Low-frequency Functional Connectivity in fMRI - Biomedical Signal Analysis: Contemporary Methods and Applications - page 266

algorithm to the partitioning of fMRI data. In a following step, a
comparison between the unsupervised algorithms implementing different
distance metrics is performed.
The TMP algorithm determines the mutual pairwise similarity be-
tween the PTCs, which leads to an important issue in fMRI data analy-
sis: What is the underlying basic similarity measure between the PTCs?
Two approaches described in the exploratory data analysis part are em-
ployed: the TMP corr considering the correlation between the PTCs and
the TMP pred considering the prediction error.
Figure 9.1 visualizes the computed distance matrices for subject #1
and for N = 25 clusters based on both the correlation and the prediction
error methods. The first row shows the unsorted distance matrices and
the second row shows the results obtained after application of the TMP
algorithm, resulting in a display of the distance matrix, where the rows
and columns appear in an ordered fashion. The emerging block-diagonal
structure reflects the characteristic of the TMP algorithm to cluster
PTCs based on their mutual dependency (i.e., their pairwise distance).
By taking the average value of all PTCs belonging to a certain
cluster, a cluster-representative PTC is obtained. Figure 9.2 shows a
comparison of the segmentation results obtained by the unsupervised
clustering methods for subject #1. The cc-cluster describes a method
based on the threshold segmentation of the correlation map. This map
assigns to each pixel the Pearson correlation coecient between the PTC
and the stimulus function. The threshold was chosen as Δ = 0 . 6, and
thus every pixel with a correlation of its PTC exceeding 0.6 is considered
to be activated and is white on the map. For the clustering methods, all
the clusters with an average correlation of PTCs above the threshold of
Δ=0 . 6 are collected and their pixels are plotted white on the map.
The average value of all PTCs belonging to a certain segmentation
determines a segmentation-specific PTC shown under the assignment
maps. A high correlation of these representative PTCs with the stimulus
function cc =0 . 75 is found exceeding for all methods.
It is important to perform a quantitative analysis of the relative
performance of the introduced exploratory data analysis techniques for
all four subjects. To do so, the proposed algorithms are compared for 9,
16, and 25 clusters in terms of ROC analysis using a correlation map with
a chosen threshold of 0.6 as the reference. The ROC performances for the
four subjects are shown in figure 9.3. The figure illustrates the average