Neural Representation Learning for Compact and Efficient Modeling of Monte Carlo Phase Space Data.

Area of Science:

• Medical Physics
• Computational Biology
• Radiotherapy

Background:

• Monte Carlo (MC) simulations are the gold standard for radiation therapy dose calculations.
• Large phase space (PHSP) files generated by MC simulations hinder clinical implementation due to storage and computational demands.

Purpose of the Study:

• To develop and evaluate NeRP-MC, the first neural representation learning approach for modeling particle distributions in PHSP data.
• To assess NeRP-MC's ability to model particle distributions using minimal training data.

Main Methods:

• Investigated proton PHSP modeling at 242 and 140 MeV using TOPAS to generate reference PHSP files.
• Trained a multi-layer perceptron with Fourier feature encoding to predict particle energies from spatial and momentum data.
• Evaluated NeRP-MC in three scenarios: compact energy modeling, sparse data modeling, and replacement of PHSP with parametric Gaussian distributions and network-predicted energies.

Main Results:

• The NeRP-MC model requires only 600 KB storage, a significant reduction from the original 3 GB PHSP file.
• NeRP-MC predicted 25 million particle energies in under 0.5 seconds on an NVIDIA A100 GPU.
• In-water dose distributions showed excellent agreement with reference data, with gamma pass rates exceeding 99% at 3%/2 mm and 90% at 1%/1 mm.

Conclusions:

• NeRP-MC enables compact modeling and fast prediction of particle energies, offering a potential replacement for large-scale PHSP files.
• This approach can substantially reduce computational and storage requirements for MC simulations in radiation therapy.
• NeRP-MC promises to advance MC simulation efficiency while maintaining crucial dosimetric accuracy for clinical applications.