Synthetic-Data-Driven Diffusion Models for Endovascular Intervention Guidance

Endovascular intervention suffers from a critical data scarcity problem: real-patient catheter trajectories are expensive and annotation-limited. We construct a high-fidelity simulation environment (stEVE) based on SOFA FEM with Cosserat beam guidewire models and biplane DSA X-ray projection, generating large-scale synthetic procedural data at scale. We then train a diffusion-based trajectory planner on this synthetic corpus and transfer it to real vascular geometries via a lightweight domain adaptation stage. Evaluated on CAS and aortic arch anatomies, our model achieves navigation success rates competitive with expert RL policies while generalizing to unseen vessel morphologies.