Position-Sensitive Silicon Photomultiplier Array with Enhanced Position Reconstruction by means of a Deep Neural Network

TL;DR

This work addresses sub-millimeter position reconstruction for a 2x2 LG-SiPM tile using a Deep Neural Network (DNN) in place of or in addition to a conventional linear reconstruction. By feeding a 6-channel readout (one per LG-SiPM channel) into a DNN, the authors learn to correct nonlinearity and distortions that arise from device nonuniformities, achieving substantially higher granularity and reduced systematic bias in the reconstructed positions. Qualitative and quantitative results show that nonlinear DNNs converge quickly (within a few epochs) and outperform the linear reconstruction across multiple data-splitting schemes, yielding mean resolutions around 66–79 μm and mean shifts reduced by factors up to 7.8, with granularity improvements up to ~3.5x (from ~687 μm to as low as ~198 μm in favorable splits). The approach elevates the practical performance of LG-SiPM-based detectors, enabling sub-millimeter localization and a dramatic increase in distinguishable light-spot regions, which is significant for compact, high-resolution gamma-ray imaging applications.

Abstract

Single-photon sensitive detectors like Silicon Photomultipliers are widely used in many medical imaging applications. By using detectors with position resolutions, it is possible to build compact photodetector readouts with reduced number of channels, but still preserving position resolution and gamma-rays imaging capabilities. In this work, we present the advantage of using a Deep Neural Networks (DNNs) light position reconstruction applied to a 2x2 array of linearly-graded SiPMs (LG-SiPMs), to minimize the distortions on the reconstructed event maps. Our approach significantly enhances both the resolution and linearity of position detection compared to the nominal reconstruction formula based on the device architecture. Remarkably, the DNN-based reconstruction boosts the number of resolved areas (pixels) by a factor of at least 5.7, allowing a higher level of precision and performance in light detection.

Figures (8)

• Figure 1: Left: front view of the SiPM tile, mounted in a compact module including signal front-end amplifiers, composed of 4 LG-SiPM. Right: SiPM module mounted on linear stages and with optical fiber placed in front.
• Figure 2: Illustration of the 3 splittings used for training and testing. The gray scale indicates the splitting ratio for each motor position.
• Figure 3: Total charge distribution (black) of the test and train sample. A lower threshold (red) on $Q$ is applied to filter the data.
• Figure 4: Schematic of the Deep Neural Network architecture used. The DDNs consist of an input layer for the amplitude $Q_i$ connected to a series of $N_{\rm{layers}}$ 64-input dense hidden layers with hyperbolic tangent activation functions. The output layer is connected to the last hidden layer by a linear activation function to reconstruct the coordinates.
• Figure 5: Reconstructed position of the test sample for the splitting technique (A). Each image corresponds to a number of layers used in the DNN (in reading order: $N_{\rm{layers}} = 0,~1,~\ldots,~5$).