HYBRID ATTENTION-BASED DEEP LEARNING FRAMEWORK FOR AUTOMATED LUNG DISEASE DETECTION FROM CHEST X-RAY IMAGES

Abstract

Chest radiography is the most readily available and affordable imaging modality for screening the lungs, but manual interpretation can be time-intensive and subjective, especially in resource-limited areas. CNNs are good at capturing local texture cues but do poorly on modeling long-range dependencies between spatially distributed pathological regions, while transformer-based models are adept at capturing context across the whole image but do poorly on modeling fine-grained local detail. To tackle the problem of multi-class chest X-ray image classification of lung diseases across five classes (Normal, Pneumonia, Tuberculosis, COVID-19, and Pulmonary Fibrosis), this paper suggests HA-LungNet, a hybrid attention-based deep learning framework that uses a Convolutional Block Attention Module (CBAM) and a lightweight transformer encoder, which is connected to a CNN backbone by a gated cross-attention fusion bridge. The local descriptors input to an EfficientNet-CBAM stream are fused adaptively with globally contextualized transformer features, and Grad-CAM visualization is incorporated to facilitate clinical interpretation. The proposed framework is evaluated with a combined benchmark consisting of ChestX-ray14, COVID-19 Radiography Database, and Shenzhen TB and Montgomery TB datasets, which achieves 96.4% accuracy, 95.3% sensitivity, 97.0% specificity, and an AUC of 0.984, outperforming the best CNN-only baseline by 3.8 percentage points and the CNN+CBAM hybrid baseline by 2.8 percentage points; ablation results confirm each component's contribution. The framework provides an efficient and interpretable automated solution to screen lung disease in low-resource healthcare settings that is clinically applicable.