(Tuesday, 7/16/2019) 1:45 PM - 3:45 PM
Room: Stars at Night Ballroom 1
Purpose: When VMAT plans fail the mandated phantom quality assurance (QA), time consuming replanning and repeated QA are required, affecting overall workflow. Early identification of such plans allows adjustments prior to QA, improving the treatment and process. We developed a machine learning (ML) method that predicts the passing rate a priori in order to reduce the reliance on manual phantom testing.
Methods: All current work on individual factors that influence the QA are focused on IMRT and our proposed project extends the 16 available metrics in the literature to the VMAT setting, by applying essential concepts behind the derivations. For example, the more irregularly the machine moves, the harder it is to deliver the planned dose accurately, which is a derivative of ‘Plan Irregularity’. Our integrated ML method consists of the following: (a) generalized linear model to predict failing rates, (b) Decision Trees (CART) to improve interpretability, (c) bagged CART to reduce variance and improve prediction accuracy, (d) Boosting to identify strong predictors amongst weak predictors (e.g., different complexity measures), (e) Stochastic Gradient Boosting to prevent overfitting, (f) Stochastic Gradient Boosted Tree with 5-fold cross-validation to reduce misclassification.We used VMAT plan data for brain, prostate, cervix, lung and abdominal sites, which were obtained from the Pinnacle planning system. This included treatment dose, plan and VMAT treatment fields. Data is divided into training and testing samples.
Results: Our ML method exhibits a prediction accuracy of at least 95% for passing rates more than 0.9 on testing data. We also observed that the metric for Mean Field Area has the highest predictive power, followed by Mean Asymmetry Distance.
Conclusion: Most complexity metrics from IMRT QA can be translated for VMAT. Our integrated ML method was applied to predict passing rates at high accuracy, exhibiting potential for streamlining the QA process.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
