Classifying sex with volume-matched brain MRI.

Area of Science:

• Neuroimaging
• Machine Learning
• Human Anatomy

Background:

• Previous studies on sex differences in brain structure yielded small effect sizes.
• Multivariate methods show promise in sex classification but often neglect brain volume control.
• Lack of volume control raises questions about the validity of previous sex classification accuracies.

Purpose of the Study:

• To determine the accuracy of sex classification from human brain gray matter properties while controlling for intracranial volume.
• To assess the robustness of machine learning classifiers in cross-cohort predictions.
• To investigate the impact of training set size and identify relevant brain regions for sex classification.

Main Methods:

• Utilized MRI data from two population-based cohorts: Study of Health in Pomerania (SHIP) and Human Connectome Project (HCP).
• Applied logistic regression and a 3D convolutional neural network (CNN) for sex classification.
• Matched individuals on intracranial volume and performed cross-cohort validation.

Main Results:

• Logistic regression achieved >92% accuracy in distinguishing sexes with matched intracranial volume in 1166 individuals.
• The logistic regression model maintained 85% accuracy on an unseen cohort without retraining.
• Classifier accuracy increased with training set size, surpassing 3000 individuals.
• No single brain region dominated classification; important features were distributed across the brain.

Conclusions:

• Accurate sex classification from brain gray matter is achievable when controlling for overall brain volume.
• Machine learning models demonstrate significant cross-cohort generalizability and benefit from large training datasets.
• Sex classification relies on distributed brain features rather than localized differences.