Synergistic Multi-Task Learning for a Unified Framework of Intelligent Coronary Artery Disease Reporting and Data System

Abstract

The latest clinical guideline of the Coronary Artery Disease Reporting and Data System (CAD-RADS) emphasizes comprehensive CAD risk evaluation, driving the development of automated diagnosis technologies toward a unified multi-task framework. Previous task-specific architectures, relying on varied pre- and post-processing, showed redundancy and prediction inconsistencies. To address this, we first proposed synergistic multitask learning and constructed a unified CAD-RADS framework. It provides a collaborative diagnosis of the CAD-RADS level, coronary artery calcium, the segment involvement score, and abnormality modifiers based on the patient's CT Angiography (CTA) volume. On the one hand, we integrate offset features across multiple scales to learn distinct attention distributions for different tasks in the latent space, thereby meeting the representation requirements for task customization within a unified architecture. On the other hand, we employ ExpectationMaximization (EM)-driven iterative optimization to interactively learn a compact basis consensus among tasks, balancing them and promoting semantic complementarity. Experimental results based on CTA volumes from 1,068 patients demonstrate our framework outperforms state-of-the-art methods, advancing the clinical application of intelligent CAD-RADS.

Ähnliche Arbeiten