Predicting Individual Traits From T1-weighted Anatomical MRI Using the Xception CNN Architecture

Abstract

Modeling individual traits is a long-standing goal of neuroscientific research, as it allows us to gain a more profound understanding of the relationship between brain structure and individual variability. In this article, we used the Keras-Tuner library to evaluate the performance of a tuned Xception convolutional neural network (CNN) architecture in predicting sex and age from a sample of 4,049 T1-weighted anatomical MRI scans originating from 1,594 participants. In addition, we used the same tuning procedure to predict the big five inventory (BFI) personality traits for 415 participants (represented by 1,253 scans), and compared the results with those generated by applying transfer learning (TL) based on the models for sex and age. To minimize the effects of preprocessing procedures, scans were subjected exclusively to brain extraction and linear registration with the 2 mm MNI152 template. Our results suggest that CNNs trained with hyperparameter optimization could be used as an effective and accessible tool for predicting subject traits from anatomical MRI scans, and that TL shows potential for application across target domains. While BFI scores were not found not be predictable from T1-weighted scans, further research is required to assess other preprocessing and prediction workflows.