Abstract WP305: Comparison Of Automated Brain Extraction Methods In Over 5000 Computed Tomography Scans

Introduction: Automated Brain Extraction Tool (BET) has been extensively studied for Magnetic Resonance Imaging (MRI), but significantly fewer tools exist for Computed Tomography (CT) scans. The performance of these limited tools, especially in stroke populations, remains largely unexplored. By removing skull, air, and neck voxels that skew intensities, skull stripping improves registration, volumetric quantification, and deep learning features. However, pathologies, surgical alterations, and acquisition artifacts can compromise reliability. To address these limitations, we developed Robust-CTBET for improved performance in challenging CT cases and benchmarked it alongside six existing algorithms. Methods: We developed Robust-CTBET using soft-tissue windowing to 0–100 HU, mild 3D smoothing (σ=1 mm), two-pass template-guided neck excision with rigid registration, re-centering by intensity centroid, and a CT-tuned BET (~0.1 fractional threshold) to yield a single brain component. We benchmarked it against six tools: HD-CTBET, CTbet-Docker, CTBET, Brainchop, SynthStrip, and classical CT_BET on 5,062 non-contrast MISTIE 3 scans, using task-oriented QC across registration, volumetrics, and deep learning. Failure was defined based on predefined task-specific criteria with detailed evaluation protocols. In addition to overall rates, failure was assessed in three subcategories: artifact-heavy (n=53), post-craniotomy (n=40), and Computed Tomography Angiography (CTA, n=10). Results: Overall: Robust-CTBET outperformed others with a 5× lower failure rate (0.3%) than CTbet-Docker (1.5%), followed by HD-CTBET at 5.2%. Classical CT_BET had the highest failure rate (31.7%). Artifact-heavy: Robust-CTBET maintained a low failure rate (3.8%); HD-CTBET and CTbet-Docker showed elevated volumetric failures (5.7%). Post-craniotomy: Only Robust-CTBET and HD-CTBET achieved 0% failure rate across all tasks. CTA: Robust-CTBET, HD-CTBET, CTbet-Docker, and SynthStrip achieved perfect performance; Brainchop had 10% volumetric failures. Conclusion: Multiple CT skull-stripping algorithms achieve high accuracy in diverse stroke imaging contexts, but no method is uniformly optimal. Robust-CTBET, developed in this study, demonstrated consistent performance across all tasks and subgroups. These results provide the first large-scale, task-specific benchmark for CT brain extraction, addressing a methodological gap in neuroimaging pipelines and guiding the selection of algorithms for clinical datasets.