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 high mortality rate, long-term neurological complications, and the critical importance of timely diagnosis and accurate disease-severity assessment. This study presents an advanced Attention-Enhanced Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN–BiLSTM) framework for multimodal brain-imaging analysis, early ischemic stroke detection, and automated disease-severity classification. The proposed framework integrates computed tomography (CT), diffusion-weighted imaging (DWI), fluid-attenuated inversion recovery (FLAIR), and apparent diffusion coefficient (ADC) imaging to extract complementary anatomical, structural, and pathological information. The experimental dataset comprises 12,500 brain-imaging examinations collected from 2,500 patients, including 1,800 ischemic stroke patients and 700 non-stroke controls. Stroke patients are categorized into mild, moderate, and severe groups using the National Institutes of Health Stroke Scale (NIHSS). A comprehensive preprocessing pipeline incorporates image registration, skull stripping, intensity normalization, noise reduction, and data augmentation to enhance image quality and improve model generalization. The proposed architecture employs CNN layers for hierarchical spatial feature extraction, BiLSTM networks for modelling inter-slice contextual dependencies, and attention mechanisms for identifying diagnostically significant regions. An adaptive multimodal fusion strategy combines complementary imaging representations to strengthen diagnostic discrimination and severity prediction. The dataset is divided into training (70%), validation (15%), and testing (15%) subsets using patient-level separation to minimize information leakage. The reported experimental results demonstrate 98.2% diagnostic accuracy, 97.6% precision, 97.1% recall, 97.3% F1-score, and a receiver operating characteristic area under the curve (ROC-AUC) of 0.991 for ischemic stroke detection. Furthermore, disease-severity classification achieves 95.4% accuracy and 94.6% macro-F1-score, demonstrating improved predictive performance compared with conventional CNN, ResNet50, and CNN–LSTM architectures. Ablation analysis indicates that attention-enhanced multimodal fusion contributes to stronger classification performance, while Grad-CAM visualizations provide interpretable representations of diagnostically relevant brain regions. The findings highlight the potential of the proposed framework to support automated stroke screening, clinical decision-making, and personalized disease-severity assessment. Further multicenter validation is necessary to establish clinical reliability, robustness, and practical applicability across diverse populations.