ISDR-Net: Interpretable Self-Supervised Differentiable Rendering Network for monocular dynamic sensor-head pose tracking and registration.

Abstract

Magnetoencephalography (MEG) offers high temporal and spatial resolution for clinical and neuroscience applications. Traditional sensor registration methods depend on complex point cloud reconstruction, which is error-prone, labor-intensive, and lacks adaptability. Recently, registration methods based on multi-view images have improved efficiency but still face challenges in dynamic responsiveness and speed. We propose a Interpretable Self-Supervised Differentiable Rendering Network (ISDR-Net) that enables monocular dynamic sensor-head pose tracking and registration. By integrating geometry-guided differentiable rendering with a transparent, unrolled optimization process, ISDR-Net ensures interpretability, computational efficiency, and adaptability. A coarse-to-fine optimization strategy rapidly initializes alignment, while adaptive refinement driven by keyframes and motion cues robustly tracks subtle head movements. Experimental results show that ISDR-Net aligns with the sub-millimeter registration accuracy of existing multi-view methods using only monocular single-frame images, with significantly reduced computational cost and stable performance under dynamic conditions. These results highlight the potential of ISDR-Net to support practical deployment of next-generation OPM-MEG systems in naturalistic environments.