From Pneumonia to Multi-Disease: Interpretable and Uncertainty-Aware Semi-Supervised Learning Strategies for Chest X-Ray Classification

Abstract

Pneumonia is a deadly respiratory disease that causes millions of deaths each year worldwide. Chest X-ray imaging is one of the most widely used and affordable tools available for screening pneumonia. However, accurate diagnosis can often be complicated because pneumonia shares similar radiographic features with other respiratory illnesses such as COVID-19 and tuberculosis, leading to potential confusion and misdiagnosis in high-volume clinical environments. To address these issues, this study proposes an automated approach for chest X-ray disease identification by comparing supervised and semi-supervised learning techniques across two publicly available datasets. Experiments were conducted under both binary (Normal vs. Pneumonia) and multiclass (COVID-19, Normal, Pneumonia, Tuberculosis) classification settings. The proposed framework integrates MobileNet-based architectures with modern pseudo-labeling methods, including FixMatch, FlexMatch, and FreeMatch. In supervised experiments, the models achieved up to 98% accuracy in binary classification and 97% in multiclass classification, with MobileNetV3Small selected for its lightweight architecture and deployment efficiency. Under limited labeled data settings, FreeMatch demonstrated the most stable SSL performance, achieving 97% accuracy using only 20% labeled data in binary classification and 30% labeled data in multiclass classification, while maintaining lower variance and improved class balance compared to competing methods. Model interpretability was evaluated using Grad-CAM, confirming that predictions were primarily focused on clinically relevant lung regions. In addition, predictive reliability was assessed using Monte Carlo Dropout, Expected Calibration Error (ECE), and Test-Time Augmentation, demonstrating improved confidence calibration and robustness. Overall, the proposed framework shows that lightweight deep learning models combined with semi-supervised learning, uncertainty estimation, and explainability can provide accurate, efficient, and clinically reliable solutions for automated chest X-ray diagnosis.