JRSBRT 10.3, p. 165-174 (PDF)

After payment has been processed for your order of a digital copy (PDF) of this article, you will see a download link on your completed order page and also receive an email containing a download link. The links, which will enable you to download one copy of the article, will expire after 24 hours.

Development and evaluation of a unified machine-learning model for predicting normal brain toxicity in single isocenter multi-target stereotactic radiosurgery
Zhuoyun Huang, Jingtong Zhao, Ke Lu, John S Ginn, Zhenyu Yang, Yibo Xie, Yongbok Kim, Trey C Mullikin and Chunhao Wang

Purpose: To develop and evaluate a machine learning model for normal brain dose outcome in linear-accelerator-based single isocenter multiple target (SIMT) stereotactic radiosurgery (SRS).

Methods: SIMT SRS cases collected totaled 231, and 181/50 cases were randomly assigned for model training/testing. The model characterizes SIMT cases using seven geometric and statistical features describing prescription, target burden, and inter-target spatial relationships. Gradient boosted tree (GBT) regression was employed to predict three dosimetric outcomes: overall V_50%, V_60%, and V_66.7%. Five hyperparameters were optimized through grid search and 10-fold cross-validation.

Four models were trained: the first model (M_a,s) predicts all three outcomes as a single output, and the second model (M_a,m) uses three GBTs to report each outcome. In the third (M_c,s) and fourth (M_c,m) models, unsupervised k-means cluster partitioned training samples into geometrically distinct groups based on statistical features, and the models were trained using clustered samples for single or multiple outputs, respectively. Model performance was evaluated using normalized mean absolute error (MAE), mean prediction uncertainty (MPU), and R².

Results: Ma,s achieved the best overall performance, with an average MAE of 6.12 cm³, MPU of 7.81 cm³ and R² = 0.88. Clustering slightly improved relative error in M_c,m but increased absolute errors and reduced overall performance in M_c,s. Each prediction required under 5 s. An Eclipse Scripting Application Programming Interface script was developed to integrate the model into the EclipseTM treatment planning system.

Conclusions: The machine-learning framework was developed and validated for predicting SIMT SRS normal-brain dose. By providing accurate and reliable dosimetric predictions, the model supports efficient treatment planning and reduces plan quality variability.

Keywords: SRS, SIMT, machine learning, GBT, plan quality