Do Deep Learning Algorithms Accurately Segment Intracerebral Hemorrhages on Noncontrast Computed Tomography? A Systematic Review and Meta-Analysis.

Background: Stroke, a major global health issue, is broadly categorized into ischemic and hemorrhagic types. The volume of hemorrhage on noncontrast computed tomography guides the treatment options and hints at prognosis. Conventional approaches to calculate intracerebral hemorrhage (ICH) volume, like the ABC/2 method, typically rely on an assumed standard shape and might be inaccurate. Advances in deep learning have significantly improved noncontrast computed tomography's capabilities in ICH volume estimation. This study conducts a comprehensive systematic review and meta-analysis to evaluate the precision of deep learning algorithms in delineating ICH on noncontrast computed tomography. Methods: A systematic review and meta-analysis, adhering to Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, was conducted on literature from 2000 to October 2023. Studies were selected on the basis of strict inclusion and exclusion criteria. Performance evaluation was done using the Dice Similarity Coefficient, and the Prediction Model Risk of Bias Assessment Tool was used for quality assessment. Statistical analysis was carried out using Stata 17.0. Results: The review included 28 studies, mainly retrospective cohorts, with a focus on convolutional neural network architectures, particularly U-Net variants. A meta-analysis of 14 studies revealed a combined Dice Similarity Coefficient of 0.85 (95% CI, 0.82-0.88). Performance was consistent across various methodologies but varied on the basis of ICH pathogenesis, with spontaneous ICH having higher accuracy. Conclusion: Deep learning models are highly effective in segmenting ICH on noncontrast computed tomography, demonstrating potential improvements in clinical neuroimaging. Despite their efficacy, challenges in segmenting smaller hemorrhages remain. The findings suggest that deep learning could reduce health care professional workloads and enhance patient care, although further research is needed to address limitations and extend clinical utility.