Prediction of the individual multileaf collimator positional deviations during dynamic IMRT delivery priori with artificial neural network.

Purposes: Multileaf collimator (MLC) positional accuracy during dynamic intensity modulation radiotherapy (IMRT) delivery is crucial for safe and accurate patient treatment. The deviations of individual leaf positions from its intended positions can lead to errors in the dose delivered to the patient and hence may adversely affect the treatment outcome. In this study, we propose a state-of-the-art machine learning (ML) method based on an artificial neural network (ANN) for accurately predicting the MLC leaf positional deviations during the dynamic IMRT treatment delivery priori using log file data.

Methods: Data of ten patients treated with sliding window dynamic IMRT delivery were retrospectively retrieved from a single-institution database. The patients' plans were redelivered with no patient on the couch using a Varian linear accelerator equipped with a Millennium 120 HD MLC system. Then the machine recorded log files data, a total of over 400 files containing 360 800 control points, were collected. A total of 14 parameters were extracted from the planning data in the log files such as leaf planned positions, dose fraction, leaf velocity, leaf moving status, leaf gap, and others. Next, we developed a feed-forward ANN architecture mapping the input parameters with the output to predict the MLC leaf positional deviations during the delivery priori. The proposed model was trained on 70% of the total data using the delivered leaf positional data as a target response. The trained model was then validated and tested on 30% of the available data. The model accuracy was evaluated using the mean squared error (MSE), regression plot, and error histogram.

Results: The deviations between the individual MLC planned and delivered positions can reach up to a few millimeters, with a maximum deviation of 1.2 mm. The predicted leaf positions at control points closely matched the delivered positions for all MLC leaves during the treatment delivery. The ANN model achieved a maximum MSE of 0.0001 mm (root MSE of 0.0097 mm) in predicting the leaf positions at control points of test data for each leaf. The correlation coefficient, that measures the goodness of fit, was perfect (R = 0.999) in all plots indicating an excellent agreement between the predicted and delivered MLC positions for the training, validation, and test data.

Conclusions: We successfully demonstrated a proposed ANN-based method capable of accurately predicting the individual MLC leaf positional deviations during the dynamic IMRT delivery priori. Our ML model based on ANN outperformed the reported accuracy in the literature of various ML models. The results of this study could be extended to actual application in the dose calculation/optimization, hence enhancing the gamma passing rate for patient-specific IMRT quality assurance.