(61d) Deep Learning for Precise Breast Cancer Diagnosis and Risk Stratification: A Multi- Institutional Study

Annually, about a million breast biopsies are performed in the United States, out of which typically a quarter receive a cancer diagnosis^1,2. The remaining benign or in-situ (stage 0) patient cases that are combined into the category of intraductal lesions of the breast are susceptible to over or under-diagnosis. Current standards for separating benign intraductal lesions from ductal carcinoma in-situ (DCIS) rely on visual inspection of morphometric features using molecularly stained slides. This involves manual interpretation of architectural and cytological patterns in the epithelial compartments of the tissue. This can often be time-consuming and lead to discordant measurements³. Intraductal proliferations are primarily composed of Usual Ductal Hyperplasia (UDH), Atypical Ductal Hyperplasia (ADH), and Ductal Carcinoma In -Situ (DCIS). The treatment profiles for all these three categories are different, making it important to diagnose these lesions precisely. In ADH cases, the affected area is excised, whereas the UDH lesion does not require any further management. DCIS is treated by either excision or complete mastectomy followed by radiation/neoadjuvant therapy.

Deep learning models can enable the identification of global and local textural patterns that are indicative of an altered epithelial profile^4,5. This allows the mapping of different architectural patterns of the epithelium into precise risk and disease categories. Recently, there have been some advances in digital approaches for breast cancer analysis. However, there are only a few studies that investigated pre-invasive lesions and early-stage tumors. One of the recent studies⁶ reports a 70% accuracy to distinguish benign samples from DCIS. One of the limitations of this study is that the proposed model works only on selected regions of interest, making it difficult to fully extend to large slide sections. Here, we present a deep neural network-based digital approach for analyzing stained slides of large surgical specimens to separate benign lesions (normal, usual hyperplasia and atypical hyperplasia combined) from DCIS. About 20000 images were curated for training, 8800 for calibration, and 8800 for independent validation. A pre-trained convolutional neural network (illustrated in the figure) was modified and retrained to adapt to the desired classification task. After inspecting the model accuracies, hyperparameters were optimized for diagnostic modeling. After achieving the desired accuracy, the final model was then projected on large patient areas to validate the model's spatial performance. First, we investigate the relevant features extracted by the developed model for classifying image patches into different classes. In the initial layers, the model primarily extracts the nuclei's outer edges. The subsequent layers estimate the finer edges in the nucleoli and the stromal fibers interspersed in the epithelial nuclei. In summary, the model is automatically segmenting the nuclei and extracting different nuclear features for the diagnostic task. High accuracy of ~85% (~20-40% higher than the current standards and reports) in the independent validation step suggests that the model is generalizable and translatable to a large variety of patient samples. Finally, the model was used to classify different lesions extracted from the set of external validation patients. The model identifies areas of DCIS and separates them from the benign population (normal + UDH + ADH).

This is the first study to report this level of accuracy on whole slide image sections. Additionally, since an entire image's spatial patterns are estimated, stromal alterations are also indirectly incorporated in the model, making it more comprehensive for diagnosis. Precise stratification of the intraductal lesions presented here can help overcome the problem of under or over-diagnosis associated with these lesions. The proposed pipeline could also assist pathologists in borderline cases or differential diagnosis such as high-grade hyperplasia and low-grade ductal carcinoma in-situ (DCIS). It will address the long-standing need for precise triaging in a quantitative manner, improving patient outcomes. The developed digital toolbox will be of interest to both healthcare clinics for risk stratification as well as early detection. Optimized diagnostic protocols in this direction will affect approximately a million biopsy samples performed annually, especially benign cases.

References

1. R.L. Siegel, K.D. Miller, A. Jemal. "Cancer statistics, 2019". CA. Cancer J. Clin. 2019. 69(1): 7–34.
2. "U.S. Breast Cancer Statistics | Breastcancer.org". https://www.breastcancer.org/symptoms/understand_bc/statistics [Mar 7 2016].
3. J.G. Elmore, G.M. Longton, P.A. Carney, B.M. Geller, T. Onega, A.N.A. Tosteson, et al. "Diagnostic Concordance Among Pathologists Interpreting Breast Biopsy Specimens". JAMA. 2015. 313(11): 1122–1132.
4. G. Litjens, C.I. Sánchez, N. Timofeeva, M. Hermsen, I. Nagtegaal, I. Kovacs, et al. "Deep learning as a tool for increased accuracy and efficiency of histopathological diagnosis". Sci. Rep. 2016. 6(May): 1–11.
5. B. Gecer, S. Aksoy, E. Mercan, L.G. Shapiro, D.L. Weaver, J.G. Elmore. "Detection and classification of cancer in whole slide breast histopathology images using deep convolutional networks". Pattern Recognit. 2018. 84: 345–356.
6. E. Mercan, S. Mehta, J. Bartlett, L.G. Shapiro, D.L. Weaver, J.G. Elmore. "Assessment of Machine Learning of Breast Pathology Structures for Automated Differentiation of Breast Cancer and High-Risk Proliferative Lesions". JAMA Netw. Open. 2019. 2(8).