Artificial Intelligence for Assisting Cancer Diagnosis and Treatment in The Era of Precision Medicine

Artificial Intelligence for Assisting Cancer Diagnosis and Treatment in The Era of Precision Medicine

Dnyaneshwari A. Gunjal*, Chaitali R. Rajput, Vaishnavi P. Wani,

Madhuri S. Pardeshi

Department of Pharmaceutics Sumantai Institute of Pharmacy,Bambrud kh.Pachora, Dist. Jalgaon, Maharashtra , India

*Correspondence: dnyaneshwarigunjal48@gmail.com

DOI: https://doi.org/10.71431/IJRPAS.2025.4201

INTRODUCTION

McCarthy and his associates coined the term "intelligence artificielle (IA)," sometimes known as "intelligence machine," while working in a Dartmouth studio in the summer of 1956. The definition of artificial intelligence (IA) is simply a computer program that can learn and recognize patterns and relationships between input and output data, then effectively use these insights to make decisions about new input data. The two main methods used to determine artificial intelligence are machine learning (ML) and deep learning (DL), which are occasionally used interchangeably respectively. [1]

In the field of computer science, machine learning (ML) is a subfield of artificial intelligence (IA), while deep learning (DL) is a particular subfield that focuses on artificial neural networks. Thanks to advancements in big data, algorithms, computer power, and internet technology, deep learning has experienced unprecedented success in a variety of fields over the past ten years, including facial recognition, image classification, voice recognition, automatic translation, and healthcare assistance. Given the enormous number of cancer diagnoses made worldwide each year, there is a lot of interest in the use of artificial intelligence in oncology. This includes, among other things, doing precise cancer diagnostics using[2]

Inspired by brain neural architecture, DL uses deep neural networks (DNNs) to develop sophisticated models with multiple hidden layers to Progresses in the application of AI in oncology in this review . We also analyze various types of data and develop prediction outputs. Unlike conventional ML techniques, which require careful engineering to design a feature extractor that transforms raw data (such as the pixel values of an image) into relevant discriminatory features before data input, DL algorithms feed the machine with raw data with which it can automatically learn the optimal deep features that best fit the task through a training process . This ability likely explains the fact that DL algorithms have been consistently improved in many common AI tasks, such as image recognition, pattern recognition, speech recognition, and natural language processing. Consequently, a majority of AI research within the oncology field involves the utilization of DL.[3]

Among DNN models, convolutional neural networks (CNNs) are the most popular DL architectures. They have been used for cancer lesion detection, recognition, segmentation and the classification of medical images. The architecture of a typical CNN  is structured by stacking three main layers: convolutional layers, pooling layers, and fully‐connected layers. In doing this, CNNs transform the original images layer by layer from pixel values to the final prediction scores. The convolutional layers involve combining input data (feature map) with convolutional kernels (filters) to form a transformed feature map. The filters in the convolutional layers are automatically adjusted based on learned parameters to extract the most useful features for a specific task. Yet, there is a drawback; it is difficult to tell what features are learned by the CNNs, which is known as the “black box”.[4]

Over the past five years, large amounts of researches have applied DL to cancer diagnosis, precision medicine, radiotherapy, and cancer research. Moreover, the American Food and Drug Administration (FDA) have approved a number of AI algorithms related to oncology and published a fast‐track approval plan for AI medical algorithms in 2018. Here, we provided an overview of the recent and enormous highlight the limitations, challenges, and future implications of AI‐powered cancer care.[5]

Figure 1: Publication statistics of deep learning by cancer area over the past five years, searched on PubMed. A. Publication statistics of deep learning by cancer diagnosis, precision medicine, radiotherapy, and cancer research. B. Publication statistics of deep learning for different cancer sites

Table 1: Summary of FDA‐approved artificial intelligence devices in the field of oncology

Figure 2: Applications of AI in cancer diagnosis, treatment and research. OARs, organs at risk

CANCER SCREENING, DIAGNOSIS, CLASSIFICATION, AND GRADING

Cancer screening for early detection, accurate cancer diagnosis, classification and grading are the key determinants of treatment decisions and patient outcomes. Over the past few years, there is increasing interest in the applications of AI in these critical areas, sometimes with performance equivalent to human experts and advantages in scalability and time‐saving. More importantly, AI has shown its potential in solving challenging problems that humans simply cannot do.[6]

