Statistical Inference of Auto-correlated Eigenvalues with Applications to Diffusion Tensor Imaging

TL;DR

This paper develops a hierarchical statistical model for the inference of eigenvalues in diffusion tensor imaging, accounting for spatial auto-correlation and eigenvalue randomness, demonstrated through simulations and real data.

Contribution

It introduces a novel hierarchical model and estimation method for auto-correlated eigenvalues in DTI, addressing a gap in current statistical approaches.

Findings

The proposed model outperforms existing methods in simulation studies.

Application to IXI dataset validates the model's effectiveness.

Model captures spatial auto-correlation and eigenvalue variability accurately.

Abstract

Diffusion tensor imaging (DTI) is a prevalent neuroimaging tool in analyzing the anatomical structure. The distinguishing feature of DTI is that the voxel-wise variable is a 3x3 positive definite matrix other than a scalar, describing the diffusion process at the voxel. Recently, several statistical methods have been proposed to analyze the DTI data. This paper focuses on the statistical inference of eigenvalues of DTI because it provides more transparent clinical interpretations. However, the statistical inference of eigenvalues is statistically challenging because few treat these responses as random eigenvalues. In our paper, we rely on the distribution of the Wishart matrix's eigenvalues to model the random eigenvalues. A hierarchical model which captures the eigenvalues' randomness and spatial auto-correlation is proposed to infer the local covariate effects. The Monte-Carlo Expectation-Maximization algorithm is implemented for parameter estimation. Both simulation studies and application to IXI data-set are used to demonstrate our proposal. The results show that our proposal is more proper in analyzing auto-correlated random eigenvalues compared to alternatives.