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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Magnetoencephalography (MEG) requires precise sensor registration for accurate clinical and neuroscience applications.
• Traditional sensor registration methods are complex, error-prone, and lack adaptability.
• Existing multi-view image-based methods improve efficiency but struggle with dynamic tracking.

Purpose of the Study:

• To develop a novel, efficient, and adaptable method for dynamic sensor-head pose tracking and registration in MEG.
• To enable sub-millimeter registration accuracy using monocular imaging for next-generation OPM-MEG systems.

Main Methods:

• Introduced the Interpretable Self-Supervised Differentiable Rendering Network (ISDR-Net).
• Integrated geometry-guided differentiable rendering with an unrolled optimization process for interpretability and efficiency.
• Employed a coarse-to-fine optimization strategy with adaptive keyframe-based refinement for robust dynamic tracking.

Main Results:

• ISDR-Net achieved sub-millimeter registration accuracy comparable to multi-view methods using only monocular images.
• Demonstrated significantly reduced computational cost and stable performance under dynamic conditions.
• Validated the network's adaptability and interpretability through its transparent optimization process.

Conclusions:

• ISDR-Net offers a computationally efficient and accurate solution for MEG sensor registration.
• The proposed method overcomes limitations of traditional and multi-view approaches in dynamic scenarios.
• Highlights the potential for practical deployment of ISDR-Net in next-generation OPM-MEG systems for naturalistic environments.