Shifting the paradigm in intracranial aneurysm detection with deep learning: A diagnostic accuracy meta-analysis and meta-regression.

Background deep learning has emerged as a promising approach for detecting intracranial aneurysms across multiple imaging modalities. To clarify its diagnostic value, a systematic review and meta-analysis was undertaken. Methods the review adhered to PRISMA standards and was prospectively registered in PROSPERO (CRD420251151933). Eligible studies included original peer-reviewed investigations reporting the diagnostic accuracy of deep learning systems, alone or in combination with radiologists, for the detection of intracranial aneurysms on CTA, MRA, or DSA. Both ruptured and unruptured lesions were considered. Data extraction followed a standardized protocol, and risk of bias was appraised using the QUADAS framework. Pooled estimates were generated under random-effects models, with heterogeneity explored through subgroup analyses and meta-regression. Results thirty-four studies comprising 49,748 patients and 6,751 aneurysms met inclusion criteria. The pooled mean age was 57.5 ± 1.9 years, and the average aneurysm size was 5.0 ± 0.6 mm. The sROC model demonstrated a sensitivity of 89% [95%CI: 86%–92%], specificity of 89% [95%CI: 84%–93%], and an AUC of 0.95 [95%CI: 0.93–0.96]. The diagnostic odds ratio (DOR) was 69 [95%CI: 37–129], with LR + 8.5 and LR– 0.12. Subgroup analysis showed the highest accuracy when deep learning was combined with radiologists (sensitivity 92%, specificity 95%, DOR 250) and when DSA was used as the imaging modality (sensitivity 91%, specificity 93%, DOR 153). Meta-regression identified patient age as a significant covariate for specificity (p = 0.007). Conclusion deep learning demonstrates high accuracy for aneurysm detection, particularly when integrated with radiologists and trained on DSA.