Comprehensive Output Estimation of Double Scattering Proton System With Analytical and Machine Learning Models

Area of Science:

• Medical Physics
• Radiation Oncology
• Computational Modeling

Background:

• Proton beam output in double scattering systems is variable and challenging to model accurately.
• Current treatment planning systems (TPS) struggle with precise modeling of these variations.

Purpose of the Study:

• To develop an empirical method for estimating proton beam output in double scattering systems.
• To design and validate both analytical and machine learning (ML) models for this estimation.

Main Methods:

• Generated three analytical models (polynomial, linear, logarithm-polynomial) and three ML models (Gaussian Process Regression with different kernels).
• Trained models on 1,544 clinical measurements and validated on 241 measurements.
• Determined minimum sample sizes for achieving ±3% accuracy and compared model agreement.

Main Results:

• Polynomial and ML GPR (exponential kernel) models achieved <3% deviation from measured outputs.
• Polynomial models required at least 20 samples per option; ML GPR required 400 samples for comparable accuracy.
• Independent models showed <2% deviation on the testing dataset.

Conclusions:

• Developed accurate polynomial and ML GPR models for proton output estimation (<3% deviation).
• These models serve as independent prediction and cross-checking tools for double scattering proton beams.