supported by US National Institute of Health NCI R00CA230195 and NHGRI R21HG012639, R01HG012875, National Science Foundation CBET 2144542, and grant 2022-253560 from the Chan Zuckerberg Initiative DAF.

Author information

These authors contributed equally: Joy Linyue Fan, Achille Nazaret.

Authors and Affiliations

Department of Biomedical Engineering, Columbia University, New York, NY, USA

Joy Linyue Fan & Elham Azizi

Irving Institute for Cancer Dynamics, Columbia University, New York, NY, USA

Joy Linyue Fan, Achille Nazaret & Elham Azizi

Department of Computer Science, Columbia University, New York, NY, USA

Achille Nazaret & Elham Azizi

Data Science Institute, Columbia University, New York, NY, USA

Elham Azizi

Herbert Irving Comprehensive Cancer Center, Columbia University, New York, NY, USA

You can also search for this author in PubMed   Google Scholar

Contributions

J.L.F., A.N. and E.A. wrote the manuscript. J.L.F. designed the figure.

Corresponding author

Correspondence to Elham Azizi .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Rights and permissions

Reprints and permissions

About this article

Cite this article.

Fan, J.L., Nazaret, A. & Azizi, E. A thousand and one tumors: the promise of AI for cancer biology. Nat Methods 21 , 1403–1406 (2024). https://doi.org/10.1038/s41592-024-02364-w

Download citation

Published : 09 August 2024

Issue Date : August 2024

DOI : https://doi.org/10.1038/s41592-024-02364-w

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

Embedding ai in biology.

Nature Methods (2024)

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing: Cancer newsletter — what matters in cancer research, free to your inbox weekly.

• Weird But True
• Sex & Relationships
• Viral Trends
• Human Interest
• Fashion & Beauty
• Food & Drink
• Health Care
• Men’s Health
• Women’s Health
• Mental Health
• Health & Wellness Products
• Personal Care Products

trending now in Lifestyle

2 foods could be fueling cancer epidemic in young people: docs

Mom ignores all boundaries, sleeps in daughter's dorm on her...

I'm a surgeon — you should go to the ER if you have this...

I lost 55 pounds on Mounjaro in six months — but here's why I...

Costco discontinued a fan-favorite item -- and customers aren't...

Has the ongoing mystery of Stonehenge finally been solved?...

This much caffeine can raise your heart disease risk — even if...

Dear Abby: My family accused me of stealing jewelry that was...

Wearable device with ai could allow for at-home breast cancer screenings: ‘accessible and personalized’.

To provide women at a high risk of breast cancer with more frequent screenings between mammograms, researchers at the Massachusetts Institute of Technology (MIT) are developing a wearable ultrasound scanner designed to be attached to a bra.

The goal is to help women detect breast cancer tumors in the early stages and maximize the survival rate, according to a press release on MIT’s website.

The researchers’ aim was to design a wearable “miniaturized ultrasound device” that allows for “consistent placement and orientation” to take images of breast tissue, according to lead study author Canan Dagdeviren, PhD, associate professor at MIT.

The device attaches to the bra like a patch, with a “honeycomb” pattern that has open spaces for the tracker to move through for an optimal field of view, Dagdeviren told Fox News Digital. 

“The ultrasound generates a wave that penetrates the targeted breast tissue,” he said. 

“When the ultrasound wave sees an obstacle like a tumor, it reflects [that] — and the ultrasound device captures this reflected wave and generates a black-and-white ultrasound image.”

He added, “From this image, you can find the coordination and shape of any anomaly in the soft breast tissue.”

In a study, a woman with a history of breast cysts wore the bra and patch, and the researchers scanned the breast at six different locations according to the patch design. 

The technology recorded the images of the breast tissue, which displayed cysts as small as 0.3 centimeters in diameter  —  the approximate size of tumors at early stages of the disease. 

The device is designed as a complement to traditional screenings by medical professionals rather than a replacement, Dagdeviren said.

“For personal health care and monitoring at home, this device can be considered as a replacement [for] the conventional handheld probe or ABUS system at the early stage of breast tumor imaging,” he said. 

Researchers plan to use artificial intelligence to analyze the ultrasound images and make diagnostic recommendations.

“For breast tumor diagnosis, this device can be a strong complement to existing screening methods, making long-term breast imaging efficient and convenient.”

In the future, researchers plan to use artificial intelligence to analyze the ultrasound images and make diagnostic recommendations.

“The AI integration has the potential to enhance diagnostic accuracy through data analysis, and also holds potential for early detection and personalized care pathways for other diseases,” Dr. Harvey Castro, a Dallas, Texas-based board-certified emergency medicine physician and national speaker on AI in health care, told Fox News Digital.

Castro was not involved in the research but reviewed the details of the device.

It could be three or four years before the device is available to consumers, Dagdeviren said. 

“The critical step is the portable system and wireless communication to the hospital,” he told Fox News Digital.

The company will also need to complete intensive human trials and gain FDA approval — something that Dagdeviren expects to cost around $40 million.

Dagdeviren said he does not foresee any safety risks associated with the device, as it does not use radiation.

Also, the wearable ultrasound patch can be used over and over, the release on MIT website’s noted. It could also help diagnose cancer in people who don’t have regular access to screening.

Limitations of the system

The device requires a long, flexible cable to connect the image processing system, the study’s lead author noted. 

Additionally, the system that collects all of the data is currently large and stationary.

“We are currently working on a portable system, which we hope to publish in a few months,” Dagdeviren said.

The imaging resolution is “sufficient, but not superior,” he noted, and the researchers are working on various image processing methods. 

“In our next paper and patent, all of these limitations will be eliminated in around six months,” Dagdeviren said.

Technology holds ‘potential to save lives’

“The development of this wearable device for breast cancer detection represents a significant advancement in health care technology,” Castro told Fox News Digital. 

“Its potential to save lives and extend its promise to other diseases is immense.”

“It cannot replace traditional mammograms and other preventive care from a breast cancer expert.”

Castro added, “However, careful consideration of its implementation, including rigorous testing, alignment with existing medical protocols, and ethical considerations, will be essential to its success.”

To maximize the potential of this technology, he emphasized the need to “strike a balance between innovation and responsible medical practice.”

There is also the need to ensure the privacy and security of personal health data and to consider “potential disparities in access,” especially if the device becomes a commercial product, Castro noted.

Depending on the user’s skill level, there could also be a risk of inconsistencies and inaccuracies, he pointed out.

“It embodies the convergence of technology and  health care, reflecting a future where medical care is more accessible and personalized,” Castro said. 

“However, it cannot replace traditional mammograms and other preventive care from a breast cancer expert.”

Aside from skin cancers, breast cancer is the most common cancer in U.S. women, representing one in three of all female cancer diagnoses each year, according to the American Cancer Society.

When breast cancer is caught early, the five-year survival rate is 99%. If it is detected in advanced stages, the survival rate is only 25%.

Advertisement

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• HHS Author Manuscripts

Cancer screening and early detection in the 21 st century

Jennifer loud.

