This study investigated the multi-site reliability of resting state fMRI (rs-fMRI) using Default Mode Network (DMN) connectivity and graph theory measures like mean shortest path distance, clustering coefficient, modularity, transitivity and global mean strength. Test-retest and between-site reliability for all metrics were calculated by variance component analysis using restricted maximum likelihood (REML) estimates. Test-retest reliability was found to be poor to fair and between-site reliability was consistently poor for all metrics.
1) Participants: 24 healthy volunteers (13 male and 11 female; ages:22-41) participated in a total of four imaging sessions (0-21 days apart) at two sites. (Site 1) Two imaging sessions were conducted using a 3T General Electric Signa HDx and a 16-channel brain array (Nova Medical). (Site 2) Two imaging sessions were conducted using a 3T GE Discovery MR750 and a 32-channel brain array (Nova Medical). Both imaging sessions at any given site were conducted on the same day. An additional 15 healthy volunteers (6 male and 9 female; ages:15-29) participated in one imaging session at Site 1.
2) MR Data Acquisition: Each imaging session (independent of site) consisted of a high-resolution (1mm isotropic) T1-weighted scan for registration purposes and two rs-fMRI scans with common imaging parameters across sites (TR/TE=2000/26ms; flip=35deg; 34 slices; FOV=20cm; 3.125x3.125x3.80mm voxel size, 290 volumes).
3) Data Processing: rs-fMRI data were processed following the pipeline from a recent study1 using AFNI2 and FSL3,4. Complete datasets of four rs-fMRI scans were thus obtained for 17 participants after motion censoring. Using the atlas5, the surviving brain volumes were divided into 278 contiguous regions-of-interest (ROIs).
4) Data Analysis: The functional connectivity (FC) matrix for each rs-fMRI scan was computed by cross-correlating the mean time series of each of the 278 ROIs. Resultant FC matrices were treated as weighted undirected graphs to calculate network measures at different threshold values of correlation. Self-connections, anti-correlations and all values less than the threshold were set to zero and the largest component of the undirected graph of ROIs was used for computing network measures.
DMN connectivity was calculated using a spherical seed region (18mm radius) placed at the posterior cingulate/precuneus (PCC; MNI coordinates 0,50,28). The mean time series of the seed region was correlated with each voxel time series to obtain a whole-brain correlation map. Using the independent dataset from 15 individuals (with same imaging parameters) the average whole-brain correlation map was used as a connectivity mask for the DMN to avoid circularity issues, and to obtain high quality ROIs. For averaging purposes, clusters of at least 500 voxels were observed for a correlation threshold of |r| > 0.13, resulting in four ROIs, largely located in PCC, medial frontal cortex (MFC), right and left temporal cortex (RTC and LTC respectively). The average Pearson correlation coefficient was obtained within each of these ROIs for each rs-fMRI scan and participant.
For all gray matter voxels, the temporal signal to noise ratio (tSNR) was also computed.
The reliability of each metric was calculated by variance component analysis according to the model6 in equation (A) of Figure 1, using restricted maximum likelihood (REML) estimates. All factors are treated as random effects, with subjects and sites fully crossed, and sessions nested within site and subject combinations. The residual term accounts for the variance due to runs within each session, day-to-day variation at each site, and other unexplained factors. Test-retest and between-site reliability was estimated by equations (B) and (C) of Figure 1.