MRExtrap: Linear Prediction of Brain Aging in Autoencoder Latent Space of MRI Scans

Jaivardhan Kapoor, Jakob H. Macke, Christian F. Baumgartner

Keywords: Brain MRI, Generative Modeling, Longitudinal Modeling, Autoencoders, Brain Aging

Abstract: Longitudinal generative modeling of high-resolution 3D Magnetic-Resonance-Imaging (MRI) scans can reveal disease progression patterns in neurological disorders such as Alzheimer's disease. We introduce a novel approach called MRExtrap for simulating aging in brain MRI volumes given previously observed MRIs, by performing linear regression in the latent space of an autoencoder. We show that well-trained convolutional autoencoders can yield latent representations that exhibit linearity with respect to the regional brain volumes when interpolated, decoded, and segmented. We exploit this structure by training a linear progression model in the latent space of the autoencoder to predict trajectories of latent representations based on the age of the subject. On the ADNI dataset, we show that predicted MRIs align closely with held-out longitudinal scans, enabling accurate modeling of age-related structural brain changes.

Submission Number: 100