2103 - Design and Evaluation of a Unified Machine Learning Model for Predicting Normal Brain Toxicity in Single-Isocenter-Multi-Target Stereotactic Radiosurgery

Purpose/Objective(s): This study presents a unified machine learning framework to predict three normal brain toxicity indicators, V50%, V60%, and V66.7%, in LINAC-based single-isocenter multi-target (SIMT) stereotactic radiosurgery (SRS). It addresses the need for rapid, multi-metric toxicity prediction to provide a reliable reference for fractionation selection and supports efficient high-quality SRS planning for complex cases.

Materials/Methods: This model was developed using 134 LINAC-based SIMT plans generated by a single experienced planner. These plans were divided into a training set of 117 plans (2–53 targets per plan, median = 6) and an independent test set of 27 plans (2–17 targets per plan, median = 6). The proposed model analyzes each SIMT plan using 6 statistical features: prescription level, target count, target volumes, target surface areas, target equivalent spherical surface areas, and target surface-to-surface distances. Gradient Boosted Trees (GBT) regression was adopted to generate the three indicators simultaneously, with the relative weights of the three indicators and other hyperparameters such as tree depth, learning rate, and feature splits optimized through grid search. A 10-fold cross-validation was adopted. For comparison purposes, three other GBT models, one for each normal brain toxicity indicator, were trained using the same input design. Within the test group, prediction performances were evaluated using the mean absolute difference (MAD), mean percentage difference (MPD), and mean prediction uncertainty (MPU) that was normalized to the total target volumes. Wilcoxon signed-rank test was adopted to examine the significance of potential differences.

Results: As shown in Table 1, the unified model outperformed individual models in terms of MAD for both V50% and V66.7%, with slightly worse results for V60% (p = 0.990). However, the unified model demonstrated improvements across all three normal brain toxicity indicators, achieving significantly lower MPD (p<0.001) and reduced MPU, indicating its accuracy and robustness. Notably, hyperparameter analysis revealed relatively smaller weights (~0.50) of both V50% and V66.7% in reference to V60% within the model penalty, suggesting the strong correlation among three indicators. The unified prediction model took less than one second to predict each SIMT case, confirming the model’s readiness for clinical use.

Conclusion: The unified model offers an efficient and accurate approach for predicting normal brain toxicity in SIMT SRS, serving as a decision aid for toxicity assessment and improving SRS planning efficiency. Future studies using larger, multi-institutional datasets will be crucial to further enhance model accuracy and evaluate the impact of different clinical practices on performance.

Abstract 2103 - Table 1