J Med Internet Res. 2026 Sep 21;28:e95041. doi: 10.2196/95041.
ABSTRACT
BACKGROUND: Traditionally, the number and location of cerebral microbleeds (CMBs) are manually calculated based on magnetic resonance imaging (MRI) characteristics such as shape, size, and signal features. Although accurate, manual detection requires expert interpretation and is costly. Therefore, it is necessary to explore an effective auxiliary detection method. In recent years, deep learning (DL) has been increasingly used in the detection of cerebral hemorrhage. Some studies have explored image-based DL models for diagnosing CMBs. Nevertheless, systematic evidence regarding their diagnostic accuracy is lacking.
OBJECTIVE: This review aimed to assess the accuracy of DL models in detecting CMBs and inform the development of intelligent detection tools.
METHODS: This study was reported in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines and was prospectively registered in PROSPERO (registration ID CRD42024628447). IEEE, Web of Science, Embase, the Cochrane Library, and PubMed were comprehensively searched up to November 1, 2024, and the database search was subsequently updated on July 5, 2026, to collect publicly published original studies on DL for detecting CMBs. The risk of bias of eligible studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 tool. Subgroup analyses were performed according to the level of analysis (lesion level and patient level) and the method of obtaining the diagnostic 4-fold table at the lesion level (direct extraction and reconstruction).
RESULTS: At the patient level, 5 studies were included, all of which developed models based on MRI. The meta-analysis results suggested that the sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), and diagnostic odds ratio (DOR) were 0.89 (95% CI 0.76-0.96), 0.86 (95% CI 0.77-0.92), 6.3 (95% CI 3.5-11.6), 0.13 (95% CI 0.05-0.32), and 50 (95% CI 12-212), respectively. At the lesion level, the sensitivity, specificity, PLR, NLR, and DOR were 0.96 (95% CI 0.93-0.98), 0.98 (95% CI 0.94-0.99), 39.5 (95% CI 16.8-92.7), 0.04 (95% CI 0.02-0.07), and 1061 (95% CI 293-3848), respectively. In the subgroup of direct extraction of the diagnostic 4-fold tables at the lesion level, the sensitivity, specificity, PLR, NLR, and DOR were 0.98 (95% CI 0.95-0.99), 0.98 (95% CI 0.90-0.99), 40.2 (95% CI 9.3-79.9), 0.02 (95% CI 0.01-0.05), and 1738 (95% CI 174-17,385), respectively. In the subgroup of reconstruction of the diagnostic 4-fold tables at the lesion level, the sensitivity, specificity, PLR, NLR, and DOR were 0.95 (95% CI 0.89-0.97), 0.98 (95% CI 0.96-0.99), 41.2 (95% CI 21.2-79.9), 0.06 (95% CI 0.03-0.12), and 736 (95% CI 224-2418), respectively.
CONCLUSIONS: DL models based on MRI appear to show favorable diagnostic performance in detecting CMBs. Given the small number of included studies, more multicenter studies are warranted to facilitate the development of more generalizable detection tools.
TRIAL REGISTRATION: PROSPERO CRD42024628447; https://www.crd.york.ac.uk/PROSPERO/view/CRD42024628447.
PMID:42767630 | DOI:10.2196/95041