Vertex Correspondence and Self-Intersection Reduction in Cortical Surface Reconstruction.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Mesh-based cortical surface reconstruction is crucial for brain morphology analysis.
• Accurate vertex correspondence is vital for group-level comparisons but traditionally requires extensive post-processing.
• Deep learning methods have advanced reconstruction but often neglect vertex correspondence optimization.

Purpose of the Study:

• To introduce Vox2Cortex with Correspondence (V2CC) for improved inter- and intra-subject vertex correspondence in cortical surface reconstruction.
• To address and reduce mesh self-intersections, particularly major types, during the reconstruction process.
• To enhance the suitability of reconstructed meshes for direct group comparisons and atlas-based parcellation.

Main Methods:

• Modified Vox2Cortex by replacing Chamfer loss with L1 loss on registered surfaces.
• Introduced a novel Self-Proximity loss to mitigate major mesh self-intersections by adjusting non-neighboring vertices.
• Categorized mesh self-intersections into minor and major types for targeted analysis.

Main Results:

• V2CC significantly improves inter- and intra-subject vertex correspondence compared to existing deep learning methods.
• The proposed Self-Proximity loss effectively reduces major mesh self-intersections.
• Achieved accurate correspondence and reduced self-intersections to below 1% for both pial and white matter surfaces, improving parcellation accuracy.

Conclusions:

• V2CC offers a robust solution for accurate vertex correspondence in cortical surface reconstruction.
• The method enhances the utility of neuroimaging data for group studies and atlas-based analyses.
• This work addresses key limitations in current deep learning approaches for brain surface analysis.