Room: Exhibit Hall | Forum 2
Purpose: The commissioning data is fundamental throughout the lifetime of a linear accelerator (LINAC) and is used for verification of proper LINAC operation, in the continuing quality assurance process and as input to the treatment planning software. Despite the critical importance of the data, there is no existing method nor commercial tool for data verification. Here we propose a novel machine learning-based method for reliably modeling the beam data of LINACs and show its significant promise in LINAC commissioning.
Methods: Beam data acquired during commissioning/annual QA (n=20) was extracted from Varian TrueBeams from multiple institutions. The data sets included percentage depth dose (PDD) and profiles across different energies, field sizes, and depths. Applying data augmentation resulted in a total of 1000 datasets. 85% of the datasets were used for training a multivariate linear regression model, with the rest employed for validation. The prediction of PDDs and profiles took place with or without Ridge Regularization. The accuracy of the predictions and number of sampling points needed for reliable predictions were evaluated using percentage relative error (pRE).
Results: Accurate prediction of PDDs and profiles were found for all beam energies investigated. Predictions of PDDs for field sizes of 4Ã—4 cm2 and 30Ã—30 cm2 was achieved with a pRE of <0.8% for all energies. Similar accuracy was found for profile prediction. Only marginal increase in prediction accuracy was achieved with increased number of sampling points beyond three. Reduction of up to 50% was found for models trained without Ridge Regularization.
Conclusion: Through this novel method we have shown accurate and reproducible generation of high precision data for LINAC commissioning for routine radiation therapy. This method has the potential to greatly reduce the burden on the medical physicists during LINAC commissioning while increasing the safety and precision when introducing new LINACs into the clinic.