Room: Exhibit Hall
Purpose: To investigate the performance of convolutional neural networks (CNN) ability to predict intensity modulated and volume modulated arc radiation therapy (IMRT/VMAT) quality assurance (QA) gamma passing rates.
Methods: A total of 797 treatment plans were delivered by accelerators from several commercial accelerators vendors (Accuray TomoTherapy, Elekta Versa HD, Siemens Artiste and Varian TrueBeam/Truebeam STX) on and measured using an ArcCHECK (Sun Nuclear) QA phantom. The treated plans were from a variety disease sites and various energies. The dose distributions projected onto the QA phantom were used as input for the to the CNN. The CNN was trained to minimize the mean absolute error (MAE) of the global gamma passing rates based on 3% local dose and 3mm distance-to-agreement with a 90% dose threshold. Dropout, batch normalization, usage of synthetic data (created by applying coordinate transformations to real data), and 10-fold cross validation were used for CNN model training.
Results: Initial results focused on 475 VMAT plans of from the full data set. After 10-fold cross validation, the MAE was reduced to 0.94 and 93% of the predicted rates has had less than 3% error. A total of 57% of the outliers were from one specific accelerator, had lower than average passing rates, and had an intermittent issue with a connecting cable which was identified and replaced.
Conclusion: The generated CNN were able to accurately predict VMAT QA gamma passing rates independent of plan features such as disease site, energy and complexity of the plan. An IMRT/VMAT QA prediction algorithm could be used during the planning process to enable the focusing of resources to plans more likely to fail and potentially allow the creation of tighter, patient-specific tolerance levels as opposed to widely-used global levels.