Unified Brain Surface and Volume Registration.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Accurate registration of 3D brain MRI scans is crucial for cross-subject neuroscientific analysis.
• Current methods often treat volumetric and surface-based registration separately, causing inconsistencies.
• These inconsistencies limit the reliability of downstream neuroimaging analyses.

Purpose of the Study:

• To introduce NeurAlign, a deep learning framework for joint volumetric and surface-based registration of 3D brain MRI scans.
• To achieve consistent and anatomically accurate alignment of both cortical and subcortical brain regions.
• To overcome limitations of traditional, separate registration approaches.

Main Methods:

• Developed a deep learning framework, NeurAlign, utilizing a unified volume-and-surface-based representation.
• Employed an intermediate spherical coordinate space to integrate surface topology with volumetric anatomy.
• Ensured geometric coherence between volume and surface domains through integrated spherical registration.

Main Results:

• NeurAlign consistently outperformed classical and machine learning-based registration methods on in-domain and out-of-domain datasets.
• Achieved improvements in Dice score by up to 7 points while maintaining regular deformation fields.
• Demonstrated significantly faster inference times (orders of magnitude) and simpler usage compared to standard methods.

Conclusions:

• NeurAlign provides superior accuracy and consistency for joint cortical and subcortical registration.
• The framework offers a faster, simpler, and more accurate alternative for brain MRI registration.
• NeurAlign establishes a new benchmark for aligning 3D brain MRI data in neuroscientific research.