DeepHemorrhageNet

An Imbalance Aware Deep Learning Model for Clinical Haemorrhagic Stroke Detection

Abstract

Notable for its high mortality, and substantial clinical costs, Hemorrhagic stroke poses a severe global health threat.  Our study proposes DeepHemorrhageNet, an imbalance-aware deep learning framework for early detection and risk assessment of hemorrhagic stroke, addressing key challenges including class imbalance in clinical data, subtle lesion detection, and integration of multimodal information. The framework consists of three core components: (1) A hybrid CNN-Transformer encoder-decoder architecture with Attention Gate (AG) mechanism, designed to extract multi-scale spatial features and model global context for precise segmentation of hemorrhagic lesions and classification of hemorrhage types; (2) An imbalance-aware module integrating Class-Balanced Focal Loss, Online Hard Example Mining (OHEM), and feature-space oversampling, which mitigates the impact of long-tailed data distribution; (3) An Enhanced Video Masked Autoencoder (E-VMAE) and Early Predictive Behavioral Recognition with Physiological Signals (EPBR-PS) algorithm for real-time detection of stroke-related abnormal behaviors by fusing temporal masking, local convolutional attention, and physiological signals. We also use a multimodal clinical dataset and behavioral video datasets for model training and validation.  To prepare the data, preprocessing steps included Borderline SMOTE for class imbalance mitigation, VIF-based multicollinearity reduction, and standardized CT/image/video processing. Consequently, Experimental results showed that DeepHemorrhageNet outperformed state-of-the-art models on public datasets, achieving 83.2% mAP on THUMOS14, 54.6% mAP on ActivityNet1.3, and high accuracy (91.45% on test set) for behavioral recognition. For hematoma expansion, the prediction of logistic regression model achieved 84.8% accuracy, 82.1% recall, and 0.91 AUC on the test set. Additionally, a web-based prototype system integrating real-time monitoring, heart rate estimation, and risk assessment was developed, providing clinical decision support by alerting on high-risk events. In conclusion, DeepHemorrhageNet effectively combines multimodal data to enhance early detection and risk assessment of hemorrhagic stroke, demonstrating strong robustness, generalization, and clinical utility for improving patient outcomes and reducing healthcare burdens.