Graph Convolutions on Spectral Embeddings for Cortical Surface Parcellation.

Area of Science:

• Neuroscience
• Computer Science
• Medical Imaging

Background:

• Neuronal cell bodies are primarily in the cerebral cortex, a complex surface crucial for brain function.
• Analyzing cortical surface data is challenging due to high geometric variability.
• Current methods like spherical inflations (e.g., FreeSurfer) are computationally expensive and time-consuming.

Purpose of the Study:

• To develop a novel approach for learning and exploiting surface data directly across multiple surface domains.
• To enable fast and accurate processing of brain surfaces by overcoming limitations in comparing data across different geometries.
• To improve brain parcellation accuracy and speed.

Main Methods:

• Leveraging spectral graph matching to transfer surface data across aligned spectral domains.
• Developing direct learning of surface data using graph convolutions.
• Exploiting spectral filters over intrinsic representations of surface neighborhoods.
• Applying the approach to brain parcellation.

Main Results:

• Demonstrated significant improvement in labeling accuracy compared to Euclidean approaches.
• Achieved drastic speed improvements over conventional brain surface processing methods.
• Validated the algorithm over 101 manually labeled brain surfaces.

Conclusions:

• The novel spectral graph matching approach enables direct learning and comparison of surface data across different brain geometries.
• This method offers a faster and more accurate alternative for brain surface analysis and parcellation.
• The findings pave the way for new families of efficient algorithms for processing complex brain surface data.