Abstract
Coronary artery stenosis detection by invasive coronary angiography plays a pivotal role in computer-aided diagnosis and treatment. However, it faces the challenge of stenotic morphology confusion stemming from coronary-background similarity, varied morphology, and small-area stenoses. Furthermore, existing automated methods ignore long-temporal information mining. To address these limitations, this paper proposes a long-term temporal enhanced You Only Look Once (YOLO) method for automatic stenosis detection and assessment in invasive coronary angiography. Our approach integrates long-term temporal information and spatial information for stenosis detection with state-space models and YOLOv8. First, a spatial-aware backbone based on a dynamic Transformer and C2f Convolution of YOLOv8 combines the local and global feature extraction to distinguish the coronary regions from the background. Second, a spatial-temporal multi-level fusion neck integrates the long-term temporal and spatial features to handle varied stenotic morphology. Third, a detail-aware detection head leverages low-level information for accurate identification of small stenoses. Extensive experiments on 350 invasive coronary angiography (ICA) video sequences demonstrate the model's superior performance over seven state-of-the-art methods, particularly in detecting small stenoses (<50%), which were previously underexplored.