Geometric deep learning on brain shape predicts sex and age

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• The relationship between human brain shape and function is complex and not fully understood.
• Cortical gyrification analysis offers insights into brain structure and pathologies.
• Previous surface-based methods relied on manual feature definition, limiting their utility.

Purpose of the Study:

• To develop a novel, data-driven deep learning method for analyzing brain shape.
• To eliminate the need for manual feature definition in cortical folding analysis.
• To establish a foundation for precision medicine applications using brain shape analysis.

Main Methods:

• A novel geometric deep learning approach using convolutional neural networks (CNNs) adapted for cortical surface data.
• Analysis of MRI data from 6410 healthy subjects (ages 6-89).
• Training two graph convolutional neural networks (gCNNs) for sex and age prediction, utilizing Class Activation Maps (CAM) and Regression Activation Maps (RAM).

Main Results:

• The gCNN accurately predicted subject's sex with 87.99% average accuracy.
• The gCNN predicted subject's age with a Pearson's correlation coefficient of 0.93 and an average absolute error of 4.58 years.
• Activation maps identified influential brain regions for prediction, highlighting the method's interpretability.

Conclusions:

• The proposed shape-based convolutional classifier is a novel, data-driven method for analyzing brain shape.
• This approach can define biomedically relevant features from brain surface data at population and individual levels.
• The method provides a critical foundation for future precision medicine applications in neurology and psychiatry.