Interpretable Deep Learning Models for Analysis of Longitudinal 3D Mammography Screenings

点击翻译按钮获取中文摘要

Project Summary Mammography screening for breast cancer has clear, substantial benefits, including significantly reduced breast cancer mortality and improved treatment options for early detected cancers. However, regular mammography screenings subject women to several potential harms, including high false positive rates, with over 60% of women experiencing a false positive finding after 10 years of annual screening; high false negative rates, with more can- cers missed in dense breasts which obscure tumor appearance; and high recall rates, causing undue anxiety and unnecessary, potentially invasive workup for women with a false positive screen. New 3D mammography technology called digital breast tomosynthesis (DBT) has shown increased cancer detection and decreased re- call rates, but radiologists require longer interpretation time and may lack experience. The clinical workflow could potentially be enhanced with computer aided detection systems. However, current methods only focus on a single mammogram exam, ignoring crucial decision-making information that a radiologist would consider, such as prior mammograms, patient demographics, and personal history. Conversely, established breast cancer risk models rely only on patient demographics and personal/family history, excluding mammographic history. Toward the overarching goal of reducing the harms and increasing the benefits of mammography screening, we propose to increase accuracy of breast cancer detection and predict future cancer development from serial 3D mammogram screenings using a novel deep learning model that jointly incorporates spatial, temporal, and non-imaging clinical information. Our method adopts attention-based neural networks, i.e., Transformers, which learn complex depen- dencies between different elements in a sequence and automatically attend to the most relevant information. In addition to the potential for improved performance, the attention mechanism provides built-in model interpretation to better understand the inputs that are important for the model’s predictions, instilling user confidence in the model and facilitating extraction of mammographic biomarkers for breast cancer detection and development. Our specific aims are to: 1) develop a powerful deep learning model for simultaneously leveraging spatial, temporal, and non-imaging clinical information from DBT exams; 2) create a new tool to detect breast cancer from lon- gitudinal DBT screenings; and 3) develop a new model for predicting development of breast cancer based on longitudinal DBT studies and extract 3D mammographic biomarkers associated with cancer development. Be- yond the direct benefit of improved breast cancer detection and risk estimation, this work could reduce radiologist reading time and workload, inform new individualized screening protocols, further our understanding of the role of breast architecture in cancer risk, and guide development and monitoring of preventive treatments. Finally, the developed deep learning methodology will have wide applicability to spatiotemporal analysis in other medical conditions and imaging domains.