Comparing tumor burden-based classification with traditional methods for knowledge-based VMAT planning in multiple brain metastases

Appl Radiat Isot. 2026 Jun 4;236:112749. doi: 10.1016/j.apradiso.2026.112749. Online ahead of print.

ABSTRACT

PURPOSE: This study systematically compares the dose prediction performance of three knowledge-based RapidPlan models-Broad-Spectrum Model (Model-BS), K-means Clustering-Based Classification Model (Model-CA), and Volume Stratification Model (Model-V)-in volumetric modulated arc therapy (VMAT) planning for multiple brain metastases (multi-BM).

METHODS: A total of 138 multi-BM patients (with 2 to 16 lesions) treated between 2021 and 2024 were retrospectively included. Among these, 122 patients were randomly selected to construct Model-BS (general knowledge-based), Model-CA (unsupervised clustering based on lesion count and total volume features), and Model-V (volume stratification based on total volume thresholds). The remaining 16 patients were designated as the test group and received automated plans (BS-plans, CA-plans, and V-plans) as well as manual plans (MP-plans). All plans used single-isocenter technique, with a prescription dose of 30 Gy/5Fr. Dose metric parameters, including conformity index (CI), gradient index (GI), and organ-at-risk (OAR) doses, were compared.

RESULTS: All knowledge-based plans (RP-plans) showed comparable or slightly improved target coverage compared to MP-plans. Compared to Model-BS, both Model-CA and Model-V significantly reduced high-dose volumes (V₂₄ Gy-V₁₂ Gy) in the Brain and normal brain tissue (Brain-PTV) (p < 0.05). Model-V plans had the lowest gradient index (GI = 5.55), significantly outperforming CA-plans (GI = 5.81) and BS-plans (GI = 5.61); however, the difference compared to MP-plans (GI = 5.89) did not reach statistical significance (p = 0.134). Model-V also demonstrated the highest conformity index (CI = 0.924). However, Model-V was slightly less effective in protecting some OARs, such as the optic nerve Dmax, compared to Model-CA.

CONCLUSION: The classification knowledge-based model, built on tumor lesion count and total volume, significantly improves the dose distribution in multi-BM VMAT planning. Among the models, Model-V achieved the optimal dose gradient (GI), while Model-CA offered the best overall balance between target conformity, dose gradient fall-off, and OAR sparing.

PMID:42258912 | DOI:10.1016/j.apradiso.2026.112749