Magnetic resonance imaging-based proton dose calculation for pelvic tumors using deep learning

<i>Objective</i>. Magnetic resonance imaging (MRI)-only proton therapy combines high soft tissue contrast with high-precision dose distributions. However, conventional dose calculation is impossible on MRI due to missing electron density information. This work investigated the feasibility of two fully deep learning (DL)-based MRI-only proton dose calculation pipelines for the pelvic region and their robustness to MRI intensity distortions.<i>Approach</i>. Two MRI-only proton dose calculation pipelines were established: A) the two-step pipeline converts MRI to synthetic computed tomography (sCT) and predicts proton dose distributions on sCT; B) the direct pipeline predicts proton dose distributions directly on MRI. MRI-CT pairs from 120 pelvis patients were considered. For modeling, 31 727 random pencil beams (PBs) and 13 430 PBs from 6 treatment plans (TPs) were calculated using Monte Carlo (MC) simulations. Performance of the pipelines was measured by comparing predicted and MC-simulated doses in terms of gamma pass rate (3 mm, 3%, dose threshold of 10%) and average relative error (ARE) for, both, individual PBs and TPs. For further understanding, an experiment was conducted to manually introduce intensity distortions to the input image and observe its influence on the predicted dose.<i>Main results</i>. Both pipelines showed high gamma pass rates (>99.2%). The two-step pipeline showed ARE of 0.11% and 2.63% for individual PBs and TP (planning target volume), respectively. For the direct prediction pipeline, larger ARE of 0.16% and up to 6.11% were observed for individual PBs and TP, respectively. The model predicting dose using MRI directly was robust against added MRI intensity distortions.<i>Significance</i>. DL-based MRI-only proton dose calculation was feasible in the pelvic region. The direct pipeline showed potential to learn the mapping between MRI image pattern and proton dose distribution, though, improvement in terms of information usage is warranted. The two-step pipeline is capable to predict proton dose distributions with low errors.