Robust Clustering of Positron Emission Tomography Data

Positron Emission Tomographic (PET) data is, in general, difficult to segment due to low signal to noise ratio and low resolution. In recent years, however, it has been demonstrated that clustering can be used as an important preprocessing step prior to parametric estimation from dynamic PET data. Here accuracy and efficiency of standard and some novel clustering methods for this dynamic data for both a scalar biologically relevant measure of the total uptake of tracer in the tissue, and the discrete time activity measurements of this activity in the tissue, are assessed. Inter, intra and inter-intra cluster measures are used for contrasting the results from different clustering algorithms, and for the two data sets. These measures are also used for assessing the reliability of the algorithms for determining the optimal number of distinct clusters in the data sets. All data used for the validation of the methods is practically obtained PET data and thus both heavily contaminated by noise and obtained at non equal time intervals, ranging from 2 seconds to 30 minutes. The analysis demonstrates that modified two-stage clustering algorithms, presented by Guo et al, 2003, are as reliable as the more expensive classical algorithms, hence validating use of these fast cheap algorithms for analyses of dynamic PET data.