Room: ePoster Forums
Purpose: The purpose of this study was to predict the actual position of MLC position during patient treatment without dry-run.
Methods: Clinac (Varian, CA, USA) was used for generating the machine log-file called the Dynalog. The process of predicting the actual position of the Dynalog file from the RT-plan consists of three folds. First, we implemented an algorithm to calculate the expected position of the Dynalog recorded in 50-ms increments from the control-point of the RT-plan recorded at 2-degree intervals for volumetric arc therapy (VMAT). Second, the deep-learning network trained to predict the actual position of the Dynalog by inputs: the expected position, gantry angle, collimator angle, and neighboring MLCs' expected position with seven time-stamps. Third, we reconstruct the 2D-fluence map using the predicted actual position of MLCs.The deep-learning network based on Long-Short-Term-Memory (LSTM) was trained using Dynalog. The first 70% of the total treatment time was used for training the network, and the remaining 30% treatment time data was used as a test set for evaluating the accuracy of the network. The accuracy of the prediction was evaluated using mean absolute error (MAE), root-mean-square error (RMSE). In addition, the 2D-fluence map of predicted actual position was reconstructed and compared with the true actual position of the Dynalog using gamma analysis.
Results: The MAE, RMSE was 0.412, 0.912 mm between the planned position and the actual position, repectively. The gamma passing rate was 3.12% (SD = 1.13) higher than planned 2D-fluence map at gamma criterion with 1%/2 mm.
Conclusion: In this study, a feasible method of the pretreatment QA without dry-run was suggested by predicting the actual movement of the MLCs from the radiation treatment plan.
Funding Support, Disclosures, and Conflict of Interest: This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIP) (No. NRF-2017R1C1B2011257)