Resting-state functional connectivity MRI (rs-fcMRI) commonly measures brain network activity in the absence of a behavioural task using blood oxygenation level dependent (BOLD) contrast. Typically computed from the temporal dependency of BOLD signal patterns across different brain regions, rs-fcMRI has been used extensively to investigate various neuroscience and clinical questions ¹. It has been shown that rs-fcfMRI can benefit from faster imaging times offered by simultaneous multi-slice (a.k.a. multiband, or “MB”) slice-acceleration ^2,3 that enables “groups” of slices to be acquired at the same time. However, this slice-grouping may incur aliasing artifacts, primarily from motion ⁴ and “physiological noise”. The latter spurious fluctuations arise from oscillatory respiration and cardiac signals, as well as respiration volume per time (RVT) ⁵ and cardiac rate variation (CRV) ⁶. It is possible that aliasing of physiological noise may introduce unwanted signal components across slices acquired within a given group, with a deleterious effect on functional conductivity estimates. Consequently, two hypotheses are investigated in this work: 1) physiological noise and associated aliasing effects introduce a slice-group effect in rs-fcMRI maps; and 2) this slice-group effect can be mitigated by physiological noise correction by retrospective gating.

All images were acquired using a Siemens TIM Trio 3 T MRI System (Siemens, Erlangen, Germany) employing the 32-channel phased-array head coil receiver. Twelve subjects were each scanned with a functional single-shot MB gradient-echo (GE) echo-planar imaging (EPI) ¹ sequence (TR=323ms, TE=30 ms, flip angle=40°, 15 slices, 3.44 mm × 3.44 mm × 4.6 mm, 2230 volumes, acceleration factor=3, phase encoding shift factor=2), and a T1-weighted anatomical sequence (MPRAGE, TR = 2400 ms, TE = 2.43 ms, FOV = 256 mm, TI = 1000 ms, readout bandwidth = 180 Hz/px, voxel size = 1 mm × 1 mm × 1 mm). In the MB-EPI scan, the 15 slices were grouped as 1-6-11, 2-7-12, 3-8-13, 4-9-14, 5-10-15, and acquired in this order. For each subject, a pneumatic belt (BioPac, Goleta, USA) and the vendor supplied pulse oximeter were used to record the respiratory and cardiac signals, respectively. Subjects were instructed to close their eyes and relax. The rs-fcfMRI processing was carried out using the FMRIB Software Library v5 (FSL5) (Analysis Group, FMRIB, Oxford, UK) ⁷.

The preprocessing pipeline included motion correction, brain extraction, spatial smoothing, and regression of six motion parameters. Five different physiological noise correction schemes were performed on the data of each subject using MATLAB (MathWorks, Natick, MA): a) no correction; b) image-based retrospective correction of cardiac and respiratory waveforms (RETROICOR) ⁸; c) RETROICOR+RVTcor; d) RETROICOR+CRVcor; and e) RETROICOR+RVTcor+CRVcor, where RVTcor and CRVcor refer to the inclusion of these physiological noise signals in the retrospective correction algorithm. Thus, five post-processed fMRI datasets were considered using a multi-regression model, focusing on rs-fcMRI maps of the default mode network (DMN), as it has extensive brain coverage. A 4 mm-radius spherical seed was generated over the posterior cingulate cortex (PCC) using the coordinates reported by reported by Van Dijk et al ⁹. For each of the five corrected datasets, the spatial average of the signal from the seed was correlated with all other voxels to generate the connectivity maps. All statistical connectivity maps were then corrected using mixture modelling ¹⁰ and the resulting Z-maps were then thresholded at Z>2.3 (p<0.01). Comparing the rs-fcMRI maps to an appropriate connectivity template for the DMN ¹¹, the number of false positive and false negative activations were separately obtained for each slice, each physiological noise correction scheme, and each subject. For each subject and correction scheme, the mean number of false positives and false negatives per slice was then determined. Two repeated-measures analyses of variance (rANOVAs) were subsequently performed using MATLAB to compare the effect of the correction schemes on the mean number of false positives and false negatives. Using the known MB-EPI slice-grouping, the slice-group effect on the number of false positives and false negatives of each slice was independently studied via linear mixed effect (LME) analysis in R (R Foundation for Statistical Computing, Vienna, Austria).

Fig. 1 illustrates the rs-fcMRI maps for a representative subject, corresponding to the five physiological noise correction schemes (a-e), and the DMN template (f) ¹¹. Using the template as reference, the number of false positives gradually reduces in progression from correction scheme a) to c), with little difference observed between schemes (c-e). Similar results were observed across the cohort, except that the number of false positives continues to decrease across all five correction schemes. The rANOVA results showed a strong trend of physiological correction reducing the number of false positives, that approached statistical significance (p=0.09). No effects were found for false negatives (p=0.81).

Table 1 shows the p-values of the slice-group LME modeling for different physiological noise correction schemes. According to the LME model without physiological noise correction, the number of false positives demonstrates a significant slice-group effect, whereas the number of false negatives does not. Additionally, the slice-group effect (both false positives and false negatives) incrementally diminishes (i.e. p-value increases) progressing from correction scheme a) to scheme e). Specifically, for false positives, RETROICOR was effective at suppressing the slice-group effect so that it was no longer statistically significant.

This work shows that in the absence of physiological noise correction, a significant slice-group effect is present in the number of false positives in MB-EPI rs-fcMRI maps. Due to the type of template comparison that was employed, the number of false negatives was large and did not show a significant slice-group effect, although a trend was observed. The work also suggests that slice-group biases can be significantly reduced by correcting for physiological effects in MB-EPI data. In this regard, appropriate physiological correction is potentially even more important for MB-EPI than for regular EPI acquisitions.