Quantitative Assessment of the Impact of Geometric Distortions and Their Correction on fMRI Data Analyses

Abstract

Functional magnetic resonance imaging (fMRI) data is typically collected with gradient-echo echo-planar imaging (GE-EPI) sequences, which are particularly prone to the susceptibility artifact as a result of B₀ field inhomogeneity. The component derived from in-plane spin dephasing induces pixel intensity variations and, more critically, geometric distortions. Despite the physical mechanisms underlying the susceptibility artifact being well established, a systematic investigation on the impact of the associated geometric distortions, and the direct comparison of different approaches to tackle them, on fMRI data analyses is missing. Here, we compared two different distortion correction approaches, by acquiring additional: (1) EPI data with reversed phase encoding direction (TOPUP), and (2) standard (and undistorted) GE data at two different echo times (GRE). We first characterized the geometric distortions and the correction approaches based on the estimated ΔB₀ field offset and voxel shift maps, and then conducted three types of analyses on the distorted and corrected fMRI data: (1) registration into structural data, (2) identification of resting-state networks (RSNs), and (3) mapping of task-related brain regions of interest. GRE estimated the largest voxel shifts and more positively impacted the quality of the analyses, in terms of the (significantly lower) cost function of the registration, the (higher) spatial overlap between the RSNs and appropriate templates, and the (significantly higher) sensitivity of the task-related mapping based on the Z-score values of the associated activation maps, although also evident when considering TOPUP. fMRI data should thus be corrected for geometric distortions, with the choice of the approach having a modest, albeit positive, impact on the fMRI analyses.

Keywords: B0 field mapping; fMRI; geometric distortions and correction; neuroimaging; susceptibility artifact.

FIGURE 1

Schematic diagram of the processing pipeline. The fMRI data of each functional run (collected with the AP phase encoding direction) is first submitted to motion and slice timing correction. Then, geometric distortions are corrected using TOPUP (requires one AP and one PA functional image) or GRE (requires a magnitude and a displacement field converted to radian/s) approaches. Subsequently, additional pre-processing steps are performed on the corrected and uncorrected data, followed by the three different data analyses: (1) registration into structural data, (2) identification of resting-state networks, and (3) mapping of task-related brain regions of interest.

FIGURE 2

Illustration of the geometric distortions associated with the susceptibility artifact in the fMRI data from three representative participants during the first run of the biological motion perception task. For the first two participants (A,B), distortions highlighted by the yellow arrows can be observed, mostly representing the compression of voxels in areas of tissue/air boundaries. These were highly attenuated after correction with either TOPUP or GRE correction approaches. Participant (C) does not present clear distortions. The red traces define the contour of the GRE-corrected images, as this approach yielded the best results, and was then considered here as reference for visualization purposes.

FIGURE 3

(Top) Voxel shift maps (VSMs), in the MNI space, averaged across participants and runs, for the (left) TOPUP and (right) GRE correction approaches. The scale was set to be symmetric for easing the interpretation of the maps (where green colors always represent no shifts). (Bottom) Histograms of the associated VSMs.

FIGURE 4

(A) Voxel-wise difference maps between PA and AP images for the uncorrected case (left), and after correcting the data for the distortions using TOPUP (top-right) and GRE (bottom-right). (B) Voxel-wise difference maps of the corrected PA images (left) and corrected AP images (right), between correction methods.

FIGURE 5

Registration into the structural image of (top) uncorrected, (middle) TOPUP corrected, and (bottom) GRE corrected fMRI data of the first participant illustrated in Figure 2. The red and green traces define the contours of the brain-extracted structural image and its segmented ventricles (I and II), respectively; these can be considered as the ground truth. Besides the squashed voxels in areas of tissue/air boundaries (right yellow arrow), it is clear that the registration of uncorrected data fails to capture the spatial distribution of the ventricles I and II (left yellow arrow); a critical region in ventricle I is zoomed in the yellow, purple and blue boxes. The BBR cost function values are also shown.

FIGURE 6

Ten group RSNs identified on (left) uncorrected, (middle) TOPUP corrected, and (right) GRE-corrected fMRI data for the first run of the biological motion perception task. The RSN templates (in red–yellow) from Smith et al. (2009) are superimposed with the spatial independent components (in blue–light blue) selected for each RSN template, according to their Dice coefficient (also shown above each RSN).

FIGURE 7

Group activation maps for the localizer (top) and the biological motion (bottom) runs, from uncorrected, TOPUP corrected and GRE-corrected fMRI data.