Deep learning based linear energy transfer calculation for proton therapy.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Traditional methods for calculating linear energy transfer (LET), crucial for relative biological effectiveness (RBE), face accuracy-vs-speed trade-offs.
• Monte Carlo simulations offer accuracy but are computationally intensive, hindering real-time dose optimization.
• Analytical approximations are faster but lack precision, impacting treatment planning.

Purpose of the Study:

• To develop and prototype a deep learning model for calculating dose-averaged LET (LETd).
• To enable real-time biological dose evaluation and LET optimization in proton therapy planning.
• To utilize patient anatomy and dose-to-water (DW) data for accurate LETd distribution generation.

Main Methods:

• A 3D Cascaded UNet deep learning model was developed.
• The model processed CT images and dose-to-water (DW) data to generate LETd distributions.
• 275 prostate proton stereotactic body radiotherapy plans were used for training, validation, and testing.

Main Results:

• The deep learning model accurately inferred LETd distributions, with calculations taking approximately 100 ms per field on an NVidia A100 GPU.
• The model achieved an average Mean Absolute Error (MAE) of 0.94 ± 0.14 MeV cm⁻¹ and a 97.4% ± 1.3% gamma passing rate against Monte Carlo simulations.
• Discrepancies were minimal, primarily observed at field edges with high dose gradients and low counting statistics.

Conclusions:

• Deep learning models can efficiently and accurately calculate LETd as a fast-forward method.
• The developed model shows significant potential for optimizing the relative biological effectiveness (RBE) in proton treatment plans.
• Future work will focus on enhancing model performance and assessing its adaptability to diverse clinical settings.