Deep Learning-Based Multi-Class Diagnosis of Lung Diseases from Chest X-Ray Images

Lung disease is characterized as a severe global health disorder, especially in low-resource areas where medical subspecialty care is not accessible. Chest X-ray imaging remains a primary diagnostic method for identifying lung conditions due to its speed and wide accessibility. Its widespread application notwithstanding, interpretation of chest X-ray images is often difficult since most pulmonary disorders have overlapping radiographic characteristics like opacities, consolidations, and infiltrative patterns. This resemblance frequently results in radiographic overlap among cases such as bacterial pneumonia, viral pneumonia, tuberculosis, and COVID-19, thereby making classification a challenge even for veteran radiologists. In this research, we implement a deep learning-based multiclass classification model for detecting lung diseases from chest X-ray images. The model effectively classifies five categories: viral pneumonia, bacterial pneumonia, tuberculosis, COVID19, and normal lungs. To warrant this, we built a strong dataset by combining a series of publicly available sources and applying data augmentation for enhanced variability and robustness. We performed preprocessing and feature selection operations like resizing, normalization, augmentation (rotation, flip, zoom), duplicate removal, and mutual information-based filtering of features prior to training and testing three deep learning frameworks, i.e., VGG19, MobileNetV3-Large, and MobileNetV3-Small, with comparable parameters. Among them, MobileNetV3-Small showed the greatest performance, 85.18% accuracy and 84.86% F1-score showing the best precision-recall balance. The system showed promising potential for use in clinical practice where timely and accurate judgment is paramount. Future updates to be explored include the inclusion of more advanced training methodology and explainable AI techniques to increase transparency, alongside real-world evaluation on a variety of clinical populations.