Boundary-Aware Transformer-UNet for Coronary Vessel Segmentation

Coronary CT angiography(CTA) has become the gold standard for diagnosing heart disease. Precisely segmentation of coronary stenosis is helpful for hemodynamic analysis. Neural Networks such as U-Net have ruled medical image segmentation field. However, Transformer-like encoder can focus on global information rather than focusing on local information with CNNs. Because of the complicated background of coronary angiography, methods above still do not work in certain images. In this paper, we proposed a Swin-Transformer based segmentation network called BATrans-Net with modules that can optimize boundary segmentation results. Experimental results on our coronary dataset show that our BATrans-Net performs more precisely than other networks.