김중선, 이준상 교수 논문발표

김중선 교수(연세대학교), 이준상 교수(연세대학교)가 Anatomical Prior Knowledge-Informed Neural Network for Segmentation of Coronary Artery and Ascending Aorta 라는 제목의 논문을 ELSEVIER에 발표하였습니다.

해당내용 확인하기 ◀ 클릭

Abstract

With the increase in the number of patients suffering from coronary artery disease, the demand for early diagnosis through coronary computed tomography angiography (CCTA) is increasing. Image segmentation is considered as a prerequisite for quantitatively analyzing medical images; thus, should be automated to properly manage the vast amounts of data. Although several studies have presented deep learning-based medical image segmentation, most of them have not considered anatomical knowledge in the domain.In this study, we propose an anatomical prior knowledge-informed segmentation method for the segmentation of both the ascending aorta and coronary artery using CCTA. Based on the observations from CCTA, we reflect on the anatomical knowledge at each stage of the proposed strategy. During image preprocessing, we start by eliminating the pulmonary vessels to suppress potential false positives. In addition, we adopt a multi-class segmentation approach considering the significant class imbalance between the ascending aorta and coronary artery. Finally, we propose and integrate a novel loss function specific to cardiovascular image segmentation.Through elaborate experiments, the proposed method is shown to be superior to conventional approaches, achieving Dice scores of 0.980 and 0.788 for the ascending aorta and coronary artery, respectively; showing the best performance among compared methods. We consider this work as an effective prerequisite for achieving patient-specific hemodynamic modeling and ultimately, is the first step towards realizing precision medicine.