Unsupervised 3D Coronary Angiography Segmentation Based on Generative Adversarial Networks

Yaoxian Yang, Liqing Gao, Jiabao Wen

Abstract: Coronary artery disease (CAD) remains a leading cause of mortality globally. Digital Subtraction Angiography (DSA) is the gold standard for visualizing coronary arteries, but the presence of noise, artifacts, and low contrast poses challenges for automatic segmentation. This paper proposes an unsupervised method for vessel segmentation based on Generative Adversarial Networks (GANs), integrating 3D priors from EG3D and leveraging a style-based generator for enhanced vessel detail and realism. Without any segmentation annotations, our method accurately separates foreground vessels from background, achieving superior detail retention compared to 2D unsupervised approaches. Experiments on real DSA datasets demonstrate the effectiveness and generalizability of our model.

External IDs:doi:10.1007/978-981-95-5634-2_22