Crystal Area Segmentation for a Scintillation Detector based on Convolutional Neural Network

Abstract

Crystal area segmentation is one of the critical procedures for decoding the detector module coupled with scintillation crystal. However, the blurring effect makes the decoding procedure challenging. For precise decoding, we propose a crystal area segmentation method based on convolutional neural network (CNN). The method is divided into training stage and evaluation stage. In the training stage, data set was extracted from five flood maps in blocks. These blocks went over preprocessing with bandpass filter (BPF) and thresholding. Then the processed blocks were used to train and test the CNN. In evaluation stage, flood map from 2 positron emission tomography (PET) scanners were tested. The method showed 99.5% and 99.4% of peak detection accuracy for each test samples while existing method achieved 91.1% and 95.4%. The proposed algorithm detected center peaks almost perfectly and improved detectability of boundary peaks. Also, the whole decoding process was done in short amount of time. However, the algorithm proposed in this paper only considered the spatial information of the peaks in flood map. In further studies we will develop improved algorithm with using both spatial and energy information to develop more precise and practical decoding algorithm.