An end-to-end joint learning framework of artery-specific coronary calcium scoring in non-contrast cardiac CT

Abstract

Accurate identification and quantification of coronary artery calcification play an import role in early diagnosis of coronary heart disease and atherosclerosis. In this paper, we have proposed an end-to-end joint learning framework (CAC-Net) for artery-specific coronary calcification identification in non-contrast cardiac CT. Unlike the previous methods, the framework establish direct mapping relationship between input CT and calcification, consequently, it can directly yield accurate results based on the given CT scans in testing process. In this framework, the intra-slice calcification features are collected by an U-DenseNet module, which is the combination of Dense Convolutional Network (DenseNet) and U-Net. Subsequently, 3D U-Net is performed to extract the inter-slice calcification feature. Joint learning of 2D and 3D module brings rich semantic features, which are beneficial to artery-specific calcification identification. In our experiment, 169 non-contrast CT exams collected from two centers are used to validate the performance of our framework. By the cross validation, we have achieved a sensitivity of 0.905, a PPV of 0.966 for calcification number and a sensitivity of 0.933, a PPV of 0.960 and a F1 score of 0.946 for calcification volume, respectively. The intra-class correlation coefficient are 0.986 for Agatston score and 0.982 for volume score. The quantitative results indicate that our method can be used as a reliable clinical diagnostic tool for coronary calcification identification.