A Surface-based deep learning approach for cortical shape analysis.

Area of Science:

• Neuroimaging
• Deep Learning
• Computational Neuroscience

Background:

• Deep learning models can predict clinical factors from human brain images.
• Brain shape metrics derived from MRI offer valuable information for clinical predictions.

Purpose of the Study:

• To apply a spherical harmonics-based convolutional neural network (SPHARM-Net) for predicting age, sex, and Alzheimer's disease (AD) diagnosis.
• To evaluate SPHARM-Net's performance on MRI-derived brain shape metrics.

Main Methods:

• Utilized SPHARM-Net, a convolutional neural network employing spherical harmonic transforms for rotational equivariance.
• Extracted brain features including vertex-wise cortical curvature, convexity, thickness, and surface area from MRI scans.
• Tested on large datasets: UK Biobank (N=32,979) for sex and age, and ADNI (N=1,213) for AD classification.

Main Results:

• Achieved high accuracy in sex classification (accuracy=0.91, balanced accuracy=0.91, AUC=0.97).
• Demonstrated strong performance in age regression (average absolute error=2.97 years, R-squared=0.77).
• Showed promising results for AD classification (accuracy=0.86, balanced accuracy=0.83, AUC=0.9).

Conclusions:

• SPHARM-Net shows promising preliminary performance for predicting age, sex, and AD from brain MRI shape metrics.
• The method is effective for established benchmarking tasks in neuroimaging.
• Future research will explore comparisons with other shape-based methods and applications to mood disorder classification.