Exploring Gender Differences Based on Functional Magnetic Resonance Images by Deep Learning and Gradient-weighted Class Activation Mapping

Previous studies have shown gender differences in cognitive function and the prevalence of some neurological diseases. Understanding gender differences in functional connectivity (FC) will provide a new perspective for understanding these differences. We wanted to develop an effective model for the gender prediction and identify the major source of gender differences in functional connectivity. In this study we used FC generated from functional Magnetic Resonance Imaging (fMRI) from the Human Connectome Project in 1038 healthy subjects. We apply deep learning to test whether functional connectivity networks can effectively predict gender and identify the main sources of gender differences in FC. The accuracy of predicting gender by using the functional connectivity network has reached more than 90%, and we found that there are significant gender differences in the FC of interhemispheric brain areas and the FC in the right hemisphere, and in addition to the previously discovered default mode network (DMN), frontoparietal network (FPN), somatomotor network (SMN), there were gender differences in visual network (VIN), auditory network (AUN), and dorsal attention network (DAN) as well, and the gender differences within each resting-state network (RSN) were also large. Our results suggest that gender differences exist in functional connectivity networks and certain subnetworks can be focused on when studying gender-related cognitive functions and neurological disorders.