Towards Interpretable AI for Longitudinal Disease Monitoring and Clinical Reporting from Chest X-Rays

Abstract

Chest radiography (CXR) plays a pivotal role in diagnostic imaging for monitoring disease progression and evaluating treatment effectiveness. Despite notable advancements in machine learning, disease progression monitoring remains relatively underexplored. Challenges arise from the specificity of biomarkers that detect change, which vary in their mechanisms, manifestations, and progression rates across diseases, alongside individual variability in response to illness and the complexity of incorporating multimodal longitudinal data. Monitoring disease progression in chest imaging involves intricate tasks, such as anatomical motion estimation and image registration, which require the spatial alignment of sequential X-rays and modeling temporal dynamics. This thesis addresses these challenges by harnessing artificial intelligence techniques for effective disease progression monitoring using non-co-registered sequential CXRs. We investigate three research directions: 1) learning a disease progression model with local and global information, 2) explainable hierarchical learnable differences for disease progression monitoring, and 3) retrieval-augmented longitudinal disease report generation. The overarching goal of this thesis is to develop models that not only accurately track disease progression but also provide interpretable insights about the patient's condition. The three directions in this thesis form a unified strategy for building interpretable, temporally aware AI models that enhance disease monitoring and reporting. Together, these contributions advance early detection and informed treatment decisions, with the potential to significantly improve patient outcomes at scale.