Identification, Characterization, and Mitigation of a Novel Multiband fMRI Signal Artifact
Poster No:
1874
Submission Type:
Abstract Submission
Authors:
Philip Tubiolo1, John Williams2, Jared Van Snellenberg2
Institutions:
1Stony Brook University, Stony Brook, NY, 2Stony Brook University Renaissance School of Medicine, Stony Brook, NY
First Author:
Co-Author(s):
Introduction:
Simultaneous multi-slice (multiband; MB) fMRI is an acceleration technique that achieves improved spatiotemporal resolution, leading to its adoption by both individual laboratories and large consortia such as the Human Connectome Project (HCP), Adolescent Brain Cognitive Development (ABCD), and UK Biobank studies. We discovered that unprocessed MB-fMRI images display strongly elevated correlations between the average timeseries in simultaneously acquired (SA) slices, as compared to non-SA (nSA) slices. This indicates the existence of a consequential shared signal between simultaneous slices that cannot be explained by true neural activation patterns. Here, we employ two public datasets (HCP and ABCD) and two in-house datasets to develop an estimation and removal method for this signal, which we term Multiband Artifact Regression in Simultaneous Slices (MARSS).
Methods:
We explored resting-state (RS) and working memory (WM) task data from: 1) HCP Young Adult, acquired using a Siemens Connectome Skyra (N=25; MB factor 8); 2) Stony Brook University (SBU) using a Siemens Prisma Fit (N=10; MB factor 6); and 3) New York State Psychiatric Institute (NYSPI) using a GE MR750 (N=10; MB factor 6). We additionally utilized RS data from the ABCD study (N=35; all MB factor 6), which includes data from Siemens Prisma (N=25), GE MR750 (N=3), and Philips Achieva (N=7) scanners.
Data from SBU and NYSPI datasets (in-house) were preprocessed before and after RoMASS using the HCP minimal preprocessing pipeline, with task modeling conducted in SPM12 and Permutation of Linear Models (PALM).
MARSS estimates artifact signal separately for each slice as the mean signal across all SA slices (excluding the current slice), orthogonalized with respect to the image's global signal (excluding all SA slices) and motion parameter (MP) estimates (and their squares, derivatives, and squared derivatives). The estimated artifact signal for each slice is then regressed onto each voxel in that slice (in a regression that includes parameters to control for global signal and MPs), and removed from the voxel timeseries via subtraction, to create an artifact-corrected image.
Data from SBU and NYSPI datasets (in-house) were preprocessed before and after RoMASS using the HCP minimal preprocessing pipeline, with task modeling conducted in SPM12 and Permutation of Linear Models (PALM).
MARSS estimates artifact signal separately for each slice as the mean signal across all SA slices (excluding the current slice), orthogonalized with respect to the image's global signal (excluding all SA slices) and motion parameter (MP) estimates (and their squares, derivatives, and squared derivatives). The estimated artifact signal for each slice is then regressed onto each voxel in that slice (in a regression that includes parameters to control for global signal and MPs), and removed from the voxel timeseries via subtraction, to create an artifact-corrected image.
Results:
Significantly elevated Pearson correlations between mean slice signals were observed in SA slices as compared to nSA slices in all datasets, and MARSS significantly reduced these correlations in all datasets (all P < 0.05). Mean, whole-brain temporal SNR (tSNR) increased following MARSS in all datasets (all P<0.05), and the cortical coefficient of variation decreased in 97% and 83% of voxels in the in-house SBU and NYSPI datasets, respectively, resulting in net increases of 1050 and 3838 voxels being included in volume-to-surface mapping during preprocessing.
Figure 1 shows the spatial distribution of the artifact estimated by MARSS, calculated as the mean absolute value (over time) of the estimated artifact timeseries, following normalization to MNI space. Single-subject difference images with and without MARSS in un-preprocessed task betas are shown in Figure 2, along with the power spectral density of spatial frequency along the slice direction, demonstrating that MARSS removes spectral power at the fundamental and harmonic frequencies of the spacing between SA slices. In preprocessed group-level data, half of all greyordinates show a t-statistic change after MARSS by a magnitude of at least 0.169 and 0.145 in SBU and NYSPI datasets, respectively, while 10% of t-statistics change by at least 0.490 and 0.415.
Figure 1 shows the spatial distribution of the artifact estimated by MARSS, calculated as the mean absolute value (over time) of the estimated artifact timeseries, following normalization to MNI space. Single-subject difference images with and without MARSS in un-preprocessed task betas are shown in Figure 2, along with the power spectral density of spatial frequency along the slice direction, demonstrating that MARSS removes spectral power at the fundamental and harmonic frequencies of the spacing between SA slices. In preprocessed group-level data, half of all greyordinates show a t-statistic change after MARSS by a magnitude of at least 0.169 and 0.145 in SBU and NYSPI datasets, respectively, while 10% of t-statistics change by at least 0.490 and 0.415.
·Figure 1
·Figure 2
Conclusions:
These results show that application of MARSS prior to preprocessing: 1) reduces artifactually elevated correlations between the mean signal in each slice of MB fMRI datasets, 2) substantially improves tSNR and the quality of volume-to-surface mapping, and 3) causes dramatic changes in group-level t-statistics that are due to removal of alterations in task betas due to the MB artifact. We recommend that all MB fMRI datasets be cleaned using MARSS prior to standard preprocessing and analysis.
Modeling and Analysis Methods:
Methods Development 1
Motion Correction and Preprocessing 2
Keywords:
Data analysis
Design and Analysis
FUNCTIONAL MRI
Informatics
MRI
Statistical Methods
1|2Indicates the priority used for review