Spatial Attention-guided Hybrid Deep Learning with Sharpened Cosine Similarity for Accurate Chest X-ray Interpretation.

INTRODUCTION

Life-threatening respiratory conditions such as COVID-19 and pneumonia demand rapid and accurate diagnosis. Chest X-rays (CXR) are widely used due to their accessibility and cost-effectiveness, but interpreting them remains clinically challenging, especially with overlapping radiological features.

METHODS

The proposed VSAG-HDL Net, a novel hybrid deep learning framework designed to enhance the accuracy and interpretability of CXR-based diagnosis. The architecture integrates a Variational Spatial Attention Fusion U-Net (VSA-FU-Net) for lesion segmentation and a Sharpened Cosine Similarity (SCS) Network for disease classification. A dataset of 21,165 CXR images from the Radiography Database was used. Segmentation performance was evaluated using Dice Similarity Coefficient (DSC) and Intersection over Union (IoU), while classification performance was assessed via accuracy metrics.

RESULTS

The VSA-FU-Net achieved a DSC of 90% and an IoU of 95%, indicating high precision in localizing lesions of varying shapes and sizes. The classification module reached an overall accuracy of 95.5%, outperforming traditional CNN-based methods such as CoroDet (+4.3%), CovXNet (+5.3%), and ShuffleNet (+3.9%). Although slightly less accurate than DenseNet+VIT (–2.0%) and DenseNet+VIT+GAP (–2.3%), the proposed framework offers competitive accuracy with significantly reduced model complexity.

DISCUSSION: The elimination of redundant feature extraction and the integration of spatial attention enhance both the diagnostic performance and computational efficiency, making the framework suitable for real-world clinical settings.

CONCLUSION

VSAG-HDL Net provides a robust, interpretable, and resource-efficient solution for chest disease detection in CXR. Its clinical integration can support early and accurate diagnostic decision-making, particularly in resource-limited environments.