Program Information
Comparison of Regression Methods for Knowledge-Based Planning
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
Presentations
SU-I-GPD-T-341 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: Knowledge-based planning (KBP) has been proved to be a powerful tool for plan guidance. It estimates the best achievable organ-at-risk (OAR) dose volume histogram (DVH) based on previous cases of similar anatomy features planned by experienced planners. Currently, regression-based models mostly use the stepwise multiple regression to model the connection between anatomical features and expected DVH outcome. The purpose of this study is to investigate the performance of alternative regression methods.
Methods: The three alternative regression methods studied are ridge regression, least absolute shrinkage selection (lasso), and elastic net regression. To evaluate model prediction accuracy, we use 155 prostate cases. We randomly choose 110 cases as the training set, and the remaining 45 cases are used as the testing set. Ten-fold cross validation is used within the training dataset to tune the parameters for each method. The principle components of 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. The training set is bootstrapped (random sampling with replacement) 100 times to construct 100 models for each method, and the final evaluation is an average of the wRMSE of bootstrapped models.
Results: For bladder prediction, lasso and elastic net generalize better than alternative methods. For rectum prediction, stepwise regression yields the least prediction error when training set size is smaller than 70; lasso and elastic net produce similar and better results when the training set size gets larger.
Conclusion: Overall, elastic net performs reasonably well for all the different prediction tasks with different settings. However, it appears no method is a clear winner. In order to get more consistent prediction performance, we are developing an ensemble learning method to combine these regression methods.
Funding Support, Disclosures, and Conflict of Interest: Supported by NIH under grant #R01CA201212 and a master research grant by Varian Medical Systems.
Contact Email: