Transformer dose engine achieved near-Monte Carlo accuracy for carbon ion therapy

KEY POINTS

• The training dataset used head and neck CT images from 28 patients, eight beam orientations, and 121 carbon-ion energies from 120 to 400 MeV/u. In total, 871,200 Monte Carlo-labeled pencil-beam blocks were generated, with six separate patients reserved for testing.
• Two models were compared under identical input and training conditions: the Transformer-based Ci-DoTA and the recurrent Ci-DoLSTM. Both accepted a three-dimensional CT block and explicit beam energy, predicted the individual pencil-beam dose, and accumulated weighted beams into a complete plan.
• At the pencil-beam level, mean gamma pass rates using 3%/3 mm criteria were 99.85% for Ci-DoTA and 99.06% for Ci-DoLSTM. Under the stricter 1%/1 mm criteria, agreement was 98.01% versus 94.29%, favouring the Transformer.
• Across all six test plans, Ci-DoTA achieved full-plan gamma pass rates of 99.04–99.76% at 3%/3 mm and 92.88–94.42% at 1%/1 mm relative to Geant4/GATE Monte Carlo calculation.
• Per-beam inference required 14.39 milliseconds with Ci-DoTA and 11.96 milliseconds with Ci-DoLSTM. The LSTM was slightly faster, but its loss of accuracy became more apparent under stringent comparison criteria.
• Residual errors clustered around lateral penumbrae, distal fall-off regions, and interfaces with abrupt density change. CT Hounsfield-unit gradient intensity was negatively associated with 1%/1 mm agreement, with a breakpoint of 625.53 HU/mm, but explained only about 30% of variance.
• The system predicted physical dose rather than a complete clinically weighted biological dose, was trained for one machine and institution, and lacked external testing across different beam models, anatomical sites, and Monte Carlo implementations.

CLINICAL TAKEAWAY

SOURCE

Medical Physics