Artificial intelligence-assisted detection of challenging ischemic stroke on diffusion-weighted imaging: a reader study.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Seoul National University Bundang Hospital, Seoul National University College of Medicine, Seongnam, Republic of Korea.
- Artificial Intelligence Research Center, JLK Inc., Seoul, Republic of Korea.
- Department of Neurology, Seoul National University Bundang Hospital, Seoul National University College of Medicine, Seongnam, Republic of Korea.
- Center for Artificial Intelligence in Healthcare, Seoul National University Bundang Hospital, Seongnam, Republic of Korea.
Abstract
To evaluate the effect of artificial intelligence (AI) assistance on the diagnostic performance of human readers for detecting challenging acute ischemic stroke (AIS) lesions on diffusion-weighted MRI. This retrospective, single-center, randomized crossover study included 3,986 patients (mean age, 68 ± 14 years; 2,383 men) who underwent initial and follow-up diffusion-weighted imaging (DWI) between February 2017 and November 2021. From this cohort, 250 cases (130 challenging AIS, 120 controls) were selected for a multi-reader performance study. Five readers interpreted cases with and without AI assistance. Diagnostic performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and Dice similarity coefficient (DSC). AI achieved a sensitivity of 96.0% and identified 79.6% (43 of 54) of false-negative stroke cases from clinical reports. AI-assisted reading significantly improved AUC from 0.85 (95% CI: 0.82-0.90) to 0.93 (95% CI: 0.90-0.95; <i>p</i> < 0.01), pooled sensitivity from 74.6% (95% CI: 69.8-79.4) to 90.6% (95% CI: 87.4-93.7; <i>p</i> < 0.01), and lesion segmentation accuracy (DSC) from 0.523 to 0.742 (<i>p</i> < 0.01). Specificity slightly decreased from 88.8% (95% CI: 85.1-92.3) to 84.0% (95% CI: 78.5-89.5; <i>p</i> = 0.05). Reader confidence also improved with AI support, especially in challenging cases. AI assistance significantly improved diagnostic performance and lesion segmentation accuracy in detecting small and hyperacute AIS lesions on DWI.