Room: Karl Dean Ballroom A1
Purpose: To accurately predict the multi-leaf collimator (MLC) positions that would be during the dynamic intensity modulated radiation therapy (IMRT) treatment delivery using machine learning approach.
Methods: IMRT treatment planned data consisting of 18 parameters were used for this purpose. Delivered leaf positions data were used as a target response for training the algorithm. Data were extracted from a Varian linac machine log (dynalog) files. A neural network (NN) architecture with 120 units, representing Varian Millennium 120 HD MLC system, was built to predict the leaf positions at control points during sliding window dynamic IMRT treatment delivery. The developed model was trained on 632 (70%), validated on 135 (15%), and tested on 135 (15%) control point data. The prediction accuracy was evaluated using the mean squared error (MSE) and the regression plot.
Results: The predicted leaves positions closely matched the leaves positions during the treatment delivery. A maximum MSE of 0.0001 mmÂ² was achieved in predicting the MLC positions on the test data for each leaf. The correlation coefficient, R, measures of the goodness of fit, was perfect (R =1) in all plots indicating an excellent agreement between the predicted and delivered MLC positional for the training, validation and test data.
Conclusion: We have demonstrated that the developed model is capable of predicting the MLC positions that would be executed during treatment delivery with a very high degree of accuracy. Integrating the MLC predicted positions into the treatment planning system (TPS) and making the necessary improvements would certainly enhance the IMRT plan QA passing rates. Moreover, the method could be further developed to be implemented for virtual patient-specific IMRT/VMAT QA with dose calculated on patient anatomy rather than water phantom.
Funding Support, Disclosures, and Conflict of Interest: The authors would like to thank King Fahad Specialist Hospital - Dammam, Saudi Arabia, for providing us with the log files used in this study.