Clinical Genetics Branch, DCEG, NCI, NIH 9609 Medical Center Drive Rockville, Maryland 20850-9772

Jeanne Murphy

Breast and Gynecologic Cancer Research Group, DCP, NCI, NIH 9609 Medical Center Drive Rockville, Maryland 20892-9712

To review the trends in and principles of cancer screening and early detection.

Data Sources

Journal articles, United States Preventive Services Task Force (U SPSTF) publications, professional organization position statements, evidence-based summaries

Cancer screening has contributed to decreasing the morbidity and mortality of cancer. Efforts to improve the selection of candidates for cancer screening, to understand the biological basis of carcinogenesis, and the development of new technologies for cancer screening will allow for improvements in the cancer screening over time.

Implications for Nursing Practice

Nurses are well-positioned to lead the implementation of cancer screening recommendations in the 21 st Century through their practice, research, educational efforts and advocacy.

The goal of cancer screening and early detection is to cure cancer by detecting the malignancy, or its precursor lesion, at an early stage prior to the onset of symptoms, when treatment of cancer is most effective. Indeed, overall cancer mortality has decreased by 25% from 1990 to 2015 for the United States U.S.), with even greater declines in the mortality rates for colorectal cancer (47% among men and 44% among women) and, breast cancer (39% among women). A portion of this decrease can be attributed to the introduction of high-quality cancer screening for colorectal and breast cancer. 1 The most successful cancer screening programs lead to the identification of precursor lesions (e.g., cervical intra-epithelial neoplasia (CIN) with cervical cancer screening and colonic polyps with colorectal cancer screening) where the treatment of the precursor lesion leads to a decrease in the incidence of invasive cancer over time. The guiding principles of screening for disease were proposed in 1968 by Wilson and Jungner 2 of the World Health Organization ( Table 1 ). Not all cancer screening recommendations meet each of these guiding principles; historically there has been a balance between the identification of early or precursor lesions and the avoidance of overdiagnosis which may lead to overtreatment ( Table 2 ).

Wilson and Jungner Criteria for disease screening 2

Potential negative outcomes of cancer screening.

Application of Cancer Screening Principles

U.S. population screening for cervical cancer serves as an exemplar of a successfully designed and implemented screening program that has been modified as the biological mechanism of the carcinogenesis of cervical cancer is more clearly elucidated and methods for primary prevention (i.e., HPV vaccination) are developed. Cervical cancer screening programs in particular adhere to several of Wilson and Junger’s principles, most importantly, that the natural history of the disease be understood and that it be an important health problem. Chronic human papilloma virus (HPV) infection is the underlying etiologic agent of the carcinogenesis of cervical cancer. Chronic HPV leads to a precancerous lesion (i.e., cervical intra-epithelial neoplasia) which can be visualized, after the detection of a positive cytology (through Pap testing), with colposcopy. The removal of the precancerous lesion using colposcopy successfully led to an overall decrease in the incidence of cervical cancers, even though there was over treatment of some early lesions. Cervical cancer screening represents an example of the use of an accurate screening test (i.e., PAP, colposcopy and now HPV testing) with adequate sensitive, specificity and positive and negative predictive value (PPV and PNV) leading to the identification of a high risk population, a pre-cancer or a cancer ( Tables 3 and ​ and4). 4 ). Population screening for colon cancer also conformed to many of Wilson and Jungner’s principles and led to improvements in overall survival of individuals who adopted screening recommendations. 1 A key feature of both cervical and colon cancer screening is the ability to directly access the tissue of interest and apply an adequate screening test. Population screening for cervical cancer reduced the incidence and mortality rates from cervical cancer and led to enthusiasm that screening programs for other cancers, or pre-cancers, would be equally successful. However, screening, detection and removal of pre-cancer or early cancer in other cancer types has not always been as successful.

Characteristics of an accurate screening test.

Performance characteristics of a screening test

A major assumption about the natural history of carcinogenesis is based on the models of carcinogenesis of colorectal cancer proposed by Vogelstein et.al. 3 The model predicted a slow-growing, linear progression from a pre-cancer to a localized cancer that would occur at a rate of time that was amendable to cancer screening, similar to the pattern of carcinogenesis observed in cervical cancer. It also assumed that there was similarity within cancer types, such that all prostate or breast cancers behaved similarly. Based on that assumption, population-based screening programs for other solid tumors were developed including breast and prostate cancer screening. However, outcomes from multiple screening programs between 1980–2010 demonstrated that breast and prostate cancers are a heterogeneous group of diseases that do not necessarily conform to the pattern of carcinogenesis as initially proposed in the Vogelstein model. 4 After population screening was introduced for breast and prostate cancer and outcomes documented overtime, lessons learned ( Table 5 ) included that

Lessons learned from population screening for breast, prostate and colon cancer

• Breast and prostate cancers were not uniform in their biology (they are heterogeneous)
• Not all early lesions (i.e., ductal carcinoma insitu or indolent prostate cancer) lead to invasive cancer
• Early detection does not always lead to improvements in overall survival, and
• There is risk to individuals when introducing screening interventions in otherwise healthy populations, including overdiagnosis and overtreatment ( Table 2 )

In addition, other cancer screening techniques rely on indirect methods to screen for cancer such as radiographic imaging (e.g., mammography) or measuring a biomarker associated with cancer (e.g., CA-125 or PSA), rather than direct visualization and access to the target organ as in colorectal and cervical cancer screening. These indirect methods of cancer screening led to compromised screening efficacy due a decrease in performance characteristics of the screening technique [(including false positives and false negatives ( Table 6 )] and an increase in overdiagnosis and overtreatment. 4 As more evidence of screening efficacy accumulates, changes in cancer screening recommendations and practice continue to occur. Prostate cancer screening guidelines changed to include shared decision-making as it became evident that the risk-to-benefit ratio of routine prostate cancer screening in men over the age of 50 was unfavorable; routine prostate cancer screening led to overdiagnosis of indolent cancer without a survival benefit while placing men at greater risk of injury related to the treatment of indolent prostate cancer. 5

Possible test outcomes of cancer screening

Improving the Precision of Candidates for Cancer Screening

Ideally, cancer screening is undertaken when the risk of cancer is high enough to justify the risk of overdiagnosis and overtreatment in an otherwise healthy population. 6 Cancer screening in healthy populations balances patient tolerance of risk, personal attitudes and the choice of a screening program most likely to have net benefit to the individual. In low-to-average risk populations, the recommended age to begin routine cancer screening is the age at which the risk of cancer begins to rise (e.g., 50 years for colorectal cancer screening) and when the tumor develops slowly. Slow tumor progression allows for the identification of a malignancy (or pre-malignancy) at an early stage which reduces the incidence of late stage cancer. For instance, the optimal screening interval for colorectal cancer screening with colonoscopy in the general population is 10 years, which allows for the removal of the pre-cancerous lesion, the adenomatous polyp, thereby reducing colon cancer. Cancer screening does not work as effectively for rapidly growing tumors or those that disseminate early, as they tend to occur between screening intervals and present with symptoms.

