ComptonSANet: A Self-Attention Network for Direct Compton Imaging

摘要: In recent years, deep learning methods have been applied to Compton imaging. However, they still face limitations in imaging efficiency and multi-source reconstruction accuracy. This paper proposes a deep neural network based on self-attention mechanisms, named ComptonSANet, which enables end-to-end imaging from Compton scattering events to radiation source locations. ComptonSANet embeds each Compton event into high-dimensional features and employs a self-attention mechanism to capture the physical correlations among events, enabling efficient image reconstruction. The model was validated using a Compton camera with a dual-layer detector. Experimental results demonstrate that ComptonSANet outperforms traditional SBP and MLEM algorithms on both simulated and measured datasets. In single-source simulation tests, it achieves a PSNR of 29.65 dB, representing significant improvements of 16.19 dB over SBP and 4.08 dB over MLEM.