Brain image completion by Bayesian tensor decomposition

Abstract

The MRI data is naturally represented by a three-order tensor. The reconstruction of MRI image from sparse observations is a challenging task, which has many potential applications for data compression, feature extraction and classifications. In this paper, we employ Bayesian Tucker decomposition to learn the low-rank representations of MRI data from partially observed voxels. By specifying the sparsity priors over factor matrices and core tensor, the multilinear ranks can be automatically determined via variational Bayesian inference, which thus avoids the difficulty in tuning parameters empirically. We apply the proposed method to MRI image with 50–80% missing voxels, the experimental results demonstrate that our method can effectively recover the whole MRI image with high predictive performance.