Integrating exposure history is commonly used to improve the identification of individuals at higher risk of cancer than the general population. 4 Targeting smokers with a 30 pack-year for low-dose chest tomography (CT) to screen for early lung cancer and identifying women with HPV infection to define a high risk population at risk of cervical cancer demonstrate efforts to use risk stratification in order to offer screening to individuals most likely to benefit and reduce screening in low risk individuals.

Risk-prediction models attempt to identify individuals at higher risk of cancer than the general population. The Breast Cancer Risk Assessment Tool 7 was one of the first tools aimed at identifying women who could benefit from breast cancer chemoprevention trials and accounts for clinical risk factors (i.e., family history, personal history, breast biopsy) as well as hormonal exposures (i.e., age of menarche). More recent risk-prediction models incorporate exposures (i.e., radiation exposure), breast density as well as biomarkers (i.e., single nucleotide polymorphisms) in an effort to improve risk-stratification. 8

The contribution of genetics and genomics to risk-stratification has steadily progressed since the identification of the germline p53 mutation in Li Fraumeni Syndrome. 9 , 10 The ability to identify individuals who carry a germline mutation associated with a hereditary cancer syndrome greatly improves risk-stratification and helps identify those individuals who may benefit from more frequent cancer screening and other preventive procedures. For example, individuals at high risk of cancer due to inherited cancer susceptibility (such as carrier of a BRCA1 or BRCA2 mutation) undergo aggressive cancer screening for the tumors associated with the syndrome and may also consider prophylactic surgery to reduce their risk of cancer. Within a family with a known BRCA1 mutation, those family members who did not inherit the mutation do not need to undergo intensive screening nor do they need to consider prophylactic surgery to prevent cancer. As the expense of genetic sequencing decrease, there is an increase in the use of genetic testing panels and other genomic technologies for risk stratification. However, important clinical challenges exist with these technologies regarding the classification of the identified genetic variants, reporting of the variants or unknown significance and how to handle incidental findings. 11 Multiple organizations have developed standards and guidelines for interpreting sequence variants and conclude that clinical genetic tests should be performed in Clinical Laboratory Improvement Amendments (CLIA)-approved labs and the results should be interpreted by a board certified clinical molecular geneticist, a molecular genetic pathologist or the equivalent. 12

When it is not so Simple to Screen: ovarian cancer

Ovarian cancer is rare, with incidence of 11.9 per 100,000, and a 5-year survival rate of only 46%. 13 It is also the most lethal of all cancers of the female reproductive system. 14 Recent evidence suggests that high-grade serous ovarian cancer, the most common and dangerous type, actually arises from malignant cells in the fimbriated end of the fallopian tube. 15 Much of this lethality is due to the difficulty of diagnosis because ovarian cancer’s vague symptoms include bloating, abdominal fullness and pain, and fatigue. 16 This leads to delayed detection, with 60% of cases diagnosed at a late stage with distant metastasis. 13 The median age at ovarian cancer diagnosis is 63, and is more common among women with a family history. Since 1975, 5-year survival has increased from 33.7% in 1975 to 46% in 2008. 13

Given its lethality, it is essential to develop effective screening strategies for ovarian cancer in order to intervene earlier in the process of disease. The challenge of ovarian cancer screening lies with the site. Unlike the uterine cervix, whose cells can be sampled directly through cervical cytology or by testing for human papillomavirus, 17 the ovaries and fallopian tubes lie deep in the pelvis, making them inaccessible to routine evaluation. This is especially problematic for asymptomatic women with germline mutations in BRCA1 or BRCA2 that place them at much higher risk of ovarian cancer (lifetime risk of 10%–25% for BRCA1 or BRCA2 vs 1.7% for the general population), but it is also problematic for the general population. 16 Ovarian cancer risk in a high-risk population can be determined through taking a careful family history, and this is a reasonable and inexpensive “Precision Public Health” intervention. 18 Population-based genetic testing for hereditary breast and ovarian cancer, called for by Mary Claire King in an opinion piece published as she accepted the 2014 Lasker Award from the National Institutes of Health 19 may identify more women who can benefit from targeted ovarian cancer screening strategies, though there is no consensus for this recommendation to date.

For asymptomatic, low risk women, strategies for ovarian cancer screening have included direct examination through bimanual examination during pelvic examination, and visualization through transvaginal (TV) ultrasonography and Doppler studies. 20 , 21 Both approaches attempt to evaluate the ovaries for abnormal, possibly cancerous, masses. Despite its recommended use, bimanual examination suffers from low sensitivity for both adnexal masses in general 22 and for ovarian cancer specifically, 23 and is associated with harms from false positive results resulting in unnecessary surgical biopsies. 24 Currently some have begun to question the inclusion of the bimanual examination in primary care guidelines as a screening test for ovarian masses. 25 Similarly, a one-time transvaginal ultrasound of asymptomatic women did not result in reduction in ovarian cancer mortality in the United Kingdom Collaborative Trial of Ovarian Cancer Screening (UKTOCS) and is not recommended as a stand-alone screening test for ovarian cancer. 26

Serum biomarkers such as CA-125 and others have been tested for efficacy in screening for ovarian cancer. CA-125, also known as MUC16, is a large glycoprotein membrane marker from the MUC family found on ovarian cancer cells, but it is not specific to them. 27 , 28 Serum levels of CA-125 are elevated in ovarian cancer and many non-cancerous conditions such as ovarian cysts and liver cirrhosis, and also in non-ovarian malignancies. 29 CA-125 as a standalone screening test is relatively insensitive for ovarian cancer, finding only about 60% of women with ovarian cancer. 30 Other serum biomarkers such as human epididymis protein (HE4) and human chorionic gonadotropin (HCG) have been tested in combination with CA-125 to improve performance characteristics of serum biomarker screening for ovarian cancer as standalone serum screening tests, 29 , 31 though evidence suggests that CA-125 is the most robust biomarker of the group. 32

The most promising approach for ovarian cancer screening is a strategy combining serum CA-125, with or without other biomarkers, and TV ultrasound. The UKTOCS in the UK 26 and the Prostate, Lung, Colorectal and Ovarian Cancer Screening (PLCO) trial in the US 33 tested similar strategies. Despite its promise, this co-testing strategy has not resulted in overall reduction in mortality due to ovarian cancer. 26 , 33 The UK trial tested a proprietary algorithm named ROCA ® that adjusted the biomarker level cut-off for normal results based on women’s clinical characteristics and the TV ultrasound result. 31 The promotion of ROCA ® (Abcodia, Cambridgeshire, UK) serum testing with TV ultrasound ran afoul of the Food and Drug Administration (FDA) for the claim that the ROCA ® test detects ovarian cancer early and reduced mortality. In late 2016 FDA issued a warning against using commercial screening tests for ovarian cancer, saying that, especially for women at high risk for hereditary ovarian cancer, “women and their doctors may not take appropriate actions to reduce their future risk if they rely on a result that shows no cancer currently present.” 34 FDA further stipulated that they did not recommend the use of ovarian cancer screening tests in the general population. 34

The history of ovarian cancer screening is a cautionary tale for nurses in considering the use of screening tests in low risk populations. It also highlights the importance of understanding the potential for harm with using what may prove, with more evidence, to be effective screening strategies that save lives.

