Imputing Longitudinal Infant Brain MRI Features Using a Self-Supervised Transformer Model

Motivation: Morphological changes of infant brain provide critical information of brain development and disease progress. However, inevitable data loss of follow-up visit hinders the longitudinal study of neural development. Goal(s): To develop a robust deep learning method which can impute missing MRI in longitudinal studies. Approach: A transformer-based model was proposed with self-supervised learning framework and a time-level imputation loss. Results: The proposed model demonstrated superior performance with reduced MAE by 55.7%, 55.8%, 28.1%, 12.9%, and 14.6% for four cortical features compared to the benchmarks of BRITS, USGAN, TIMESNET, SAITS, and original Transformer. It also enhanced downstream longitudinal prediction task by 19.1% in MSE. Impact: The proposed model is able to impute the missing data in longitudinal studies of infants, which may enrich the information along development trajectory and downstream analyses.