Fig 1.
PCA is first applied to each subject for dimension reduction, then the dimension-reduced matrices are stacked together and PGICA is applied; as a comparison, a second dimension reduction step is usually required by fastICA algorithm. PGICA can potentially be used in combination with any of the existing back reconstruction methods such as dual regression and GIG-ICA.
Fig 2.
True signals for the simulation examples.
Each component is a two dimensional array where the pixels in a square have higher intensities than the rest of the array. A random noise is added to each of the components at all pixels.
Fig 3.
Map of available source in simulation 3.
Fig 4.
Histograms of age (left), IQ (middle), and SRS (right) for participants in ABIDE plotted and colored by disease diagnosis and overlaid, where blue corresponds to typically developed (TD) controls and red corresponds to ASD individuals.
Fig 5.
Boxplots (for both fastICA and PGICA) of the average correlations (log-transformed) of the true signals with the estimated signals from simulation 1 on the left and simulation 2 on the right.
Table 1.
Summary measures of the correlations in the two simulation examples.
Table 2.
Compare fastICA/InfoMax ICA/PGICA accuracies.
Fig 6.
Axial, sagittal, and frontal (left to right) planes of the auditory, control, default mode and visual networks (from top) estimated using 301 fMRI scans from the 1,000 Functional Connectomes Project dataset.
The thresholded maps are overlaid on a greyscale MNI template brain. The 90th slice is shown from the MNI template in each of the plots. The colors correspond to the intensities in the estimated brain networks where white: high intensity to red: low intensities.
Fig 7.
3D view of auditory, control, default mode and visual networks (from top).
Table 3.
Speed increase of PGICA.
Fig 8.
Axial, sagittal, and frontal (left to right) planes of the default mode, auditory and visual networks (from top) estimated using 779 fMRI scans from the ABIDE dataset.
The thresholded maps are overlaid on a greyscale MNI template brain. The colors correspond to the intensities in the estimated brain networks where white: high intensity to red: low intensities.