Improving the Infrastructure for Cancer Screening

Continued progress to reduce death rates from cancer in the United States will only be achieved if there is broad commitment to understanding the determinants of cancer, including access to care, affordability, and social and environmental factors associated with cancer risk. 1 National cancer registries, linked to cancer screening programs, can support detailed cohort studies to improve outcomes research leading to quality improvements in cancer screening programs. Indeed, the Breast Cancer Screening Consortium 4 has linked data from regional mammography registries to increase the diversity of their sample populations and the American College of Radiology’s national lung cancer screening aims to develop outcomes-based research in support of quality improvements. Such efforts support evidence-based practices and will allow for continuous process improvement in outcomes of cancer screening and research methodologies.

The selection of ideal candidates to screen or not screen is an understudied area ripe for future research. As individuals age and acquire co-morbidities (competing risks), the balance between risk and benefit of screening may shift in favor of increased risk with limited- to no-benefit. One risk prediction model, e-Prognosis ( http:eprognosis.ucsf.edu/ ) uses age and specific health measures to predict overall survival at different ages. Future research will address the utility of these tools across all cancer screening recommendations to identify those who will benefit most from screening and those most likely to be harmed. 35

The translation of cancer screening research into effective public health policy requires nurses to be cognizant of the multiple levels of policy complexity. 36 As evidence of screening efficacy is demonstrated through research, healthcare legislation requires insurance coverage for screening recommendations developed by the United States Preventive Services Task Force (USPSTF). Changes in screening recommendations by the USPSTF can ignite professional, public and political controversy as evidenced by the debate surrounding the revised 2009 Task Force recommendation for breast cancer screening. 37 All healthcare providers should plan to effectively communicate the scientific underpinnings of new research and the potential for cultural, political and policy implications. A well-developed communication plan incorporates a review of the research, the basis of the recommendation and the implications of the research for all stakeholders (including the public, politicians and policymakers). Nurses play an essential role in the dissemination of research and the evaluation and implementation of new cancer screening programs to the public and other stakeholders.

Cancer screening practice in the 21 st century will integrate genomics, risk prediction, patient preferences and improvements in health care delivery systems into patient care services. Essential nursing functions will continue to be in high demand as the aging population of the United States increases and more individuals have access to care ( Table 7 ). Nurses will lead the transformation of cancer care in all healthcare settings and work to ensure that all patients receive high quality cancer care. 38 Cancer screening recommendations have been shown to significantly decrease the mortality from certain cancers (i.e., cervical and colorectal), while more modestly decreasing mortality of others. At every point of care, and every level of practice, nurses will improve cancer screening through their interactions with patients and families to increase understanding of the rationale for and importance of adherence to cancer screening recommendations. As always, nurses will continue to follow the evidence for practice to maintain nursing practice at the state-of-the-art of cancer screening and advocate in support of public policies that expand access to care.

Nursing actions in support of cancer screening

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Contributor Information

Jennifer Loud, Clinical Genetics Branch, DCEG, NCI, NIH 9609 Medical Center Drive Rockville, Maryland 20850-9772.

Jeanne Murphy, Breast and Gynecologic Cancer Research Group, DCP, NCI, NIH 9609 Medical Center Drive Rockville, Maryland 20892-9712.

Official websites use .gov

A .gov website belongs to an official government organization in the United States.

Secure .gov websites use HTTPS

A lock ( ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites.

Health Care Use Among Cancer Patients With Diabetes, National Health and Nutrition Examination Survey, 2017–2020

ORIGINAL RESEARCH — Volume 21 — August 8, 2024

Ara Jo, PhD 1 ; Sarina Parikh, BHS 2 ,3 ; Nathalie Sawczuk, MHA 1 ; Kea Turner, PhD, MPH, MA 4 ,5 ; Young-Rock Hong, PhD, MPH 1 ( View author affiliations )

Suggested citation for this article: Jo A, Parikh S, Sawczuk N, Turner K, Hong Y. Health Care Use Among Cancer Patients With Diabetes, National Health and Nutrition Examination Survey, 2017–2020. Prev Chronic Dis 2024;21:240066. DOI: http://dx.doi.org/10.5888/pcd21.240066 .

PEER REVIEWED

Introduction

Acknowledgments, author information.

What is already known on this topic?

Cancer patients with multiple chronic diseases have unplanned hospitalizations because of a lack of appropriate care management. Multiple chronic diseases among people with cancer are associated with worse clinical outcomes and survivorship than among people with cancer only.

What is added by this report?

Patients with cancer and prediabetes had higher levels of health care use than patients with cancer only. A diagnosis of type 2 diabetes did not significantly affect health care use among patients with cancer.

What are the implications for public health practice?

Optimal care coordination and early management of prediabetes among patients with cancer via primary care may contribute to improving cancer survivorship.

Diabetes is a common comorbidity among people with cancer. The objective of our study was to examine patterns of health care use among patients with cancer and either type 2 diabetes or prediabetes.

We used data from the National Health and Nutrition Examination Survey (NHANES) for 2017–2020. The study population included US adults aged 18 years or older who were diagnosed with any cancer and type 2 diabetes or prediabetes (established by self-report and/or hemoglobin A 1c measurement). We used Poisson and multivariate logistic regression models to determine the effect of comorbidity on health care use, defined as health care visits and overnight stays in a hospital.

Of 905 cancer patients representing 27,180,715 people in the US, 24.4% had a type 2 diabetes diagnosis, and 25.8% had a prediabetes diagnosis. Patients with cancer and prediabetes had a significantly higher rate of health care visits (incidence rate ratio = 1.11; 95% CI, 1.01–1.22; P = .03) than patients with cancer only. We found no significant association between having cancer and type 2 diabetes and the number of health care visits or overnight hospital stays compared with patients with cancer only.

More emphasis should be placed on optimal care coordination among people with cancer and other conditions, such as diabetes and prediabetes, to reduce the impact of comorbidity on health care use. Interventions integrated with technology to provide timely access to education on preventing or managing diabetes and prediabetes among cancer patients are warranted.

Diabetes is a common comorbidity among people with cancer. As patients with cancer live longer due to advances in cancer treatment, rates of chronic conditions, such as diabetes, are expected to rise among people with cancer. People with type 2 diabetes (hereinafter, diabetes) have a substantially higher risk of cancer incidence and death, leading to poorer survivorship compared with people without diabetes (1,2). For example, people with diabetes, compared with people who do not have diabetes, have double the risk for liver and pancreatic cancers and have a higher risk of developing bladder, colon, and breast cancers (3). In addition, as cancer incidence and death rates have risen consistently over time, the comorbidity of cancer with other chronic diseases has gained attention (4,5). Despite these clinical outcomes, the research is limited on care delivery for people with cancer and other comorbidities.

