3D Event Reconstruction in Radiation Detectors using Convolutional Neural Networks
- Resource Type
- Conference
- Authors
- Banerjee, Srutarshi; Ballester, Manuel; Rodrigues, Miesher; Vija, Alexander Hans; Katsaggelos, Aggelos K.
- Source
- 2022 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC) Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC), 2022 IEEE. :1-3 Nov, 2022
- Subject
- Components, Circuits, Devices and Systems
Computing and Processing
Engineered Materials, Dielectrics and Plasmas
Nuclear Engineering
Photonics and Electrooptics
Signal Processing and Analysis
Solid modeling
Three-dimensional displays
Voltage
Reconstruction algorithms
Feature extraction
Sensors
Convolutional neural networks
- Language
- ISSN
- 2577-0829
Room Temperature Semiconductor Detectors (RTSDs) such as CdZnTe are being widely used in diverse applications, to provide interaction position of γ-rays and excellent energy resolution. These sensors are often pixelated. The full potential of these RTSDs are explored while developing advanced single-polarity charge sensing reconstruction algorithms. Current position and energy reconstruction algorithms rely on physical models. Deep learning (DL) techniques can perform event reconstruction with better energy resolution and improved position information than conventional non-DL methods. Here we present a DL approach based on Convolutional Neural Networks (CNN) to identify the 3D interaction position and energy deposition of the γ-rays in a single pixel CdZnTe detector. The network is trained in a supervised fashion with input-output data pairs. The input data to the DL model consists of voltage signals at the electrodes from the superposition of signals generated from each hit of an incident event on the single pixel detector and the output data consists of the 3D position and energy deposited in those hits. The DL model consists of 5 stages of convolutional layers. Every convolutional layer is followed by successive batch normalization and max-pooling layers. These layers extract features from the input data. After the feature extraction, there are 2 stages of fully connected layers. The 3D hit positions and energies deposited by those events are output of this model.