Deep Learning Models for Intracranial Aneurysms under 3mm: A Meta-Analysis

Introduction Intracranial aneurysms (IAs) smaller than 3 mm represent a persistent diagnostic challenge. Despite advances in CT angiography (CTA) and digital subtraction angiography (DSA), these lesions often escape detection due to their subtle morphology and image noise. Deep learning (DL) algorithms have shown high overall diagnostic accuracy for IAs, but their performance in the small aneurysm subgroup is less clear. Accurate detection of these lesions is crucial, as even small aneurysms carry rupture risk. This meta‐analysis evaluates DL models’ diagnostic accuracy in identifying IAs <3 mm. Methods A systematic review of meta‐analyses and multicenter trials was performed, extracting subgroup performance metrics for aneurysms <3 mm. Eligible studies included both retrospective and prospective cohorts validated against DSA or expert consensus. Data on sensitivity, specificity, and AUC were collected. A pooled analysis was conducted using random‐effects modeling to compare overall diagnostic accuracy versus small aneurysm detection performance. Results Across four representative studies (Gu et al. 2022; Din et al. 2023; Hu et al. 2024; You et al. 2024), overall diagnostic performance of DL models was strong, with pooled sensitivity 0.93 (95% CI: 0.89‐0.96) and specificity 0.95 (95% CI: 0.92‐0.97). However, sensitivity for aneurysms <3 mm declined substantially, ranging from 0.71 to 0.82 across studies (pooled estimate: 0.76, 95% CI: 0.70‐0.81). While large aneurysms were detected with >95% sensitivity, false negatives were disproportionately concentrated among <3 mm lesions. Models such as VA‐Unet demonstrated recall of only ∼74% for <3 mm aneurysms compared with >96% for larger lesions. Moreover, false‐positive rates were higher in this subgroup, with an average of 2‐3 per case, reflecting overcalling of vascular irregularities. Prospective trials confirmed these limitations: Hu et al. (2024) observed clinician‐AI combined AUC improvement (0.909 vs. 0.787 without AI), but sensitivity for <3 mm aneurysms remained inferior. Subgroup analysis also revealed that bifurcation aneurysms and stenotic vessels further reduced model performance, compounding the challenge. Conclusion Deep learning models achieve high overall accuracy for intracranial aneurysm detection on CTA; however, performance diminishes significantly for aneurysms <3 mm. Sensitivity drops to ∼75%, with higher false‐positive rates, underscoring the “small aneurysm problem.” While AI enhances clinician performance and reduces reading times, its limitations in detecting small lesions highlight the need for targeted algorithmic refinement. Future directions should prioritize balanced training datasets enriched with small aneurysm cases, multimodal imaging fusion, and adaptive thresholding to improve sensitivity without inflating false positives. Addressing this gap is critical to ensure that AI tools can reliably aid in early detection and management of high‐risk but subtle lesions.