Diagnostic Accuracy of Machine and Deep Learning Methods in Intracranial Hemorrhage Detection: A Systematic Review and Meta-Analysis.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, University Medical Center Groningen, University of Groningen, 9700 AB Groningen, The Netherlands.
- Department of Neurology, University Medical Center Groningen, University of Groningen, 9700 AB Groningen, The Netherlands.
- Department of Imaging, Radboud University Medical Center, 6500 HB Nijmegen, The Netherlands.
- Department of Nuclear Medicine and Molecular Imaging, University Medical Center Groningen, University of Groningen, 9700 AB Groningen, The Netherlands.
- Department of Epidemiology, University Medical Center Groningen, University of Groningen, 9700 AB Groningen, The Netherlands.
Abstract
<b>Background/Objectives</b>: In this systematic review we investigated the diagnostic accuracy of machine and deep learning algorithms for the detection of intracranial hemorrhage (ICH) and its subtypes such as subarachnoid, intraventricular, intraparenchymal, subdural and epidural hemorrhage. In addition, we investigated their generalizability by comparing methods tested on private vs. public datasets, and FDA- vs. non-FDA-approved algorithms. <b>Methods</b>: PubMed, Scopus and Embase were searched between January 2010 and December 2025. Eligible studies included patients with acute neurological deficit that received non-contrast CT head scans and implemented machine and deep learning algorithms for ICH detection, that were tested on private and/or public datasets, received FDA approval or not, and used clinicians' evaluation as reference standard. Pooled sensitivity and specificity were incorporated for meta-analysis. <b>Results</b>: A total of 89 publications received full assessment and 51 studies were included in the meta-analysis. The pooled sensitivity and specificity for ICH detection were 94.5% (95% CI: 92.4-96.0) and 95.3% (95% CI: 93.1-96.9), respectively. For AI methods tested on private and public datasets, the pooled sensitivity reached 94.4% (95% CI: 91.2-96.5) and 94.5% (95% CI: 91.8-96.4), with a pooled specificity in 95.6% (95% CI: 92.9-97.2) and 95.1% (95% CI: 90.1-97.5), respectively. FDA- and non-FDA-approved AI algorithms showed a pooled sensitivity of 90.7% (95% CI: 87.1-93.4) and specificity of 95.8% (95% CI: 90.3-98.3) in the former group, and 95.5% (95% CI: 93.4-96.9) and 95.7% (95% CI: 92.4-97.6) for the latter group. <b>Conclusions</b>: Study results demonstrated that the diagnostic accuracy of the developed AI algorithms for ICH detection is equivalent to that of experienced clinicians. Methods tested on private and public datasets, and FDA- vs. non-FDA-approved, differ slightly in ICH diagnosis, showing encouraging signs for their incorporation in decision support systems in stroke patient pathways.