Room: Stars at Night Ballroom 2-3
Purpose: To estimate patient-specific PTV margins for liver SBRT with motion-tracking based on a dry-run session prior to planning (mock treatment), and provide warnings when relatively large errors are predicted during treatment.
Methods: Using log file data for 118 courses (104 patients), 22 features from their first Synchrony model (emulating the mock treatment) including motion-compensation errors are calculated. A support vector machine (SVM) classifier is trained using mock data to predict root-mean-square errors for the subsequent treatment models to facilitate patient-specific margins. For the warning system, the only data used is what is available to the CyberKnife Synchrony system up until the most recent X-ray image is acquired. These data include all manual user interactions such as removing invalid model points. 56 features are calculated including statistics concerning morphological LED time-trace features. A warning system is created via a linear SVM classifier, trained with mock data and X-ray data to predict high errors (â‰¥ 3 mm) during the treatment. The classifiers are evaluated by confusion matrix and receiver operating characteristic metric.
Results: The mock model gives an accuracy of 89% Â± 9% (5-fold cross-validation) to predict if root-mean-square errors for subsequent treatments are above 2 mm for all courses. The variation of errors, linearity between internal target motion and external marker motion, and target motion amplitudes from mock data are important for classifying two groups (errors â‰¥ 2 mm and < 2 mm). For 84% of the time, the warning system can give warnings of an incoming error above 3 mm evaluated via group 5-fold cross-validation.
Conclusion: The mock data can be used to predict errors to provide patient-specific margins prior to planning. It is possible to build a warning system to predict when errors above a predefined threshold will occur during the treatment.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the Ontario Consortium for Adaptive Interventions Radiation Oncology, a research grant from Accuray Inc., CA, Ontario Graduate Scholarship, and John Lyndhurst Kingston Scholarship.
Statistical Analysis, Stereotactic Radiosurgery, Respiration
TH- Dataset analysis/biomathematics: Machine learning techniques