People with cancer and comorbidities, compared with those who have cancer and no comorbidities, have greater unplanned use of health care services, including higher rates of unplanned hospital readmissions (6,7) and revisits to the emergency department (8). One study showed that among people with cancer and comorbidities, diabetes was the top reason for emergency department revisits (24% of all revisit encounters) (8). Another study found that the average length of hospital stay among people with cancer and diabetes was significantly longer than among patients with no comorbidity (9). In that study, the average length of a hospital stay among patients with colorectal cancer and diabetes who underwent surgery was almost 17 days, which is 3 days longer than among patients with cancer only (9). Furthermore, health care costs are of critical concern. A national study, which used 5 years of data from the Medical Expenditure Panel Survey (2010–2014), found that cancer patients spent on average 4 times more in annual health expenditures than noncancer patients (10). Early initiation of chronic disease prevention and management with a primary care physician can mitigate this financial burden.

Many patients with cancer face the challenges of comanaging cancer and chronic diseases. In a qualitative study conducted in 2021 and 2022 at 3 New York City hospitals among 15 women with breast cancer and either diabetes or prediabetes, participants reported a lack of information and education on managing chronic diseases and the burden of co-management with different providers (11). In addition, patients tended to prioritize cancer treatment over diabetes management with their primary care physician (11). These struggles may be more detrimental for patients who are at a higher-than-average risk of developing diabetes. For example, a national cohort study in Korea found that a diagnosis of cancer increased the risk of subsequent diabetes (12). A case-cohort study in Israel that investigated the association between hormone therapy and diabetes risk among 2,246 female breast cancer survivors found that 48% of diabetes incidence could have been prevented had patients not received hormone therapy (13). Early implementation of a diabetes prevention strategy, particularly for patients with cancer and prediabetes, elevated blood glucose, or active engagement with a primary care physician during cancer treatment, could prevent comorbidity and improve survivorship. Furthermore, cancer treatments such as chemotherapy, radiation, or immunotherapy are associated with a higher prevalence of prediabetes (14).

Comorbidities or complications associated with cancer are linked to increased health care costs and various kinds of health care use, including ambulatory care visits and emergency department visits (15,16). However, evidence that focuses on the effects of specific kinds of comorbidity, such as diabetes, on health care use is limited. One study that used data from a statewide electronic health record database from 2007 to 2017 in the US found a significant association of having both diabetes and colorectal cancer with emergency department visits but did not examine other outcomes, such as hospitalization, which is a major driver of health care costs (17). Furthermore, little is known about how patterns of health care use differ across stages of diabetes. Addressing these gaps may help to improve the delivery of effective clinical care and preventive services for people with cancer and diabetes.

The objective of this study was to examine the association of health care use patterns among patients with cancer, stratified by diagnosis of diabetes or prediabetes. Findings from the current study may guide research to develop an optimal coordinated care model for early detection of prediabetes or diabetes and to enhance cancer survivorship for people with cancer and comorbidities.

Our study used a cross-sectional design and data from the National Health and Nutrition Examination Survey (NHANES) for the 3-year cycle of 2017–2020, before the pandemic. NHANES has been conducted since 1960 and is designed to assess the health and nutritional status of adults and children in the US. It collects nationally representative data through clinical examinations, selected medical and laboratory tests, and self-reported data. NHANES uses a stratified, multistage probability sample design and recommends using weights, stratification, and cluster variables to account for the complex sample design (18). Thus, we applied these variables to the statistical analyses to generate population estimates.

Study population

Our study population comprised adults aged 18 years or older who were diagnosed with any cancer and had physician-diagnosed diabetes or prediabetes. Those with a cancer history were identified by using the question, “Have you ever been told by a doctor or other health professional that you had cancer or a malignancy of any kind?” Physician-diagnosed diabetes and prediabetes were identified through self-report on the NHANES questionnaire. In addition, to reduce the risk of recall bias, we used NHANES laboratory results of the hemoglobin A 1c (HbA 1c ) test. We excluded data on undiagnosed diabetes because the sample size was too small for generating population estimates. We classified people into 3 categories: 1) those with a cancer history only, 2) those with any cancer history and prediabetes, and 3) those with any cancer history and diabetes. We excluded records that had missing data for these variables.

A primary outcome was the number of visits to a physician’s office, a clinic, or “some other place” in the previous 12 months. This visit did not include hospitalizations, emergency department visits, home visits, or telephone calls. A secondary outcome was an overnight stay in a hospital in the previous 12 months. It excluded overnight stays in the emergency department.

Independent variable

A primary independent variable was comorbidity status. We categorized the study population into 3 groups: 1) cancer only, 2) cancer and prediabetes, and 3) cancer and diabetes. Control variables were demographic characteristics (age, sex, and race and ethnicity), education, body mass index (BMI), and having a usual source of care (yes or no). We treated age as a continuous variable. Sex was a dichotomous variable (male or female). We categorized race and ethnicity into 4 categories: 1) Hispanic or Latino, 2) non-Hispanic Black, 3) non-Hispanic White, and 4) Other (American Indian or Alaska Native, Asian, and Native Hawaiian or Pacific Islander) or multiracial. We converted education into a dichotomous variable (less than high school and high school graduate or above). Financial status was measured by the ratio of income to poverty (total family income divided by the poverty threshold) and dichotomized into 2 levels: 1) poor (ratio <1) and 2) rich (ratio ≥1). Health status was measured by self-reported general health condition and grouped into 2 levels: 1) fair or above (excellent, very good, good, or fair) and 2) poor. BMI was categorized into 3 levels: 1) normal (BMI, 18.5–24.9), 2) overweight (25.0–29.9), and 3) obese (≥30.0). Health insurance status was categorized into 2 levels: 1) yes, insured, and 2) no, uninsured. We excluded underweight people due to a high risk of mortality and little relevance to our study. Lastly, we treated usual source of care as a dichotomous variable (has a usual source or does not have a usual source of care). We counted the number of other chronic diseases reported by the survey respondent, such as arthritis, cancer (if the respondent has ≥1 cancers), cardiovascular diseases (eg, congestive heart failure, coronary heart disease, angina, or stroke), chronic kidney disease, depression, hypertension, and pulmonary diseases (eg, emphysema, chronic bronchitis, or asthma). We categorized these data into 4 groups: 1) no other comorbidity, 2) 1 additional comorbidity, 3) 2 additional comorbidities, and 4) ≥3 additional comorbidities.

Statistical analysis

We conducted a descriptive analysis of the baseline characteristics of the 3 groups of NHANES respondents (cancer only, cancer and prediabetes, and cancer and diabetes). We used χ 2 tests and t tests to determine significant differences between groups, with P < .05 considered significant. We used a Poisson regression model to determine the effect of comorbidity status (cancer only, cancer and prediabetes, and cancer and diabetes) on the number of health care visits in the previous 12 months. We used a multivariate logistic regression model to examine the risk of an overnight hospital stay associated with comorbidity status. We conducted both unadjusted and adjusted models. The Poisson regression model produced incident rate ratios (IRRs) and 95% CIs, and the multivariate logistic regression model produced odds ratios (ORs) and 95% CIs. The Pearson χ 2 test was used to evaluate the goodness-of-fit for the Poisson regression model, and the Akaike Information Criterion (AIC) was used to evaluate the goodness-of-fit for the multivariate logistic regression model. We used SAS version 9.4 (SAS Institute, Inc) for all analyses. This study was exempted from the University of Florida Institutional Review Board review because of the use of publicly available data. We followed the STROBE statement in conducting methods and reporting results (19).

