Evaluating a Protocol-Agnostic Machine Learning Model for Automated Thoracic Radiotherapy Treatment Planning
Abstract
Purpose
To evaluate a protocol-agnostic machine learning (PlanAI) (Sun Nuclear, Mirion Medical Company, Florida, USA)1 model for automated treatment planning in external-beam radiotherapy for thoracic cancer.
Methods
Lung cancer patients (N=14) were replanned using the PlanAI. PlanAI uses shape relationship features between organs-at-risk (OARs) and the targets to predict expected and best achievable dose–volume histograms (DVHs) and derives objectives from the best achievable DVHs. The predicted DVHs were imported into RayStation 2023B (Stockholm, Sweden) for treatment planning. Two workflows were studied: (1) PlanAI-only optimization and (2) PlanAI + User, in which a planner, blinded to the clinically approved plan, adjusted objective weights only to meet clinical goals. Plan quality was assessed using OARs’ DVH and the Paddick Conformity Index (PCI)2; planning times were recorded. Two-sided paired t-tests were used with alpha=0.05.
Results
Clinical goals were achieved with the PlanAI + User workflow, and planning target volume (PTV) coverage was equivalent to clinically approved plans. Mean PCI (±SD) was 0.69 ± 0.27 (clinically approved), 0.66 ± 0.28 (PlanAI + User; p=0.79 vs clinically approved), and 0.66 ± 0.27 (PlanAI-only; p=0.73 vs clinically approved), indicating no significant differences in dose conformity. The PlanAI-only optimization reached an initial solution in 3.3 ± 2.3 minutes, while the PlanAI + User workflow required 45.7 ± 25.3 minutes to finalize clinically acceptable plans (p<0.05). OARs’ DVH indicated PlanAI reduces low-dose spills to select OARs, particularly the esophagus, lungs, stomach, and pericardium, relative to clinical plans.
Conclusion
The PlanAI rapidly generates achievable objective sets, producing plans that meet clinical goals with quality comparable to clinically approved plans, while markedly reducing planning time. Refining objective protocols, expanding libraries, and incorporating blinded physician review will enhance clinical reliability. Thereby enabling physicians to leverage anatomy and dose constraints for predictive planning that reduces replanning, adaptive therapy and improves treatment.