Bayesian Estimation for Fast Sequential Diffeomorphic Image Variability

Abstract

In this paper, we analyze the diffeomorphic image variability using a Bayesian method to estimate the low-dimensional feature space in a series of images. We first develop a fast sequential diffeomorphic image registration for atlas building (FSDAB) to reduce the computation time. To analyze image variability, we propose a fast Bayesian version of the principal geodesic analysis (PGA) model that avoids the trivial expectation maximization (EM) framework. The sparsity BPGA model can automatically select the relevant dimensions by driving unnecessary principal geodesics to zero. To show the applicability of our model, we use 2D synthetic data and the 3D MRIs. Our results indicate that the automatically selected dimensions from our model can reconstruct unobserved testing images with lower error, and our model can show the shape deformations that corresponds to an increase of time.