This is an early access version, the complete PDF, HTML, and XML versions will be available soon.
Optimization of Automated Radiotherapy Planning for Head and Neck Cancers and Brain Tumors Using Knowledge-Based Planning Models
1
Lower Silesian Center for Oncology, Hematology and Pulmonology, 53-439 Wroclaw, Poland
2
1st Radiation and Clinical Oncology Department, Maria Sklodowska-Curie National Research Institute of Oncology, 44-102 Gliwice, Poland
3
Department of Radiotherapy Lower Silesian Oncology, Pulmonology and Hematology Center, 53-439 Wroclaw, Poland
4
Department of Oncology, Medical University, 53-413 Wroclaw, Poland
*
Author to whom correspondence should be addressed.
Cancers 2026, 18(14), 2216; https://doi.org/10.3390/cancers18142216
Submission received: 10 June 2026
/
Revised: 3 July 2026
/
Accepted: 7 July 2026
/
Published: 9 July 2026
(This article belongs to the Special Issue Next-Generation Automation in Radiation Oncology: Innovations in Planning, Informatics, and Clinical Workflow)
Simple Summary
Developing automated radiation treatment plans for complex head and neck cancers and brain tumors remains a major challenge due to anatomical variability and differences in clinical planning practice. This study evaluates a single institutional knowledge-based planning model designed to standardize and streamline treatment planning across connected radiotherapy sites. The findings show that conventional model-quality statistics do not always predict clinical usefulness. With careful curation of the training data, the automated system maintained target coverage while reducing doses to several organs at risk, including the oral cavity and larynx. The model should be interpreted as a planning standardization and decision-support tool that reduces manual variability, while final plan approval remains the responsibility of experienced physicists and radiation oncologists.
Abstract
Objectives: Our objective was to develop and evaluate a locally trained knowledge-based planning (KBP) model for head and neck (H&N), brain, and central nervous system malignancies using RapidPlan, and to determine whether standard statistical metrics such as the coefficient of determination (R2) and outlier frequency are definitive predictors of clinical utility. Methods: An institutional dataset of 594 plans was retrospectively curated into a 497-plan training set. Performance was evaluated in 370 paired plan comparisons generated with identical beam geometry. Training–validation overlap was explicitly quantified at both plan and patient levels, and a plan-level held-out sensitivity analysis was performed. Additional analyses included monitor units (MUs), subgroup assessment, clinically relevant OAR threshold achievement, and 95% confidence intervals for paired differences. Results: The validation set included 370 plans from 289 patients. At the plan level, 303 validation plans overlapped with the training model, and 67 were held-out cases; at the patient level, no fully patient-independent validation cohort was available. RapidPlan maintained target coverage while reducing OAR doses, including oral cavity Dmean (−7.62%Rx; 95% CI: −8.91 to −6.32; p < 0.001) and larynx Dmean (−7.57%Rx; 95% CI: −9.14 to −6.00; p < 0.001). The same direction of benefit was observed in the plan-level held-out subset. MU did not increase with RapidPlan and decreased from 764.2 ± 275.5 to 695.8 ± 210.3 MU (Delta = −68.4 MU; 95% CI: −89.4 to −47.4; p < 0.001). Conclusions: A high R2 was not required for clinically useful optimization objectives in this heterogeneous cohort. However, the retrospective design and patient-level overlap limit claims of full generalizability. The model should therefore be interpreted as a clinically useful standardization and decision-support tool requiring expert review rather than as a replacement for the judgment of physicists and radiation oncologists.
