Segmentation of Dynamic PET scans using Cluster Analysis

The purpose of this study was to investigate the possibility to segment a dynamic positron emission tomography (PET) image based on the behaviour of different tissue types to the same injected tracer. Another aim was to find out if the optimum number of clusters for K-means clustering could be found. Finally, the possibility of improving the signal to noise ratio (SNR) was investigated. The K-means cluster analysis clusters the data via an iterative process. This process uses the time activity of each voxel to assign this voxel to one of the K clusters. The number of clusters, K, must be pre-selected. The result of the analysis is a clustered image where each voxel is assigned to the cluster from which the centroid most closely resembles the time activity of that voxel. The study was carried out with the use of a self-written script within the program Matlab. This script was used to analyse the PET data by first performing the K-means cluster analysis with the use of the built in function kmeans. After this the mean squared error (MSE) was determined to look at the possibility of estimating the optimum number of clusters, K. The within cluster variance and the summed deviation were also calculated for the same purpose. For the SNR improvement a different type of clustering, soft clustering, was employed. This method uses the distance of each voxel to the different cluster centroids to represent each voxel by a linear combination of all centroids. It was found that the K-means cluster analysis gives a clear segmentation of the dynamic PET image. The optimum number of clusters can be estimated with the use of the MSE or the within cluster variance by looking at the stabilising region of the resulting graphs. Clues have been found that, when the surroundings are not being taken into account, the optimum number of clusters could be calculated by looking at the summed within cluster deviation. This should be subject of further research. As a last result, the SNR of the dynamic PET images have been improved with the use of the results of the cluster analysis. This results can be used to represent the original dataset with a much smaller new dataset.