Figure 1.
Histogram of the average CC matrix of all datasets. The majority of CC values are below 0.2 with a relatively small number of values above 0.4. Negative values exist but are relatively few and only take on small values.
Figure 2.
Schematics of the iterative clustering pipeline.
Schematics for iterative hierarchical clustering framework based on resulting cluster size. Small clusters are excluded and large clusters are clustered further iteratively using a smaller cluster count.
Figure 3.
Hierarchical clustering result.
Clustering results from the selected cut levels. Bilaterally symmetric clusters are displayed in the same figures using two different colors (blue and red). The large cluster went through further iteration is marked with a green box.
Table 1.
Cluster count for hierarchical clustering.
Table 2.
Hierarchical clustering results.
Figure 4.
Comparison between the hierarchical clustering and ICA results.
Set of hierarchical clusters (left column) and ICA components (middle column) with good matching. The right column displays their corresponding overlays with overlapping areas showing in orange. The group ICA t-maps were binarized with a threshold at p<0.001 for ease of comparison with clustering results.
Table 3.
Comparison between the ICA and clustering results.
Figure 5.
Cluster size from all clusters produced by applying the proposed framework to the acquired resting-state fMRI datasets. It is clear that many clusters have less than 50 voxels.
Figure 6.
Intra-cluster connections of the clusters.
Scatter plot of the normalized intra-cluster connections (the ratio between the number of connections with CC≥0.3 and the number of possible connections) as a function of the cluster size. Locally weighted scatterplot smoothing (LOESS) regression analysis (dashed curve) shows a distinct change in trends in the data. To emphasize this change, linear regression (shown in blue) was used to extrapolate the initial distinct trend of the LOESS curve. Linear regression occupied 63 of the smallest clusters (shown in blue crosses) and intersects the x-axis at 50.02. This observation was used to determine the minimum cluster size threshold.
Figure 7.
Unilateral networks from ICA analysis.
Symmetrical one-sided networks found amongst ICA results. Both in their entirety cannot be obtained through hierarchical clustering due to the spatial overlap between them.
Figure 8.
The default mode network in its entirety found at cut number k = 18 from the first iteration. This cluster was found through exhaustive search over all cut numbers.