Self-Supervised X-Ray Coronary Angiography Segmentation with Vessel-Aware Synthesis Learning

Percutaneous coronary intervention (PCI) re quires precise, real-time coronary vessel segmentation to ensure procedural safety and efficacy. Self-supervised vascular segmentation technology has attracted widespread attention because it does not rely on artificial standards and has good segmentation effects. However, existing self supervised methods focus primarily on the topological structure of vessels, neglecting texture and geometric de tails. To address this, this paper uses mixed data of real information and synthetic information for training, and simulates the relationship between the two brightness attenuation values of X-rays after passing through the background and blood vessels respectively during the data mixing process. This method utilizes vessel texture synthesis and geometric contour bending in self-supervised vesselseg mentation, and more accurately simulates the relationship between the brightness attenuation of X-rays after passing through the background and blood vessels respectively through the product of ray attenuation rates (instead of the traditional attenuation value addition), thereby solving the object model defects of the existing self-supervised vessel segmentation framework. Experiments have shown that Vessel Aware Synthesis Learning can significantly re duce segmentation errors (Jaccard increased by 9.6%) and capture more vascular details (Recall increased by 13.8%). The proposed method can effectively eliminate the reliance on labor-intensive manual annotations. This development improves AI-driven medical image analysis and offers a practical approach to reducing the cost of creating reliable AI tools for medical imaging by decreasing the need for extensive annotations. Source code is available at https: //github.com/bigOrange123456789/SSCVS.git.