Development of an Ultra-Fast and Setup-Robust Deep-Learning Rbe Dose Engine for Carbon Ion Therapy Using a Conditional Mednext Architecture
Abstract
Purpose
Accurate voxel-wise RBE modeling for carbon ion radiation therapy (CIRT), such as the Mayo Clinic Florida Microdosimetric Kinetic Model (MCF-MKM), is computationally prohibitive for time-constrained clinical workflows. Monte Carlo (MC) calculations can require hours per beam, limiting rapid biological dose verification and adaptive review. This delay prevents routine evaluation of setup-perturbed scenarios and hinders practical implementation of microdosimetry-informed decision support. We developed an ultra-fast deep-learning (DL) RBE dose engine designed to remain accurate under clinically relevant setup uncertainties.
Methods
A 4-level, 25-layer 3D MedNeXt network was trained to predict a 20-bin discretized microdosimetric representation from planning CT. Robustness was enforced using conditional inputs: CT was concatenated with three constant 3D parameter maps encoding rigid setup perturbations about the treatment isocenter (4-channel input). Translations were sampled on a grid of −15, −10, −5, 0, 5, 10, 15 mm in both X and Y, and in-plane rotations were sampled at −5, −3, 0, 3° (±5° range), yielding 196 (=7×7×4) perturbations per patient. Data included 10 patients: 9 for training (1,764 samples) and 1 independent patient contributing 196 perturbed cases split into 98 validation and 98 independent testing. Predicted bins were converted to voxel-wise radiobiological parameters using MCF-MKM, and dose-based RBE (RBE_D) was used to generate the final RBE-weighted dose. Training used L1 loss with AdamW and mixed precision. Accuracy was assessed via 3D gamma analysis (3%/3 mm) between DL-predicted and MC-derived RBE-weighted doses.
Results
The model generated all 20 bins in 3.93 ± 0.48 s per case. Across the 196 validation/testing cases, the mean 3%/3 mm gamma passing rate was 95.51 ± 1.04%, maintaining agreement under translations and rotations, including high-gradient/high-LET regions.
Conclusion
Conditional MedNeXt enables near-real-time, setup-robust RBE_D-weighted dose estimation for CIRT, reducing reliance on repeated MC recalculation and supporting adaptive and QA workflows.