Three-dimensional dose monitoring for proton pencil beam based on dual-panel PET with neural network

The development of dual-panel positron emission tomography (PET) detectors for proton therapy has made range uncertainty assessment and dose monitoring more feasible, as they are cost-effective and easy to integrate into proton therapy systems. However, image artifacts caused by limited detection angles and noise issues arising from insufficient sampling at typical clinical dose remain major challenges, particularly in three-dimensional (3D) dose monitoring. In this paper, we proposed a transformer-based architecture to predict the dose distribution from deduced proton-induced positron emitter activity. Mean Relative Error (MRE) is used to evaluated the performance of the model. Furthermore, we theoretically investigate how missing angles and clinical dose variations affect activity and dose reconstruction. The results show that as the clinical dose increases, the quality of PET reconstruction improves until the dose prediction performance reaches a saturation point, which we define as the minimum clinically detectable dose. Notably, this minimum detectable dose varies by anatomical region — 0.2 Gy for the head, and 0.4 Gy for both the abdomen and thorax. At this dose level, the proposed framework achieves a MRE of approximately 0.99% across all anatomical sites. This study shows that our proposed prediction model enables dual-panel PET to effectively predict 3D proton spot dose, and lays the foundation for us to move further reconstructing the range and dose under diverse clinical therapy configurations.