Fig 1.
The computational framework of group sparse representation on the whole-brain fMRI signals from two groups of subjects (GF: Female, GM: Male).
(a) Extracting whole-brain fMRI data from subjects x (subscript represents the label of subject, e.g., x). (b) FMRI data matrices (Sx) from all the subjects are aggregated (S). (c) Coefficient matrix A with the same spatial information and group correspondence of S, which is decomposed into 2 matrices (AGF, AGM) corresponding to two groups (GF, GM), and each group is made of sub-matrices corresponding to the sparse representation for each subject.
Table 1.
Affective ratings of the movie and head motion parameter of females and male.
Fig 2.
The computational framework of group-wise statistical analysis.
Coefficient matrix A is composed of two groups of subjects (GF: k female subjects, GM: l male subjects). Each row in group T-test represents a component network, which is then mapped backed to brain volume color coding with z-scores and called z-score map (nF1 = … = nFk = nM1 = … = nMl = nx). Each row of sub-coefficient matrices Ax (AFp or AMq) representing individual coefficient spatial map of all the subjects are set as input of SPM12 for two-sample t-tests.
Fig 3.
Representative brain networks (z-score maps) identified by group sparse representation method.
Networks identified by both sparse representation and ICA methods are highlighted by rectangle frames (color code shared with Fig 4).
Fig 4.
The representative brain networks (z-score maps) identified by (a) tensor ICA method, and (b) spatial concatenation group ICA.
Networks identified by both sparse representation and ICA methods are highlighted by rectangle frames (color code shared with Fig 3).
Fig 5.
Clusters showing gender difference.
(a) 14 clusters detected by group sparse representation that show significantly higher activation in females than males: clusters #1–7 (P<0.005 for voxel height and FDR-corrected P<0.005 for cluster extent); cluster #8–14 (P<0.005 for voxel height and FDR-corrected P<0.05 for cluster extent). Clusters detected by (b) tensor ICA and (c) spatial concatenation group ICA that show significantly higher activation in females than males (P<0.01 for voxel height and P<0.01 for cluster extent). Clusters belonging to same brain region identified by all methods are highlighted in colors.
Table 2.
Brain regions with greater activation in females than males as detected by group sparse representation (sorted by p-value in ascending order).
The Network Index refers to the index of dictionary atom generated by group sparse representation algorithm (corresponding to the index in Fig 3).
Fig 6.
Test-retest reliability of brain networks.
(a) Brain maps of the voxel-wise ICCs of matching networks identified by group sparse representation. (b) Average voxel-wise ICCs and (c) scan-wise ICCs of networks detected by all the methods (Sparse representation/tensor ICA/spatial concatenation group ICA/temporal concatenation group ICA: Visual: #143/#3/#28/#30; Auditory: #72/#7/#11/#1; Dorsal attention: #28/#11/#26/#32; Default mode-salience: #81/#2/#10/#11). Error bars signify variance.
Fig 7.
Time courses of representative networks identified by group sparse representation, tensor ICA and spatial concatenation group ICA.
Pearson’s correlation between the time courses of group sparse representation and ICA is labeled on the panel of the corresponding ICA method (upper left corner).