What question did this study set out to answer?

To compare the volumetric accuracy of AI-assisted methods versus fully automated techniques in quantifying hemorrhage volumes in spontaneous intracerebral hemorrhage.

April 27, 2026Open Access

AI-assisted versus fully automated volumetry for hemorrhage quantification in spontaneous ICH with ventricular extension: A multi-center study

Key Points

To compare the volumetric accuracy of AI-assisted methods versus fully automated techniques in quantifying hemorrhage volumes in spontaneous intracerebral hemorrhage.
Retrospective multicenter diagnostic accuracy study involving 189 sICH patients with IVH from seven institutions.
CT scans analyzed using two AI platforms, Vendor A and Vendor B, against a reference standard established through AI-assisted manual correction by neuroradiologists.
Agreement assessed using intraclass correlation coefficients and Bland-Altman analysis with predefined clinical thresholds.
Both AI systems had high ICC for IPH and good-to-excellent ICC for IVH compared to the reference standard.
Bland-Altman analysis showed significant individual errors exceeding clinical thresholds for both AI platforms.
AI-assisted manual correction achieved ICC >0.99 for both compartments, completing corrections in approximately one minute.

Abstract

Background: Spontaneous intracerebral hemorrhage (sICH) with intraventricular hemorrhage (IVH) extension is a neurological emergency associated with high mortality, where separate quantification of intraparenchymal hemorrhage (IPH) and IVH volumes is essential for risk stratification and treatment decisions.While commercial artificial intelligence (AI) tools increasingly promise to automate this task, head-to-head comparisons of their separate volumetric accuracy and the clinical utility of AI-assisted correction workflows remain poorly characterized.Methods: In this retrospective multicenter diagnostic accuracy study, 189 sICH patients with IVH extension from seven institutions in China were included.Non-contrast CT scans were analyzed by two commercial AI platforms: Vendor A (uAI-HematomaCare, U-Net-based) and Vendor B (Strokedoc, Trans-UNet-based).A blinded AI-assisted manual correction workflow performed by two senior neuroradiologists established the reference standard.Agreement was assessed using intraclass correlation coefficients (ICC) and Bland-Altman analysis, with predefined clinical thresholds of 6 mL for IPH and 2 mL for IVH.Processing times were compared.Results: Both AI platforms achieved excellent ICC for IPH (Vendor A: 0.979; Vendor B: 0.991) and good-toexcellent ICC for IVH (Vendor A: 0.855; Vendor B: 0.935) versus the reference standard.However, Bland-Altman analysis revealed that 95% limits of agreement for both AI systems exceeded predefined clinical thresholds for both compartments (IPH: Vendor A -9.96-9.08 mL; Vendor B -6.84-5.04mL; IVH: Vendor A -7.02-7.52 mL; Vendor B -5.18-4.39mL), indicating clinically significant individual-level errors.In contrast, the AI-assisted manual correction workflow achieved near-perfect inter-rater reproducibility (ICC >0.99 for both compartments) with 95% limits of agreement entirely within acceptable thresholds, completing corrections in approximately one minute.Automated processing was 65-75% faster than manual correction.Conclusions: Fully automated AI volumetry for sICH with IVH demonstrates high group-level correlation but may produce individual errors exceeding clinically acceptable ranges for treatment decisions.An 'AI-assisted human correction' collaborative model achieves clinical-grade accuracy within approximately one minute and represents the optimal current practice pathway for integrating AI into acute stroke precise volume measurement.

Mark Helpful

Bookmark

Relay

View Full Paper