xAI: An Explainable AI Model for the Diagnosis of COPD from CXR Images

Chronic Obstructive Pulmonary Disease (COPD) is a major health concern worldwide, the third leading cause of death. However, it often goes unnoticed until it reaches severe stages. Although spirometry tests are the definitive method for COPD diagnosis, their accessibility is limited in under-resourced regions. In contrast, Chest X-rays (CXRs) are more universally available, suggesting their potential utility as preliminary screening tools for COPD. This study applies deep learning algorithms to the large-scale VinDR-CXR dataset to identify early-stage COPD using CXRs. The research employs the ChestX-ray14 dataset for model pre-training and then utilizes VinDR-CXR for model development and validation. Through transfer learning, it was observed that the Xception base model with a recall of 98.2% obtained via fine-tuning was better than the ResNet50 model. Grad-CAM (Gradient Class Activation Mapping) heatmaps and SHAP (SHapley Additive exPlanations) offer explainability for true positive cases on the Xception model from both datasets. The findings underscore the potential of DL models in facilitating early COPD detection via CXRs, especially where spirometry remains less accessible. Further exploration into multimodal-based COPD diagnosis is recommended for subsequent research.