Room: Karl Dean Ballroom C
Purpose: To improve knowledge model prediction accuracy by incorporating case-based reasoning to handle novel anatomies that are of the same type but vary beyond the original training samples.
Methods: A total of 105 pelvic IMRT cases were retrospectively analyzed. Among them, eighty cases are prostate cases while the other twenty-five are prostate-plus-LN cases. A multiple stepwise regression model was trained using 85 cases (80 prostate and 5 prostate plus LN). The remaining 20 prostate-plus-LN cases were validation cases. The proposed workflow started with evaluating feature novelty of new cases against 5 training prostate-plus-LN cases via leverage. Case with larger leverage value than all 5 training cases was flagged as novel/outlier and invoked case-based reasoning. The case database was composed of a dose atlas constructed using 5 training cases. Prostate-plus-LN anatomy was parameterized using three anatomical features: topological connectivity, nodal separation and nodal length. The matched atlas case was identified with L2-norm similarity. The atlas and query case were then linked via deformable registration and the atlas dose was warped towards query anatomy as dose guidance. Case-based dose prediction was compared against the regression model prediction using Sum of Residual (SR) as a dose-volume histogram (DVH) accuracy metric. Two organs-at-risk, bladder and rectum, were considered for modeling. One-sided Wilcoxon Signed-Rank test was performed to compare DVH accuracy.
Results: A total of 13 cases were identified as novel and triggered case-based reasoning. Mean SR of all case-based prediction and regression prediction for bladder of 13 outliers was 0.035Â±0.022 and 0.062Â±0.031, respectively (p<0.0287). For rectum, the mean SR was 0.024Â±0.021 and 0.037Â±0.020 for case-based prediction and regression prediction (p=0.0732).
Conclusion: Case-based prediction using dose atlas improved over regression model for novel cases. We expect that incorporation of case-based reasoning could improve overall knowledge model prediction accuracy. Further validation with larger datasets is warranted.
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