Breast Cancer Risk Prediction Using Deep Learning

HomeRadiologyVol. 301, No. 3 PreviousNext Reviews and CommentaryFree AccessEditorialBreast Cancer Risk Prediction Using Deep LearningMin Sun Bae , Hyug-Gi KimMin Sun Bae , Hyug-Gi KimAuthor AffiliationsFrom the Department of Radiology, Inha University Hospital and School of Medicine, 27 Inhang-ro, Jung-gu, Incheon 22332, South Korea (M.S.B.); and Department of Radiology, Kyung Hee University Hospital, Seoul, South Korea (H.G.K.)Address correspondence to M.S.B. (e-mail: [email protected]).Min Sun Bae Hyug-Gi KimPublished Online:Sep 7 2021https://doi.org/10.1148/radiol.2021211446MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also the article by Zhu et al in this issue.Dr Min Sun Bae is an associate professor of radiology at Inha University Hospital and School of Medicine and serves on the editorial board in the breast imaging section of the Korean Journal of Radiology. Her research interests focus on improving screening outcomes and applying deep learning methods to mammography and breast MRI.Download as PowerPointOpen in Image Viewer Dr Hyug-Gi Kim is a senior researcher of radiology at Kyung Hee University Hospital. His research interests focus on development of medical customized deep learning models and applying deep learning methods to medical data.Download as PowerPointOpen in Image Viewer Mammography is the only screening test that has been proven to reduce breast cancer mortality in randomized controlled trials. Although the starting age and screening interval differ between countries, mammography is the standard of care for breast cancer screening in most countries. The sensitivity of mammography, however, depends in part on tumor size, conspicuity, and breast tissue composition. There is an increasing awareness of the need for supplemental screening in subgroups of certain women for whom mammography is less effective (1). The addition of breast MRI or US to mammography increases the detection rate of small node-negative cancer in women at higher-than-average risk for breast cancer. Even in women with dense breasts at average risk, supplemental screening tests (eg, MRI, US, and digital breast tomosynthesis) in addition to mammography increase cancer detection over that with mammography alone (1).The risk of developing breast cancer varies among individuals. Breast cancer risk can be predicted using existing risk assessment tools, such as the Tyrer-Cuzick and Gail models. These models were developed based on populations with information available on known risk factors (2). It has been shown that mammographic breast density is an independent risk factor for breast cancer and that there is a strong association between mammographic parenchymal patterns and breast cancer risk (3). The more complex and denser the breast parenchyma, the higher the risk of subsequent breast cancer. Image-based risk assessment models might enable more accurate risk prediction at the individual level.Recently, researchers have shown that the mammography-based deep learning (DL) models improve the performance for predicting breast cancer risk (4,5). Yala et al (4) found that a DL model using mammographic images outperformed the Tyrer-Cuzick model (area under the receiver operating characteristics curve [AUC], 0.68 vs 0.62, respectively; P < .01). A hybrid DL model incorporating both mammographic images and traditional risk factors showed the best performance (AUC, 0.70). Dembrower et al (5) demonstrated that the performance of a DL mammography-based model was better than that of the best model based on breast density (AUC, 0.65 vs 0.60, respectively; P < .001).Breast cancer is a heterogeneous disease and consists of intrinsic molecular subtypes that have distinct biologic and clinical features. Indolent breast cancers are more likely to be detected at screening, whereas more aggressive cancers are diagnosed in the interval between two screening rounds (interval cancers). The detection rate of interval cancers is one of the key indicators that monitor the performance of breast cancer screening programs. Screen-detected breast cancers are often of a smaller size, at an earlier stage, and have positive hormone receptor status. Interval breast cancers have the characteristics of rapid tumor growth, being at a more advanced stage, and have a poor prognosis (6). It is not known, however, if the risk estimates for breast cancer differ in women who later develop screen-detected versus interval cancers.In this issue of Radiology, Zhu and colleagues (7) investigated