Geometric Brain Surface Network For Brain Cortical Parcellation.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Surface-based analyses of brain imaging data rely on accurate brain atlases for assessing structural and functional changes.
• Traditional methods for cortical surface parcellation are often time-consuming and may require manual intervention.
• Recent deep learning methods face computational challenges and may need post-processing for refinement.

Purpose of the Study:

• To develop an accurate, fully-automatic cortical surface parcellation method that operates directly on original brain surfaces.
• To introduce the Deep Brain Parcellation Network (DBPN), an end-to-end deep learning model for this task.

Main Methods:

• DBPN utilizes intrinsic and extrinsic graph convolution kernels to analyze vertex neighborhood topology.
• The network dynamically encodes topological information into node features.
• A two-stage deep network architecture, comprising a coarse parcellation network and a refinement network, is employed.

Main Results:

• DBPN dynamically deciphers graph topology and constructs a non-linear mapping to parcellation labels.
• The model was evaluated on a large public dataset.
• DBPN demonstrated superior performance compared to state-of-the-art baseline methods in terms of accuracy and efficiency.

Conclusions:

• DBPN offers an accurate and efficient solution for fully-automatic cortical surface parcellation.
• The proposed method overcomes limitations of traditional and existing deep learning approaches.
• DBPN holds significant potential for advancing surface-based brain imaging analyses.