Prompt gamma emission prediction using a long short-term memory network.

Area of Science:

• Medical Physics
• Computational Physics
• Radiotherapy Physics

Background:

• Proton therapy offers precise dose delivery but requires accurate monitoring.
• Prompt gamma (PG) emission imaging is a promising real-time method for verifying proton range.
• Predicting PG emission is crucial for enhancing the accuracy and safety of proton therapy.

Purpose of the Study:

• To develop and validate a Long Short-Term Memory (LSTM)-based method for predicting prompt gamma (PG) emission in proton therapy.
• To evaluate the performance of different LSTM model configurations using computed tomography (CT) data from prostate cancer patients.
• To assess the prediction accuracy in terms of gamma passing rate and range shift across a wide range of proton energies.

Main Methods:

• Utilized CT scans from 33 prostate cancer patients to generate 10^7 proton pencil beam (PB) histories for Monte Carlo (MC) simulations.
• Extracted 3D relative stopping power (RSP), PG, and dose data for training and validation using LSTM networks.
• Trained and tested three LSTM models: RSP-input/PG-output, RSP/dose-input/PG-output (single-energy), and RSP/dose-input/PG-output (multi-energy).

Main Results:

• The multi-energy LSTM model (input RSP/dose, output PG) demonstrated superior performance.
• Achieved a mean gamma passing rate of 98.5% (92.8% worst-case) for PBs between 125-210 MeV.
• Exhibited a mean absolute PG range shift of 0.15 mm (1.1 mm maximum) with prediction times under 130 ms per PB.

Conclusions:

• Developed a sub-second LSTM-based PG emission prediction method with high accuracy for prostate cancer patients.
• The multi-energy model effectively predicts PG emissions across a broad spectrum of proton energies.
• This method holds significant potential for real-time quality assurance and range verification in clinical proton therapy.