Deep learning-based generation of direct stopping power ratio maps for MR-only proton therapy of primary brain tumor patients.

Background and purpose: A magnetic resonance-only proton therapy (MRoPT) workflow requires synthetic computed tomography (CT) to derive the stopping power ratio (SPR) for treatment planning. Here, we introduce three U-Net models to predict synthetic SPR (sSPR) maps from magnetic resonance imaging (MRI) scans of primary brain tumor patients. Materials and methods: SPR maps and heterogenous MRI data were collected from 140 patients. Three 2D-U-Net models were independently developed with axial, coronal, and sagittal MRI scan slices, the combination of which resulted in 2.5D sSPR.To evaluate the sSPR and associated dose maps, the mean absolute error (MAE), dose volume histogram-parameter differences, e.g. D2% and D98% of the clinical target volume (CTV), and the mean and maximum dose (Dmean and Dmax, respectively) for organs at risk (OAR; brainstem, optic chiasm, optic nerves, lacrimal glands and whole brain) were computed. Finally, gamma and range shifts analyses were performed. Results: The head cohort mean MAE was 0.062 ± 0.009 for the 2.5D sSPR. For the CTV, D2% and D98% differences ranged from -0.60% to 0.75%. Regarding the OAR, the Dmean and Dmax differences, corrected for relative biological effectiveness, were within ±2.70Gy (RBE). The mean pass rate for the local gamma criterion of 2%/2 mm with 10% dose threshold was 76.19 ± 4.67%. Averages of absolute range shifts ranged from 0.50 mm to 1.79 mm. Conclusions: Generated sSPR showed minor discrepancies from SPR. This study is a step closer to an MRoPT workflow for brain tumor patients, albeit the residual errors full impact needs to be more investigated in future work.