Machine Learning-Based Prediction of Elekta MLC Motion with Dosimetric Validation for Virtual Patient-Specific QA.

Accurate multi-leaf collimator (MLC) motion prediction is a prerequisite for precise dose delivery in advanced techniques such as IMRT and VMAT. Traditional patient-specific quality assurance (QA) methods remain resource-intensive and prone to physical measurement uncertainties. This study aimed to develop machine learning (ML) models to predict delivered MLC positions using kinematic parameters extracted from DICOM-RT plans for the Elekta Versa HD system. A dataset comprising 200 patient plans was constructed by pairing planned MLC positions, velocities, and accelerations with corresponding delivered values parsed from unstructured trajectory logs. Four regression models, including linear regression (LR), were trained to evaluate the deterministic nature of the Elekta servo-mechanism. LR demonstrated superior prediction accuracy, achieving the lowest mean absolute error (MAE) of 0.145 mm, empirically confirming the fundamentally linear relationship between planned and delivered trajectories. Subsequent dosimetric validation using ArcCHECK measurements on 17 clinical plans revealed that LR-corrected plans achieved statistically significant improvements in gamma passing rates, with a mean increase of 2.24% under the stringent 1%/1 mm criterion (p < 0.001). These results indicate that the LR model successfully captures systematic mechanical signatures, such as inertial effects. This study demonstrates that a computationally efficient LR model can accurately predict Elekta MLC performance, providing a robust foundation for implementing ML-based virtual QA. This approach is particularly valuable for time-sensitive workflows like adaptive radiotherapy (ART), as it significantly reduces reliance on physical QA resources.