Revisiting single-point-source localization in Compton cameras through detector-level statistical cues revealed by ComptonNet analysis

Abstract

Compton cameras are widely used for gamma-ray imaging owing to their high sensitivity, wide field of view, and broad energy coverage. Recent deep-learning models, such as ComptonNet, have demonstrated robust source localization directly from raw detector events, even under sparse-photon conditions. However, the decision process of such models remains unclear, limiting interpretability and further improvement. In this study, we perform a systematic analysis of the latent representations of ComptonNet and uncover the detector-level statistical cues, including interaction-position asymmetries in the scatterer and absorber, together with the scatterer-to-absorber event ratio, for single-point-source localization, under the constraints of fixed detector geometry and single-source conditions. While currently limited to single sources, this analysis reveals fundamental properties of the network's decision process. Guided by these insights, we develop a simple rule-based estimator and a compact model, Posi-Net. Posi-Net achieves localization accuracy comparable to or better than ComptonNet, while improving interpretability and memory consumption.