@article{
  author = {Long, Dr. zhengtao; },
  title = {Unsupervised Physics-Guided Deep Learning for Sparse-Data Compton Imaging},
  keywords = {Compton camera;Deep learning;Image reconstruction;Physics-guided;},
  abstract = {In response to the instability of Compton-camera reconstruction under ultra-low-count conditions, the substantial loss of usable information caused by conventional strict event screening, and the widespread reliance of existing deep-learning methods on image-label supervision, this study proposes SACN, a physics-constrained reconstruction framework for sparse Compton imaging. This method takes the set of encodable raw events as a unified input and achieves a unified representation of different interaction patterns through event-type identifiers and missing masks. During training, a differentiable Compton forward model is embedded into the optimization objective. Physical likelihood constraints are imposed on events for which Compton geometric consistency can be defined. Combined with geometric pre-localization and a conditional neural implicit field, the method reconstructs a continuous three-dimensional source distribution without image labels. Using a Geant4 simulation platform for a dual-layer GAGG:Ce Compton camera, we compared two data-usage strategies, namely strictly screened-event reconstruction and all-event-utilization reconstruction, under a unified physical simulation setting in single-source, double-source, and multi-energy three-source scenarios. Results show that, under the current simulation setting, SACN provides higher localization accuracy, better artifact suppression, and higher reconstruction efficiency under low-count conditions. Compared with the reference methods, the localization error is reduced by 64.2%--85.1%, the contrast-to-noise ratio is improved by 23.4%--189.1%, and the reconstruction time in the high-statistics single-source scenario is shortened by 99.96% relative to MLEM. In addition, the model is trained only with single-energy Cs-137 data, yet still maintains relatively stable performance for unseen isotopes and complex multi-source scenarios. This indicates a certain degree of cross-energy transferability under the current detector model and simulation conditions. This study provides a potentially useful physics-constrained and image-label-free reconstruction route for low-dose radionuclide imaging and weak-source detection.},
  doi = {10.12074/202604.00155},
  url = {https://chinaxiv.org/abs/202604.00155},
  timestamp = {2026-06-01},
}