Program Information
Outlier Identification in Radiation Therapy Knowledge Modeling Using One-Class Support Vector Machine and Robust Regression
Y Sheng1*, Y Ge2 , J Zhang1 , T Xie1 , F Yin1 , J Kirkpatrick1 , Q Wu1 , (1) Duke University Medical Center, Durham, NC,(2) University of North Carolina at Charlotte, Charlotte, NC
Presentations
SU-H4-GePD-J(A)-2 (Sunday, July 30, 2017) 4:30 PM - 5:00 PM Room: Joint Imaging-Therapy ePoster Lounge - A
Purpose: To develop an effective method for identifying geometric and dosimetric outliers in radiation therapy knowledge modeling in order to ensure proper application and update of models.
Methods: A total of 126 prostate cases, 126 prostate plus lymph node cases and 103 prostate bed cases were included. They were all clinical IMRT cases. Ten additional prostate cases re-planned with dynamic-arc were introduced to simulate dosimetric outliers. The bladder and rectum were used as the organs-at-risk (OARs).Isometric mapping was performed to convert the high dimensional feature vectors into a 2D space in which the cluster frontier was generated using the one-class support vector machine. Cases outside the frontier were flagged as geometric outliers and removed. The remaining cases were used to perform the robust regression and the robust standard deviation of the residuals was calculated. Dosimetric outliers were identified based on 1% false discovery rate and were thereafter removed.All remaining inlier cases were 10-fold cross-validated and the model built with inliers was validated on outlier cases. The weighted sum of absolute residual (WSAR) was recorded for inliers and outliers to evaluate dose-volume histogram prediction accuracy. Wilcoxon Rank-Sum test was performed to compare two groups.
Results: Eighteen and 20 geometric, 23 and 18 dosimetric outliers were identified for bladder and rectum, respectively. All dynamic-arc cases were identified for both OARs. For bladder, the mean and standard deviation of WSAR were 0.026 and 0.014 for inliers, 0.073 and 0.054 for outliers, respectively. For rectum, they were 0.027 and 0.013 for inliers, 0.094 and 0.087 for outliers. The WSARs were significantly different between inliers and outliers for both organs (p<0.05).
Conclusion: We established a workflow for identifying geometric and dosimetric outliers. Results demonstrated the effectiveness of identifying both types of outliers and enhancing the model quality for clinical implementation.
Funding Support, Disclosures, and Conflict of Interest: This work is partially supported by NIH under grant #R01CA201212 and a master research grant by Varian Medical Systems.
Contact Email: