Performance of Automated Algorithm in Large and Medium Vessel Occlusion Detection: A Real-World Experience

Aakanksha Sriwastwa ¹, Yasmin N Aziz ², Kara Weiss ³

⁰From the Department of Radiology (A.S., L.L.-L.W., A.S.V.), University of Cincinnati Medical Center, Cincinnati, Ohio sriwasaa@ucmail.uc.edu.

Ajnr. American Journal of Neuroradiology +

|

September 26, 2024

View abstract on PubMed

Summary

An AI algorithm accurately detected anterior large vessel occlusions (LVO) in stroke patients but showed lower performance for posterior LVO and medium vessel occlusions (MeVO). Awareness of these limitations is key for timely stroke diagnosis and triage.

Area Of Science

Neuroscience
Medical Imaging Analysis
Artificial Intelligence in Medicine

Background

Rapid detection of large (LVO) and medium vessel occlusions (MeVO) is critical for acute ischemic stroke management.
Commercial AI software for automated LVO detection is available, but its performance across different occlusion types and expanding mechanical thrombectomy indications requires further understanding.

Purpose Of The Study

To evaluate the performance of a commercial, fully automated AI algorithm (Viz.ai) for detecting LVO and MeVO in code stroke patients.
To assess the algorithm's accuracy in identifying anterior LVO (aLVO), posterior LVO (pLVO), and MeVO.

Main Methods

Retrospective analysis of consecutive code stroke patients (March 2020-February 2023) with CTA processed by Viz.ai.
Comparison of AI detection against a reference standard of radiology reports from 12 board-certified radiologists, adhering to STARD guidelines.
Primary outcome: accuracy for aLVO; Secondary outcomes: accuracy for all LVO, aLVO with M2-MCA occlusion, and aLVO with all MeVO.

Main Results

Of 3576 technically sufficient patients, 616 (17.2%) had vessel occlusions.
The AI algorithm achieved high sensitivity and specificity for anterior LVO (aLVO): 91% and 93%, respectively.
Performance was lower for all LVO (73% sensitivity), aLVO with M2-MCA occlusion (74% sensitivity), and aLVO with all MeVO (65% sensitivity), while specificity remained high (≥92%) across these categories.

Conclusions

The automated AI algorithm demonstrates high accuracy for anterior LVO detection but has limitations in identifying posterior LVO and MeVO.
Healthcare teams must recognize the discrepancies between AI findings and actual LVO/MeVO rates for optimal patient triage and timely intervention.