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

Abstract

Manual evaluation of non-contrast CT scans (NCTS) for detecting subdural hematoma (SDH) is time consuming, potentially inaccurate, and subjective to the expert analyzing them. In recent years, two deep learning (DL) algorithms have been popularly studied in this respect, namely convolutional neural networks (CNN) and U-Net architectures, the latter being a specialized type of CNN. We performed the first meta-analysis comparing various DL models for SDH detection. MEDLINE, Cochrane, Scopus, and Embase databases were searched from inception through December 2025. Studies evaluating ML model performance on an independent test dataset were included. The main outcome measures were sensitivity, specificity, diagnostic odds ratio (DOR), accuracy, and precision of CNN, U-Net, and hybrid DL models. Univariate meta-regression analyses were performed. 30 testing datasets incorporating 67,266 NCTS were included. U-Net demonstrated significantly higher sensitivity (0.916;p = 0.04) and precision (0.983;p = 0.001) while high specificity, DOR, and accuracy values were consistently observed across all DL techniques. Internal testing (p = 0.05) was a borderline significant predictor of high specificity while recent publication year (p < 0.001), U-Net architecture (p = 0.035), and 3D models (p = 0.022) emerged as significant moderators of high precision. The U-Net architecture was also a borderline significant predictor of high DOR (p = 0.049). While this single arm meta-analysis depicts potential superiority of U-Net models with respect to sensitivity and precision, these findings are based off only 4 pooled U-Net datasets in comparison to the 22 pooled for CNN architectures. Future well-powered studies evaluating the U-Net model are necessary to ensure a fair comparison of U-Net architectures to other DL designs before reaching to any definitive conclusions in this respect.