Clustering Analysis for Neurotransmitter Response Profiles of Dynamic PET data

In this paper we investigated clustering as a potential aid in reducing false positive rates when modelling neurotransmitter activation using the 1p-ntPET model. The aim was to investigate whether clustering time-activity curves (TACs) before kinetic modelling improves specificity for detecting activation states. We used two popular unsupervised clustering algorithms, k-means and Gaussian Mixture Models (GMM), which were applied prior to voxel-wise kinetic modelling. We generated statistically independent sets of TACs corresponding to [C-11]raclopride kinetics and we investigated the impact of: (a) the level of noise in the TACs, (b) the activation magnitude and (c) the ratio between the number of TACs with and without activation on the classification accuracy of each clustering approach. The classification performance was assessed by calculating the sensitivity and specificity for each clustering approach and was compared against conventional single-voxel modeling. Results showed that applying clustering before voxel-wise kinetic modeling improves classification of TACs between active and non active states, at least for moderate to less noisy TACs. The single voxel modelling results in better specificity at higher noise levels and clustering assures better results at noise levels from 1k counts to 1M counts.