Smart-LungNet for Lung Disease Classification

This study proposes Smart-LungNet, an automated deep learning framework designed to classify lung conditions into three categories: Normal, Lung Opacity, and Viral Pneumonia. Utilizing the Lung X-Ray Image Dataset of 3,475 images, we evaluated several pre-trained architectures, including ResNet18, DenseNet121, and MobileNetV2. MobileNetV2 was selected as the baseline due to its balance of efficiency and performance (88.5% accuracy). We enhanced this model by unfreezing all layers for fine-tuning and integrating a Squeeze-and-Excitation (SE) block after the initial convolutional layer to improve channel-wise feature attention. The proposed Smart-LungNet achieved a testing accuracy of 89.85% and an F1- score of 89.84%, outperforming ResNet18, DenseNet 121 and MobileNetV2. So, Smart-LungNet can help effectively to aid radiologists in the timely diagnosis of lung pathologies.