The unweighted sample size was 905, representing 27,180,715 people in the US. Of these cancer patients, 24.4% (weighted percentage) had a type 2 diabetes diagnosis, and 25.8% (weighted percentage) had a prediabetes diagnosis ( Table 1 ). The mean age of the total study population was 63.9 years. People with cancer and diabetes (mean age, 68.8 y) and people with cancer and prediabetes (mean age, 66.7 y) were older, on average, than people with cancer only (mean, 59.9 y). The percentage of people with less than a high school diploma was significantly larger among people with cancer and diabetes (10.2%) and cancer and prediabetes (9.4%) than people with cancer only (5.2%). The percentage of people who had a BMI in the obese range was significantly larger among people with cancer and diabetes (63.3%) and cancer and prediabetes (43.9%) than people with cancer only (30.7%). The percentage of people with 3 or more additional comorbidities was significantly larger among people with cancer and diabetes (51.0%) and cancer and prediabetes (30.3%) than among people with cancer only (17.1%). Regardless of comorbidity status, more than 95% of people had health insurance. The percentage of people with a usual source of care was significantly larger among people with cancer and diabetes (98.3%) and cancer and prediabetes (97.2%) than among people with cancer only (91.6%).

In the unadjusted Poisson regression model, the IRR for the number of health care visits in the previous 12 months was significantly higher among people with cancer and diabetes (IRR = 1.19; 95% CI, 1.12–1.27; P < .001) than among people with cancer only ( Table 2 ). However, after controlling for covariates, the comorbidity of cancer and diabetes was not significantly associated with increases in the number of health care visits (IRR = 1.04; 95% CI, 0.94–1.15; P = .44). After controlling for covariates, the comorbidity of cancer and prediabetes was associated with increases in the number of health care visits in the previous 12 months (IRR = 1.11; 95% CI, 1.01–1.22; P = .03). The results of the goodness-of-fit test for both unadjusted and adjusted models were not significant, indicating that neither model fit the data well.

In the multivariate logistic regression, the unadjusted model showed that people with diabetes and cancer were 2.5 times more likely than people with cancer only to stay overnight in a hospital (OR = 2.55; 95% CI, 1.54–4.21). However, after controlling for covariates, this association was not significant (OR = 1.57; 95% CI, 0.82–3.02). Moreover, we found no significant association in comorbidity with prediabetes for the risk of an overnight stay in a hospital in either the unadjusted or adjusted model ( Table 3 ). The goodness-of-fit test for the adjusted model had a lower AIC value than the unadjusted model, indicating a better fitting model.

The objective of our study was to examine patterns of health care use among people with cancer and either prediabetes or diabetes. In our nationally representative sample, patients with cancer and diabetes had 19% more health care visits than people with cancer only according to the unadjusted regression model, and patients with cancer and prediabetes had 11% more health care visits than people with cancer only according to the adjusted regression model. Future studies may be needed to test strategies to improve care coordination and early initiation of preventive care strategies for people with cancer at risk of developing prediabetes and diabetes.

Having diabetes and cancer increased the risk for an overnight stay in a hospital in the unadjusted regression models, whereas having prediabetes and cancer increased the number of health care visits in the adjusted regression model only. These findings indicate that different stages of diabetes may drive different health care needs. In the qualitative study conducted in 2021 and 2022 at 3 New York City hospitals among 15 women with breast cancer and either diabetes or prediabetes, 7 participants reported glucose levels of more than 200 mg/dL (normal is 70–90 mg/dL) and 9 participants indicated a lack of glucose control during cancer treatment (11). In addition, as cancer treatment tends to be prioritized over other treatment, diabetes prevention and management led by a primary care physician may be paused (20). Medication adherence for chronic diseases may also decline due to the priority of cancer treatment (21,22). In addition, many cancer patients with comorbidities may not receive self-management education or guidelines for preventive care, negatively affecting cancer survivorship (23). Moreover, our study found that patients with cancer and diabetes were 2 times more likely to be hospitalized, whereas patients with cancer and prediabetes did not have significantly higher rates of hospitalization. This finding was supported by literature showing that patients with cancer and at least 1 comorbidity were more likely than patients with no comorbidities to be hospitalized (6,24). Clinical guidelines for managing patients with cancer and prediabetes are lacking, and communication guidelines for coordinated care between oncologists and primary care physicians are limited. Because many patients with cancer tend to prioritize cancer treatment over primary care for prediabetes or diabetes, detrimental clinical outcomes and increased health care use may not be preventable without early prevention or ongoing management. In response to increases in the prevalence of prediabetes and cancer, it is important to develop a systematic preventive care model for early-stage chronic diseases (eg, prediabetes, prehypertension) that includes collaboration between oncologists and primary care physicians. Such a model could be a cost-effective strategy for improving cancer survivorship.

Our study also found that more than 80% of comorbid people were overweight or obese (compared with 67.5% among those with cancer only). It is well established that obesity is significantly associated with cancer incidence and mortality (25) and is a risk factor for cancer and chronic diseases (eg, diabetes, prediabetes) (26,27). Excessive body fat causes chronic inflammation that may be attributed to cancer treatment–associated adverse outcomes (25). Thus, it is important to control overweight and obesity during cancer treatment. A combination of diet and exercise was identified as a more effective intervention for weight loss than a standard of care for patients with cancer (28). Clinicians need to provide self-management guidelines for lifestyle changes when a cancer diagnosis is first made, especially among overweight or obese patients. In the qualitative study among 15 women with breast cancer and either diabetes or prediabetes, participants indicated not receiving guidance on self-management or having a designated clinician who continuously monitored them (11). One in-depth patient interview found that a patient searched for diet or exercise information on Google (11). This research suggests a need for self-management guidelines provided by clinicians for controlling overweight or obesity and monitoring chronic disease progression.

Educational attainment was significantly associated with comorbidity status. Among patients with less than a high school diploma, the percentage of patients with a comorbidity was twice the percentage of patients with no comorbidity (9.4% and 10.2% vs 5.2%). Education may be key to health behaviors and the prevention of adverse outcomes. It is well established that education inequality is associated with cancer survivorship (29,30). For example, a study in The Netherlands showed that among patients with cancer and comorbidity, those with a low level of education (equivalent to primary school) had a 3 times higher risk of death than those with a university degree (29). A study of education differentials in cancer deaths in Lithuania found an inverse educational gradient for selected cancer sites among men and women, noting that substantial shares of cancer deaths (8% to 35%) could have been avoided or postponed (30). Increasing access to resources for patients with low levels of education may help to minimize the number of comorbidities that can arise and ultimately improve their cancer survivorship. Particularly, providing more resources may benefit from developing effective and structured communication strategies with providers.