Cancer screening and early detection

Cancer screening has contributed to decreasing the mortality of some common cancers . The most successful examples are the identification of precancerous lesions (e.g., cervical intra‐epithelial neoplasia [CIN] for cervical cancer screening, and adenomatous polyps for colorectal cancer screening) where the treatment leads to a decrease in the incidence of invasive cancer . Given the requirement for high throughput technology and a fast turnaround, automation is being used to improve the efficiency of cancer screening.[7]

For cervical cancer screening, Wentzensen et al. developed a DL classifier for p16/Ki‐67 dual‐stained (DS) cytology slides trained on biopsy‐based gold standards. In independent testing, AI‐based DS had equal sensitivity and substantially higher specificity compared with a Pap smear and manual interpretation of DS. Most importantly, AI‐based DS reduced unnecessary colposcopies by one‐third compared with Pap smears (41.9% vs. 60.1%, P < 0.001), while it had a similar performance in identifying high‐grade CIN, which indicates immediate treatment. For colorectal cancer screening, a prospective randomized controlled trial including 1,058 patients showed that AI‐assisted colonoscopy significantly increased adenoma detection rates and the mean number of adenomas found per patient compared with conventional colonoscopy (29.1% vs. 20.3%), which was attributed to a higher number of diminutive adenomas found . This is particularly important because a 1% increase in the adenoma detection rate is associated with a 3% decrease in colorectal cancer incidence.[8]

Automated nodule detection and classification on low‐dose computed tomography (CT) and mammography for lung and breast cancer screening have attracted significant attention. Several successful CNN‐based models have achieved classification accuracies of 80% to 95% , which shows their transformative potential in lung cancer screening. Radial et al. proposed a DL algorithm that uses patients current and prior low‐dose CT scans to predict the risk of lung cancer with outstanding results (area under the curve [AUC] of receiver operating characteristic = 0.944). Improvement in breast cancer screening with AI mammography has also been verified in preclinical studies , as well as in clinical settings . McKinney et al.  established an AI system for breast cancer screening using an ensemble of three CNN‐based models. A reduction in the numbers of false positives and false negatives was observed compared with the original decisions made in the course of clinical practice. In an independent study by six radiologists, the AUC for the AI system was 11.5% higher than the average AUC achieved by the 6 radiologists. Notably, this AI system has the ability to generalize from the training data to multicenter data.[9]

An emerging area for the early detection of cancers is liquid biopsies for circulating tumor DNA (ct DNA) or cell‐free DNA (cf DNA) obtained via a simple blood test. These are particularly important for cancer types that currently have no effective screening method. In a promising work, Cohen et al. developed Cancer SEEK for the early detection and prediction of eight cancer types using CT DNA. With Cancer SEEK, samples are first classified as cancer‐positive using a logistic regression model applied to 16 gene mutations and the expression levels of 8 plasma proteins. The cancer type is then predicted using a random forest classifier, with accuracies ranging from 39% to 84%. Although liquid biopsies are promising for early cancer detection, so far, they have been limited to traditional ML algorithms . As data acquisition from liquid biopsies increases, we anticipate that DL models will eliminate the need for manual selection and duration of discriminatory features, as well as allowing for the combination of multiple data types to enhance early cancer detection.[10]

Cancer diagnosis, classification, and grading

CNN-based deep learning models have been widely reported to accurately diagnose cancers, classify cancer subtypes, and identify cancer grades using radiology (e.g., CT and MRI), histopathology (e.g., whole slide imaging [WSI]), and endoscopy images (e.g., esophagogastroduodenoscopy and colonoscopy). The majority of these models show accuracies at least comparable to those of professionals.[11]

CNN-based deep learning algorithms have demonstrated remarkable precision in detecting cancerous growths on histopathology slides. The best CNN algorithm (a Google Net architecture-based model) produced an AUC of 0.994 in an international competition (CAMELYON16) for diagnosing breast cancer metastasis in lymph nodes using WSI with hematoxylin-eosin (HE) staining. This algorithm outperformed the best pathologist, who had an AUC of 0.884, and did so more quickly. It is very difficult to diagnose cancer, hence DL algorithms have also been used to forecast the origin of unknown primary tumors.[12]

