(Monday, 7/15/2019) 3:45 PM - 4:15 PM
Room: Exhibit Hall | Forum 7
Purpose: To improve knowledge model robustness by retaining novel anatomy cases in rapid learning fashion.
Methods: A total of 105 pelvic IMRT cases were included. Among them, eighty cases are prostate cases while the other twenty-five are prostate-plus-LN cases. Prostate-plus-LN cases served as novel anatomy as compared to prostate cases. A knowledge model using case-by-case atlas-based method was used as benchmark for evaluating the performance of a stepwise regression based model with different amount of training cases. Two scenarios were simulated, namely Scarce Scenario with 80 prostate and 5 prostate-plus-LN cases as training data, and Ample Scenario with 80 prostate and 20 prostate-plus-LN cases. The remaining 5 prostate-plus-LN cases served as validation cases. The simulation was performed with four-fold cross validation. The original model under Scarce Scenario (SS) and extended model with Ample Scenario (AS) was evaluated against the baseline atlas-based model for DVH prediction accuracy using sum of residual (SR) as a metric. The prediction of each validation case from SS model and AS model was compared for mean difference using the Wilcoxon Singed-Rank test against the baseline atlas-based model.
Results: For bladder, the mean (standard deviation, SD) SR for SS model, AS model and baseline model was 0.062 (0.033), 0.043 (0.045) and 0.035 (0.023), respectively. Statistical significant difference was observed between baseline model and SS model (p=0.029) while no significant difference was observed between baseline model and AS model. For rectum, the mean (standard deviation, SD) SR for SS model, AS model and baseline model was 0.037 (0.020), 0.033 (0.022) and 0.024 (0.022), respectively. No significant difference was observed between any paired group (p=0.073, 0.170).
Conclusion: Retaining novel cases in linear regression based knowledge model improves the model robustness. Significant improvement was observed for the bladder model. “Retaining� completes the feedback loop for the case based reasoning framework.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by NIH/NCI 1R01CA201212.
Keywords
Taxonomy
TH- Dataset analysis/biomathematics: Machine learning techniques
Contact Email
