Performance quantification of clustering algorithms for false positive removal in fMRI by ROC curves

Introduction: Functional magnetic resonance imaging (fMRI) is a non-invasive technique that allows the detection of specific cerebral functions in humans based on hemodynamic changes. The contrast changes are about 5%, making visual inspection impossible. Thus, statistic strategies are applied to infer which brain region is engaged in a task. However, the traditional methods like general linear model and cross-correlation utilize voxel-wise calculation, introducing a lot of false-positive data. So, in this work we tested post-processing cluster algorithms to diminish the false-positives. Methods: In this study, three clustering algorithms (the hierarchical cluster, k-means and self-organizing maps) were tested and compared for false-positive removal in the post-processing of cross-correlation analyses. Results: Our results showed that the hierarchical cluster presented the best performance to remove the false positives in fMRI, being 2.3 times more accurate than k-means, and 1.9 times more accurate than self-organizing maps. Conclusion: The hierarchical cluster presented the best performance in false-positive removal because it uses the inconsistency coefficient threshold, while k-means and self-organizing maps utilize a priori cluster number (centroids and neurons number); thus, the hierarchical cluster avoids clustering scattered voxels, as the inconsistency coefficient threshold allows only the voxels to be clustered that are at a minimum distance to some cluster.