Multi-level perception fusion with pseudo-stenosis pretraining for robust coronary artery stenosis detection

Coronary artery stenosis is a major pathological basis of coronary artery disease (CAD), and clinical evaluation relies on coronary angiography (CAG) as the gold standard. However, manual assessment suffers from high subjectivity, low efficiency, and poor inter-observer consistency, highlighting the necessity of automated detection methods. Existing studies are mostly “single-task” approaches (segmentation-only or detection-only), which fail to explicitly exploit the complementary information between the two tasks. Moreover, the scarcity of annotated samples limits generalization in complex vascular structures. To address these challenges, we propose a stenosis detection model that integrates a Multi-level Perception Fusion (MPF) framework with Pseudo-Stenosis Pretraining. The MPF framework places a vessel segmentation module before the detection module, providing vascular spatial constraints and structural cues of stenosis. By propagating detection loss to the vessel segmentation module, this module is also enhanced to highlighting the difference of stenotic regions. Meanwhile, Pseudo-Stenosis Pretraining generates morphologically realistic pseudo stenosis samples and adopts a pretraining–fine-tuning scheme to mitigate overfitting in limited data scenarios. Experimental results demonstrate that the proposed model achieves superior performance compared with state-of-the-art methods on both a private dataset and the public ARCADE dataset. On the private dataset, the model achieves F1, recall, and mean Average Precision (mAP) of 68.1, 73.7, and 72.3 in the detection task, outperforming LSKNet, Oriented-RCNN, ReDet, S2ANet, and RetinaNet. On the ARCADE public dataset, the model achieves F1, sensitivity, and area under the curve (AUC) of 60.52, 62.24, and 99.17, surpassing U-Net, PoolNet, and YOLOv8m.