JRSBRT 10.3, p. 175-185 (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.

A geometric evaluation of an artificial intelligence-based segmentation tool for brain metastases in stereotactic radiosurgery
Zhenghao Xiao, Patrick Hiepe, Yingxuan Chen, Wentao Wang, Brian Keith, Ana Mcwhinnie Fernandez, Rebecca Ljungqvist, Kevin Judy, James Evans, Christopher Farrell, David Andrews, Sophia Shah, Kelsey Carey, Wenyin Shi and Haisong Liu

Purpose: While automated segmentation based on artificial intelligence (AI) shows promise for brain metastases, volume-dependent performance patterns remain incompletely characterized. This study evaluates the performance of an AI-based auto-segmentation algorithm across clinically relevant tumor size ranges to help optimal clinical workflow integration.

Methods: We retrospectively analyzed 148 brain metastases from 34 patients treated with stereotactic radiosurgery (SRS). AI-generated contours on T₁ contrast-enhanced magnetic resonance images (MRIs) were compared to reference contours drawn by neurosurgeons then agreed by radiation oncologist. Geometric evaluation parameters include Dice similarity coefficient (DSC), Hausdorff distance, centroid distance (CD), and volume ratio (VR). Performance was stratified across four tumor volume categories (<0.05 cc, 0.05–0.2 cc, 0.2–1.0 cc, >1.0 cc) with focused analysis at 0.2 cc.

Results: For lesions ≥0.2 cc, 93.2% achieved DSC ≥0.7 compared to 51.7% for lesions <0.2 cc (p < 0.001). All lesions >1.0 cc achieved DSC ≥0.7. Median DSC improved with increasing volume: 0.63 (<0.05 cc), 0.79 (0.05–0.2 cc), 0.84 (0.2–1.0 cc), and 0.93 (>1.0 cc) (p < 0.001). Median CD remained consistent across all groups (range 0.30-0.35 mm, p = 0.298), demonstrating reliable spatial localization. AI-generated tumor volumes were slightly conservative (median VR: 0.79).

Conclusions: The AI auto-segmentation algorithm for brain metastases exhibits predictable, size-dependent performance patterns for T1 contrast-enhanced MRI lesions. The algorithm provides reliable segmentation for lesions ≥0.2 cc with minimal editing requirements and consistent spatial localization across all sizes. Smaller lesions benefit from algorithm-assisted detection with targeted physician refinement. These findings support stratified clinical implementation that leverages algorithm strengths while maintaining quality oversight, offering meaningful improved efficiency in SRS workflows.

Keywords: Elements, AI tumor segmentation, brain metastases, stereotactic radiosurgery, SRS