A Physics-Constrained Deep Learning Method for Compton Cameras 3D Imaging

Compton cameras serve as important tools for gamma-ray imaging. However, traditional 3D reconstruction methods employ stringent event selection criteria that utilize only approximately 0.05% of all detected events in typical camera configurations. This low utilization rate leads to poor image quality in low-statistics scenarios. To address these shortcomings, this paper proposes a new Physics-Constrained Deep Learning Method (PC-DLM) for 3D imaging with Compton camera based on deep learning algorithm and Transformer neural network. The algorithm is trained and validated on simulated data. Geant4 simulated dataset validation demonstrates that in the camera configuration of this paper, PC-DLM outperforms both traditional and existing deep learning algorithms in terms of localization accuracy, shape and intensity restoration for 3D reconstruction. In particular, it exhibits low MSE, high PSNR and SSIM compared to SBP, MLEM, and 3D-UNet in a low-statistics scenario with N=100. This paper demonstrates the ability of deep learning to accurately localize and recover the 3D spatial distribution of radiation sources in low-statistics Compton data.