Abstract WP253: Deep Learning for the Detection of Ischemic Stroke within 24 Hours Using Non-Contrast Computed Tomography: A Systematic Review and Meta-Analysis of Diagnostic Accuracy

Abstract

Introduction: Early diagnosis of acute ischemic stroke (AIS) is vital to improve prognosis and reduce mortality. While diffusion-weighted MRI is the gold standard, non-contrast CT (NCCT) is more accessible and enables faster assessment. The Alberta Stroke Program Early Computed Tomography Score (ASPECTS) scoring system facilitates standardized evaluation of ischemic lesions on NCCT. Deep learning (DL) holds promise for improving early AIS diagnosis by augmenting NCCT diagnostic accuracy. Objective: To evaluate the diagnostic performance of DL models in predicting the ASPECTS score and diagnosing AIS using NCCT images. Methods: This study followed PRISMA guidelines and was registered in PROSPERO (CRD420251000160). PubMed, Embase, and Scopus screened from inception to 19 May 2025. Studies included that diagnosed AIS using NCCT, with confirmation by a ground truth defined as DWI, CTA, CTP. The ground truth was used as a comparator to evaluate the performance of the DL model. The primary outcomes were HSROC curves summarizing the diagnostic performance of DL models in predicting ASPECTS scores and diagnosing AIS. As secondary outcomes, pooled sensitivity, specificity, and AUC values were estimated for ASPECTS prediction and AIS diagnosis using DL models. Results: Out of 265 screened studies, 20 met the inclusion criteria, of which 14 were eligible for meta-analysis. A total of five studies involving 3,799 patients contributed to the analysis of ASPECTS score prediction, while nine studies with 11,297 patients were included in the evaluation of AIS diagnosis prediction. The summary point of the HSROC curve corresponding to ASPECTS score prediction revealed a sensitivity of 86.7% (95% CI: 81.1–90.8), and specificity of 86.9% (95% CI: 72.7–94.3, Figure 1). In terms of diagnosing AIS, the HSROC summary point indicated a sensitivity and specificity of 80.9% (95% CI: 67.8–89.5) and 89.9% (95% CI: 81.1–94.9), respectively (Figure 1). Figures 2 and 3 present the pairwise meta-analysis results of ASPECTS prediction and AIS diagnosis, reporting sensitivity, specificity, accuracy, and pooled AUCs of 0.85 (95% CI: 0.79–0.91) and 0.85 (95% CI: 0.78–0.93), respectively. Conclusion: Deep learning models demonstrate high diagnostic accuracy in predicting ASPECTS scores and detecting acute ischemic stroke on non-contrast CT within 24 hours of onset. These findings suggest that DL may improve early AIS diagnosis, though further validation is required for clinical use.

Ähnliche Arbeiten