Atlas-Based Prediction of Plan Quality and Treatment Time for Gamma Knife Stereotactic Radiosurgery of Brain Metastases
Abstract
Purpose
To predict achievable Gamma Knife (GK) radiosurgery plan quality and treatment time through analysis of geometric features from a large atlas of previously treated brain metastases.
Methods
A retrospective dataset of 2,681 brain metastases from 989 patients treated between 2021 and 2025 on GK ICON (Elekta) using Leksell GammaPlan (v11.1.1-v11.4.2) was analyzed. Four geometric descriptors were extracted for each target: volume, sphericity, mean radial distance from target center to skull, and the ratio of minimum‑to‑maximum radial distances. Plan performance was quantified using six outputs: treatment time normalized against prescription dose and source strength, Paddick conformity index (PCI), number of shots, dosimetric margin, homogeneity index (HI), and gradient index (GI). A histogram gradient boosting classifier was trained to evaluate predictive relationships between target features and plan characteristics, using 5‑fold cross‑validation to assess model stability.
Results
Predictive performance varied by output. The model performed best for number of shots (RMSE = 3.88±0.36; R² = 0.760) and normalized treatment time (RMSE = 0.295±0.016 min/Gy; R² = 0.650). Moderate performance was observed for PCI (RMSE = 0.106±0.009; R² = 0.585) and HI (RMSE = 0.069±0.001; R² = 0.584). Predictability was limited for dosimetric margin (R² = 0.105) and GI (R² = 0.131). Permutation importance showed that volume was the dominant predictor across outputs, sphericity contributed modestly, and mean radial distance and the minimum‑to‑maximum radial distance ratio had minor effects.
Conclusion
Target geometry strongly influenced treatment time and number of shots in GK radiosurgery for brain metastases, while PCI and HI showed moderate dependence. These findings support the development of geometry‑based predictive tools that can help establish consensus plan quality and treatment time tradeoffs and inform treatment appointment scheduling for clinical practice. Future studies incorporating additional descriptors, such as distance to organs‑at‑risk, may improve the models’ accuracy.