Daily Multi‐Leaf Collimator positional error prediction using machine learning: A Dynalog‐based approach for patient specific quality assurance
Samira Aghaei Akizi, Seyed Mahmoud Reza Aghamiri, Zahra Siavashpour, Reza Ghaderi, Soharb BarimaniAbstract
Background
Accurate positioning of multi‐leaf collimators (MLCs) is critical for ensuring the dosimetric precision of intensity‐modulated radiation therapy (IMRT). Nonetheless, daily MLC positional variations can degrade treatment accuracy, emphasizing the need for robust methodologies to detect and predict these deviations.
Purpose
This study proposes a machine learning (ML) approach to predict daily MLC positional deviations using factors contributing to positional errors, enabling daily patient‐specific quality assurance (PSQA) to assess the dosimetric impact of these discrepancies on patient treatment.
Methods
DICOM‐RT and Dynalog files from 29 IMRT patients were collected, and 78 features were extracted, including aperture‐based features, mechanical parameters, and MLC check data. A LightGBM machine learning model was constructed to predict daily positional errors between planned and delivered MLC positions at each control point. For dosimetric analysis, nine independent IMRT plans were evaluated using dose volume histogram (DVH) parameters such as D 98% , and V 100% for targets and D max , D mean for Organs at Risk (OARs). Moreover, the Gamma Pass Rate (GPR) was calculated for both the delivered and predicted dose distributions, with regarding the primary planned dose distribution as the reference.
Results
The proposed LightGBM model demonstrated accurate predictions, achieving Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of 0.035 mm and 0.078 mm for the training dataset, 0.070 mm and 0.164 mm for the validation dataset, and 0.054 mm and 0.144 mm for the test dataset, respectively. Leaf velocity, with a Spearman correlation coefficient of 0.703, was the most influential feature. Evaluation on nine independent clinical cases demonstrated good agreement between delivered and predicted dosimetric deviations from the planned values, with mean differences below 0.200 Gy and approximately 1.5% for target structures, and even smaller discrepancies for OARs. The GPR (3%/3 mm criterion) was about 100% for delivered doses and 99.99% for predicted doses.
Conclusions
The proposed ML‐based LightGBM model offers an efficient daily PSQA tool, with the potential to enhance the treatment accuracy and support proactive clinical decision‐making in IMRT.