Deep Learning-Based Multi-Label Classification of Chest X-Ray Images for Thoracic Disease Detection
Abstract
Chest radiography is widely used for thoracic disease diagnosis; however, manual interpretation is time-consuming and dependent on radiologist expertise. This study develops and evaluates a deep learning-based multi-label classification framework for automated thoracic disease detection using 125,928 chest X-ray images representing fifteen disease categories and a normal category. Two transfer learning architectures, ResNet50 and DenseNet121, were evaluated under identical training conditions using Macro AUC, Micro AUC, and Subset Accuracy. Experimental results showed that ResNet50 outperformed DenseNet121, achieving a Macro AUC of 0.8132, Micro AUC of 0.8668, and Subset Accuracy of 0.5129, compared with 0.7967, 0.8514, and 0.4890, respectively, for DenseNet121. Disease-wise analysis demonstrated strong performance for COVID-19, Emphysema, Edema, Cardiomegaly, and Pneumothorax, whereas Infiltration, Pneumonia, Consolidation, and Nodule remained more challenging. A graphical user interface and heatmap visualization module were developed to improve usability and model interpretability. The results provide a consistent baseline for large-scale multi-label thoracic disease classification and establish a basis for future evaluation using modern CNN and CNN-Transformer architectures.