Deep-learning-based linac beam modelling with sparse beam data measurements.

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning

Background:

• Accurate linac beam modelling is crucial for effective radiation therapy.
• Traditional methods for beam data acquisition are time-consuming and resource-intensive.

Purpose of the Study:

• To introduce linac beam modelling network (LBMnet), a deep-learning approach for efficient linac beam modelling.
• To enhance the accuracy and efficiency of generating percentage depth dose (PDD) and beam profiles for radiation therapy.

Main Methods:

• Developed LBMnet using an auto-encoder architecture trained on sparse single-field measurements from Elekta Versa HD™ linacs.
• Employed data augmentation and pseudo-profile input for improved prediction accuracy.
• Validated predictions against measured beam data and assessed clinical dose distributions using gamma analysis (1%/1 mm and 0.5%/0.5 mm criteria).

Main Results:

• LBMnet achieved PDD prediction accuracy within ±2% up to 22 cm depth.
• Profile predictions showed maximum differences within 3% in high-dose regions, with minor deviations in penumbra.
• Clinical dose distributions demonstrated over 99% and 91% agreement with the 1%/1 mm and 0.5%/0.5 mm criteria, respectively.

Conclusions:

• LBMnet shows significant potential for improving beam modelling accuracy and efficiency in radiation therapy.
• The model provides robust predictions even with a limited dataset, offering a reliable, time-saving alternative to conventional methods.