Segmentation of dynamic PET or fMRI images based on a similarity metric

In this paper, we present a new approach for segmentation of image sequences by clustering the pixels according to their temporal behavior. The clustering metric we use is the normalized cross-correlation, also known as similarity. The main advantage of this metric is that, unlike the traditional Euclidean distance, it depends on the shape of the time signal rather than its amplitude. We model the intra-class variation among the time signals by a truncated exponential probability density distribution, and apply the expectation-maximization (EM) framework to derive two iterative clustering algorithms. Our numerical experiments using a simulated, dynamic PET brain study demonstrate that the proposed method achieves the best results when compared with several existing clustering methods.