Resting state network estimation in individual subjects

Abstract

Resting state functional magnetic resonance imaging (fMRI) has been used to study brain networks associated with both normal and pathological cognitive functions. The objective of this work is to reliably compute resting state network (RSN) topography in single participants. We trained a supervised classifier (multi-layer perceptron; MLP) to associate blood oxygen level dependent (BOLD) correlation maps corresponding to pre-defined seeds with specific RSN identities. Hard classification of maps obtained from a priori seeds was highly reliable across new participants. Interestingly, continuous estimates of RSN membership retained substantial residual error. This result is consistent with the view that RSNs are hierarchically organized, and therefore not fully separable into spatially independent components. After training on a priori seed-based maps, we propagated voxel-wise correlation maps through the MLP to produce estimates of RSN membership throughout the brain. The MLP generated RSN topography estimates in individuals consistent with previous studies, even in brain regions not represented in the training data. This method could be used in future studies to relate RSN topography to other measures of functional brain organization (e.g., task-evoked responses, stimulation mapping, and deficits associated with lesions) in individuals. The multi-layer perceptron was directly compared to two alternative voxel classification procedures, specifically, dual regression and linear discriminant analysis; the perceptron generated more spatially specific RSN maps than either alternative.

Keywords: Brain mapping; Functional connectivity; Multilayer perceptron; Resting state network; Supervised classifier; fMRI.

Copyright © 2013 Elsevier Inc. All rights reserved.

Figures

Figure 1

Seed ROIs for generation of correlation map data. Seed ROIs resulting from a meta-analysis of task foci (see section 2.3) were defined in volume space. To visualize the surface variability of volume-defined regions, 5 mm radius spherical ROIs centered on stereotactically defined coordinates were projected onto surface reconstructions for each individual. Transparent regions indicate at least 20% surface overlap of ROIs across subjects. Opaque regions indicate at least 50% overlap. Figure S1 shows both hemispheres in slice format.

Figure 2

Single-subject, voxel-wise estimation of RSNs using the trained MLP. A. Each locus produces one correlation map that ultimately results in N=7 MLP estimates of network membership at that locus. B. Each correlation map is masked to include only grey matter voxels and projected into principal component space. C and D. The masked image is passed through the neural network. See Appendix A and Figure A1 for details of MLP connections and training process. E. Output values converted to percentiles (uniform distribution over 0 to 1 interval) for surface displays. The 8th (nuisance component) MLP output is not illustrated.

Figure 3

Projection of RSNs into PCA space. A. Temporal correlation matrix: For each subject, the processed BOLD time-courses were averaged over each seed region. The resulting matrices were averaged across subjects. B. Spatial correlation matrix: For each subject, correlation maps were produced for every seed region. Matrices of spatial correlation between each seed's map were computed, and then averaged across subjects. C. Principal component analysis: PCA was performed on correlation maps, yielding the eigenvectors of the map-to-map spatial covariance matrix. Correlation maps for each seed in each subject were projected onto the PCA components, thus generating a locus in PCA space for each of the 3,675 training images. Color indicates the task analysis from which the region was derived.

Figure 4

MLP Training trajectories as reflected in the Optimization dataset. A. Total RMS error as a function of iteration number. Error decreased monotonically for all networks until reaching a global minimum. The black line represents the total RMS error across all networks. The optimal early stopping point was defined as the global minimum of the total RMS error. B. Change in RMS error for each RSN (sign inverted derivatives with respect to iteration). The plotted values have been normalized by change in mean RMS error (black curve in A). Note sequential appearance of -ΔRMS error peaks and expanded iteration scale. C: ROC curves plotted in parallel with panel A. AUC values in the Training set asymptotically approached unity (not shown), whereas the Optimization data exhibited local maxima (inset). The black line represents the mean AUC across the 7 RSNs. Iterations index is shown on a logarithmic scale in all plots to emphasize early performance.

Figure 5

RSN topographies in individual participants from Validation dataset 1. Voxel-wise MLP results are shown for 3 participants. These are the best, median, and worse performers as determined by RMS error. Voxelwise MLP output values have been converted to a percentile scale within each RSN and sampled onto each individual's cortical surface.

Figure 6

MLP SMN results obtained in Validation dataset 2 individuals. A: Five individuals were selected to represent the correspondence between SMN variability and anatomical variability in the central sulcus (see text for details). MLP SMN scores are displayed overlayed on individual MP-RAGE slices. The bright contour corresponds to the 90th percentile of voxel values. Note: high SMN scores track the shape of the central sulcus (red arrows). B: Correlation between the Talairach Y-coordinate of the centroid of MLP SMN (un-normalized) output values and the Y-coordinate centroid of the central sulcus fundus traced over the path indicated in the right inset figure. The SMN centroid was evaluated over the X–Y range indicated by the left inset figure.

Figure 7

Surface-averaged MLP results. Top: Surface-based average over 100 participants from Validation dataset 2. Middle: Standard deviation of RSN values across subjects. Bottom: Winner-take-all maps depict surfaces with patches colored according to the network with the largest value.

Figure 8

Volume-averaged MLP results 692 participants from Validation dataset 2, displayed in slices. WTA indicates the winner-take-all result (thresholded at 0.7 for the winning value). All seven networks were represented in the cerebellum despite absence of cerebellar seeds in the training data. Note left lateralized cerebral foci and right lateralized cerebellar foci for the language network (white arrows, LAN column); similarly, note right cerebral and left cerebellar foci for the ventral attention network (white arrows, VAN column).

Figure 9

Comparison of MLP to alternative methodologies. A. Selected group-average RSN topography estimates computed with dual regression (DR), linear discriminant analysis (LDA), and a multi-layer perceptron (MLP). B. RSN estimates evaluated over a priori seed ROIs. Topography estimates for each network (A) were averaged over voxels within each seed. The resulting scores were averaged over subjects and plotted for pairs of RSNs (e.g., SMN vs. VIS scores). Markers are colored based on the prior assignment of each seed. Line segments extend from the voxel-wise median score (50th percentile) to the center of mass of the ROI scores for the two RSNs defining the exhibited plane. Note that only the MLP successfully separates LAN seeds from DMN along the LAN axis. C. Inter-class correlation of RSN scores computed as the Pearson correlation coefficient between pairs of RSNs. Note that RSN scores are least correlated for the MLP, indicating more complete orthogonalization.

Figure 10

Examples of objective evaluation of methodology. A. Effect of total quantity of fMRI data on MLP performance. The plotted points (small triangles) represent total RMS error obtained with the fully trained MLP under three random sub-samplings of the Validation 1 dataset. RMS error (E) was fit to a 3-parameter rational function (red line). Asymptotic RMS error (~15.6%) was estimated from the function model. Note monotonically decreasing error with increasing data quantity. The large symbols report values for particular datasets: diamond: Training; circle: Optimization; triangle: Validation 1; square: Validation 2. The inset surface displays show the effect of available data quantity on WTA results. Note less RSN fragmentation with more data. B. Effect of Optimization ROI size on MLP performance. Each seed radius was evaluated with 5 replicates. Red line: locally linear scatterplot smoothing (LOESS, smoothing parameter of 0.5). Note clear minimum at approximately 10 mm radius.