Figure 1.

Framework of white‐matter functional network clustering analysis (a). First, the raw T1 was segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF), and then normalized to the MNI space. The subject‐level GM, WM masks were obtained based on maximum probability of each voxel and subsequently averaged to identify a group‐level WM mask. Second, raw fMRI were preprocessed, and time series of 17,716 voxels were extracted from WM mask. Third, 17,716 voxels were subsampled to 4,426 nodes to reduce computational complexity. Pearson's correlation coefficient between each voxel and each node was calculated, and this resulted in a 17,716 × 4,426 correlation matrix for each subject. A K‐means clustering analysis (distance metric correlation, 10 replicates) was performed on the averaged correlation matrix across all subjects. Finally, we randomly divided the averaged connectivity matrix (17,716 × 4,426) into four folds (17,716 × 1,106). The same clustering computation was performed on each fold separately. This resulted in four clustering results. By measuring the similarity of the clustering results between any two folds using Dice's coefficient, the stability of clusters were evaluated. (b) Calculation of functional covariance connectivity (FCC). Based on the idea of “correlation of correlations,” FCC was used to estimate the covariant relationship between two WM networks based on their correlations with multiple GM regions [Color figure can be viewed at http://wileyonlinelibrary.com]