Hybrid Deep Ensemble Feature Fusion for High-Precision Intracranial Hemorrhage Detection in Neuroimaging

Abstract

Intracranial Hemorrhage (ICH) is a neurological emergency, which is life threatening and requires prompt and sound diagnosis in order to minimize the morbidity and mortality rates. However, traditional deep learning models do not in general capture complex multi-scale structural and textural patterns within neuroimaging, leading to poor generalization. To fill this gap, this paper suggests a hybrid VGG16-VGG19 deep learning architecture that incorporates complementary hierarchical feature representations to detect ICH successfully using CT scans. The approach includes the normalization of intensities, a thorough data augmentation approach, transfer learning, and a new feature-fusion technique that fuses the multi-level descriptors obtained with both VGG versions. It is trained by forwarding the fused representation to optimized dense layers with dropout regularization, which permits a display of strong discrimination between hemorrhagic and non-hemorrhagic cases. The given framework attains 97.2% accuracy, 96.8% sensitivity, 97.6% specificity, F1-score of 97.0% and AUC of 0.985 evaluated on a curated dataset of 3,000 neuroimaging scans, surpassing the performance of individual VGG models, as well as a variety of more recent state-of-the-art frameworks. The obtained results indicate the ability of multi-scale feature fusion to capture subtle hemorrhagic patterns and reveal that the framework can be a clinically reliable and scalable computer-aided diagnostic aid in emergency neuroimaging operations.