MSMA-DenseNet201: Distinguishing Viral Pneumonia, Bacterial Pneumonia, and Healthy Cases in Pediatric Chest Radiographs

This study focuses on the challenging task of diagnosing childhood pneumonia from chest X-rays, particularly in resource-limited environments. We present an improved DenseNet201 architecture enhanced with a novel Multi-Scale Mixed Attention (MSMA) module, which integrates multiscale feature extraction, dual attention mechanisms, and residual connections to boost lesion localization and classification. Experimental results on a large pediatric chest X-ray dataset demonstrate that our approach significantly outperforms the baseline DenseNet201 and other attention-based models in terms of accuracy. We extend the classification from binary to ternary by distinguishing bacterial from viral pneumonia, achieving promising results that demonstrate the model's potential for reliable automated screening in clinical settings.