A Statistical Learning Approach to the Accurate Prediction of Multi-leaf Collimator Errors during Radiotherapy

Abstract

Purpose: Discrepancies between planned and delivered movements of multi-leaf collimators (MLCs), a linear accelerator component which modulates the radiation field shape, are an important source of errors in dose distributions during radiotherapy. It is therefore advantageous to be able to predict these errors, and account for them before the plan is delivered to the patient. In this work, we used machine learning techniques to predict these discrepancies, assessed the quality of the predictions, and examined the impact the errors have on quality assurance procedures, and patient dosimetry. Materials and Methods: 114 volumetric modulated arc therapy plans for either head and neck cancer, or prostate cancer were retrospectively chosen. Features hypothesized to be predictive of errors were calculated from the plan files, such as leaf position, velocity and acceleration, whether the leaf was moving towards or away from the isocenter of the MLC, and many others. Differences in position between synchronized plan (DICOM-RT) files and machine reported delivery (DynaLog) files acquired during QA were used as a target response for training predictive models. Model accuracy was assessed using a testing set of data which was not used for model training or validation. Predicted positions were then incorporated into the treatment planning system (TPS), and the increase in accuracy from accounting for errors was verified using gamma analysis. Results: The developed models are capable of predicting the errors to a high degree of accuracy. For moving MLC leaves, predicted positions were shown to be significantly closer to the delivered positions than were the planned positions. Integration of the predicted positions into dose calculations performed within the TPS was shown to increase gamma passing rates measured against dose distributions delivered during QA at all criteria. The proposed method increased gamma passing rates of head and neck plans with 1%/2mm criteria by 4.17% (SD = 1.54%) on average. Impact on patient dosimetry was assessed using dose volumetric histograms. In all cases, predicted dose volumetric parameters were in closer agreement with delivered parameters than were planned parameter, particularly for organs at risk at the periphery of the treatment field. Conclusions: In this study, we showed that MLC errors can be accurately predicted, and that accounting for these errors gives treatment planners a more realistic view of the dose distribution as it will truly be delivered to the patient.