Improving Across Dataset Brain Age Predictions Using Transfer Learning

Abstract

Brain age has shown it’s potential as a biomarker of healthy or accelerated neurological ageing. Utilizing deep learning methods, brain age predictions for healthy individuals have become increasingly more accurate, but seem adversely affected when applied to new scanner data or differently preprocessed scans. We thus focused on transfer learning methods for convolutional neural network based brain age prediction models. The four models were trained and evaluated on a large multi-site dataset (N = 2543) and new site longitudinal dataset (N = 5632). Next, we assessed the ability of three transfer learning approaches, namely bias correction (BC), domain adaptation (DA) and full transfer (FT), to generalize the brain age prediction performance across the datasets. Our results indicate that models using transfer learning outperform models trained from scratch in similar studies. We further show that simpler and less expensive transfer learning approaches, such as BC or DA, perform better than FT and generalize well across datasets and preprocessing procedures, with mean absolute error and mean absolute (longitudinal) difference error as low as 3.3 and 1.1 years respectively, increasing their potential to practically deliver brain age biomarker to aid in diagnosis of neurodegenerative diseases and/or monitoring of their progression.