Deep learning proton beam range estimation model for quality assurance based on two-dimensional scintillated light...

Area of Science:

• Medical Physics
• Radiotherapy Physics
• Machine Learning in Healthcare

Background:

• Optical camera systems with scintillators are used for proton beam range quality assurance (QA).
• Analytical methods struggle to derive dose distributions from scintillation images due to quenching and optical effects.

Purpose of the Study:

• To develop a deep learning method for converting scintillation light distribution (LD) into dose distribution in a water phantom.
• To predict proton beam range and spread-out Bragg peak (SOBP) using a 2D map conversion.

Main Methods:

• A 2D residual U-net deep learning model predicted 2D water dose maps from 2D scintillation LD maps.
• Monte Carlo simulations generated datasets with varied proton beam energies, field sizes, and shifts.
• Model performance was evaluated using Bragg peak fitting and gamma index analysis on simulated and predicted dose maps.

Main Results:

• The deep learning method achieved high accuracy in estimating proton beam range (0.02 mm resolution, <0.1 mm deviation) and SOBP width (0.19 mm resolution, <0.8 mm deviation).
• Simulated and predicted dose distributions showed good agreement via gamma analysis, with minor discrepancies in specific regions.

Conclusions:

• Deep learning conversion of scintillation LDs is a feasible and accurate method for estimating proton beam range and SOBP width.
• This approach enhances quality assurance in proton therapy by overcoming limitations of traditional analytical methods.