Verification of neural network based algorithm for crystal identification of PET block detector

Abstract

The neural network based self-organizing feature map (SOFM) for crystal identification of PET block detector has been proved effective on the Siemens Inveon detector. In this paper, the algorithm is verified in the system level with all the 192 block detectors (each coupled to four Photomultiplier Tubes - PMTs) on a Siemens Biograph four-ring clinical PET scanner, and is tested on an Avalanche Photo Diode (APD) detector of a BrainPET insert. The gain drifting of PMT and APD are also simulated to investigate the responses of the neurons. The verification results show that on the Biograph scanner, the modified SOFM algorithm achieves equivalent crystal identification accuracy to the currently used software method. On the APD detector, the algorithm overcomes low peak- background ratio in the position profile and performs more accurately than the method currently used. The simulations show that the neurons can track the changes of crystal peaks in real- time. This feature brings a good solution to eliminate crystal identification errors caused by gain drifting, especially the gain changes of APD detectors due to temperature variations on the BrainPET insert.