A Magnetic Source Imaging Registration Method Based on Keypoint Self-Supervised Constraints and Multi-View Differentiable Rendering

In magnetic source imaging systems, precise registration between sensor arrays and individual anatomical structures critically impacts source localization accuracy. Existing mainstream registration methods typically rely on structured light or laser scanning to acquire head point clouds for geometric fitting. While achieving high accuracy under ideal conditions, these approaches generally suffer from complex acquisition workflows, lengthy data collection times, and high equipment costs, limiting their practical application in engineering and clinical settings.This paper proposes a multi-view RGB image-based magnetic source imaging registration method. It models the registration problem as a self-supervised optimization process for rigid body parameters with clear physical meaning, eliminating the need for 3D point cloud reconstruction. The method directly establishes the relationship between the 3D head model and 2D image observations through a geometric projection model, using keypoint reprojection error as the primary constraint. It simultaneously incorporates keypoint confidence weighting and robust kernel functions to effectively suppress false detections and occlusion effects on the optimization process.Experimental results demonstrate that data acquisition is completed in just 30 seconds, achieving processing speeds 4 to 10 times faster than structured light and laser scanning methods. In simulated head model experiments, it achieves a translation error of 0.33 mm and an attitude error of 0.11 degrees of pose error, significantly outperforming structured light solutions while matching the order of magnitude of laser scanning methods. In real-human experiments, it achieved 0.85 mm translation error and 0.21 degrees pose error. By relying solely on a monocular camera, this method maintains millimeter-level and sub-degree-level registration accuracy while reducing equipment costs to 1/10 to 1/20 of traditional scanning solutions, significantly simplifying the system workflow and demonstrating strong engineering and clinical application potential. It demonstrates strong potential for both engineering and clinical applications.