Encrypted login | home

Program Information

An Ensemble Regression Method for Improved Knowledge-Based Planning Model Robustness

J Zhang

J Zhang1*, Y Ge2, T Xie1, Y Sheng1, F Yin1, Q Wu1, (1) Duke University Medical Center, Durham, NC, (2) University of North Carolina at Charlotte, Charlotte, NC


SU-H4-GePD-J(B)-5 (Sunday, July 30, 2017) 4:30 PM - 5:00 PM Room: Joint Imaging-Therapy ePoster Lounge - B

Purpose: The purpose of this study is to improve the robustness of knowledge-based planning (KBP) models in the presence of outlier cases.

Methods: A selection of generalized linear regression methods are utilized in the proposed ensemble model. The model pool consists of stepwise regression, and three penalized linear regression methods, including ridge regression, least absolute shrinkage selection (lasso), and elastic net regression. Leave-one-out cross validation is used within the training dataset to generate out-of-sample but in-training-set predictions. These predictions are subsequently used to optimize the weight for regression coefficients generated by each regression methods. Models that perform well in cross validation usually get bigger weights. Dose-volume histograms (DVH) are modeled and predicted for bladder and rectum. Prediction performance is measured with weighted root-mean-squared error (wRMSE), the weight of which is selected before testing to emphasize the prediction accuracy of high dose regions based on clinical significance. In the outlier experiment, we include 40 prostate 2PTV IMRT cases in the training set and add 10 dynamic conformal arc cases as dosimetric outliers. The ensemble model and individual models are trained with the “contaminated” training set. The model performance are evaluated on 110 prostate 2PTV IMRT cases with wRMSE. Note that the experiment is designed to test the model robustness in extreme settings.

Results: The proposed ensemble method outperforms or performs similarly to the best performing single regression model. It produces the lowest prediction error for bladder and rectum DVH. In rectum DVH predictions, we observed significantly improved wRMSE for the ensemble model compared with individual models (p<0.05). In bladder predictions, all models, except for ridge regression, perform similarly well.

Conclusion: In summary, the ensemble regression method is more reliable than alternative methods, and is more appropriate for clinical use to ensure consistently accurate DVH prediction.

Funding Support, Disclosures, and Conflict of Interest: Supported by NIH under grant #R01CA201212 and a master research grant by Varian Medical Systems.

Contact Email: