Pitfalls in the Clustering of Neuroimage Data and Improvements by Global Optimization Strategies

Abstract

In this paper, we examined three vector quantization (VQ) methods used for the unsupervised classification (clustering) of functional magnetic resonance imaging (fMRI) data. Classification means that each brain volume element (voxel), according to a given scanning raster, was assigned to one group of voxels based on similarity of the fMRI signal patterns. It was investigated how the VQ methods can isolate a cluster that describes the region involved in a particular brain function. As an example, word processing was stimulated by a word comparison task. VQ analysis methodology was verified using simulated fMRI response patterns. It was demonstrated in detail that VQ based on global rather than local optimization of the objective function yielded a higher performance. Performance was measured in statistically relevant series of VQ attempts using several indices for goodness, reliability and efficiency of VQ solutions. Furthermore, it was shown that a poor local optimization caused either an underestimation or an overestimation of the stimulus-induced brain activation. However, this was not observed if the cluster analysis was based upon a global optimization strategy.