Additionally, DL has been regularly demonstrated to be successful in diagnosing malignant disorders employing endoscopy, positron emission tomography-CT (PET-CT) scans, CT, and MRI. Yuan et al. most recently used CT scans to create a classifier that used a three-dimensional (3D) Res Net algorithm to predict occult peritoneal metastasis in colorectal cancer. The classifier's AUC of 0.922 was significantly higher than that of routine contrast-enhanced CT diagnosis (AUC = 0.791). In a different study, Ke et al. trained and tested a self-constrained 3D Dense Net that was able to differentiate between benign nasopharyngeal hyperplasia and nasopharyngeal cancer (NPC) using MRI data from 4,100 patients. The AUC of the network was reported to be 0.95–0.97. Regarding endoscopy, Luo et al. created a gastrointestinal AI diagnostic in a multicenter research.[13]

DL models are employed for more difficult cancer classifications and grading tasks in addition to dichotomous diagnosis. In order to categorize WSI for lung tissues into three classes (normal, lung adenocarcinoma, and lung squamous cell carcinoma), Coudray et al. created Deep PATH, an Inception-v3 architecture-based model, which has a reported AUC of 0.97. Additionally, with a 75% agreement between the algorithm and pathologists, the CNN was effectively taught to perform automated Gleason grading of prostate cancer. Radiology scans can also be used for cancer grading. With a reported AUC of 0.83, Zhou et al. created a DL method (based on SE Net and Dense Net) to predict liver cancer grades (low versus high) using MRI data. All things considered, these findings demonstrate the promising use of AI in cancer categorization

Technically and practically speaking, these DL-based diagnostic tools incorporate performance-tuning and performance-enhancing capabilities, which streamlines traditional computer-aided diagnosis pipelines and lowers false-positive rates. The robustness and generalizability of DL models require improvement, even if preclinical evaluations of AI tools have opened the door for clinical trials to increase the precision and effectiveness of cancer diagnosis.[14]

Predicting gene mutations in cancer

DL algorithms have also been applied to histopathology images in order to characterize the underlying genetic and epigenetic heterogeneity. A CNN was trained to predict six distinct genetic mutations using HE-stained WSI of lung cancer, with an AUC ranging from 0.733 to 0.856 based on a held-out testing population. Additionally, with AUCs >0.71, the CNN model (Inception-V3) found frequent mutations in liver cancer using WSI. DL techniques have also been built using WSI to forecast pan-cancer gene changes, localized amplifications and deletions, chromosomal arm gains and losses, and whole-genome duplications. DL models have been used to predict mutational footprints, including microsatellite instability (MSI) status and tumor mutational burden (TMB) status, which are the most significant biomarkers, in addition to predicting mutations in specific genes.

It has been investigated to use noninvasive radiological pictures, like CT or MRI scans, in addition to histopathological images to detect cancer mutations. For instance, DL models can be used to predict the EGFR mutation status in non-small cell lung cancer (NSCLC) using both CT and PET/CT images, both of which have AUCs >0.81. AUCs ranging from 0.70 to 0.84 were attained by Shboul et al. in another study when they used machine learning to predict O6-methylguanine-DNA methyltransferase methylation, isocitrate dehydrogenase mutation, 1p/19q co-deletion, alpha-thalassemia/mental retardation syndrome X-linked mutation, and telomerase reverse transcriptase mutation of low-grade gliomas with radiomics. TMB status in NSCLC has also been predicted using CT scans (AUC = 0.81). Although the results were encouraging, it is yet unclear which features the CNN models are learning to identify the status of mutations.[15]]

PATIENT PROGNOSIS, RESPONSE TO THERAPY, AND PRECISION MEDICINE

Precision medicine involves treating each patient differently. It aims to separate people into subgroups according to variations in their disease prognosis or how they respond to a specific treatment in order to customize therapeutic treatments for those who will benefit and prevent expenses and negative consequences for those who won't. DL algorithms are used to automatically extract features from medical data to build models that can accurately predict a patient's likelihood of tumor recurrence and response to treatment. Based on the prediction results, physicians can provide more precise and suitable therapies.

