SCH: Interpretable Machine Learning and Discovery in Medical Images

Duke University, Durham NC

Investigators

Abstract

Breast and lung cancer remain leading causes of cancer-related mortality. Despite advances in machine learning (ML) that have enhanced risk prediction in cancer screening, the lack of interpretability in these "black box" models limits their clinical utility. There is a significant unmet need for diagnostic tools that not only predict cancer risk but also provide actionable insights into the underlying imaging biomarkers. This work develops dimension reduction and interpretable ML architectures to reveal subtle imaging features associated with cancer risk. The hypothesis is that interpretable ML methods can be applied to mammography and chest CT images to uncover previously unrecognized imaging biomarkers to improve early cancer detection and ena ble personalized screening strategies. This proposal has four specific aims: Aim 1: Develop and validate robust dimension reduction tools to visualize high-dimensional imaging features from deep learning models, providing a platform to identify clinically relevant biomarkers. Aim 2: Establish an interpretable framework for characterizing breast tissue in mammography by developing a foundation model to capture multi dimensional texture beyond density measures and deriving a radiologist-understandable lexicon for subtle breast asymmetries linked to short-term cancer risk. Aim 3: Create clinically interpretabl e ML architectures for chest CT imaging by designing models for anomaly detection that distinguish high-risk features from normal variations, and identifying radiologist-understandable features that reduce the need for a computational model. Aim 4: Develop integrated clinical risk scoring algorithms that combine imaging biomarkers with patient data to produce user-friendly risk calculators for both breast and lung cancer. The expected outcome of this project is a suite of validated , clinically relevant tools for early cancer detection and personalized screening. If successfu l, this work can impact cancer screening practices, enhancing early detection, guiding personalized treatment decisions, and ultimately reducing cancer mortality. This work has the potential to shift current paradigms in cancer risk assessment by enabling rapid, evidence-based risk stratification that can be seamlessly incorporated into different clinical settings. RELEVANCE (See instructions): This study targets early detection of breast and lung cancer - two of the leading causes of cancer deaths. New artificial intelligence tools will be developed to clearly show what imaging features indicate high risk, which will help doctors make better decisions about screening. Ult imately, these improvements will help patients get personalized care and can save lives and reduce the burden of cancer on communities.