Entropy-regularized dual-stream attention fusion for multi-disease lung classification.

Accurate and reliable classification of lung diseases from medical imaging remains challenging due to overlapping radiological patterns, disease heterogeneity, and variations in image quality. To address these challenges, this paper proposes an Adaptive Dual-Stream Attention Fusion (ADSAF) framework for automated lung disease classification. The proposed model integrates convolutional neural networks and vision transformers to jointly capture fine-grained local textures and long-range global contextual dependencies. An adaptive attention fusion mechanism dynamically learns image-specific weighting between the two feature streams, overcoming the limitations of static fusion strategies. In addition, a Self-Attention Refinement Module enhances disease-relevant regions while suppressing background noise, improving both discriminative capability and interpretability. ADSAF is designed as a general multi-class lung disease classification framework and is validated in this study through task-specific evaluations on pneumonia and COVID-19 datasets, demonstrating strong accuracy, robustness, and generalization. Experimental results demonstrate consistent performance gains over state-of-the-art CNN, transformer, and ensemble models in terms of accuracy, F-score, sensitivity, and robustness under noisy conditions and domain shifts, achieving up to 98.3% accuracy. Grad-CAM visualizations further confirm the model’s focus on clinically meaningful lung regions, highlighting its potential for reliable clinical decision support.