Learning Disentangled Representation for Vessel-Specific Coronary Artery Calcium Scoring

Vessel-specific coronary artery calcium scoring plays a crucial role in the early diagnosis of cardiovascular disease. The key challenge of the algorithm design lies in the ambiguity of calcium category (i.e., affiliated vessel category) prediction due the lack of clear visibility of coronary arteries in calcium scoring CT images. To tackle the challenge, we propose DISCO-Net in this paper, which explicitly incorporates anatomical domain knowledge into vessel-specific calcium scoring. DISCO-Net disentangles the challenge into two simpler sub-problems: calcium category prediction and binary calcium segmentation. To handle the former problem, we design a Representative Region extractor that formulates vessel anatomical knowledge modeling as a region extraction problem which alleviates the category ambiguity issue. To address the latter issue, we design a Class-agnostic Calcium extractor that explicitly considers calcium intensity and spatial distribution characteristics. Combining both extractors, DISCO-Net solves the problem effectively. Experimental results on two public datasets and an in-house dataset demonstrate that our proposed method achieves state-of-the-art performance. Notably, our method, as a fully automated algorithm, achieves comparable performance to independent radiologist on the or-CaScore public dataset.