AI assisted detection of large vessel occlusion on non-contrast CT: multinational validation and reader study

Area of Science:

• Radiology
• Artificial Intelligence
• Neurology

Background:

• Early detection of large vessel occlusion (LVO) is crucial for stroke triage.
• Current artificial intelligence (AI) algorithms for LVO detection on non-contrast CT (NCCT) require multinational validation and evidence of clinical impact.

Purpose of the Study:

• To validate a machine learning-based LVO detection algorithm across diverse international patient cohorts.
• To evaluate the impact of AI assistance on clinician diagnostic performance in stroke triage.

Main Methods:

• Retrospective validation of an AI algorithm using independent Korean and US cohorts.
• Assessment of standalone AI performance using AUC, sensitivity, and specificity.
• Evaluation of AI's impact on clinician diagnostic performance through a multi-reader, multi-case crossover study.

Main Results:

• The AI algorithm demonstrated high standalone performance (AUC 0.963 in Korea, 0.899 in US).
• AI assistance significantly improved pooled AUC (0.718 to 0.852), sensitivity (46.6% to 63.7%), and specificity (91.9% to 94.9%).
• AI provided a net reclassification improvement (NRI) of 5.5% and a benefit-to-harm ratio (BHR) of 2.89.

Conclusions:

• Multinational validation confirmed the robustness and generalizability of the AI algorithm for LVO detection on NCCT.
• AI assistance substantially enhanced clinician sensitivity and diagnostic accuracy.
• The findings suggest AI's significant potential as a supportive tool in acute stroke triage.