Removal of Residual Motion Artifacts in fMRI using Constrained Independent Component Analysis

Introduction: Image registration of fMRI data only corrects for bulk movements while leaving secondary artifacts such as, spin history effects, motion induced dynamic field inhomogeneity changes, and interpolation errors untouched. Secondary artifacts increase variability in time-series and reduce sensitivity of activation detection. Methods such as the use estimated motion parameters as “Nuisance Explanatory Variables” (NEV) in a General Linear Model (GLM), [GLM+NEV] [1], or use of Independent Component Analysis (ICA) [2,3] to isolate and remove motion related components are commonly used to remove residual motion artifacts after image registration. Use of ICA, while very effective, is a two stage process involving, first, isolation and subsequent identification of motion components before removing them. The identification step can be tedious and time consuming as one has to manually inspect or use a similarity measure to identify/isolate components that are due to motion. However, ICA does have an advantage over use of motion parameters as NEVs in GLM. ICA is able to better isolate sources or components that are related to motion in a possibly non-linear and unpredictable fashion namely, interactions of susceptibility induced field inhomogeneity with motion, interpolation artifacts and other structured noises [2,3,4]. Thus, it is potentially a more complete method of correcting for residual motion artifacts compared to use of GLM+NEV. In this study, we used temporally constrained ICA (cICA) [5,6] to combine isolation and identification of residual motion artifact components into a single step using estimated motion parameters as references for the cICA algorithm. We also explored whether use of cICA with motion parameters as references improved quality and completeness of removal of residual motion artifacts in BOLD fMRI. Methods: Separate scans, each 200 seconds long, were acquired while four volunteers performed a visual motor task inside the scanner. Each scan consisted of five 40-second cycles of alternating task and control periods. Each volunteer was instructed to intentionally move their head during the scans. These motion corrupted images in the time series were then realigned using MCFLIRT utility of FSL [7]. Statistical parametric analysis was carried out on the realigned images in four different ways. i) GLM: Activation parameters were estimated using a GLM as implemented in FEAT utility of FSL. ii) cICA: Six estimated motion parameters were used as references for cICA algorithm to isolate components most correlated to motion. These components were then removed and the rest combined to obtain a time series of images with reduced residual motion artifacts. iii) GLM+NEV: Similar to “GLM” method however estimated motion parameters were included as NEVs in the GLM design matrix. iv) cICA+NEV: Same as in cICA but, motion parameters were included as NEV in the GLM matrix for estimation of statistical parameters. A z-threshold of > 3.5 was used to identify active pixels. Active pixels in the brain were aggregated into a histogram with ten bins over a range of zscores, 3.5 up to max z-score value. Number of active pixels at each bin was added cumulatively to generate a graph that showed the incremental number of active pixels at each bin (z-score value) and the total number of active pixels detected thus far – the final data point represents the sum of all active pixels. These cumulative graphs were compared across the four analysis methods (see figure 1). Results: FIGURE – 1 FIGURE 2