Synopsis

A novel hybrid seed-based approach to assess the test-retest reliability of ICNs is introduced that uses an exhaustive but systematic, ROI-limited search for seed pairs with maximal correlation. We examine the reproducibility of these seed pairs and their relation to four well-known ICNs. We then use this approach to assess test-retest reliability between subjects scanned at 3T and 7T field strengths.

Introduction

Previous examinations using resting-state functional MRI (rs-fMRI) to examine intrinsic connectivity networks (ICNs) have been based on a number of strategies to assess general test-retest reliability (1). Because an exhaustive search for connections across every voxel is too cumbersome, some form of dimensional reduction is required, especially as the resolution of rs-fMRI studies improves at ultra-high field strengths. Data driven methods such as independent components analysis (ICA) or graph theory can identify key networks, but physiological interpretation of these networks is problematic. Seed-based analyses rely on previously identified locations of interest from either task-based fMRI (2), group connectivity atlases (3), or data driven approaches (2,3), but restricting assessment to relatively few specific locations limits generalizability. Region-of-interest (ROI) analysis uses the average signal of large anatomical areas, missing more granular features of ICNs.

We present a novel hybrid seed-based approach to assess the test-retest reliability of ICNs. An exhaustive but systematic, ROI-limited search is performed for seed pairs with maximal correlation. We examine the reproducibility of these seed pairs and their relation to four well-known ICNs. We then use this approach to assess test-retest reliability between subjects scanned at 3T and 7T field strengths.

Methods

Healthy subjects (N=5) were scanned four times: twice at 3T (Siemens Prisma, Erlangen, Germany) and twice at 7T (Siemens Magnetron) over an average of 199 days (144 - 234). At 3T, 137 volumes were collected using single-band EPI with 31-4 mm thick axial slices (TE/TR = 29/2800 ms, 80° flip, 128x128 matrix, 256x256 mm2, FOV). At 7T, 128 volumes were collected using multi-band EPI (MB factor = 3) with 81-1.5 mm thick axial slices (TE/TR = 21/2800 ms, 70° flip, 160x160 matrix, 210x210 mm2, FOV). High resolution T1 weighted images were acquired for anatomical context

The last 124 volumes were kept for analysis. The data were brain extracted (FSL), motion-corrected and corrected for slice timing (SLOCOMO), and physiologically corrected (PESTICA). 3T data were filtered with a 2D Hamming window while the 7T data remained unfiltered. Anatomical scans were segmented into 16 ROIs (FreeSurfer). For each subject, segmentations and functional scans were affine transformed to the first 7T scan for that subject (AFNI). Subject-specific atlases were then applied for subsequent analysis.

For each scan, a 3x3 seed around each voxel from one ROI was correlated with a 3x3 seed around each voxel in another ROI. The seed pair that yielded the highest correlation was selected for test-retest analysis. This was repeated for every ROI pair (16x16 ROIs) excluding overlapping regions. Correlations were normalized to z scores using whole-brain correlation maps to the seeds (4). Reliability was assessed using intra-class correlation coefficients (ICC) of seed pairs. Reliability was also assessed using the median displacement of seed pairs between scans.

Results

Over all subjects and scans, 4252 max correlation seed pairs were found. The median ICC at 3T was 0.37 (0.033 - 0.93), while at 7T it was 0.34 (0.0017 - 0.90) (t-test showed no significant difference, p = 0.63). The median spatial displacement of seed pairs from session 1 to session 2 was 15.85 mm at 3T and 12.5 mm at 7T (t-test showed no significant difference, p < .0044).

Figure 1 plots session 1 against session 2 for selected connections within the central executive network (CEN) at 3T, while Figure 2 plots the same at 7T. The total ICC for the CEN was 0.36 for 3T and 0.29 for 7T, however the ICC for CEN nodes to non-CEN nodes was 0.88 and 0.68, respectively. The same pattern was found for the motor (0.14/0.19 intra-network, 0.58/0.73 inter-network), default mode (0.29/0.42 intra-network, 0.87/0.79 inter-network), and salience (0.41/0.39 intra-network, 0.78/0.77 inter-network) networks

Discussion

The typical 3T rs-fMRI study contains ~75,000 voxels, requiring some 5.6 billion correlations for an exhaustive, seed-based correlation analysis. This expands to ~500,000 voxels with 2.5 x 1011 correlations at 7T. The ROI-limited approach introduced here reduces that computational burden to ~90 million (3T) and ~4.8 billion (7T) correlations without a priori knowledge of node locations or non-physiological data driven approaches.

While there was little significant difference in overall seed-based reliability between 3T and 7T, there were significant differences at the network level, which showed higher reliability at 3T. This is likely due to lower SNR from higher resolution voxels at 7T (2x2x4 mm3 vs 1.2x1.2x1.5 mm3). In addition, due to a combination of wider search space and minimized spatial averaging, ICC-based reliability was lower than has been reported previously. However, the current study demonstrates strong inter-network reliability, which deserves further exploration.

Acknowledgements

References

(1) Shehzad et al., Cerebral Cortex. 2009 Oct; 19:2209-2229

(2) Thomason et al., NeuroImage. 2011; 55:165-175

(3) Gue et al., NeuroImage. 2012; 61:1471-1483

(4) Lowe et al., NeuroImage. 1998; 7:119-132