Explainable Multi-Label Chest X-Ray Disease Classification Using DualPool-DenseNet121 and Grad-CAM Visualization

Abstract

The global prevalence of lung diseases has risen markedly over the past three decades, driven by factors such as population growth, harmful environmental exposures, and increased life expectancy. Epidemiological studies have highlighted not only the growing burden of chronic respiratory disorders but also the trend toward earlier onset across age groups. Early detection and timely treatment of these conditions are essential to improving long-term clinical outcomes and reducing premature mortality. Chest radiography (X-ray) remains the most widely used and cost-effective initial diagnostic tool for respiratory diseases. However, the reliable identification of multiple thoracic abnormalities, such as cardiomegaly, emphysema, hernia, infiltration, mass, nodules, atelectasis, pneumothorax, pleural thickening, pneumonia, fibrosis, consolidation, effusion and edema, poses challenges even for experienced radiologists. Recent advances in deep learning, particularly convolutional neural networks, have improved medical image analysis and classification. In this study, we propose DualPool-DenseNet121, a modified DenseNet121 architecture combining global average and global max pooling, for multi-label classification of thoracic abnormalities on the NIH ChestX-ray14 dataset. Following removal of “No Finding” cases, our analysis used 51,759 pathological radiographs from 14,402 patients, annotated for fourteen pathological conditions, partitioned patient-wise into training, validation, and test sets with verified zero patient overlap. The pipeline incorporates standardized preprocessing, offline augmentation of under-represented classes, staged transfer learning, and per-class decision thresholds selected on the validation set. Model performance was evaluated using per-class AUC together with threshold-based clinical metrics. The proposed model achieved a macro-averaged AUC of 0.803 and a micro-averaged AUC of 0.831 on the held-out test set. Zero-shot external validation on the CheXpert validation set reached a macro-averaged AUC of 0.806 across the five overlapping pathologies with sufficient positive cases. Grad-CAM visualizations indicated that, in correctly classified cases, the model frequently attended to anatomically relevant regions. These results suggest that the proposed approach provides a reliable and reproducible retrospective benchmark for multi-label thoracic abnormality classification. External validation and expert radiological assessment are required before any clinical application can be considered.