A generalization performance study on the boosting radiotherapy dose calculation engine based on super-resolution.

Area of Science:

• Medical Physics
• Artificial Intelligence in Radiation Therapy
• Radiotherapy Planning

Background:

• Accurate and efficient dose calculation is crucial for radiation therapy planning.
• Online adaptive radiotherapy (OLART) is hindered by time-consuming dose calculations.
• There is a need for faster, high-resolution dose prediction methods.

Purpose of the Study:

• To develop a deep learning-based dose super-resolution model for rapid and accurate dose prediction.
• To enable efficient high-resolution dose distribution calculations in clinical settings.

Main Methods:

• Developed a Multi-stage Dose Super-Resolution Network (MDSR Net) using low-resolution dose data and CT images.
• Trained and tested the model on 340 VMAT plans across various disease sites, including unseen sites for generalizability.
• Compared MDSR Net performance against HD U-Net and cubic interpolation using DVH metrics and gamma analysis.

Main Results:

• MDSR Net demonstrated high accuracy with low prediction errors (0.06-0.84% for benchmark, 0.02-1.03% for generalization).
• Achieved high gamma passing rates (83.1-91.0% benchmark, 71.3-90.3% generalization), outperforming other methods (p < 0.05).
• Prediction errors showed minimal correlation with CT values but varied with dose and dose gradient.

Conclusions:

• The MDSR Net model provides robust and generalizable dose prediction, agreeing well with high-resolution distributions.
• The model offers fast and accurate dose calculations, facilitating clinical implementation of OLART.
• This deep learning approach addresses the need for efficient dose calculation in modern radiotherapy.