A Hessian-Based Deep Learning Preprocessing Method for Coronary Angiography Image Analysis

Abstract

Leveraging its high accuracy and stability, deep-learning-based coronary artery detection technology has been extensively utilized in diagnosing coronary artery diseases. However, traditional algorithms for localizing coronary stenosis often fall short when detecting stenosis in branch vessels, which can pose significant health risks due to factors like imaging angles and uneven contrast agent distribution. To tackle these challenges, we propose a preprocessing method that integrates Hessian-based vascular enhancement and image fusion as prerequisites for deep learning. This approach enhances fuzzy features in coronary angiography images, thereby increasing the neural network’s sensitivity to stenosis characteristics. We assessed the effectiveness of this method using the latest deep learning networks, such as YOLOv10, YOLOv9, and RT-DETR, across various evaluation metrics. Our results show that our method improves AP50 accuracy by 4.84% and 5.07% on RT-DETR R101 and YOLOv10-X, respectively, compared to images without special pre-processing. Furthermore, our analysis of different imaging angles on stenosis localization detection indicates that the left coronary artery zero is the most suitable for detecting stenosis with a AP50(%) value of 90.5. The experimental results have revealed that the proposed method is effective as a preprocessing technique for deep-learning-based coronary angiography image processing and enhances the model’s ability to identify stenosis in small blood vessels.