Crystal Identification in PET – Accepted for Publication in IEEE TNS

Abstract—Determining the location of crystals within detector
arrays (position profile) is a crucial part of system tuning in
Positron Emission Tomography (PET) scanners. It provides a
basis for the mapping of detected events and the most probable
crystal location to assign the event. Accurate assignments are
crucial for proper coincidence event processing and reconstructed
image resolution. In high resolution imaging systems, image
resolution is of primary importance and proper position profiles
are a critical part of system tuning. The High Resolution
Research Tomograph (HRRT) PET scanner is composed of 936
detector blocks each with 64 dual layer crystals arranged in
8×8 grids. Significant engineer time is spent fixing errors in the
automated position profile estimation software used for system
tuning. We have developed a probabilistic approach to position
profile estimation and applied it to the HRRT PET system. Our
approach is composed of a segmentation model for crosstalk
filtering, a prior over valid position profile configurations, and
a grid partitioning algorithm for crystal location finding. Our
model outperforms the manufacturer supplied position profile
estimation software, yielding a 39% decrease in mean squared
error rate as compared to a gold standard configuration in the
HRRT, while being a general solution applicable to many detector
array configurations.

Log mean squared error rates of our Grid Partitioning with noise and crosstalk segmentation model (Seg+GridPart) versus the Siemen’s peak finding model shows a 39% decrease (t = 4:25; p < 0:008) in MSE. Our full model is compared to the Gaussian mixture model (GMM) without noise segmentation, the Gaussian mixture model with segmentations (Seg+GMM), and the Gaussian mixture model with segmentations and initialized with the results form the GridPart model (Seg+GridPart+GMM) versus the gold standard configuration across 702 detector blocks (44,928 crystals).

Log mean squared error rates of our Grid Partitioning with noise and crosstalk segmentation model (Seg+GridPart) versus the Siemen’s peak finding model shows a 39% decrease (t = 4:25; p < 0:008) in MSE. Our full model is compared to the Gaussian mixture model (GMM) without noise segmentation, the Gaussian mixture model with segmentations (Seg+GMM), and the Gaussian mixture model with segmentations and initialized with the results form the GridPart model (Seg+GridPart+GMM) versus the gold standard configuration across 702 detector blocks (44,928 crystals).