Lung cancer, metastatic melanoma, and other cancers can be treated using immunotherapy medications. However, checkpoint inhibitor medication does not work for about 50% to 80% of cancer patients. Programmed death-ligand 1 (PD-L1) expression, TMB, MSI, and somatic copy number changes are examples of biomarkers of the immunogenic tumor microenvironment that are being used to predict response to immunotherapies. However, a biopsy was used to obtain these biomarker data, which is invasive, challenging to conduct longitudinally, and restricted to a specific tumor site. Moreover, biomarkers might not have much predictive power. Regardless of PD-L1 expression, immunotherapy with pembrolizumab in conjunction with conventional chemotherapy improved survival for all patients in the KEYNOTE-189 clinical study. In order to fulfill the objective of precision medicine, numerous researchers have created DL

Accurate predictive tests are necessary to guide patient selection because, in addition to immunotherapy, other treatments (such as targeted therapy and neoadjuvant chemotherapy [NAC]) have shown notable clinical success in some groups. AI and big data can be used together to meet this goal. Imaging phenotypes linked to a certain mutation can be found using AI predictive models. The benefit of this AI-based method is that it can repeatedly and noninvasively determine the mutation status. A PET/CT-based DL model for patients with NSCLC, which uses radiomic characteristics to distinguish between EGFR-mutant types and wild-type with an AUC of 0.81, validated this strategy. Furthermore, using a lot of radiomics data, DL algorithms have demonstrated effectiveness in predicting breast cancer patients' reactions to NAC.[16]

DEEP LEANING IN RADIOTHERAPY 

Half of patients receive radiotherapy, making it a crucial modality in the treatment of cancer. The development of AI algorithms and their incorporation into radiotherapy workflows are greatly aided by the image-, data-driven, and quality assurance frameworks of radiotherapy. Investigating AI to support automated treatment planning, target volume, and organs at risk (OAR) delineation in radiation has generated a lot of attention.

The precision of target volume and OAR delineation is largely dependent on the radiation oncologists' experience, and it is a labor-intensive operation. In the automated delineation of OAR in the head and neck, thorax, abdomen, and pelvic regions, CNN-based semantic segmentation has continuously been recognized as a state-of-the-art technique. OAR typically appears on CT scans, and each patient's runtime is only a few seconds long. Organs with large volumes, rigid shapes, and regular shapes, like the mandible (Dice similarity coefficient [DSC] = 0.94), parotid (DSC = 0.84), kidney (DSC = 0.96), and liver (DSC = 0.97), had relatively high segmentation accuracies based on these published studies. In contrast, organs with small volumes, movable shapes, and irregular shapes, like the optic

Automated contouring of tumor targets by DL remains a significant issue due to the wide range of tumor forms, locations, and interior morphologies. However, automatic contouring expedites the procedure and enhances radiation oncologists' uniformity. Numerous malignancies, including nasopharyngeal, cervical, colorectal, lung, and breast cancers, have been studied using automated delineation of the gross tumor volume (GTV) and clinical target volume (CTV). By using a 3D CNN model on MRI, Lin et al. initially created an automated contouring method for NPC. With an overall accuracy of 79%, they discovered adequate concordance between the AI tool and human specialists in an independent test. Additionally, the AI tool performed better than half of the radiation oncologists from seven hospitals in a multicenter test.

Automated treatment planning is a significant use of AI in radiotherapy. In order to accomplish particular dosimetry goals, radiotherapy planning is a complicated procedure that incorporates "trial-and-error" depending on physicists' subjective priorities. As a result, clinical physicists' experience has a significant impact on the quality of treatment planning. Although knowledge-based automated planning systems, like Eclipse's RapidPlan, have increased planning quality consistency, they are not ideal because they are unable to estimate realistic dose distributions for individual patients. Lately, DL-based techniques have emerged as a viable strategy for customized 3D dose optimization and prediction. For head and neck malignancies, Fan et al. initially created an automated treatment planning approach based on ResNet to accomplish voxel-by-voxel dose optimization and precise 3D dose prediction. The findings revealed

 Additional uses of AI in radiotherapy include intra- and inter-fraction motion monitoring, image reconstruction, synthetic CT creation, image registration, and the prediction of radiation-induced toxicities. In conclusion, AI may enhance radiotherapy's precision, effectiveness, and caliber. Furthermore, by using effective and efficient automated segmentation, image processing, and automated treatment planning tools based on DL, which are substantially faster than conventional methods, MRI-only radiation and real-time adaptive radiotherapy might be accomplished.[17]

DL IN CANCER RESEARCH

DL techniques have been used in many areas of cancer research, such as establishing randomized controlled trials (RCTs), generating anti-cancer therapies, and examining biological foundations. Studies have employed DL to examine the connection between genotypes and phenotypes in an effort to understand the molecular pathways underlying cancer, with numerous successes previously documented. The function of F-box/WD repeat-containing protein 7 (Fbw7) in cancer cell oxidative metabolism was identified in a recent study using DL algorithms and gene expression signatures from The Cancer Gene Atlas dataset. Additionally, Watson for Genomics found genetic changes across a range of cancer types that could have clinical implications but were missed by traditional molecular tumor boards. Finding these genetic variations identifies targets in addition to pertinent biological pathways.

