Click here to

Session: Translational Technologies and Techniques [Return to Session]

Prescription Tradeoff Decision Support for Pancreas SBRT: From Templates to Artificial Intelligence Models

Y Xie1*, R Li1, W Wang1, M Hito2, L Zhang3, W Giles1, H Stephens1, Q Wu1, F Yin1, Y Ge4, Q Wu1, Y Sheng1, (1) Duke University Medical Center, Durham, NC, (2) Princeton University, ,,(3) North Carolina State University, ,,(4) University of North Carolina at Charlotte, Charlotte, NC


TH-F-TRACK 5-5 (Thursday, 7/29/2021) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Purpose: Radiation oncologists typically employ template constraints for dose prescription and OAR sparing. However, dose prescription and OAR constraints often form competing constraint pairs which require tradeoff discussions between the planner and radiation oncologist. In this study, we designed a novel artificial intelligence (AI) tool to support radiation oncologists in making optimal tradeoff decisions. It that can estimate achievable tradeoff scenarios before treatment planning starts.

Methods: This decision support tool has an effective user interface module that integrates with a treatment planning system to facilitate the prescription decision process. It is powered by a machine learning core which is trained to learn the balance between PTV coverage and OAR sparing from a dataset of 98 pancreatic SBRT cases retrospectively selected for this study. A 10-fold cross validation was performed comparing clinical plan’s PTV coverage (V100%) with model predicted values. These comparisons were further stratified by attending physicians to gain additional insights into decision making process. Cases with large discrepancy were identified, analyzed, and re-planned to validate and assure the achievability (i.e. accuracy) of predictions.

Results: The clinical plan PTV V100% was 87.7 ± 14.5% while the model predicted value was 90.5 ± 9.6% (p =0.005). Model agreement discrepancy was observed among attending physicians. Among all 98 cases, nine cases were identified with large variation from model predictions. For the re-planned cases, an average of 15.3% improvement was achieved over the original clinical plans, indicating the AI model’s accuracy and benefits for clinical planning guidance.

Conclusion: The AI decision support assistant provides dynamic and interactive tradeoff guidance. Its decision support prior to treatment planning can save valuable efforts from both the planning team and radiation oncologist. Moreover, this tool should also work effectively as a training tool providing valuable tradeoff insights for fresh resident who are new to the treatment site.

Funding Support, Disclosures, and Conflict of Interest: This work is partially supported by NIH R01CA201212 research grant and Varian master research grant.



    Radiation Therapy, Decision Theory


    IM/TH- Informatics: Informatics in Therapy (general)

    Contact Email