Reducing Redundancy in Small Lesion Features for Multi-Shape Stenosis Detection in Coronary X-ray Angiography

The automatic detection of multi-shape coronary artery stenosis through X-ray angiography has important clinical implications for the diagnosis and treatment of coronary artery disease. However, there are several challenges in accurately detecting multi-shape stenoses, including the small size of stenoses, large size variations among multi-shape stenoses, unclear boundaries, and deformation caused by cardiac motion, such as stretching and contraction. To this end, we propose a framework for multi-shape stenosis detection at the sequence level by optimizing the representation of small lesions. Specifically, we propose a contribution-guided redundancy reduction module to suppress the feature redundancy of the background region while dynamically optimizing stenosis representations. Furthermore, to tackle the challenges of classification caused by the similarity in lesion appearance and unclear boundaries, we leverage the relationship between task-specific features and cross-enhancement to improve classification performance. Finally, a sequence relaxation strategy to extend the representation of lesions from the single-frame level to the sequence level. Experimental results indicate that the proposed method demonstrates a significant advantage over comparative methods in the task of multi-shape stenosis detection.