Deep Transfer Learning for Cerebral Cortex Using Area-Preserving Geometry Mapping

Area of Science:

• Neuroimaging
• Computational Geometry
• Machine Learning

Background:

• Deep learning application in brain image analysis is limited by small datasets.
• Existing 2D pretrained models cannot be directly applied to 3D brain images.

Purpose of the Study:

• To develop a novel framework for applying 2D pretrained deep learning models to 3D brain image analysis.
• To enable the use of powerful 2D models for analyzing complex 3D neuroimaging data.

Main Methods:

• 3D cortical meshes from MRI were reconstructed and projected into 2D planar meshes using area-preserving geometry mapping.
• 2D deep learning models pretrained on ImageNet were fine-tuned for cortical image classification.
• The framework was validated on sex classification (Human Connectome Project) and autism spectrum disorder (ASD) classification (Autism Brain Imaging Data Exchange).

Main Results:

• The proposed framework significantly improved classification performance for both sex and ASD detection using transfer learning.
• A 2-stage transfer learning strategy further boosted ASD classification accuracy by using sex classification as an intermediate task.
• The method successfully bridges the gap between 3D cortical data and 2D deep learning models.

Conclusions:

• The developed framework effectively adapts 2D pretrained models for 3D brain image analysis, overcoming sample size limitations.
• This approach offers a valuable tool for cognitive and psychiatric neuroscience research, enhancing diagnostic capabilities.
• The study demonstrates the potential of computational geometry mapping and transfer learning in neuroimaging analysis.