Abstract Body: Background
FDA clearance of fully automated artificial intelligence (AI)-based software for quantifying intracerebral hemorrhage (ICH) volumes has the potential to meaningfully impact the acute management of hemorrhagic stroke. ICH volume is a critical prognostic factor, with larger hemorrhages associated with oral anticoagulant (OAC) use typically resulting in poorer outcomes. Quantifying ICH volume in OAC-related ICH presents challenges due to variability of morphology and density. Although prior studies suggest AI models may improve the accuracy of ICH volume calculation compared to the ABC/2 method and Semi-Autonomous Segmentation (SAS), there is limited data evaluating their performance in OAC-related ICH.

Methods
A retrospective analysis was conducted on 161 adults presenting to a comprehensive stroke center from 2016-2021 with acute supratentorial OAC-related ICH. Volumes on initial de-identified CT brain were measured using SAS (Syngo.Via, Siemens) and AI (Viz ICH volume, Viz.ai). Agreement of ICH volumes between SAS and AI was assessed using the Intraclass Correlation Coefficient (ICC) for absolute agreement. A two-way mixed effects model was employed for single measurements. A Bland-Altman (BA) analysis with proportional bias assessment was performed. Data analysis was conducted using R studio.

Results
Out of 161 adults, 50 met eligibility criteria and 39 (78%) CT scans were analyzed by AI and SAS. AI software failed to correctly process 11 scans due to small ICH volumes (n=8), misclassification of lesions as subdural (n=1), or image retrieval issues (n=2). For the 39 scans that were analyzed, the median ICH volume measured by SAS was 15.89 cm³ (IQR 5.69 – 41.86 cm³) and by AI was 17.0 cm³ (IQR 5.0 – 44.5 cm³). The ICC for absolute agreement between the software platforms was 0.973 (95% CI 0.950 – 0.986), indicating excellent reliability. A BA plot revealed a mean difference (bias) of -0.861 cm³ (95% CI -3.1 – 1.37 cm³) with limits of agreement from -14.84 cm³ to 13.12 cm³, demonstrating good agreement between the two methods with no significant proportional bias.

Conclusions
There is strong agreement and reliability in OAC-related ICH volume measurements between SAS and AI. Such local validation is imperative for safe and responsible integration of AI tools into clinical workflows. Further research into limitations of AI, including failure modes and biases is necessary to inform human oversight.