RCTs must be successful in order for new cancer treatments to be adopted. Nonetheless, one of the most difficult parts of conducting trials is thought to be successfully recruiting the right patients. The process of matching potential subjects with complex qualifying criteria is time-consuming, labor-intensive, and challenging. Hassanzadeh et al. automated this procedure by using a Multi-Layer Perceptron model and natural language processing to extract relevant information from patient records. This helped them gather evidence for better decision-making on patient eligibility based on specific inclusion and exclusion criteria. It received a micro-F1 score of 84% overall. The effective implementation of RCTs also depends on the selection of top-enrolling investigators. Gligorijevic et al. suggested a DL strategy to learn from both investigator- and trial-related data in order to automate selection.[18]

CHALLENGES AND FUTURE IMPLICATIONS

Even though AI is being studied extensively in oncology, research is still needed to convert DL models into practical uses. The interpretability of algorithms, data access, medical ethics, and the generalizability of its applications are some of the obstacles to enhancing physicians' performance and adoption of clinically applied DL.

Generalizability and real‐world application

The performance of DL models tends to decline when used at multiple hospitals because to the significant variation in medical data among institutions; consequently, external validation sets might be necessary to verify their performance. Furthermore, DL's extraordinarily high parameter count raises the risk of overfitting and restricts its generalizability across other populations. More importantly, oncologists must take into account a range of information in clinical settings, such as imaging data, laboratory tests, clinical manifestations, and epidemiological histories, in order to make an accurate decision. The majority of recent research, however, has only used one kind of data (like imaging) as the input model. To mimic real clinical settings, a multimodal DL model incorporating the aforementioned information plus imaging data needs to be constructed in

Interpretability: the black‐box problem

DL has been criticized for being a “black box” that does not explain how the model generates outputs from given inputs. The large number of parameters involved makes it difficult for oncologists to understand how DL models analyze data and make decisions. However, some efforts have been made to make this black box more transparent. For example, the heat map‐like class activation algorithm, visualizes which image regions are taken into account with DL models when making decisions and to what degree. These innovative studies render DL tools more interpretable and applicable in clinical oncology settings.[19]

Data access and medical ethics

In addition to technological difficulties, DL research also encounter resource and ethical issues. Training data is crucial to DL's effectiveness and plausibility. Insufficient data could result in overfitting, which would lower performance in an external test cohort. Since medical data is frequently the property of different institutions, there aren't many data-sharing systems that connect them, which raises worries about patient information security. Thankfully, privacy-preserving distributed DL (DDL) and multicenter data-sharing agreements are starting to remove this barrier. In order to allow several parties to learn together using a deep model without explicitly sharing local datasets, DDL offers a privacy-preserving alternative. Additionally, the Cancer Imaging Archive, which compiles clinical photos from many hospitals and institutions, offers a nice illustration of[20]

CONCLUSIONS

DL is a newly developed AI method in oncology which is rapidly progressing. With the growth of high‐quality medical data and the development of algorithms, DL methods have great potential in improving the precision and efficiency of cancer diagnosis and treatment. Moreover, the positive attitude of the FDA towards AI medical devices further increases the prospect of DL's practical application in oncology. For the realization of clinical implementation, future researches should focus on the reproducibility and interpretability to make DL methods more applicable

REFERENCES

1.        McCarthy J, Minsky ML, Rochester N, Shannon CE. A proposal for the dartmouth summer research project on artificial intelligence, august 31, 1955. AI Magazine. 2006;27(4):12. [Google Scholar]

2.        Yasser E‐M, Honavar V, Hall A. Artificial Intelligence Research Laboratory. 2005.

3.        Yu KH, Beam AL, Kohane IS. Artificial intelligence in healthcare. Nat Biomed Eng. 2018;2(10):719–31. [PubMed] [Google Scholar]

