Multi-frame dynamic information fusion and vascular structure constraint for real-time enhancement of coronary angiography images

Abstract

Coronary angiography (CAG) serves as the gold standard for diagnosing coronary heart disease, but its poor image quality and high levels of noise interference have consistently affected the diagnoses of physicians and hindered the development of intelligent auxiliary diagnosis for coronary heart disease. To address these problems, we propose a real-time coronary angiography image enhancement network based on multi-frame dynamic information fusion and vascular structure constraint (MFSC-Net), which is a conditional generative adversarial network. First, we introduce multi-scale attention block (MAB) to reduce network parameters, achieving real-time image processing. The generator network includes optical flow information extraction block based on RAFT, feature extraction block (FEB), multi-frame dynamic information fusion block (MDIF), and image reconstruction block (IRB). MDIF fuses the optical flow information of key-frame with key-frame itself at the feature level, thereby enhancing image with low vascular contrast and suppressing the background. The vascular structure constraint (VSC), present in the discriminator, is divided into vascular morphology constraint (VMC) and vascular intensity constraint (VIC), which ensure the continuity, integrity, and realism of the vessels in the enhanced results. Extensive experiments based on our proprietary dataset demonstrate that the coronary angiography image enhancement effect of our proposed MFSC-Net is superior to other state-of-the-art (SOTA) methods. Additionally, our method is of significant importance for reducing surgical risks, improving diagnostic efficiency, and promoting the intelligent auxiliary diagnosis of coronary heart disease. The code is available at https://github.com/yuhui0416/MFSC-Net.