Synergizing Artificial Intelligence 3D Dose Predictor with Ethos Intelligent Optimizer for Head and Neck Adaptive Radiotherapy Treatments
J Visak*, E Inam, B Meng, D Nguyen, D Parsons, T Zhuang, D Moon, V Avkshtol, D Sher, M Lin, University of Texas Southwestern Medical Center, Dallas, TX
Presentations
SU-H300-IePD-F6-5 (Sunday, 7/10/2022) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 6
Purpose: Varian Ethos treatment-planning-system(TPS) functionally generates plans differently than conventional TPS with novel intelligent-optimization-engine(IOE). IOE accepts user-defined clinical goals offering no direct interaction with optimizer or insight in which objectives are most conflicting in real-time. This reduces planning confidence for complex head-and-neck(H&N) cases. We evaluate two machine-learning models’ ability to interact with IOE and guide planner in creation of robust and high-quality H&N plans.
Methods: Ten clinical H&N cases were selected and re-planned on Ethos using (1)commercial knowledge-based planning model with RTOG-based universal constraint template (KBP-RTOG) and (2)clinically deployed in-house artificial-intelligence-3D-dose-predictor (AI-Guided). The dose predictor is unique that physician can one-step launch the model and preview the 3D dose to aid clinical goal selection. Both models were derived from similar training data. KBP-RTOG plans were optimized until universal and DVH-predictions were satisfied and AI-guided plans until criteria was mostly met. Target coverage was normalized to 95% coverage of highest dose level PTV. High-impact organs-at-risk and target coverage were evaluated with respect to clinically treated(Benchmark) plans. Plans were compared for statistical significance using paired two-tailed student t-test.
Results: Both KBP-RTOG and AI-Guided plans demonstrated acceptable plan quality, but overall AI-Guided plans were superior. Hotspot for all plans on average were 107±1%. Both KBP-RTOG and AI-Guided plans OAR doses present no statistical difference from Benchmark plans with the exception of KBP-RTOG larynx (3.64Gy higher, p=0.03) and the contralateral parotid (2.51Gy higher, p=0.03). In total, 8 of 9 high-impact OARs were better spared in AI-Guided plans.
Conclusion: While AI-Guided plans were most optimal, commercial knowledge-based planning techniques in conjunction with RTOG universal constraints are readily shareable across clinics. Plan quality is likely superior in AI-Guided plans because physicians can preview dose and provide upfront guidance and realistic objectives to planners. KBP-RTOG potentially achieves similar plan quality with additional iterative editing of clinical goals.
Keywords
Treatment Planning, Inverse Planning, Optimization
Taxonomy
TH- External Beam- Photons: adaptive therapy
Contact Email