Two-stage training strategy combined with neural network for segmentation of internal mammary artery graft

Abstract

Cardiovascular disease is a major health condition affecting the global population, leading to 420 million diagnoses and nearly 20 million deaths annually. Coronary heart disease is one of the most prevalent cardiovascular diseases, and its diagnosis and treatment creates a large social and economic burden, necessitating a large number of imaging evaluations and repetitive manual assessments by expert radiologists. Recently, deep learning-based methods have demonstrated excellent performance in coronary artery recognition (coronary artery segmentation and centerline extraction, etc.). However, current research focuses on CTA and MRI data types, and for inexperienced cardiologists, even if the neural network gives accurate segmentation results, it is impossible to determine its specific location in the heart. In response to the above problems, we constructed a novel dataset, CardiacVessel, which contains 2D slices from CTA images through 3D reconstruction techniques, and performed pixel-level annotations of vessels. We then propose a simple and effective two-stage training strategy for internal mammary artery graft segmentation. We conduct extensive experiments to compare the performance of various CNN- and transformer-based network architectures. The proposed two-stage training strategy can significantly improve network performance.