the ability of DL models to estimate the risk of developing screen-detected and interval cancers. The authors used breast cancer screening data obtained over 9 years (2006–2014) at two centers in the United States and included nearly 23 000 mammograms from 5708 women. A convolutional neural network was trained and validated with cancer-free mammographic images from 4039 women (training set, n = 3231; validation set, n = 808). The negative mammograms were obtained, on average, 3 years before breast cancer diagnosis (standard deviation, 1.6 years). The authors present the results of the DL model tested on 1669 women, 538 of whom were subsequently diagnosed with screen-detected (n = 431) or interval (n = 107) cancers. All breast cancers were invasive. In this study, screen-detected cancer was defined as cancer that occurred within 1 year of a positive screening mammogram (Breast Imaging Reporting and Data System [BI-RADS] category 0, 3, 4, or 5), and interval cancer was defined as cancer that occurred within 1 year of a negative screening mammogram (BI-RADS category 1 or 2). The input of the model was four standard mammographic views (ie, bilateral craniocaudal and mediolateral oblique views). The output of the model was a prediction for cancer versus noncancer (matched control group) and a prediction for classifying screen-detected cancer, interval cancer, and noncancer. The DL models were adjusted for clinical factors including age, body mass index, family history of breast cancer, history of breast biopsy, and race. Breast density was assessed based on both qualitative (BI-RADS density grade) and quantitative (density measured with automated software) methods.Zhu et al (7) assessed the performance of the image-based DL model (DL analysis of mammograms only), clinical risk model (based on clinical data including breast density), and combined DL model (combined DL analysis of mammograms with clinical data including breast density). In the test set of 1669 women, the image-based DL model had better performance than the clinical risk model for comparing women with screen-detected cancers versus controls (C statistics of 0.66 and 0.62, respectively). However, the DL model did not perform better than the clinical risk model in comparing interval cancers versus controls (C statistics of 0.64 and 0.71, respectively). The C statistics of the combined DL model were 0.66 and 0.72 for predicting the risk of screen-detected and interval cancers, respectively.The study suggests that mammographic images contain indicators of breast cancer risk and that DL can contribute to risk assessment. This is consistent with other recent studies on the DL breast cancer risk models (4,5). Mammography-based DL models are based on the rich information contained within the images beyond a breast density measure, which is perhaps beyond the limit of human detection. The study also offers some interesting evidence that the mammography-based DL models can stratify risk prediction of screen-detected and interval cancers. However, as Zhu et al noted, there are limitations to the study. As DL results are not easily explainable, we may not explain why conclusions are reached. In the study by Zhu et al (7), heat maps or saliency maps could be obtained to show regions within the mammographic images identified in women at high versus low risks of developing breast cancer. This visual explanation could provide insight into what image patterns the DL models are using to predict the risk of screen-detected and interval cancers (8). There is little doubt that more DL studies are needed to produce further advances in image-based risk prediction. Other promising approaches to improve breast cancer risk prediction include genetic risk factors and blood-based markers (9). In the future, DL models incorporating mammographic images and genetic and traditional risk factors may enable more accurate and individualized risk prediction.In the era of personalized screening, it has been recognized that a "one-size-fits-all" approach cannot be applied for breast cancer screening because breast cancer is a highly heterogeneous disease (10). Furthermore, each individual woman has a different risk for breast cancer. It should also be noted that some of the current screening guidelines take into account cost-effectiveness in the general population. It is possible that women should have personalized screening intervals and tailored options of supplemental screening. The decision about breast cancer screening should be made through a shared