IcoConv : Explainable brain cortical surface analysis for ASD classification.

Area of Science:

• Neuroscience
• Computer Vision
• Medical Imaging

Background:

• Autism Spectrum Disorder (ASD) diagnosis relies on behavioral observation, with limited objective biomarkers.
• Analyzing complex 3D neuroimaging data presents computational challenges.

Purpose of the Study:

• To develop and validate a novel computational method for analyzing 3D brain shapes.
• To apply this method to differentiate individuals at high risk for ASD.

Main Methods:

• A novel approach analyzing 2D perspectives of 3D brain shapes using modified 2D Convolutional Neural Networks (CNNs).
• Utilized an icosahedron convolution operator for efficient pooling of multi-view data.
• Applied to brain attributes (cortical thickness, surface area, CSF volume) mapped onto a sphere.

Main Results:

• The method successfully performed binary classification to distinguish high-risk positive and negative ASD cases.
• Generated gradient-based explainability maps visualized on the brain surface.
• Findings align with known ASD-affected brain regions.

Conclusions:

• The novel 3D shape analysis method is effective for neuroscientific research, specifically in ASD.
• Explainability maps provide insights consistent with existing ASD knowledge.
• The approach holds potential for discovering new aspects of ASD.