Fusing Imaging and Clinical Data for COPD Analysis: A Multimodal DenseNet Approach with Grad-CAM

Abstract

In the analysis and diagnosis of lung diseases through chest X-rays with deep learning models, the presence of bones and tissues, as well as burrs and blurs that may be present on the X-rays, can cause complications. COPD, or chronic obstructive pulmonary disease, is a respiratory disease that progresses steadily. It has been a challenging disease to classify and grade. It is important that the severity and grading of the disease are done correctly and on time. This paper aims to present a method that can classify the severity of COPD using MDenseNet-Grad-CAM. This method combines Multimodal DenseNet and Gradient-weighted Class Activation Mapping. It combines X-rays with clinical parameters that can indicate the progression of the disease. This combination can help the model learn better and provide a more holistic approach. It also incorporates Grad-CAM for better interpretability. This method can provide a better approach than other models that only rely on X-rays or data. It has a better classification accuracy than the reference models. It has an accuracy rate of 98.55%. Ablation studies confirmed the contribution of each component. DenseNet alone achieved 87.55% accuracy. Adding clinical data increased accuracy to 92.11%. Integrating Grad-CAM without clinical data reached 93.65%. Combining DenseNet, clinical data, and Grad-CAM produced the best accuracy of 98.55%. Findings confirm its value for early COPD diagnosis and severity classification. The method demonstrates strong potential for clinical decision support and patient care applications.