Symptom Disentanglement in Chest X-Ray Images for Fine-Grained Progression Learning

Chest radiography is a commonly used diagnostic imaging exam for monitoring disease severity. Machine learning has made significant strides in static tasks (e.g., segmentation or diagnosis) based on a single medical image. However, disease progression monitoring based on longitudinal images remains fairly underexplored, which provides informative clues for early prognosis and timely intervention. In practice, the development of underlying disease typically accompanies with the occurrence and changes of multiple specific symptoms. Inspired by this, we propose a multi-stage framework to model the complex progression from symptom perspective. Specifically, we introduce two consecutive modules namely Symptom Disentangler (SD) and Symptom Progression Learner (SPL) to learn from static diagnosis to dynamic disease development. By explicitly extracting the symptom-specific features from a pair of chest radiographs using a set of learnable symptom-aware embeddings in SD module, the SPL module can leverage these features to obtain the symptom progression features, which will be utilized for the final progression prediction. Experimental results on the public dataset Chest ImaGenome show superior performance compared to current state-of-the-art method. Code is available at: https://github.com/zhuye98/SDPL.git.