BACKGROUND: Acute ischemic stroke is a time-sensitive medical emergency requiring rapid and accurate diagnosis to improve patient outcomes. While Diffusion-Weighted Imaging (DWI) on magnetic resonance imaging (MRI) is highly sensitive, artificial intelligence (AI) offers a potential solution to enhance diagnostic speed and accuracy. In this study we aimed to evaluate and externally validate a DWI-MRI-based deep learning model for automated stroke detection and compare its multicenter diagnostic performance with expert radiologists to determine generalizability and clinical utility. METHODS: This retrospective study involved 732 patient cases (acute ischemic stroke and non-stroke controls) from three different centers. A deep convolutional neural network (CNN) model was developed, trained, and internally validated using data from center 1 (n = 452 for training, n = 80 for validation). The model's generalizability was then tested using independent external validation datasets from center 2 (n = 100) and center 3 (n = 100). The model's diagnostic performance (sensitivity, specificity, accuracy, and area under the receiver operating characteristic (ROC) curve (AUC)) was systematically compared with that of three expert radiologists. RESULTS: The deep learning model demonstrated excellent diagnostic performance. In the internal validation, the model achieved 100% sensitivity, 100% specificity, and 100% accuracy (AUC = 1.000). Crucially, it maintained high performance on the external validation cohorts, achieving 100% sensitivity, 98% specificity, and 99% accuracy for both center 2 (AUC = 0.987) and center 3 (AUC = 0.986). This performance was comparable with the expert radiologists, who also achieved high accuracy across all datasets. Visualization techniques (gradient-weighted class activation map (Grad-CAM)) confirmed that the AI model focused on the correct pathological regions when making its classifications. CONCLUSIONS: The DWI-MRI-based deep learning model provides high and reliable diagnostic accuracy for acute ischemic stroke, with performance comparable with that of expert radiologists. Its robust performance across multicenter data highlights its potential as a dependable decision-support tool in emergency departments, especially in settings with limited specialist availability, to facilitate faster and more consistent stroke diagnoses.
Kuzan et al. (Tue,) studied this question.