Harnessing the Power of Machine Learning for Accurate and Efficient Linear Accelerator Beam Data Commissioning

Abstract

Linear accelerator (linac) commissioning is a fundamental process in the safe introduction of new linacs into the clinical workflow of any radiation oncology department. Commissioning data is vital not only for verification of proper linac operation, but also for the continuing quality assurance process and as input to the treatment planning software. Current commissioning of a new linac is a highly reparative and time-consuming task performed under time pressure but with great need for high precision. The process lacks in clear operational standards and no validation tools are commercially available. In this study, we propose a fast and robust linac beam data commissioning method to significantly save time and manpower in the process of linac commissioning. Routine linac commissioning/annual beam data, including percentage depth dose (PDD) and depth dose profile (DDP) at various beam energies and field sizes, were extracted from three medical linear accelerators at our institution. The datasets were split into training and testing data. For each beam energy, we formulated the PDD or DDP as an output of least square optimization problem, which is solved by a multivariate linear regression model with an input of a few point data measurements at a different field size. Using data augmentation technique, we use 1000 sets of data to train the linear regression model. PDD and DDP were predicted using models trained with and without Ridge Regularization and a few measured sampling points. The predicted PDD and DDP were evaluated using percentage relative error (pRE). Both PDD and DDP were accurately predicted at different beam energies and field sizes. The pRE for predictions of field sizes of 4×4 cm2 and 30×30 cm2 based on 10x10 cm2 beam data was less than 0.8% for both 6 MV and 15 MV beams. Three data points sampled along the PDD was needed to obtain this accuracy; no further increase in accuracy was found with increased number of sample points. For model trained without Ridge Regularization, the pRE was reduced by up to 50%. For model training with Ridge Regularization, data augmentation showed only marginal improvement on pRE. We propose a fast and accurate machine learning-based method to generate linac commissioning beam data for routine radiation therapy. With this method, beam data related to PDD and DDP can be accurately generated using only a few measured sampling points. We envision that this approach will significantly simplify the tedious linac commissioning procedure, save time and manpower, while increasing the safety and precision of the commissioning process.