Understanding and Modeling Human-Robot Interaction of Artificial Intelligence (AI) Tool in Real Radiation Oncology Clinic Using Deep Neural Network: A Feasibility Study Based On Three Year Prospective Clinical Data
Presentations
TH-F-BRC-6 (Thursday, 7/14/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Ballroom C
Purpose: Our institution has been using artificial intelligence (AI) based planning tool for whole breast radiation therapy (WBRT) for 3 years. We aim to evaluate long-term robustness of the AI tool and model planner's interaction with AI using a deep neural network (NN).
Methods: A total of 1151 patients have been treated since in-house AI-based planning tool was released for clinic in 2019. All patients were included in this study. The AI tool automatically generates fluence maps and creates “AI plan”. Then planner evaluates the plan and attempts manual fluence modification before physician’s approval (“final plan”). The manual-modification-value (MMV) of each beamlet is the difference between fluence maps in AI and final plan. The MMV was recorded for each planner. A deep NN using UNet3+ architecture was developed to predict MMV with AI fluence map, corresponding dose map and organ map in the beam’s eye view (BEV). Among all patients included, 526 patients treated with 6MV beam were utilized to develop the NN with 100 epochs, among which 400/80/46 were used for training/validation/testing. Network modified plans (“AI-m plan”) were compared against AI and final plans for dosimetric endpoints including breast PTV V90%(%), V95%(%), V105%(cc), and Dmax(%).
Results: AI and final plans were comparable in PTV coverage and OAR sparing (Mood’s median test), except V105% (reduced in final plan by 18.4 cc, p<0.001). Ten out of 12 planners showed significantly decreasing modification with mean MMV of 4.0%, 2.3% and 0.8% in 2019, 2020 and 2021, indicating improving acceptability(p<0.001). AI-m plans showed improvement over AI plans in breast V105%(cc)(62.4±84.7 vs 87.6±207.9), Dmax(%)(108.9±1.4 vs 109.7±1.6).
Conclusion: This study demonstrated the robustness of AI planning and steady improvements in acceptability. Utilizing deep NN enables us to understand human planner's patterns to improve AI plans, which could further strengthen AI’s adaptability in the clinic.
Funding Support, Disclosures, and Conflict of Interest: This study is partially supported by AAPM seed grant.
Keywords
Taxonomy
Contact Email