AN ADVANCED ATTENTION-ENHANCED CNN–BILSTM FRAMEWORK FOR MULTIMODAL BRAIN-IMAGING ANALYSIS, EARLY DIAGNOSIS, AND DISEASE-SEVERITY CLASSIFICATION IN PATIENTS WITH ISCHEMIC STROKE

Abstract

Ischemic stroke represents a major global healthcare challenge due to its substantial contribution to mortality, long-term neurological disability, and increasing clinical demands for rapid and accurate diagnosis. Despite advances in neuroimaging technologies, early identification of ischemic abnormalities and reliable assessment of neurological severity remain challenging because of heterogeneous lesion characteristics, subtle pathological changes, and limitations in conventional image interpretation. This study presents an advanced Attention-Enhanced Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN–BiLSTM) framework integrating medical image processing, computer vision, and multimodal deep learning for automated ischemic stroke diagnosis and disease-severity classification. The investigation utilizes a comprehensive dataset comprising 12,500 multimodal brain-imaging examinations obtained from 2,500 patients, including 1,800 ischemic stroke patients and 700 non-stroke controls. Four complementary neuroimaging modalities, namely computed tomography (CT), diffusion-weighted imaging (DWI), fluid-attenuated inversion recovery (FLAIR), and apparent diffusion coefficient (ADC), are incorporated to capture diverse anatomical and pathological characteristics. The proposed framework employs image registration, intensity normalization, noise reduction, and data augmentation, followed by CNN-based spatial feature extraction, BiLSTM-based contextual dependency modeling, and attention-guided feature refinement. Furthermore, neurological impairment is categorized into mild, moderate, and severe stages using the National Institutes of Health Stroke Scale (NIHSS). Experimental results demonstrate superior diagnostic performance, achieving 97.8% accuracy, 97.2% precision, 96.9% sensitivity, 98.1% specificity, an F1-score of 97.0%, and an area under the receiver operating characteristic curve of 0.989. The proposed architecture additionally attains 95.6% accuracy in multiclass disease-severity classification, outperforming conventional CNN, ResNet50, standalone BiLSTM, and CNN–LSTM architectures. Comparative and ablation analyses demonstrate the effectiveness of multimodal feature integration and attention-based representation learning in improving diagnostic discrimination and classification reliability. The findings establish the potential of the proposed CNN–BiLSTM framework to support accurate computer-assisted stroke identification, automated neurological severity assessment, and data-driven clinical decision-making. Overall, this research contributes a unified computational approach for multimodal neuroimaging interpretation, providing a promising foundation for intelligent stroke diagnosis and severity prediction in clinical healthcare environments, while highlighting the importance of further external validation.