Applications of deep learning in intracranial aneurysm imaging: A scoping review of detection, risk prediction, and...

Background: Intracranial aneurysms (IAs) pose a significant risk of rupture and subarachnoid hemorrhage, necessitating early, accurate detection and risk stratification. With advances in artificial intelligence, deep learning (DL) has emerged as a transformative tool in neurovascular imaging. However, the clinical translation of DL applications remains constrained by variability in model design, data sources, and validation strategies. The aim of the present study was to systematically map and evaluate the landscape of DL applications in the detection, segmentation, risk prediction, and outcome assessment of IAs, with attention to methodological rigor, clinical utility, and translational limitations. Methods: We conducted a scoping review of studies indexed in PubMed, Scopus, and Web of Science up to August 2023, following PRISMA-ScR guidelines. Eligible studies employed DL algorithms for IA-related diagnostic or prognostic tasks using radiological imaging. Data extraction included model architecture, imaging modality, validation strategy, performance metrics, and thematic focus. Study quality was assessed using the Joanna Briggs Institute (JBI) critical appraisal tools. Results: Forty-two studies met the inclusion criteria, encompassing over 10,000 patients across diverse imaging platforms and DL architectures. Convolutional neural networks (CNNs) were the most commonly used models, with reported sensitivities ranging from 73% to 99% and AUCs frequently exceeding 0.85. Despite promising results in IA detection and rupture risk prediction, only a minority of studies conducted external validation or addressed post-treatment outcomes. Major gaps include a lack of benchmarking across models, limited explainability, and regulatory or ethical frameworks. Conclusion: DL algorithms demonstrate strong diagnostic and predictive performance in IA imaging but face critical barriers to clinical integration, including interpretability challenges, dataset heterogeneity, and limited generalizability. Future research should prioritize multicenter validation, explainable AI techniques, and outcome-focused modeling to advance safe and effective deployment in neurosurgical care.