Clustering of high dimensional longitudinal imaging data

In the study of brain disease processes and aging, longitudinal imaging studies are becoming increasingly commonplace. Indeed, there are hundreds of studies collecting multi-sequence multi-modality brain images at multiple time points on hundreds of subjects over many years. A fundamental problem in this context is how to classify subjects according to their baseline and longitudinal changes in the presence of strong spatio-temporal biological and technological measurement error. We propose a fast and scalable clustering approach by defining a metric between latent trajectories of brain images. Methods were motivated by and applied to a longitudinal voxel-based morphometry study of multiple sclerosis. Results indicate that there are two distinct patterns of ventricular change that are associated with clinical outcomes.