The purpose of this study was to develop and evaluate machine learning models for predicting quality assurance (QA) outcomes of volumetric modulated arc radiation therapy (VMAT) treatment plans. A dataset of 500 VMAT treatment plans and diode-array QA measurements were collected for this study. Gamma passing rates (GPRs) were computed using a 3%/3 mm dose difference and distance-to-agreement gamma criterion with local normalization. 241 complexity metrics and plan parameters were extracted from each treatment plan and their relative importance for accurately predicting GPRs was assessed and compared using feature selection methods via forests of Extra-Trees, mutual information, and linear regression. Hyperparameters of different machine learning models – which included linear models, support vector machines (SVMs), tree-based models, and neural networks – were tuned using cross-validation on the training data (80%/20% training/testing split). Features were weakly correlated with GPRs, with the small aperture score (SAS) at 50 mm having the largest absolute Pearson correlation coefficient (0.38; p < 0.001). The SVM model, trained using the 100 most important features selected using the linear regression method, gave the lowest cross-validation testing mean absolute error (MAE) of 3.75%. This represents a significant 41.1% improvement (p < 0.001) over “random guessing” error as simulated by randomly sampling a fitted normal distribution to the testing data. These predictive models can help guide the plan optimization process to avoid solutions which are likely to result in lower GPRs during QA.
- Machine learning
- Patient-specific quality assurance
- Radiation therapy
- Treatment planning
- Volumetric modulated arc therapy