4.        Siegel RL, Miller KD, Jemal A. Cancer statistics, 2020. CA Cancer J Clin. 2020;70(1):7–30. [PubMed] [Google Scholar]

5.        Wainberg M, Merico D, Delong A, Frey BJ. Deep learning in biomedicine. Nat Biotechnol. 2018;36(9):829–38. [PubMed] [Google Scholar]

6.        Meyer P, Noblet V, Mazzara C, Lallement A. Survey on deep learning for radiotherapy. Comput Biol Med. 2018;98:126–46. [PubMed] [Google Scholar]

7.        Samuel A. Some studies in machine learning using the game of checkers. IBMJ Res Dev. 1959;3:210–29. [Google Scholar]

8.        Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, et al. Dermatologist‐level classification of skin cancer with deep neural networks. Nature. 2017;542(7639):115–8. [PMC free article] [PubMed] [Google Scholar]

9.        Krizhevsky A, Sutskever I, Hinton GE. Imagenet classification with deep convolutional neural networks. Commun ACM. 2017;60(6):84–90. [Google Scholar]

10.    Bi WL, Hosni A, Schabath MB, Giger ML, Birkbak NJ, Mehrtash A, et al. Artificial intelligence in cancer imaging: Clinical challenges and applications. CA Cancer J Clin. 2019;69(2):127–57. [PMC free article] [PubMed] [Google Scholar]

11.    Abele DR, Adams AM, Berg CD, Black WC, Clapp JD, Fagerstrom RM, et al. Reduced lung‐cancer mortality with low‐dose computed tomographic screening. N Engl J Med. 2011;365(5):395–409. [PMC free article] [PubMed] [Google Scholar]

12.    Byers T, Wender RC, Jemal A, Baskies AM, Ward EE, Brawley OW. The American Cancer Society challenge goal to reduce US cancer mortality by 50% between 1990 and 2015: Results and reflections. CA Cancer J Clin. 2016;66(5):359–69. [PubMed] [Google Scholar]

13.    Corley DA, Jensen CD, Marks AR, Zhao WK, Lee JK, Doubeni CA, et al. Adenoma detection rate and risk of colorectal cancer and death. N Engl J Med. 2014;370(14):1298–306. [PMC free article] [PubMed] [Google Scholar]

14.    Wentzensen N, Lahrmann B, Clarke MA, Kinney W, Tokugawa D, Poitras N, et al. Accuracy and efficiency of deep‐learning‐based automation of dual stain cytology in cervical cancer screening. J Natl Cancer Inst. 2021;113(1):72–9. [PMC free article] [PubMed] [Google Scholar]

15.    Wang P, Berzin TM, Glissen Brown JR, Bharadwaj S, Becq A, Xiao X, et al. Real‐time automatic detection system increases colonoscopic polyp and adenoma detection rates: a prospective randomised controlled study. Gut. 2019;68(10):1813–9. [PMC free article] [PubMed] [Google Scholar]

16.    Zhao W, Yang J, Sun Y, Li C, Wu W, Jin L, et al. 3D deep learning from CT scans predicts tumor invasiveness of subcentimeter pulmonary adenocarcinomas. Cancer Res. 2018;78(24):6881–9. [PubMed] [Google Scholar]

17.    Kang G, Liu K, Hou B, Zhang N. 3D multi‐view convolutional neural networks for lung nodule classification. PLoS One. 2017;12(11):e0188290. [PMC free article] [PubMed] [Google Scholar]

18.    Ciompi F, Chung K, van Riel SJ, Setio AAA, Gerke PK, Jacobs C, et al. Towards automatic pulmonary nodule management in lung cancer screening with deep learning. Sci Rep. 2017;7:46479. [PMC free article] [PubMed] [Google Scholar]

19.    Ardila D, Kiraly AP, Bharadwaj S, Choi B, Reicher JJ, Peng L, et al. End‐to‐end lung cancer screening with three‐dimensional deep learning on low‐dose chest computed tomography. Nat Med. 2019;25(6):954–61. [PubMed] [Google Scholar]

20.    Swiderski B, Kurek J, Osowski S, Kruk M, Barhoumi W, editors. Deep learning and non‐negative matrix factorization in recognition of mammograms. Eighth International Conference on Graphic and Image Processing (ICGIP 2016); 2017: International Society for Optics and Photonics. [Google Scholar]