Abstract 378: Diagnostic Accuracy of Artificial Intelligence for Intracranial Aneurysm Detection on CT Angiography: A Systematic Review and Meta‐Analysis of Prospective and Retrospective Studies.

Abstract

Introduction Intracranial aneurysms (IAs) are a major cause of subarachnoid hemorrhage, with high morbidity and mortality if undiagnosed. Computed tomography angiography (CTA) remains the first‐line imaging modality, but interpretation is challenging and error‐prone, particularly for small or complex aneurysms. Artificial intelligence (AI), particularly deep learning models, has emerged as a promising adjunct to improve diagnostic accuracy and efficiency. While several studies have evaluated AI in aneurysm detection, their designs and results vary. We conducted a systematic review and meta‐analysis to synthesize evidence on the diagnostic accuracy of AI for IA detection on CTA across retrospective and prospective studies. Methods A systematic search identified meta‐analyses, multicenter retrospective validations, and prospective clinical trials evaluating AI‐based detection of IAs on CTA. Eligible studies included both standalone AI performance and AI‐assisted clinician performance, with reference standards from digital subtraction angiography (DSA) or expert consensus. Data on sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were extracted. A bivariate model was applied to estimate pooled diagnostic accuracy, and subgroup analyses compared retrospective versus prospective designs. Results Twenty eligible studies, encompassing over 20,000 patients, were analyzed. At the patient level, pooled sensitivity and specificity of AI algorithms were 0.92 (95% CI: 0.85‐0.96) and 0.96 (95% CI: 0.94‐0.97), respectively. Lesion‐level pooled sensitivity was similarly high at 0.92 (95% CI: 0.87‐0.95). Meta‐analysis of reader‐assistive designs showed that AI increased clinician sensitivity by 9% (risk ratio 1.09, p < 0.001) without affecting specificity, and reduced mean reading time by 7.4 seconds per case. Prospective multicenter validation confirmed high performance: Hu et al. reported sensitivity of 0.957 and improved AUC from 0.787 to 0.909 with AI assistance, while Wei et al. demonstrated comparable AUCs between AI and radiology reports (0.93 vs. 0.91, p = 0.67). Advanced segmentation networks, such as VA‐Unet, achieved sensitivity above 96% and Dice coefficients up to 0.78, enabling robust morphological characterization. False‐positive rates varied, with some models averaging 2‐3 per case, highlighting workflow integration challenges. Conclusion AI demonstrates strong diagnostic accuracy for IA detection on CTA, with pooled sensitivities above 90% and AUCs approaching 0.95. Both retrospective and prospective studies confirm that AI enhances clinician sensitivity, accelerates interpretation, and maintains specificity. Real‐world validation underscores its clinical readiness, though false positives and dataset heterogeneity remain barriers. Standardized evaluation frameworks and multicenter prospective trials are warranted to ensure safe, scalable integration into routine neuroradiological practice.

Ähnliche Arbeiten