PulmoScan AI: An Explainable Deep Learning-Based Clinical Decision Support System for Multi-Class Lung Disease Detection Using Chest X-Ray Images

Abstract

Lung diseases such as pneumonia and tuberculosis (TB) represent a major global health burden, particularly in low- and middle-income regions with limited access to expert radiological interpretation. Early and accurate diagnosis through chest X-ray (CXR) imaging is critical, yet conventional radiological interpretation suffers from inter-observer variability and a shortage of expert radiologists in resource-limited settings. This paper proposes PulmoScan AI, a deep learning-based full-stack clinical decision support system for automated detection and classification of lung diseases from CXR images. The system employs EfficientNetB0 with transfer learning for multi-class classification, trained on a curated dataset of 11,910 images, and detects three categories: Normal, Pneumonia, and Tuberculosis. The model achieves a test accuracy of 97.9%, AUC-ROC of 0.9982, macro precision of 98.28%, macro recall of 97.70%, and macro F1-score of 97.96%, outperforming four baseline architectures (a shallow CNN, a custom CNN, MobileNetV2, and ResNet50) trained under an identical protocol; five-fold cross-validation confirms stable performance across data partitions (97.20 ± 0.39% mean accuracy). A web-based clinical decision support interface integrates real-time prediction with confidence thresholding, Grad-CAM explainability, and an occupational risk assessment module. Experimental results demonstrate the feasibility and clinical potential of the approach for deployment in health screening programs.