KEY POINTS
- This proof-of-concept study developed NeRP-MC, an implicit neural representation model for proton Monte Carlo phase-space data generated with TOPAS. Reference datasets contained 25 million primary protons at 140 MeV and 242 MeV.
- The model was evaluated for compact energy modelling, reconstruction from 1.25 million training particles representing a 20-fold reduction, and complete phase-space replacement using Gaussian spatial and angular distributions with network-predicted energies.
- Network storage was 600 kilobytes, compared with 3 gigabytes for the original ASCII phase-space file or 875 megabytes in binary format. Prediction of 25 million particle energies required less than 0.5 seconds on an NVIDIA A100 graphics processor.
- At 242 MeV, gamma pass rates were 100% at 3%/2 mm and 2%/1 mm, and 99.5%-99.6% at 1%/1 mm. At 140 MeV, corresponding pass rates were 99.9%, 96.9%-97.3%, and 89.9%-90.5%.
- Bragg peak depth was reproduced exactly in all evaluated scenarios. However, testing was limited to two energies, one nozzle, a single pencil beam, a homogeneous water phantom, and primary protons with unity weight.
CLINICAL TAKEAWAY
NeRP-MC could substantially reduce the storage and generation burden associated with proton Monte Carlo phase-space files while retaining high dosimetric agreement in simple test conditions. The findings are technically relevant for future dose-calculation, quality-assurance, and adaptive-planning workflows, but remain proof-of-concept and require validation across clinical energies, heterogeneous geometries, secondary particles, and complex beam configurations.