An Improved Gamma Interaction Position Estimation Using Deep Neural Networks For Resistor Based Multiplexing Circuit

A resistor based multiplexing circuits such as discretized positioning circuit (DPC) and symmetric charge division have been commonly used to reduce the number of output channels and the associated cost in gamma detectors. However, they can lead to a degradation of image quality as the number of pixels increases because of their nonlinear property. The purpose of this study was to develop deep neural networks (DNNs) fitting algorithm to improve the gamma positioning accuracy of a pixelated detector with a resistor based multiplexing circuit. The detector module was composed of a 12 x 12 crystal array of LYSO with one to one coupling to a 12 x 12 SiPM array and 144:4 resistor based DPC. Acquired training dataset at the center of each pixel with a collimated point source and the (x, y) position was given to each dataset for DNNs fitting target. The DNNs consisted of the first input layers of five nodes and six hidden layers of 144, 48, 36, 24 and 12 nodes at each layer. Last output layers have two nodes that give x and y coordinate. To evaluate the performance of DNNs fitting algorithm using the training dataset, the accuracy of pixel identification and peak-to-valley ratio were calculated at the center and the edges of detector. DNNs fitting algorithm improved the peak-to-valley ratio on the image and the accuracy of pixel identification was also enhanced especially at the detector edges. Average peak-to-valley ratios for x-axis and y-axis were improved by 11 times and 12 times, respectively. Average pixel identification error rates at 32 pixels were improved by 15%. Since the proposed DNN was trained by simple input vectors of 4 channel pulse-peaks, it was applicable to larger input vectors from pulse-shape features to further improve the accuracy of position estimation.