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

Abstract

This paper introduces linac beam modelling network (LBMnet), a deep-learning-based approach for efficient linac beam modelling, generating percentage depth dose (PDD) and beam profiles by predicting beam data from sparse single-field measurements, thereby enhancing beam modelling efficiency in radiation therapy. The LBMnet model, based on an auto-encoder architecture, was trained on a limited dataset obtained from three Elekta Versa HD™ linacs during commissioning and annual quality assurance processes. The dataset included PDDs and beam profiles across field sizes (1 × 1–40 × 40 cm2) of 6 MV x-ray beams. Data augmentation and pseudo-profile input were employed for improved accuracy in the high-dose and penumbral regions. The model’s predictive performance was assessed by comparing the absolute differences between the predicted and measured beam data. Additionally, a beam model was created using the beam data predicted by LBMnet. To evaluate the accuracy of this LBMnet-based beam model, gamma analysis was performed on dose distributions of ten clinical cases using the 1%/1 and 0.5%/0.5 mm criteria. LBMnet demonstrated a PDD prediction accuracy within ±2% up to a depth of 22 cm. The profile predictions showed maximum differences within 3% in the high-dose regions, except for the penumbra areas. For clinical dose distributions, gamma analysis showed over 99% and 91% agreement with the 1%/1 0.5%/0.5 mm criteria, respectively. The LBMnet model shows a strong potential for improving beam modelling accuracy and efficiency. Despite the limited dataset, the model delivered robust predictions, providing a reliable and timesaving alternative to conventional measurement-based methods.