Room: Karl Dean Ballroom B1
Purpose: IMRT QA measurements are routinely performed prior to treatment deliveries to verify dose calculation and delivery accuracy. In this work, we applied a machine learning based approach to predict EPID-based IMRT patient QA passing rates.
Methods: 129 IMRT plans with various treatment sites were planned in Eclipse and delivered with portal dosimetry on 3 TrueBeam and 2 Trilogy linacs. EPID QA results of a total of 986 beams were collected and analyzed using gamma criteria of 2%/2mm and 3%/3mm, each with a 5% threshold. The datasets applied for training the machine learning models consisted of 886 beams, with the remaining 100 beams used for the testing set. Ten-fold cross validation was utilized to prevent â€œoverfittingâ€? and to validate the model. Each beam was characterized by a set of 31 features including plan complexity and machine characteristics. Five different machine learning algorithms (KNN, NL-SVM, AdaBoost, Random Forest and XGBoost) were applied to predict the passing rates. A stacking algorithm was employed to combine the final results.
Results: The stacking algorithm demonstrated that 92% of predictions were within 3% of the measured 2%/2mm gamma passing rate with a maximum error of 4.18%, while 99% were within 1% of the 3%/3mm rate with a maximum error of 2.5%. The most important predictive features were identified: maximum aperture displacement from central axis, union area of aperture, ratio between MLC aperture area and jaw area, beam irregularity and maximum of jaw positions.
Conclusion: We have demonstrated that portal dosimetry patient QA passing rates can be accurately predicted by stacking multiple machine learning algorithms. The machine learning based approach will allow physicists to better identify the failures of IMRT QA measurements and to develop proactive QA approaches.