Multi-Task Keypoint Detection with Embedding Supervision for Robust X-ray and DRR Matching

Keypoint matching between X-ray and DRR images is challenging due to modality differences, anatomical variations, and occlusions. We propose a unified multi-task framework that integrates heatmap-based localization, visibility prediction, and embedding supervision to robustly align pelvic structures. Built upon a ResNet-FPN backbone, our model leverages soft-argmax for coordinate regression and combines global and keypoint-level embeddings for semantic alignment. Experimental results demonstrate superior performance in localization accuracy, matching robustness, and visibility estimation compared to both traditional and deep learning baselines.