Comparing crystal identification algorithms for PET block detectors

Abstract

A high resolution detector is being developed for our small animal position emission tomography (MicroPET). Each scintillation event is mapped in a two dimensional (2-D) position through the relative ratio of the output signals of the PMTs. Crystal Look-up table (CLT) used in ThuMicroPET scanner defines the matching relation between signal position of a detected event to a corresponding detector pixel location. It has a direct impact on imaging quality and brings significant influence to the gantry overall performance. This paper compares Fuzzy C-Means (FCM) Clustering Algorithm, Self Organizing Feature Map (SOFM) neural network, and Watersheds algorithm for crystal identification. The primary result shows, the watersheds method has the fast speed and higher accuracy for crystal identification but sensitive with preprocessing; and SOFM has advantage of on-line monitoring and training, with also relatively reasonable accuracy, but unluckily long time cost. FCM have lower stability and accuracy in this experiment. In addition we present a new simple but powerful parameter Interaction Ratio (IR), to quantitatively assess the performance of Crystal Identification of detectors. As proved as experiment, not only it could be quantitatively evaluate the segmentation method generating LUT, but also assess the performance of block detector itself.