Cross-Domain Transfer Learning for Vessel Segmentation in Computed Tomographic Coronary Angiographic Images

Abstract

Segmenting coronary arteries in computed tomographic angiography images is an essential procedure for coronary artery disease diagnosis. However, it still remains challenging due to the insufficient annotation data for supervised deep learning methods. To solve this problem, we propose a novel cross-domain transfer learning network to adaptively transfer knowledge learned from public liver vessel dataset for coronary artery segmentation. The signed distance map learning task is joined to enforce the network to transfer tubular structure knowledge from the liver vessel. Moreover, an adaptive feature-selection module is used to determine the optimal fine-tune strategy for every target sample. We conduct ablation experiments to demonstrate the effectiveness of the auxiliary task and module. We also compare the proposed method with other state-of-the-art transfer learning and segmentation methods. Results showed that our method achieve the best performance on accurate coronary artery segmentation. Our method achieves the best Dice score of 81.60%, an improvement of at least 1% with respect to other methods.