Learnable Pooling in Graph Convolutional Networks for Brain Surface Analysis.

Area of Science:

• Computational neuroscience
• Machine learning
• Medical image analysis

Background:

• Brain surface analysis is crucial for neuroscience but computationally challenging due to complex, non-Euclidean cortical geometry.
• Existing machine learning methods struggle with pooling strategies for surface data, often limited to fixed graphs.

Purpose of the Study:

• To propose a novel learnable graph pooling method for processing multiple brain surface-valued data.
• To enhance the analysis of non-Euclidean brain surface data using machine learning.

Main Methods:

• Developed a learnable graph pooling method utilizing graph spectral embedding for intrinsic node aggregation.
• Applied the method to process multiple surface-valued data for subject-based information extraction.

Main Results:

• Demonstrated superior performance compared to existing pooling techniques for graph convolutional networks on benchmark datasets.
• Achieved state-of-the-art results in brain surface analysis across various prediction tasks.

Conclusions:

• The proposed learnable pooling strategy offers a flexible and effective approach for brain surface analysis.
• This method advances machine learning applications in computational neuroscience, improving diagnostic and predictive capabilities.