Hybrid Bi-directional GRU-Attention U-Net with MLP Classifier for Accurate Pneumonia Segmentation and Classification
摘要
Pneumonia, a leading cause of global morbidity and mortality, particularly among children under five and elderly populations, accounted for over 740,000 deaths in 2019 alone as reported by the World Health Organization. With an estimated 450 million individuals affected annually, timely and accurate diagnosis remains a critical challenge in clinical settings. While chest X-ray imaging is the most practical and widely used modality for pneumonia detection due to its affordability and availability, conventional computer-aided diagnostic (CAD) systems and deep learning models often fall short in handling spatial dependencies, subtle lesion variations, and the temporal dynamics evident in disease progression. Addressing these limitations, this study proposes a novel hybrid deep learning framework that combines a bi-directional Gated Recurrent Unit (GRU)-enhanced U-Net with an attention mechanism for improved pneumonia segmentation, alongside a Multi-Layer Perceptron (MLP) classifier for accurate disease classification. The bi-directional GRU modules embedded within the U-Net architecture effectively capture long-range contextual information, while the attention gate selectively enhances critical lesion features. Post segmentation, a dynamic feature fusion strategy extracts high-level representations, which are fed into the MLP classifier to distinguish between normal and pneumonia-affected lung regions. Experimental validation using a Kaggle chest X-ray dataset demonstrates the superior performance of the proposed model, achieving 98% accuracy, an F1 score of 0.93, precision of 0.95, and recall of 0.92. This integrated approach not only improves diagnostic precision but also holds significant potential as a clinical decision support tool, aiding radiologists in early detection, risk stratification, and personalized treatment planning for pneumonia patients.