BEAM3R: Beam's-eye-view architecture with Mamba-3 for implicit dose reconstruction

To meet the requirements for photon and proton dose calculation in the DoseRAD2026 challenge, researchers proposed the BEAM3R framework. This architecture utilizes a Mamba-3-based state-space deep sequence model, combined with physical transmission conditions for modeling. Its core structure shares a 2D CNN encoder-decoder for processing each plane beam eye view slice: the photon model uses bidirectional Mamba-3 to capture the dose contribution of downstream materials; the proton model uses a forward core along with learned energy prefix markings and Bragg peak refinement modules. Additionally, the framework introduces axial grid alignment and implicit super-resolution representations based on subpixel phase packing, aiming to reduce interpolation artifacts and support high-resolution reconstruction. In the preliminary DoseRAD2026 test set, the local gamma throughput for CT to photon and proton models was 96.8% and 96.0%, respectively, with average absolute errors of 0.0041 and 0.0079 at the hierarchical plan level. However, the performance of the models based on synthetic CT decreased, and the gamma throughput for photons and protons…