Optimal coordinated care is crucial to mitigate the burden of comorbidities on health care use and costs among patients with cancer. Despite the growing need for increased care coordination between primary care physicians and oncologists, no standardized care coordination model exists for managing the comorbidity of cancer and chronic diseases such as prediabetes or diabetes (31). Additionally, the involvement of primary care physicians in cancer care is limited, especially during active cancer treatment (32). Previous research identified some barriers to effective cancer care coordination, including inadequate communication between oncologists and primary care providers and between patients and primary care providers; geographic limitations; and limited interoperability of the electronic health record among health care providers (32,33). Fortunately, the recent rapid technological evolution has provided new opportunities to reduce these barriers. Studies conducted at the Johns Hopkins Primary Care for Cancer Survivors clinic in 2015 and the Duke Cancer Institute during 2020–2021 found that comorbid patients were more likely to use telehealth for cancer and primary care, and telehealth improved outcomes such as patient satisfaction and survivorship (34,35). Using artificial intelligence in the care coordination process and communication will become pivotal to improving an efficient and effective care coordination model. An optimal care coordination model integrated with technology can be achieved by using standardized communication channels among health care providers and between health care providers and patients and the interoperability of electronic health records. Moreover, appropriate data privacy and security regulation will be essential to ensure patient trust in the care coordination model. To leverage these benefits, standardized clinical guidelines for managing comorbidities in patients with cancer should be developed. These guidelines would provide clear recommendations on integrating care coordination.

Limitations

Our study has several limitations. First, the diagnosis information obtained from a self-reported survey may be subject to recall bias, and we could not determine the exact timing of the diagnosis of diabetes or prediabetes and cancer. Second, our study used cross-sectional data, which prevented us from following disease progression over time and examining the effects of various treatments. A study that uses longitudinal data is needed to understand the effect of comorbidity on health care use among cancer patients. Third, we could not identify the reasons for health care use because of a lack of data. A study that incorporates electronic health records may identify patient-centered health care needs for those with comorbidities. Lastly, while the study identified patients with cancer who had undiagnosed diabetes, the sample size was too small to generate population estimates. Studies that use larger data sets could examine the role of undiagnosed diabetes on cancer prognosis and outcomes.

Among people with cancer, diabetes was significantly associated with an increased risk of an overnight hospital stay, whereas prediabetes was significantly associated with an increase in the number of health care visits. Our findings suggest that it may be beneficial to prioritize preventive measures (eg, screening) to prevent prediabetes from progressing to diabetes in patients with cancer and develop optimal coordinated care, which could help alleviate the strain on the health care system and improve oncology care.

The authors received no external financial support for the research, authorship, or publication of this article. The authors declared no potential conflicts of interest with respect to the research, authorship, or publication of this article. No copyrighted material, surveys, instruments, or tools were used in the research described in this article.

Corresponding Author: Ara Jo, PhD, Department of Health Services Research, Management and Policy, University of Florida, Health Sciences Center, PO Box 100195, Gainesville, FL 32610-0195 ( [email protected] ).

Author Affiliations: 1 Department of Health Services Research, Management and Policy, University of Florida, Gainesville. 2 College of Public Health and Health Professions, University of Florida, Gainesville. 3 Now with School of Dental Medicine, University of Pennsylvania, Philadelphia. 4 Department of Health Outcomes and Behavior, Moffitt Cancer Center, Tampa, Florida. 5 Department of Gastrointestinal Oncology, Moffitt Cancer Center, Tampa, Florida.

