Healthcare Analytics · Published 2026-06-26 · DOI 10.1016/j.health.2026.100476
Joshua Pinem, Widi Astuti, Adiwijaya
The diagnosis of lung diseases from chest X-ray (CXR) images remains challenging due to inter-observer variability among radiologists and delays in clinical decision-making. Although Vision Transformer (ViT)-based architectures have shown strong performance in medical image classification, their clinical adoption is hindered by insufficiently structured feature representations for visually similar diseases, limited explainability, and the lack of quantitative uncertainty estimation. This study proposes XConViT, a reliable Vision Transformer framework designed to support clinical decision-making in multi-class lung disease classification. XConViT is exclusively designed as a clinical decision support system (CDSS) operating under a human-in-the-loop paradigm, and is not intended for autonomous medical diagnosis to replace the critical role of radiologists. XConViT integrates Supervised Contrastive Learning (SupCon) to enhance inter-class feature separability, Monte Carlo Dropout-based uncertainty estimation to provide calibrated predictive confidence, and Explainable Artificial Intelligence (XAI) techniques to improve transparency. Visual explanations are generated using Grad-CAM and attention rollout, enabling clinicians to better understand and validate the model’s predictions. The proposed approach was evaluated on two publicly available datasets, COVID-Qu-Ex and Pneumonia Chest X-ray. Experimental results demonstrate that XConViT consistently outperforms baseline models and state-of-the-art ViT variants, achieving accuracies of 0.9487 and 0.8739 on the respective datasets, corresponding to an improvement of approximately 5% over conventional ViT models. Furthermore, XConViT exhibits superior probabilistic calibration, as indicated by a reduced Expected Calibration Error, and supports selective prediction, yielding an additional 4%–5% accuracy gain when considering only high-confidence samples. By jointly delivering accurate, calibrated, and explainable predictions, XConViT contributes a practical and trustworthy deep learning framework for CXR-based lung disease classification. The novelty of this work lies in the unified integration of Supervised Contrastive Learning, uncertainty estimation, and XAI within a single ViT architecture, advancing the readiness of deep learning models for real-world clinical deployment.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →
Pinem, J., Astuti, W., Adiwijaya (2026). A contrastive and uncertainty-aware analytics framework for reliable lung disease detection from chest imaging. Healthcare Analytics. https://doi.org/10.1016/j.health.2026.100476