Connectivity-based Cortical Parcellation via Contrastive Learning on Spatial-Graph Convolution.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Cortical parcellation, crucial for neuroscience, has been advanced by diffusion imaging and tractography.
• Previous methods face limitations due to simplistic computational schemes and inadequate brain imaging data representation.

Purpose of the Study:

• To develop and evaluate a novel cortical parcellation framework using tractography-derived structural brain connectivity.
• To address limitations of existing methods by employing advanced spatial-graph representation learning.

Main Methods:

• The Spatial-graph Convolution Parcellation (SGCP) framework utilizes a two-stage deep learning approach.
• Stage one involves self-supervised contrastive learning with a spatial-graph convolution network encoder for data embedding.
• Stage two employs a supervised classifier for voxel-wise classification to achieve parcellation.

Main Results:

• SGCP was evaluated on parcellating 5 brain regions using a 15-subject DWI dataset.
• The framework demonstrated superior performance compared to traditional and other deep learning-based parcellation methods.
• Consistent high performance was observed across all tested parcellation tasks.

Conclusions:

• The proposed SGCP framework offers a robust and effective solution for cortical parcellation.
• Its strong performance suggests potential as a generalizable tool for studying human brain composition using connectivity data.
• SGCP advances the application of representation learning in neuroscientific research and medical image analysis.