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

Abstract

AbstractSingle-isocenter multiple-target (SIMT) stereotactic radiosurgery (SRS) has become an increasingly common treatment approach for patients with multiple intracranial lesions. While SIMT improves treatment efficiency, it also introduces planning challenges related to normal brain dose, particularly when multiple targets are spatially close. Excessive normal brain dose is associated with an increased risk of radiation-induced toxicity and often necessitates repeated plan revisions during clinical planning. In current practice, assessment of normal brain dose relies heavily on planner experience and iterative re-optimization, which can prolong planning time and introduce variability in plan quality. In this thesis, a machine learning–based framework was developed to predict normal-brain dosimetric metrics prior to plan optimization for LINAC-based SIMT SRS. A retrospective cohort of 231 SIMT SRS plans was collected, with 181 cases used for model training and 50 cases reserved for independent testing. Each case was characterized using geometric and statistical features describing prescription dose, target burden, and inter-target spatial relationships. Gradient Boosted Trees regression models were trained to predict three clinically relevant normal brain dose metrics: V50%, V60%, and V66.7%. Model performance was evaluated using mean absolute error, prediction uncertainty, and coefficient of determination. In addition, an Eclipse Scripting Application Programming Interface (ESAPI)–based tool was developed to enable model integration directly within the treatment planning system. The developed models demonstrated accurate and reliable prediction of normal brain dose metrics across a range of SIMT geometries. Predictions were generated within seconds, making the approach suitable for clinical workflow integration. This work demonstrated that machine learning–based dose prediction can serve as a practical decision-support tool during SIMT SRS planning, with the potential to improve planning efficiency and reduce variability in plan quality.