Estimation of the Biological Age of the Human Brain Using Multitask Self-Supervised Learning

Assessing brain health and its changes is crucial with the aging population. Predicting the biological age of the brain is a promising biomarker for evaluating overall health, made possible with machine learning. As efficient training strategies are necessary to accurately predict human brain age, we propose a self-supervised multitask model for predicting brain age from the ADNI dataset's 2D magnetic resonance imaging (MRI). This model outperforms single-task models and improves brain age prediction by providing more general features through the self-supervised task of rotation class prediction as second task in multi-tasking. We have also examined how self-supervised pretraining implemented by the DINO framework (knowledge distillation with no labels) affects our network. We discovered that these pretraining weights had a considerable impact and indicated that our proposed model, performed equally well as state-of-the-art models, with a Mean Absolute Error (MAE) of 3.27 years.