Unifying Brain Age Prediction and Age-Conditioned Template Generation with a Deterministic Autoencoder

Age-related morphological brain changes are known to be different in healthy and disease-affected aging. Biological brain age estimation from MRI scans is a common way to quantify this effect whereas differences between biological and chronological age indicate degenerative processes. The ability to visualize and analyze the morphological age-related changes in the image space directly is essential to improve the understanding of brain aging. In this work, we propose a novel deep learning based approach to unify biological brain age estimation and age-conditioned template creation in a single, consistent model. We achieve this by developing a deterministic autoencoder that successfully disentangles age-related morphological changes and subject-specific variations. This allows its use as a brain age regressor as well as a generative brain aging model. The proposed approach demonstrates accurate biological brain age prediction, and realistic generation of age-conditioned brain templates and simulated age-specific brain images when applied to a database of more than 2000 subjects.