Machine Learning Based Prediction for Patient-Specific Quality Assurance: Understand the Plan Complexity for VMAT
X Zhang*, Y Zhang, B Liu, N Yue, K Nie, Rutgers Cancer Institute of New Jersey, New Brunswick, NJ
Presentations
PO-GePV-T-174 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: To evaluate complexity metrics of volumetric modulated arc therapy (VMAT) plans and predict the portal dosimetry based patient-specific quality assurance (QA) results with machine learning.
Methods: A hundred thirty consecutive VMAT plans treated for various treatment sites from Sep 2020-Feb 2021 were included in the study. Each beam was delivered on TrueBeam Linacs with portal dosimetry and the results were analyzed using 2%/2mm global dose/distance-to-agreement (DTA) gamma passing rate (GPR). Total of 313 beams, including 60 beams exceeding the GRP tolerance, were used to build the machine learning model. Forty-six features of both the plan complexity and machine performance were collected for model building. Random forest algorithm was applied in the feature selection process. With selected features, conventional logistic regression and machine-learning based method, were used to create the classification model that can predict potential failure. In addition, five-fold cross-validation was employed to evaluate the model performance.
Results: The area under the curve (AUC) and sensitivities of the machine-learning based model and logistic regression model were 0.88 and 85%, 0.76 and 78%, respectively. MLC modulation complexity, plan irregularity and jaw tracking are the most important features to predict the potential QA failure.
Conclusion: Portal dosimetry based QA failure might be potentially predicted by machine-learning based model for VMAT plans. This information may be integrated into the plan optimization to generate more robust plan prior to the physical QA check.
ePosters
Keywords
Modeling, Quality Assurance, Portal Imaging
Taxonomy
Not Applicable / None Entered.
Contact Email