Accurate segmentation and labeling of coronary artery segments in X-ray angiography with an improved UNet-based cGAN architecture

Abstract

X-ray coronary angiography (XCA) is the gold standard for the diagnosis and treatment of coronary artery disease (CAD). Accurate segmentation and labeling of coronary artery segments is critical in the CAD diagnostic process. This study introduces UCNet, an instance segmentation method that combines conditional generative adversarial networks (cGAN) with an improved UNet architecture, to improve the labeling and segmentation of coronary segments in XCA images. By leveraging binary segmentation images of coronary vessels as condition variables, our approach facilitates data generation based on specific criteria. To accurately identify and delineate each coronary segment, we propose a novel segment loss function that utilizes the intersection between predicted masks and ground truth for each segment, thereby improving the accuracy of instance segmentation. In addition, to mitigate class imbalance among vessel segments, we incorporate focal loss and multi-class dice loss to improve the detection of underrepresented segments. Evaluation of UCNet on the ARCADE Challenge datasets at MICCAI 2023 shows an average F1 score of 84.43% across 20 coronary segments. This segmentation performance is superior to state-of-the-art coronary segment labeling methods, despite being trained on a smaller amount of labeled data. Furthermore, our improved UNet significantly outperforms six mainstream U-shaped architectures (including UNet, UNet++, nnUNet, AttentionUNet, SwinUNet, and TransUNet) for vessel labeling and boundary segmentation in terms of accuracy, sensitivity, specificity, precision, intersection over union (IoU), and F1 scores. These results confirm the effectiveness and practicality of our proposed method.