Improved performance in fuzzy clustering of functional MRI datasets by effective processing strategies

The artificial datasets simulated a transversal brain slice (128x128 pixels) with 35 time-points series over a time invariant texture. Three datasets, each with the same regions of activation (49 pixels), were created by adding 3% Gaussian noise and with functional contrast-to-noise (CNR) levels of 1.33, 1.66, and 2.0; values common in fMRI experiments of the human brain. Three hybrid datasets were also analyzed, consisting of 25 activation pixels overlaid on an in vivo single slice MRI (64x64 pixels). Simulated 140 time-points series were added with the same effective CNR used on the artificial datasets, 1.33, 1.66 and 2.0. Data sets are available at http://www.ci.tuwien.ac.at/research/oenb/oenb_data.html. For each cluster analysis run, two association coefficients were calculated, a weighted Jaccard coefficient (JC) and the correlation coefficient (CC) as described in [2]. EROICA [3], part of the EvIdent software package, is an EDA approach specifically designed to improve the efficacy of fuzzy c-means clustering for functional MRI datasets. Using spectral peak (SP), a frequency-domain solution to finding periodic signals buried in noise, EROICA takes advantage of the fact that most functional MRI experiments result in periodic (or nearly periodic) activation time-courses (TC) to separate the original region of interest into a “noisy” set and a “potentially interesting” set. Fuzzy clustering is performed on the latter subset, which is likely to contain periodic activations. With clustering EDA methods such as fuzzy c-means and neural gas, users specify the expected number of clusters in the dataset. EROICA implements cluster merging, where clusters with similar TC centroids are amalgamated, thus the user can safely specify a large number of initial clusters and not impair computational performance as similar clusters are merged. For all datasets, analysis runs were performed with 5, 10 and 20 as the initial number of clusters using EROICA, fuzzy c-means and neural gas. No pre-processing of the time courses, other than normalization, was performed using the latter two methods.