Abstract WP306: Comparative Efficacy of Traditional Algorithms and Deep Learning Models for Stroke Detection Using MRI Imaging: A Systematic Review and Meta-Analysis

Abstract

Introduction: Stroke remains a leading cause of morbidity and mortality globally, with a 23% relative annual increase in incidence worldwide and a staggering 87% rise in the United States alone. Magnetic Resonance Imaging (MRI) is the most widely used modality for stroke diagnosis; however, it is susceptible to missed diagnoses due to factors such as operator dependency and subtle imaging features. Artificial intelligence (AI)-based models, particularly Traditional Algorithms (TA) and Deep Learning (DL) techniques, offer significant promise in improving diagnostic accuracy by enhancing the detection and prediction of stroke from MRI images. These AI models can help address the challenges of timely and accurate stroke diagnosis. Methods: A systematic search was conducted on SCOPUS, Embase, PubMed, CENTRAL, and IEEE Xplore up to August 2025. The risk of bias was assessed using QUADAS 2.0. Results: A total of 592 studies were screened, with 10 meeting the inclusion criteria, involving 16,449 patients (13,888 using Traditional Algorithms and 2,561 using Deep Learning Models). The performance of Traditional Algorithms and Deep Learning Models in stroke detection was compared across multiple metrics. For accuracy, Traditional Algorithms achieved 88.94% (95% CI: 84.97–94.90), while Deep Learning Models performed slightly better with an accuracy of 88.98% (95% CI: 80.90–97.08). In terms of precision, Traditional Algorithms showed a value of 88.12% (95% CI: 85.73), whereas Deep Learning Models exhibited a lower precision of 77.34% (95% CI: 65.66–89.03). Regarding prediction capability, Traditional Algorithms had a prediction rate of 87.20% (95% CI: 81.52–92.89), surpassing the Deep Learning Models at 83.86% (95% CI: 81.47–86.24). Conclusion: Both Traditional Algorithms and Deep Learning Models demonstrate high accuracy in stroke detection using MRI, with Traditional Algorithms outperforming in precision and prediction. Future research should focus on optimizing Deep Learning Models to address these performance gaps. Risk of Bias was low.

Ähnliche Arbeiten