Reduction of confounding effects with voxel-wise Gaussian process regression in structural MRI

Abstract

We propose to use Gaussian process regression to remove confounds from gray matter (GM) density maps in order to improve performance in automated detection of neurodegenrative diseases. Age, total intracranial volume, sex, and acquisition site were included as design variables. Based on data from the control populations, a Gaussian process regression model was learned for each voxel. This model was used to compute maps of expected GM densities based on the subject's characteristics. For classification, the maps of expected GM densities were subtracted from the observed GM densities, thereby reducing confounding effects. The performance with and without subtraction of confounding effects were evaluated in four classification tasks: (1) patients with mild cognitive impairment (MCI) that did convert to Alzheimer's disease (AD) versus stable MCI patients, (2) patients with AD versus age-matched controls, (3) pre-manifest patients with Huntington's disease (HD) versus controls, and (4) manifest HD patients versus age-matched controls. The proposed method improved the classification performance in most cases, and never caused a significant decrease. The performance was similar to that obtained after reduction of confounding effects with kernel linear regression.