decision-making process, enabling women to make an informed decision with their physicians based on individual risk assessment (10). Further work is therefore needed to strengthen the risk prediction models to support personalized approaches to breast cancer screening.Disclosures of Conflicts of Interest: M.S.B. disclosed no relevant relationships. H.G.K. disclosed no relevant relationships.References1. Comstock CE, Gatsonis C, Newstead GM, et al. Comparison of abbreviated breast MRI vs digital breast tomosynthesis for breast cancer detection among women with dense breasts undergoing screening. JAMA 2020;323(8):746–756. Crossref, Medline, Google Scholar2. Cintolo-Gonzalez JA, Braun D, Blackford AL, et al. Breast cancer risk models: a comprehensive overview of existing models, validation, and clinical applications. Breast Cancer Res Treat 2017;164(2):263–28.[Published correction appears in Breast Cancer Res Treat 2017;164(3):745.]. Crossref, Medline, Google Scholar3. Gastounioti A, Conant EF, Kontos D. Beyond breast density: a review on the advancing role of parenchymal texture analysis in breast cancer risk assessment. Breast Cancer Res 2016;18(1):91. Crossref, Medline, Google Scholar4. Yala A, Lehman C, Schuster T, Portnoi T, Barzilay R. A deep learning mammography-based model for improved breast cancer risk prediction. Radiology 2019;292(1):60–66. Link, Google Scholar5. Dembrower K, Liu Y, Azizpour H, et al. Comparison of a deep learning risk score and standard mammographic density score for breast cancer risk prediction. Radiology 2020;294(2):265–272. Link, Google Scholar6. Porter PL, El-Bastawissi AY, Mandelson MT, et al. Breast tumor characteristics as predictors of mammographic detection: comparison of interval- and screen-detected cancers. J Natl Cancer Inst 1999;91(23):2020–2028. Crossref, Medline, Google Scholar7. Zhu X, Wolfgruber TK, Leong L, et al. Deep learning predicts interval and screening-detected cancer from screening mammograms: a case-case-control study in 6369 women. Radiology 2021.https://doi.org/10.1148/radiol.2021203758. Published online September 7, 2021. Link, Google Scholar8. Norgeot B, Quer G, Beaulieu-Jones BK, et al. Minimum information about clinical artificial intelligence modeling: the MI-CLAIM checklist. Nat Med 2020;26(9):1320–1324. Crossref, Medline, Google Scholar9. Pashayan N, Antoniou AC, Ivanus U, et al. Personalized early detection and prevention of breast cancer: ENVISION consensus statement. Nat Rev Clin Oncol 2020;17(11):687–70.[Published correction appears in Nat Rev Clin Oncol 2020;17(11):716.]. Crossref, Medline, Google Scholar10. Saccarelli CR, Bitencourt AGV, Morris EA. Is it the era for personalized screening? Radiol Clin North Am 2021;59(1):129–138. Crossref, Medline, Google ScholarArticle HistoryReceived: June 7 2021Revision requested: June 22 2021Revision received: June 24 2021Accepted: June 25 2021Published online: Sept 07 2021Published in print: Dec 2021 FiguresReferencesRelatedDetailsCited By2022 International Conference on Electronics and Renewable Systems (ICEARS)RRemya, NHema Rajini2022Accompanying This ArticleDeep Learning Predicts Interval and Screening-detected Cancer from Screening Mammograms: A Case-Case-Control Study in 6369 WomenSep 7 2021RadiologyRecommended Articles BI-RADS Category 3 Is a Safe and Effective Alternative to Biopsy or Surgical ExcisionRadiology2020Volume: 296Issue: 1pp. 42-43BI-RADS Terminology for Mammography Reports: What Residents Need to KnowRadioGraphics2019Volume: 39Issue: 2pp. 319-320Cancer Yield and Patterns of Follow-up for BI-RADS Category 3 after Screening Mammography Recall in the National Mammography DatabaseRadiology2020Volume: 296Issue: 1pp. 32-41Additional Breast Cancer Detection at Digital Screening Mammography through Quality Assurance Sessions between Technologists and RadiologistsRadiology2020Volume: 294Issue: 3pp. 509-517AI to Dismiss Normal Breast MRI Scans and Reduce WorkloadRadiology2021Volume: 302Issue: 1pp. 37-38See More RSNA Education Exhibits Quality Assessment of the BI-RADS 3 Classification and Utility of a BI-RADS 3 AuditDigital Posters2019Ready for BI-RADS Mammography Assessment Categories? Take a Quiz!Digital Posters2018Probably Benign Assessment on Breast MRI: Appropriate and Inappropriate Utilization Digital Posters2018 RSNA Case Collection BI-RADS 4C - High suspicion for malignancyRSNA Case Collection2022BI-RADS 2: Dystrophic CalcificationsRSNA Case Collection2022BI-RADS 2 - Steatocytoma multiplex RSNA Case Collection2022 Vol. 301, No. 3 Metrics Altmetric Score PDF download