Enhanced subject-specific resting-state network detection and extraction with fast fMRI

Abstract

Resting-state networks have become an important tool for the study of brain function. An ultra-fast imaging technique that allows to measure brain function, called Magnetic Resonance Encephalography (MREG), achieves an order of magnitude higher temporal resolution than standard echo-planar imaging (EPI). This new sequence helps to correct physiological artifacts and improves the sensitivity of the fMRI analysis. In this study, EPI is compared with MREG in terms of capability to extract resting-state networks. Healthy controls underwent two consecutive resting-state scans, one with EPI and the other with MREG. Subject-level independent component analyses (ICA) were performed separately for each of the two datasets. Using Stanford FIND atlas parcels as network templates, the presence of ICA maps corresponding to each network was quantified in each subject. The number of detected individual networks was significantly higher in the MREG data set than for EPI. Moreover, using short time segments of MREG data, such as 50 seconds, one can still detect and track consistent networks. Fast fMRI thus results in an increased capability to extract distinct functional regions at the individual subject level for the same scan times, and also allow the extraction of consistent networks within shorter time intervals than when using EPI, which is notably relevant for the analysis of dynamic functional connectivity fluctuations. Hum Brain Mapp 38:817-830, 2017. © 2016 Wiley Periodicals, Inc.

Keywords: ICA; MREG; fast fMRI; high frequency fluctuations; resting state networks.

Figure 1

Random Templates: Example templates created by randomly selected FIND parcels and their Dice score distribution (the number of bins is 100 for each plot, equally spaced within the range of calculated Dice values). The template sizes (# of voxels) are shown below each map. Smaller templates (for example #3 and #7) tended to result in lower distributions of Dice coefficients. [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 2

Dice thresholds for the detection of FIND templates: a) The box plots show the distributions of Dice values as a function of the sizes of the random templates. Each bin corresponds to a 500‐voxel range of template sizes with a corrected 95% confidence interval. The corresponding threshold values for each FIND template (upper bound of the confidence interval) are shown in Table I. [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 3

Number of templates: Each box plot shows the number of networks detected across the subjects as a function of the number of extracted independent components in the analysis of the 10‐minute MREG data. Asterisks indicate statistically significant differences in the number of detected RSNs (P < 0.05 corrected). All of the indicated differences were highly significant (P < 0.001). [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 4

Number of networks, EPI vs MREG: The graph shows the mean number of detected ICs for each network, with error bars representing the standard deviation across the subjects. The plots show a comparison of the two imaging techniques according to a) duration of the analyzed time interval, b) number of time points. [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 5

Percentage of subjects with detected networks according to interval duration or number of volumes: a) in MREG data, with the number of volumes matched with EPI, b) in EPI data, c) in MREG data, with the interval duration matched with EPI. Number of volumes and seconds are shown on the x‐axis vs. the 14 different networks on the y‐axis. The average percentages across all networks are shown in the bottom row. [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 6

dDMN networks detected in MREG data: Default‐mode network components detected in 3 different subjects; a), b) and c). The 1st row of each subplot shows the dorsal DMN component representing the whole network (using the red‐yellow colormap on the right), while the 2nd row shows the smaller Ics (partial sub‐networks using the colormaps shown on the bottom) that showed significant Dice similarity with the dorsal DMN template. All ICs are thresholded at z > 2. [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 7

EPI networks: ICs corresponding to each RSN for different analyzed durations. Each row shows results from a single subject and column shows maps with maximum Dice index with the corresponding template. The red background indicates ICs whose Dice index was not statistically significant. On the leftmost column, FIND templates are displayed in green. [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 8

MREG networks: ICs corresponding to each RSN for different analyzed durations. Each row shows results from the same single subjects as in Figure 7. Each column shows maps with maximum Dice index with the corresponding RSN template. The red background indicates ICs whose Dice was not statistically significant. On the leftmost column, FIND templates are displayed in green. [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 9

Tracking the right executive control network (RECN): The length of each window is 50 seconds and axial maps show the highest Dice index match with the RECN template in a single subject. Fluctuations of Dice indices over time are plotted in blue and the threshold used for statistical significance is indicated in red. In 4 of the 12 windows (red background), the components with maximum Dice index do not reach statistical significance. [Color figure can be viewed at http://wileyonlinelibrary.com]