Abstract:
The purpose of this study was to develop the virtual patient-specific VMAT QA based on the extracted features from the multileaf collimator (MLC) patterns and fluence map. The machine learning (ML) approach was used to develop the prediction (or regression model) and classification models. The input of models applied Multileaf collimator (MLC) patterns and fluence map as the key features, whereas gamma passing rates (GPR) results from patient-specific VMAT QA of electronic portal imaging devices (EPID) as the label or the response class for testing these models. Sensitivity and specificity scores were calculated for models’ accuracy. The highest sensitivity score was observed at the regression model, with 81.25% sensitivity, and the highest specificity score was also observed at the classification model with 66.67% specificity. The models’ accuracy can be improved by increasing the population in the training data set. This study indicated the virtual patient-specific VMAT QA (model-based prediction) using ML approach has feasibility for determining the VMAT plan risk that could fail the tolerance without actual QA measurement. Moreover, the virtual patient-specific VMAT QA shows the potential to corporate within the VMAT treatment planning process.
