Despite its great potential in studying brain anatomy and structure, diffusion magnetic resonance imaging (dMRI) is marred by artefacts more than any other commonly used MRI technique. In this paper we present a non-parametric framework for detecting and correcting dMRI outliers (signal loss) caused by subject motion.

Signal loss (dropout) affecting a whole slice, or a large connected region of a slice, is frequently observed in diffusion weighted images, leading to a set of unusable measurements. This is caused by bulk (subject or physiological) motion during the diffusion encoding part of the imaging sequence. We suggest a method to detect slices affected by signal loss and replace them by a non-parametric prediction, in order to minimise their impact on subsequent analysis. The outlier detection and replacement, as well as correction of other dMRI distortions (susceptibility-induced distortions, eddy currents (EC) and subject motion) are performed within a single framework, allowing the use of an integrated approach for distortion correction. Highly realistic simulations have been used to evaluate the method with respect to its ability to detect outliers (types 1 and 2 errors), the impact of outliers on retrospective correction of movement and distortion and the impact on estimation of commonly used diffusion tensor metrics, such as fractional anisotropy (FA) and mean diffusivity (MD). Data from a large imaging project studying older adults (the Whitehall Imaging sub-study) was used to demonstrate the utility of the method when applied to datasets with severe subject movement.

The results indicate high sensitivity and specificity for detecting outliers and that their deleterious effects on FA and MD can be almost completely corrected.