Similarity-based Deep Learning and Computer Vision Methods for Registration of Coronary Angiogram-Fluoroscopy

Abstract

Cardiovascular disease (CVD) is the second leading cause of death in Korea and the leading cause of death worldwide. CVD primarily affects the heart and major arteries, including the coronary arteries, which supply blood to the heart muscle. Coronary artery disease, a type of CVD, involves the buildup of cholesterol and fatty deposits in artery walls, leading to narrowed or blocked arteries that hinder blood flow. This condition can result in myocardial infarction, where heart muscle cells die due to lack of blood, or angina pectoris, characterized by arteries narrowed by over 70% without tissue death. In severe cases, these conditions can cause sudden death from a heart attack. Percutaneous Coronary Intervention (PCI) is a crucial treatment and diagnostic method for this disease. PCI involves using guidewires and stents to reopen blood vessels and restore normal blood flow, requiring high precision and expertise. During PCI, practitioners must manage dual-monitor imaging angiography and real-time fluoroscopy, which involves significant radiation exposure due to contrast agents. To address these challenges, the dynamic coronary roadmap (DCR) have been developed. By using ECG-gating, DCR matches angiographic images with real-time fluoroscopic images to assist surgeons. Recently, deep learning-based registration methods have been developed to cover patients with irregular heartbeats without using electrocardiograms. They used transfer learning to match images, tailored specifically to each patient. However, they need manual labeling of blood vessels before live fluoroscopy. Moreover, the transfer learning takes about ten minutes, which may not be practical in real clinical setting. In order to reduce the transfer learning time, this thesis proposes a new efficient transfer learning method that utilizes blood vessel similarities in both angiograms of a pre-trained model and the patient of interest. Similarities are computed by principal component analysis and cosine similarity. In addition, a residual U-Net architecture, employing residual blocks and leaky ReLU activation, is proposed to accelerate the learning process. These made about twice fast transfer learning (from 10 to 5 minutes). To achieve even faster registration, this thesis proposes an alternative traditional computer vision method. In this method, blood vessels are segmented by conventional computer vision techniques such as Frangi filters, etc. and careful inlet matching between the diagnostic catheter in the angiogram and the guide catheter in the fluoroscopy. Then an overlap similarity is used to best match the guidewire and the segmented blood vessel with final refinement of two registered images by RANSAC algorithm. This conventional method can register two X-ray images within one minute with comparable registration errors with previous methods across diverse clinical data. These technologies could be used as appropriate visual guidance during coronary interventions.