Abstract
Iodinated contrast agents are widely utilized in numerous interventionalprocedures, yet posing substantial health risks to patients. This paperpresents CAS-GAN, a novel GAN framework that serves as a "virtual contrastagent" to synthesize X-ray angiographies via disentanglement representationlearning and vessel semantic guidance, thereby reducing the reliance oniodinated contrast agents during interventional procedures. Specifically, ourapproach disentangles X-ray angiographies into background and vesselcomponents, leveraging medical prior knowledge. A specialized predictor thenlearns to map the interrelationships between these components. Additionally, avessel semantic-guided generator and a corresponding loss function areintroduced to enhance the visual fidelity of generated images. Experimentalresults on the XCAD dataset demonstrate the state-of-the-art performance of ourCAS-GAN, achieving a FID of 5.87 and a MMD of 0.016. These promising resultshighlight CAS-GAN's potential for clinical applications.