Real-world evaluation of the accuracy of the Viz.AI automated intracranial hemorrhage volume calculation tool

Background Appropriate management of spontaneous intracerebral hemorrhage (ICH) and intraventricular hemorrhage (IVH) requires rapid, accurate volume estimation. Viz.AI has developed an artificial intelligence (AI)-powered ICH calculation tool that may improve existing methods. Methods Adult patients presenting to a large healthcare system between December 2015 and December 2021 with spontaneous ICH greater than 10mL and within 72 hours since ictus were analyzed for hematoma volume. mABC/2 (modified ABC/2) was measured by a board-certified neurosurgeon. Semi-autonomous segmentation (SAS) was performed by a trained medical student on 3D Slicer, adjudicated by a board-certified neurosurgeon and used as a surrogate ground-truth volume. Results 139 CTs met inclusion criteria. Mean ICH volume via ground-truth SAS was 47.69±27.19 mL. Mean ICH volume difference between SAS and AI and SAS and mABC/2 was 4.77±4.06 mL and 8.36±9.48 mL, respectively (p<0.01). Bland-Altman plots yielded AI and SAS limits of agreement between −4.45 and 13.18 mL, and mABC/2 and SAS limits of agreement between −21.35 and 27.02 mL. The average absolute difference between IVH volume yielded by SAS and AI was 3.26±3.55 mL. Bland-Altman plots yielded IVH volume limits of agreement between −7.48 mLand 10.47 mL. AI was 94.6% sensitive and 94.0% specific for detecting IVH in the presence of ICH. The average time-to-volume for SAS and AI was 424±208 and 151±49.7 s (P<0.01), respectively. Conclusion Viz.AI is more accurate than mABC/2, and more rapid than SAS. The combination of speed and accuracy makes Viz.AI viable for clinical decision-making and clinical trial use.