MultiClass Classification of Chest Diseases using CXR Images with DenseNet201+CNN and Grad CAM Visualization

Accurately identifying lung diseases is really important for treating them well, especially now with COVID-19 making respiratory problems worse worldwide. In response to the global impact, this paper has presented a deep learning framework for the multiclass classification of chest diseases, covering COVID-19, pulmonary fibrosis, tuberculosis, lung opacity, and normal cases. A modified DenseNet201 combined with a custom CNN has been introduced in this paper for a more effective classification of lung diseases. The proposed DenseNet201+CNN network surpasses key transfer learning models (ResNet50, VGG16, VGG19, DenseNet201) achieving a testing accuracy of 95.43%. This study employs DenseNet201 due to its superior feature extraction capabilities and efficiency in handling vanishing gradients. Contributions of this study include a 5-class classification system with a sizable combined dataset, addressing class imbalance through augmentation and rigorous model evaluation comparing with four other transfer learning models. The incorporation of Explainable AI (Artificial Intelligence), specifically Gradient Weighted Class Activation Mapping (Grad-CAM), enhances data visualization, ensuring transparency in the model's decision-making process. This study strives to advance automated disease diagnosis, potentially transforming early detection approaches for critical lung conditions.