Abstract
Percutaneous coronary angiography remains the diagnostic gold standard for coronary artery disease. However, the complex and high-volume nature of the imaging data renders the clinical interpretation of coronary lesions a time-consuming, labor-intensive, and inherently subjective process. This retrospective study collected and preprocessed Coronary artery angiography (CAG) image data from 408 patients with acute myocardial infarction (AMI). An improved YOLOv4 algorithm was developed, validated on standard VOC datasets, and subsequently calibrated via transfer learning on the CAG training set for automated lesion detection and classification. The model-derived lesion characteristics were then statistically correlated with the occurrence of Major Adverse Cardiovascular Events (MACEs) during patient follow-up. The improved model achieved a post-modification mean Average Precision (mAP) of 84.72% (95% CI: 83.44-85.99%) on the VOC dataset. For coronary lesion detection, the model yielded an overall mean Average Precision (mAP) of 55.01%. Importantly, lesion characteristics automatically detected by the model-specifically completely occluded lesions (Log-rank p = 0.003) and multibranching lesions (Log-rank p = 0.033)-demonstrated a significant association with the cumulative incidence of MACEs. The innovative, improved YOLOv4 model exhibits robust performance in effectively and accurately detecting and classifying coronary lesions within AMI patient angiography imagery. This study provides a valuable AI-assisted diagnostic tool and offers preliminary insights for long-term prognostic assessment by seamlessly integrating deep learning-derived anatomical features with MACEs prediction.