A Comparative Study of Respiratory Diseases Classification Using Grad-CAM-Based DenseNet Architectures

Respiratory diseases such as COVID-19, tuberculosis, and pneumonia remain major global health concerns, and CXR imaging plays a crucial role in their early detection and diagnosis. However, manual interpretation of chest radiographs is time-consuming and subject to variability among clinicians. Deep learning offers a promising solution to support automated diagnosis, although challenges remain regarding optimal model selection and interpretability. The contribution of this study is a comparative evaluation of DenseNet121, DenseNet169, and DenseNet201 architectures, combined with Grad-CAM to enhance transparency in decision-making. Two openly accessible collections of chest radiograph images were employed. Dataset-1 consisted of 6,432 images of Normal, Pneumonia, and COVID-19, while Dataset-2 included 15,421 images of Normal, Pneumonia, and Tuberculosis. Each model was trained for 50 epochs under four optimizers, namely Adam, Adamax, and SGD. Performance was assessed using metrics evaluation and Grad-CAM was utilized to depict the areas that significantly shaped the model’s predictions. The results demonstrated that DenseNet169 consistently achieved the most balanced performance across datasets and optimizers. On Dataset-1 with Adam optimization, it reached an accuracy of 97.46%, precision of 96.05%, recall of 96.24%, F1-score of 96.10%, and specificity of 97.68%. On Dataset-2, it achieved 97.11% accuracy, 96.17% precision, 95.67% recall, 95.69% F1-score, and 97.83% specificity. These outcomes confirm that DenseNet169 is particularly well-suited for screening applications where sensitivity is critical. Grad-CAM depictions additionally confirmed that the model concentrated on diagnostically pertinent pulmonary regions, thereby strengthening clinical trust. In conclusion, DenseNet169 proved to be the most robust and reliable architecture for respiratory disease categorization, while Grad-CAM enhanced model interpretability. These results emphasize the promise of DenseNet-driven strategies as supportive instruments in medical image analysis and indicate opportunities for continued enhancement in clinical practice.