Emergency Brain Hemorrhage Region Detection in Non-Contrast CT Imaging Using Ensemble Deep Learning Architectures

Abstract
This research proposes an ensemble-based deep learning framework for accurate emergency brain hemorrhage detection using non-contrast computed tomography (NCCT) imaging. The system integrates high-capacity convolutional backbones (U-Net, DenseNet, EfficientNet) with spatial attention modules (CBAM and SE blocks) and multi-scale feature fusion to enhance both pixel-level lesion localization and classification precision. Evaluated on multi-institutional datasets including RSNA, CQ500, and local NCCT scans, the model consistently achieved hemorrhage detection accuracies exceeding 94.5%, with segmentation F1-scores peaking at 0.902 and Dice scores above 0.89. The ensemble approach improved sensitivity to subtle or mixed-type hemorrhages while reducing false positives by 21–26% compared to single-architecture baselines. Attention-guided encoding and hybrid decoder fusion contributed to a 4.1–5.1% gain in accuracy across heterogeneous clinical conditions. Real-time inference performance was maintained with an average prediction latency under 500ms per full CT volume, supporting deployment in high-pressure emergency care environments. The model’s modular design allows for easy integration into PACS systems, enhancing radiologist workflows and enabling scalable, interpretable, and triage-ready AI-assisted diagnosis.

Author
Sazan Kamal Sulaiman

DOI
https://doi.org/10.1109/ICCR64138.2025.11292534

ISSN
2836-8142

Publish Date: 19-Dec-2025