Meta-analysis of deep learning for subdural hematoma detection on CT finds U-Net models may have higher sensitivity and precision

In a meta-analysis of 30 test datasets (67,266 non-contrast CTs), U-Net deep learning models had higher sensitivity (0.916) and precision (0.983) for subdural hematoma detection vs other architectures. However, only 4 U-Net datasets were pooled vs 22 for CNN, so conclusions are…

• Sensitivity, specificity, and accuracy were consistently high across all deep learning techniques.
• The potential superiority of U-Net is limited by the small number of pooled U-Net datasets (4) compared with CNN (22).

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.