• Pearson-Stuttard J, Papadimitriou N, Markozannes G, Cividini S, Kakourou A, Gill D, et al. . Type 2 diabetes and cancer: an umbrella review of observational and Mendelian randomization studies. Cancer Epidemiol Biomarkers Prev . 2021;30(6):1218–1228. PubMed doi:10.1158/1055-9965.EPI-20-1245
• Ranc K, Jørgensen ME, Friis S, Carstensen B. Mortality after cancer among patients with diabetes mellitus: effect of diabetes duration and treatment. Diabetologia . 2014;57(5):927–934. PubMed doi:10.1007/s00125-014-3186-z
• Rahman I, Athar MT, Islam M. Type 2 diabetes, obesity, and cancer share some common and critical pathways. Front Oncol . 2021;10:600824. PubMed doi:10.3389/fonc.2020.600824
• Centers for Disease Control and Prevention. An update on cancer deaths in the United States. Published February 28, 2022. Accessed July 11, 2023. https://stacks.cdc.gov/view/cdc/119728
• National Cancer Institute. Cancer statistics. Published April 2, 2015. Accessed July 10, 2023. https://www.cancer.gov/about-cancer/understanding/statistics
• Bur AM, Brant JA, Mulvey CL, Nicolli EA, Brody RM, Fischer JP, et al. . Association of clinical risk factors and postoperative complications with unplanned hospital readmission after head and neck cancer surgery. JAMA Otolaryngol Head Neck Surg . 2016;142(12):1184–1190. PubMed doi:10.1001/jamaoto.2016.2807
• Goel AN, Raghavan G, St John MA, Long JL. Risk factors, causes, and costs of hospital readmission after head and neck cancer surgery reconstruction. JAMA Facial Plast Surg . 2019;21(2):137–145. PubMed doi:10.1001/jamafacial.2018.1197
• Nene RV, Brennan JJ, Castillo EM, Tran P, Hsia RY, Coyne CJ. Cancer-related emergency department visits: comparing characteristics and outcomes. West J Emerg Med . 2021;22(5):1117–1123. PubMed doi:10.5811/westjem.2021.5.51118
• Zhang X, Wang X, Wang M, Gu J, Guo H, Yang Y, et al. . Effect of comorbidity assessed by the Charlson Comorbidity Index on the length of stay, costs, and mortality among colorectal cancer patients undergoing colorectal surgery. Curr Med Res Opin . 2023;39(2):187–195. PubMed doi:10.1080/03007995.2022.2139053
• Park J, Look KA. Health care expenditure burden of cancer care in the United States. Inquiry . 2019;56:46958019880696. PubMed doi:10.1177/0046958019880696
• Pinheiro LC, Cho J, Rothman J, Zeng C, Wilson M, Kern LM, et al. . Diabetes and cancer co-management: patient-reported challenges, needs, and priorities. Support Care Cancer . 2023;31(2):145. PubMed doi:10.1007/s00520-023-07604-x
• Hwangbo Y, Kang D, Kang M, Kim S, Lee EK, Kim YA, et al. . Incidence of diabetes after cancer development: a Korean national cohort study. JAMA Oncol . 2018;4(8):1099–1105. PubMed doi:10.1001/jamaoncol.2018.1684
• Hamood R, Hamood H, Merhasin I, Keinan-Boker L. Diabetes after hormone therapy in breast cancer survivors: a case‐cohort study. J Clin Oncol . 2018;36(20):2061–2069. PubMed doi:10.1200/JCO.2017.76.3524
• Ose DJ, Viskochil R, Holowatyj AN, Larson M, Wilson D, Dunson WA, et al. . Understanding the prevalence of prediabetes and diabetes in patients with cancer in clinical practice: a real-world cohort study. J Natl Compr Canc Netw . 2021;19(6):709–718. PubMed doi:10.6004/jnccn.2020.7653
• May P, Garrido MM, Aldridge MD, Cassel JB, Kelley AS, Meier DE, et al. . Prospective cohort study of hospitalized adults with advanced cancer: associations between complications, comorbidity, and utilization. J Hosp Med . 2017;12(6):407–413. PubMed doi:10.12788/jhm.2745
• Rotbain EC, Rostgaard K, Andersen MA, Da Cunha-Bang C, Niemann CU, Frederiksen H, et al. . Healthcare utilization and comorbidity in chronic lymphocytic leukemia. Clin Epidemiol . 2021;13:1155–1165. PubMed doi:10.2147/CLEP.S337495
• Storey S, Zhang Z, Luo X, Von Ah D, Metzger M, Zhang J, et al. . Association of comorbid diabetes with clinical outcomes and healthcare utilization in colorectal cancer survivors. Oncol Nurs Forum . 2021;48(2):195–206. PubMed doi:10.1188/21.ONF.195-206
• Centers for Disease Control and Prevention. NHANES survey methods and analytic guidelines. Accessed August 11, 2023. https://wwwn.cdc.gov/nchs/nhanes/analyticguidelines.aspx
• von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP; STROBE Initiative. Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. BMJ . 2007;335(7624):806–808. PubMed doi:10.1136/bmj.39335.541782.AD
• Cho J, Nilo D, Sterling MR, Kern LM, Safford MM, Pinheiro LC. Eliciting primary care and oncology provider perspectives on diabetes management during active cancer treatment. Support Care Cancer . 2021;29(11):6881–6890. PubMed doi:10.1007/s00520-021-06264-z
• Dashputre AA, Gatwood KS, Schmidt J, Gatwood J. Impact of oral oncolytic initiation on medication adherence for pre-existing comorbid chronic conditions. J Oncol Pharm Pract . 2020;26(4):835–845. PubMed doi:10.1177/1078155219875206
• Gatwood J, Dashputre A, Rajpurohit A, Gatwood K, Mackler E, Wallace L, et al. . Medication adherence among adults with comorbid chronic conditions initiating oral anticancer agent therapy for multiple myeloma. JCO Oncol Pract . 2022;18(9):e1475–e1483. PubMed doi:10.1200/OP.22.00008
• Earle CC, Neville BA. Under use of necessary care among cancer survivors. Cancer . 2004;101(8):1712–1719. PubMed doi:10.1002/cncr.20560
• Christodoulou I, Ukert B, Vavuranakis MA, Kum HC, Giannouchos TV. Adult cancer-related emergency department utilization: an analysis of trends and outcomes from emergency departments in Maryland and New York. JCO Oncol Pract . 2023;19(5):e683–e695. PubMed doi:10.1200/OP.22.00525
• Pati S, Irfan W, Jameel A, Ahmed S, Shahid RK. Obesity and cancer: a current overview of epidemiology, pathogenesis, outcomes, and management. Cancers (Basel) . 2023;15(2):485. PubMed doi:10.3390/cancers15020485
• Leach CR, Weaver KE, Aziz NM, Alfano CM, Bellizzi KM, Kent EE, et al. . The complex health profile of long-term cancer survivors: prevalence and predictors of comorbid conditions. J Cancer Surviv . 2015;9(2):239–251. PubMed doi:10.1007/s11764-014-0403-1
• Gallagher EJ, LeRoith D. Obesity and diabetes: the increased risk of cancer and cancer-related mortality. Physiol Rev . 2015;95(3):727–748. PubMed doi:10.1152/physrev.00030.2014
• LeVasseur N, Cheng W, Mazzarello S, Clemons M, Vandermeer L, Jones L, et al. . Optimising weight-loss interventions in cancer patients — a systematic review and network meta-analysis. PLoS One . 2021;16(2):e0245794. PubMed doi:10.1371/journal.pone.0245794
• Aarts MJ, Kamphuis CBM, Louwman MJ, Coebergh JWW, Mackenbach JP, van Lenthe FJ. Educational inequalities in cancer survival: a role for comorbidities and health behaviours? J Epidemiol Community Health . 2013;67(4):365–373. PubMed doi:10.1136/jech-2012-201404
• Jasilionis D, Smailyte G, Vincerzevskiene I, Shkolnikov VM. Educational differentials in cancer mortality and avoidable deaths in Lithuania, 2001–2009: a census-linked study. Int J Public Health . 2015;60(8):919–926. PubMed doi:10.1007/s00038-015-0745-0
• Hohmann NS, McDaniel CC, Mason SW, Cheung WY, Williams MS, Salvador C, et al. . Patient perspectives on primary care and oncology care coordination in the context of multiple chronic conditions: a systematic review. Res Social Adm Pharm . 2020;16(8):1003–1016. PubMed doi:10.1016/j.sapharm.2019.11.014
• Petrovic B, Morgan S, Afkham A, O’Brien MA, Sussman J, Fitch M, et al. . Cancer specialist perspectives on implementing an online communication system with primary care providers. Ann Fam Med . 2022;20(20 Suppl 1):3246. PubMed
• Walsh J, Harrison JD, Young JM, Butow PN, Solomon MJ, Masya L. What are the current barriers to effective cancer care coordination? A qualitative study. BMC Health Serv Res . 2010;10(1):132. PubMed doi:10.1186/1472-6963-10-132
• Choi Y, Parrillo E, Wenzel J, Grabinski VF, Kabani A, Peairs KS. Optimizing cancer survivorship in primary care: patient experiences from the Johns Hopkins Primary Care for Cancer Survivors clinic. J Cancer Surviv . 2023;17(5):1286–1294. PubMed doi:10.1007/s11764-022-01166-3
• Golla V, Frascino N, Wilson LE, Oakes M, Oeffinger KC, Check DK, et al. . Telehealth patterns in primary and onco-primary care. Cancer Survivorship Research & Care. 2023;1(1):2288879.

a All values are weighted percentages, unless otherwise indicated. b Determined by t test for continuous variable and χ 2 tests for categorical variables. c Includes American Indian or Alaska Native, Asian, and Native Hawaiian or Pacific Islander, and multiracial. d Measured by the ratio of income to poverty (total family income divided by the poverty threshold) and dichotomized into 2 levels: 1) poor (ratio < 1) and 2) rich (ratio ≥ 1). e Visits to a physician’s office, a clinic, or some other place in the previous 12 months, not including hospitalizations, emergency department visits, home visits, or telephone calls. f Excludes overnight stays in the emergency department.

Abbreviation: IRR, incidence rate ratio. a Visits to a physician’s office, a clinic, or some other place in the previous 12 months, not including hospitalizations, emergency department visits, home visits, or telephone calls. b Controlled for age, sex, race and ethnicity, education, poverty-to-income ratio, body mass index, number of additional comorbidities, and health insurance.

a Excludes overnight stays in the emergency department. b An odds ratio with a 95% CI that includes 1 indicates no significant effect on risk. c Controlled for age, sex, race and ethnicity, education, poverty-to-income ratio, body mass index, number of comorbidities, and health insurance.

The opinions expressed by authors contributing to this journal do not necessarily reflect the opinions of the U.S. Department of Health and Human Services, the Public Health Service, the Centers for Disease Control and Prevention, or the authors’ affiliated institutions.