An Efficient and Interpretable Deep Learning Framework for Pneumonia Detection and Severity Assessment from Chest Radiographs

Abstract

Pneumonia presents considerable challenges to healthcare systems worldwide, necessitating swift and accurate diagnostic methods. This study introduces an integrated artificial intelligence approach combining a convolutional neural network classification with infection region segmentation to improve pneumonia diagnosis and assess severity. Our dual-function model was crafted using the QaTa-COVID-19 and COVID-19 Radiography datasets, enabling strong differentiation between COVID-19 pneumonia, non-COVID pneumonia, and standard chest X-rays. The proposed framework combines VGG16 convolutional features with DenseNet connectivity patterns, achieving a classification accuracy of 98.71% while using only about 8 million parameters, significantly surpassing traditional models like VGG16, ResNet50, and DenseNet121. We incorporated a UNet-based segmentation component to assess disease severity by effectively outlining infection regions. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations improve  interpretability of model, providing clinicians with insights into decision-making. Our results indicate that this hybrid method enhances diagnostic precision, infection quantification, and algorithmic transparency. Future research will explore multimodal data fusion and investigate transformer architectures to boost pneumonia detection capabilities. This study contributes to developing an automated, explainable diagnostic framework that effectively aids clinical decision-making in pneumonia management.