Geometric Convolutional Neural Network for Analyzing Surface-Based Neuroimaging Data.

Area of Science:

• Neuroscience
• Machine Learning
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) excel at 2D image analysis but struggle with complex 3D brain surface data.
• Representing the cerebral cortex as a semi-regular geometric mesh poses challenges for conventional deep learning methods.

Purpose of the Study:

• To develop a geometric CNN (gCNN) for effective pattern recognition on mesh-based brain surface data.
• To adapt CNNs for surface-based neuroimaging analysis, enhancing machine learning applications in brain research.

Main Methods:

• Proposed a gCNN integrating data sampling and reshaping for mesh surface compatibility with standard CNN toolboxes.
• Implemented gCNN for sex classification using cortical thickness maps from the Human Connectome Project (HCP).
• Compared gCNN performance against Support Vector Machine (SVM) and 2D CNN.

Main Results:

• gCNN achieved significantly higher classification accuracy than SVM and 2D CNN for cortical thickness maps.
• Demonstrated position invariance of local features, allowing pre-trained model reuse without significant performance degradation.
• The brain-like architecture and surface-based data handling contribute to gCNN's superior performance.

Conclusions:

• gCNN offers a powerful new tool for surface-based machine learning in neuroscience.
• Facilitates advanced scientific investigations and clinical applications by enabling deep learning on brain surface data.
• The gCNN approach bridges the gap between conventional CNNs and complex neuroimaging data analysis.