Vascular anatomy-aware self-supervised pre-training for X-ray angiogram analysis
VasoMIM introduces a self-supervised pre-training framework for X-ray angiogram analysis by integrating anatomical knowledge. It uses anatomy-guided masking and consistency loss to train models to learn robust vascular semantics directly from large datasets like XA-170K. This creates a foundation model that significantly outperforms existing methods across multiple clinical tasks without needing extensive manual annotations.