Exhibit Hall | Forum 4
Purpose: To benchmark the lung SBRT VMAT QA using weighted gamma passing rates (wGPRs) and predict the results using a random-forest model with histogram-based features.
Methods: The selected 983 ArcCHECK measurements of lung SBRT Eclipse plans were previously performed on a Varian TrueBeam machine with 6MV-FFF beams and HD120-MLC. GPRs were calculated using SNC Patient software, and wGPRs were computed with optimized weights (70% of 3%/2mm and 30% of 3%/1mm). These wGPRs were first corrected for trending errors using the local normalization technique and then converted into statistically meaningful scores close to Normal distribution through power transform. Total six integrated histograms were calculated: two histograms from the verification dose (dose-volume and gradient-volume histograms) plus four histograms from the DICOM Plan (gantry angle-MU, gantry speed-MU, MLC gap-MU, and MLC speed-MU histograms). Total 634 histogram indices were extracted, then reduced to the final 34 features by removing collinearity. Our random-forest model was trained using 10-fold cross-validation on the training data (80% of the dataset) and recursive feature elimination (RFE) technique to classify the scores with calibrated prediction probabilities. The model performance was evaluated on the testing data (remaining 20%) by average precision (AP) and ROC area-under-curve (AUC).
Results: The model performance on the testing data was 0.75 for the AP and 0.94 for the ROC AUC. The most important feature was the MLC gap standard deviation, which moderately correlated with the score (0.65). The correlation coefficient between the calibrated probability and the score was -0.70.
Conclusion: The wGPR is more suitable for lung SBRT VMAT QA benchmarking, which is more stringent and less noisy than the commonly used 3%/2mm and 3%/1mm criteria. The trained Random-forest model can accurately predict the results without measurements, which can be implemented in the clinic with proper commissioning and periodical QA to reduce the QA workload.