Centerline-supervision multi-task learning network for coronary angiography segmentation

With convolutional neural networks’ remarkable performance in computer vision, more and more studies are applying deep learning to vessel image segmentation tasks. This work focuses on the task of coronary X-ray angiography segmentation, which is critical in the diagnosis of cardiovascular disorders. For vascular segmentation in coronary X-ray angiography images, we propose a novel deep learning model based on the UNet backbone. We first equip a channel attention module in skip-connections to improve pixel-wise segmentation accuracy by emphasizing the effective channel in low-level features. A centerline auxiliary supervision module is also introduced at the network’s end to provide prior knowledge of vessel connectivity and thick vessels, utilizing the existing binary segmentation annotations efficaciously. Consequently, the network can devote more attention to pixels that ensure vessel tree connectivity and high confidence. Extensive experiments demonstrate the effectiveness of the two modules in improving the performance of the model. We compared our results to the recently proposed networks and revealed that these two modules can be added to other U-shaped networks to enhance performance. In our experiments, our method produced the best results in terms of sensitivity and dice score, with 82.48 and 85.28, respectively.