Program Information
Development of Knowledge Models for IMRT Treatment Planning Utilizing a Rapid Learning Approach
L Yuan1*, Y Ge2 , Y Sheng3 , F Yin4 , Q J Wu5 , (1) Duke University Medical Center, Durham, NC, (2) University of North Carolina at Charlotte, Charlotte, NC, (3) Duke University, Durham, NC, (4) Duke University Medical Center, Durham, NC, (5) Duke University Medical Center, Durham, NC
Presentations
MO-G-304-3 (Monday, July 13, 2015) 5:15 PM - 6:00 PM Room: 304
Purpose: We present an implementation of the knowledge-based planning method utilizing a rapid learning approach which enables the accurate evolution of models as new data are collected during daily clinical practice.
Methods: The efficacy of knowledge-based planning depends on the accuracy of the knowledge models. Existing methods for knowledge modeling are all based on batch training approach using a static set of training cases. In this study, we investigate a rapid learning approach to train the models using plans from daily clinical practice. A base knowledge model is first established from an existing database. Then each subsequent case continuously evaluates and updates the knowledge model. At each step, evaluation criteria including the studentized residual and the leverage are calculated. They are used to identify outliers from the current model and their causes. Different learning actions are triggered by these criteria. The efficiency and accuracy of the learning approach is quantified by the learning curve. As an initial validation, 100 cases in the pelvic region, including low-to-intermediate risk (n=40), high risk (n=20) prostate, and anorectal (n=40) cases are used. The model-predicted generalized equivalent uniform dose (gEUD) for bladder and rectum are compared with clinical values by a bootstrap validation method.
Results: The Median Absolute Differences between the predicted OAR gEUD and the clinical values in all three types of cases gradually decrease as increasing number of training cases are added in training. The knowledge models learned by this method reach comparable level of prediction accuracy at 2.1% of prescription dose as in batch training mode but with less training cases.
Conclusion: The rapid learning approach facilitates incremental learning of knowledge models from clinical cases of multiple cancer types with improved efficiency. This approach will realize effective implementation of knowledge-based planning in clinics where new cases and cancer types continue to accumulate.
Funding Support, Disclosures, and Conflict of Interest: Partially supported by NIH/NCI under grant #R21CA161389 and a master research grant by Varian Medical Systems.
Contact Email: