Transformer Network with Self-Supervised Learning for Stenosis Detection in CT Angiography

Abstract

Coronary artery stenosis is a common coronary artery disease (CAD) that may pose high risk to the life of patients. However, the poor imaging quality at lesions causes difficulties for automatic detection of stenosis in cardiac CT angiography. Previous supervised learning methods improve the robustness of detection by introducing networks with strong context modeling capabilities such as RNN and Transformer, yet requiring large-scale dataset for a high performance. In this paper, we propose a novel self-supervised Transformer network for stenosis detection in multi-planar reformatted (MPR) images reconstructed with the centerlines of the coronary arteries. A Transformer with cross-shaped attention, which can capture the global information of coronary branches efficiently in the MPR images, is introduced into the proposed network. Moreover, an auxiliary self-supervised learning task that encourages the Transformer network to learn spatial relations within an image is introduced. Extensive experiments are conducted on a dataset of 78 patients annotated by experienced radiologists. The results illustrate that the proposed method achieved better results in F1 (0.79) than other state-of-the-art methods.