3585 - Protocol-Agnostic Machine Learning for Thoracic Radiotherapy Planning: Clinical Results and Workflow Benefits
Presenter(s)
J. Patel1, H. Feely2, and T. R. McNutt1; 1Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, 2Johns Hopkins Medicine, Baltimore, MD
Purpose/Objective(s): To evaluate a protocol-agnostic, anatomy-based machine learning model (PlanAI, Sun Nuclear) that predicts patient-specific achievable dose-volume histogram (DVH) objectives for thoracic radiotherapy, and to determine its impact on plan quality, organ-at-risk (OAR) sparing, and planning workflow efficiency in thoracic cases.
Materials/Methods: Nineteen previously treated hypofractionated thoracic cancer patients were retrospectively replanned. PlanAI uses random forest ensembles to predict achievable DVH objectives for targets and OARs based on spatial relationships among target volumes and OARs. Predicted optimization goals were exported to RayStation and applied in two workflows: 1. fully automated PlanAI optimization without human intervention and 2. PlanAI+User in which a planner, blinded to the original clinical plan, adjusted objective weights only as needed to meet clinical goals. Plan metrics included PTV coverage, Paddick conformity index, OAR DVH parameters (esophagus, lung, lungs–GTV, pericardium, spinal canal, stomach, trachea), and total planning time for the two PlanAI workflows. Paired statistical tests compared PlanAI, PlanAI+User, and clinical plans; p<0.05 was considered significant.
Results: PlanAI produced clinically acceptable plans with reduced optimization time compared with PlanAI+User (mean 4.34 ± 1.56 min vs 41.3 ± 24.6 min; p<0.001). PlanAI and PlanAI+User plans demonstrated improved OAR sparing versus original clinical plans. PlanAI yielded statistically significant mean dose reductions versus clinical plans in esophagus (9 Gy), lungs (8 Gy), and trachea (6 Gy) (all p<0.05). The largest relative improvements occurred at intermediate organ volumes. Paddick conformity index were comparable among PlanAI, PlanAI+User, and clinical plans (0.66, 0.68, and 0.69, respectively; p>0.05), with target coverage of 95% of the PTV. Although PlanAI provided rapid dose sparing, PlanAI+User or clinical approaches were occasionally necessary to meet specific OAR constraints (e.g., esophagus or trachea V10% - V30% doses) or physician-mandated clinical goals, necessitating manual adjustments and longer planning time. Importantly, in all plans where clinical goals were not met, PlanAI predicted objective functions that were above the clinical goals, demonstrating the model’s ability to identify when objectives are unlikely to be achievable and thereby mitigate unnecessary plan revisions by informing realistic, achievable goals.
Conclusion: PlanAI can rapidly generate protocol-agnostic thoracic radiotherapy plans that preserve target coverage while improving OAR sparing in most cases, substantially decreasing automated optimization time and standardizing achievable DVH goals. By providing patient-specific achievable constraints and flagging unattainable clinical goals, PlanAI helps clinicians anticipate planning tradeoffs, address feasibility, reduce unnecessary optimization effort, and streamline workflows.