Weakly Supervised Cerebellar Cortical Surface Parcellation with Self-Visual Representation Learning.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Brain Anatomy

Background:

• The cerebellum, crucial for motor control, has a complex, convoluted structure often oversimplified by traditional volumetric analysis.
• Deep sulci and high folding in the cerebellar cortex pose challenges for accurate structural and functional analysis.

Purpose of the Study:

• To develop and validate a novel cortical surface-based analysis for the cerebellum.
• To accurately characterize the highly folded cerebellar cortex and enable detailed regional analysis.
• To overcome limitations of conventional methods in capturing localized cerebellar changes.

Main Methods:

• Reconstruction of geometrically accurate and topologically correct cerebellar cortical surfaces.
• A weakly supervised graph convolutional neural network for automatic parcellation of cerebellar surface regions.
• A two-step learning approach: contrastive self-learning for surface patch representation and mapping to parcellation labels.
• Directly processing original cerebellar cortical surfaces without registration or spherical mapping.

Main Results:

• The proposed method successfully reconstructs and parcellates the cerebellar cortical surface.
• Experimental validation using the Baby Connectome Project data demonstrated superior accuracy and effectiveness compared to existing methods.
• The learning-based model accurately handles the complex geometry of the cerebellar cortex.

Conclusions:

• Cortical surface-based analysis offers a more accurate approach to studying the cerebellum's complex structure.
• The novel graph convolutional neural network method provides effective automatic parcellation of cerebellar regions.
• This technique enhances the ability to detect localized functional and structural changes in the cerebellum.