Session Item

The aim of this study is to compare and evaluate the performance of various machine and deep learning models to predict the delivered multileaf collimator (MLC) positions for volumetric-modulated arc therapy (VMAT) plans delivered on an Elekta linear accelerator.

In this study, 100 log files (70 for training and 30 for testing) containing 8,000 control points for VMAT treatment plans delivered on an Elekta linear accelerator were retrospectively obtained from a single institution. From the log files, eight planned parameters were extracted: gantry angle, collimator angle, Y1 and Y2 jaw positions, leaf gap, leaf position, leaf velocity, and leaf acceleration. These parameters were used as inputs to the models and the delivered leaf position from the log files were used as the target. The regression models examined were the linear regression, support vector, random forest, extreme gradient boosting (XGBoost), and artificial neural network (ANN). The models were trained with data from Y1 and Y2 banks and tested on leaves from both banks. Validation of the model performance was done using the mean absolute error (MAE), root mean square error (RMSE), the coefficient of determination R², and fitted line plots showing the relationship between the predicted and delivered positions.

The MAE during validation for support vector, linear regression, random forest, XGBoost, and ANN were 0.210, 0.205, 0.188, 0.220, and 0.179 mm, respectively. The RMSE during validation for support vector, linear regression, random forest, XGBoost, and ANN were 0.122, 0.121, 0.285, 0.335, and 0.096 mm, respectively. The MAE and RMSE achieved by each model during testing are reported in the same order as the metrics reported for validation (support vector, linear regression, random forest, XGBoost, and ANN). The maximum MAE achieved during testing on leaves from Y1 bank were 0.335, 0.337, 0.300, 0.332, and 0.296 mm, and the RMSE were 0.474, 0.478, 0.452, 0.479, and 0.441 mm. The maximum MAE achieved during testing on leaves from Y2 bank were 0.518, 0.516, 0.532, 0542, and 0.508 mm, and the RMSE were 0.721, 0.726, 0.745, 0.667, and 0.709 mm. All models achieved a R² value of 0.999 on the training, validation, and testing datasets. Table 1 summarizes the results for the training, validation, and testing datasets. Fitted-line plots for each model showing the relationship between the predicted and delivered positions for one leaf is shown in Figure 1. The results show that the models’ performance on leaves from the Y2 bank were worse than their performance on leaves from the Y1 bank. 

Although the models show higher errors on leaves from the Y2 bank, the ANN still slightly outperformed the other models in predicting the leaf positions for VMAT plans. Including more data and with further model tuning, these errors can be reduced further, thus improving the model accuracy.