Multi-task slice-plane-driven learning for coronary stenosis and plaque classification in CCTA

Abstract

Coronary artery disease (CAD) remains a leading cause of morbidity and mortality worldwide. It is primarily caused by the gradual accumulation of atherosclerotic plaques in the coronary arteries, leading to luminal narrowing or occlusion, which poses a serious threat to human health. Accurate diagnosis of stenosis severity and plaque characteristics is essential for effective treatment planning. Coronary computed tomography angiography (CCTA), as a non-invasive and high-resolution imaging modality, has been widely adopted for early screening and assessment of CAD. However, traditional interpretation of CCTA relies heavily on manual reading or rule-based semi-automated tools, which are time-consuming and subject to inter-observer variability. In this work, we propose a multi-task, slice-plane-driven deep learning framework to simultaneously classify coronary artery stenosis severity and plaque types. Mimicking clinical multi-view reading procedures, the proposed method constructs a multi-plane reconstruction module to extract anatomical features from the coronal, sagittal, and axial planes. Based on a unified backbone, we develop four model variants—TransformerNet, ConvNet, ConvNet-SCL, and ConvNet-WSCL—to explore the impact of different encoding strategies and loss functions on model performance. Evaluated on a multicenter dataset, the proposed methods outperform existing approaches across multiple classification levels (lesion-level, vessel-level, and patient-level) for both stenosis and plaque analysis. Specifically, ConvNetSCL showed the highest accuracy in stenosis severity classification, with 88.05% accuracy at the lesion level and 83.33% accuracy at the patient level. For plaque type classification, ConvNet-WSCL achieved the best performance, with 84.57% accuracy at the lesion level, demonstrating the effectiveness of the proposed supervised contrastive learning strategies.