The use of machine learning models for subdural hematoma detection: a single-arm meta-analysis

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Manual evaluation of non-contrast CT scans (NCTS) for subdural hematoma (SDH) detection is time-consuming and subjective.
• Deep learning (DL) models, including Convolutional Neural Networks (CNN) and U-Net architectures, are increasingly studied for SDH detection.

Purpose of the Study:

• To conduct the first meta-analysis comparing the performance of various DL models for SDH detection on NCTS.
• To evaluate sensitivity, specificity, diagnostic odds ratio (DOR), accuracy, and precision of CNN, U-Net, and hybrid DL models.

Main Methods:

• A systematic search of MEDLINE, Cochrane, Scopus, and Embase databases was performed up to December 2025.
• Included studies evaluated ML model performance on independent test datasets, focusing on CNN, U-Net, and hybrid DL models.
• Meta-regression analyses were conducted to identify predictors of model performance.

Main Results:

• 30 testing datasets (67,266 NCTS) were included in the meta-analysis.
• U-Net models demonstrated significantly higher sensitivity (0.916) and precision (0.983) compared to other DL techniques.
• High specificity, DOR, and accuracy were consistently observed across all DL models; U-Net architecture was a significant moderator of high precision.

Conclusions:

• The U-Net architecture shows potential superiority in sensitivity and precision for SDH detection on NCTS.
• However, findings are based on a limited number of U-Net datasets compared to CNNs.
• Further well-powered studies are required for a definitive comparison of U-Net architectures against other DL designs.