KEY POINTS
- This retrospective model-development study used magnetic resonance imaging and direct stopping power ratio datasets from 140 primary brain tumor patients treated with adjuvant proton therapy in the ProtoChoice-Hirn trial.
- Three two-dimensional U-Net models were trained using axial, coronal, and sagittal slices; their voxel-wise median produced a 2.5-dimensional synthetic stopping power ratio map. The test set included 21 patients.
- The 2.5-dimensional approach performed best, with head mean absolute error of 0.062 ± 0.009, bone mean absolute error of 0.124 ± 0.025, structural similarity index of 0.89 ± 0.02, and bone Dice similarity coefficient of 0.86 ± 0.03.
- Dose recalculation showed clinical target volume D2% and D98% differences ranging from -0.60% to 0.75%. Organ-at-risk mean and maximum dose differences were within ±2.70 Gy(RBE).
- Gamma agreement was high in high-dose regions, with 99.13% local 2%/2 millimetres pass rate using a 90% dose threshold, but lower with a 10% dose threshold at 76.19% ± 4.67%. Mean absolute range shifts ranged from 0.50 to 1.79 millimetres.
CLINICAL TAKEAWAY
Direct generation of synthetic stopping power ratio maps from magnetic resonance imaging may help move brain tumor proton therapy closer to a magnetic resonance-only workflow by avoiding computed tomography-to-magnetic resonance registration and computed tomography number conversion steps. However, residual range errors, limited test-set size, non-simultaneous imaging, and low-dose gamma discrepancies mean this remains technical workflow-development evidence rather than implementation-ready clinical proof.