Disentangling Factors of Morphological Variation in an Invertible Brain Aging Model.

Neuroimaging-based machine learning models are widely employed to analyze healthy brain aging and its pathological deviations. This includes regression models that estimate a brain's biological age using structural MR images, generative models that capture the conditional distribution of aging-related brain morphology changes, and hybrid generative-inferential models that handle both tasks. Generative models are useful when systematically analyzing the influence that different semantic factors have on brain morphology. Within this context, this paper builds upon a recently proposed normalizing flow-based, generative-inferential brain aging model (iBAM) that uses supervision to disentangle age and age-unrelated identity information of a subject's brain morphology in its structured latent space. We analyze the effects adding sex as an additional supervised factor to iBAM has on the latent space when using real data. Moreover, we propose to learn an identity part that is ordered with respect to the amount of morphological variability covered. Our results on T1 images from more than 5000 healthy subjects show that iBAM is able to successfully disentangle age and sex from the identity information using supervision. Furthermore, the identity part is ordered, which aids efficient exploration and summarization of inter-subject variations.