Unbiased Diffeomorphic Mapping of Longitudinal Data with Simultaneous Subject Specific Template Estimation

Longitudinal mapping techniques for neuroimage analysis in computational anatomy have an important potential for bias associated to the order of input scans. Geodesic trajectories which pass from a template onto a baseline image and then through each follow up image have been shown to overestimate atrophy rate in the entorhinal cortex, while the reverse is true for trajectories pass through the data in the opposite order. We propose a method to remove this source of bias by inserting a patient specific template into a time series at a specific point to be estimated from the data, and simultaneously producing a time varying mapping connecting each image in the series. We demonstrate the efficacy of this method using segmentations of the entorhinal and surrounding cortex in subjects with early Alzheimer’s disease.