A Multivariate Approach for Denoising of T2 Relaxation Decay Curves in Myelin Water Fraction Imaging

Accurately measuring brain myelin content in vivo holds great promise for studying and understanding of number of brain diseases. A surrogate marker for myelin, myelin water fraction (MWF), can be measured with MR relaxation techniques that sample the T2 decay with multiple echoes. Low signalto-noise ratios in later echoes, sub-optimal flip angles, and the fact that robust decomposition into multiple exponential curves is notoriously difficult, all conspire to reduce the accuracy of MWF estimates. The resulting maps are typically spatially noisy -- despite the fact that adjacent white matter voxels are usually assumed to have similar myelin measures in vivo. Here we propose a spatio-temporal filtering process prior to the standard fitting based on a combination of multivariate empirical mode decomposition (MEMD) and multiset canonical correlation analysis (MCCA) to decompose and find the most robust temporal decay pattern among voxels that have similar overall decay curves and across the white matter. Based on enhanced spatial smoothness measures, increased test-retest reliability within subjects, and decreased Coefficient of Variation of MWF scores, we suggest that the proposed approach provides enhanced accuracy of the ultimately-computed MWF maps.