Representative scale-invariant characteristics of male and female brains in magnetic resonance images

Abstract

This paper investigates the link between sex and the human brain from anatomical MRI data, where a primary confound is the size difference between male and female groups. Anatomy is characterized by the 3D scale-invariant feature transform (SIFT), where features are salient image regions that are automatically identified and normalized according local size or scale. Experiments use T1-w MRI data of 422 healthy unrelated subjects from the Human Connectome Project (HCP) dataset (211 males, 211 females, 22–36 years of age). We found that brain sex may be predicted via image-to-image feature matching with 91.9% accuracy, that classification is driven largely by weakly-informative features distributed throughout the brain and shared by unique subsets of subjects, and that a pair of features identified in the right and left thalamic regions of 97% of subjects predicts sex with 74% accuracy. Misclassified subjects exhibit features typical of the opposite sex in one or both hemispheres.