Multi-Model Framework for Chest X-ray Interpretation with Clinically Calibrated Tuberculosis Detection

DOI

https://www.doi.org/10.59256/ijire.20260704009

Published In

Volume 7, Issue 4 (2026)

Published Online

July-August 2026

Pages

56-70

Article Type

Original Article

Check for updates

Chest X-ray (CXR) interpretation requires both comprehensive thoracic pathology assessment and disease-specific screening under heterogeneous imaging conditions. Existing deep learning approaches primarily focus on either multi-label abnormality classification or isolated disease detection, with limited emphasis on clinically deployable integration across these tasks. In this work, we propose a modular chest X-ray interpretation framework that combines general thoracic pathology classification with specialized tuberculosis (TB) screening within a unified inference pipeline. The system integrates DenseNet- and Vision Transformer (ViT)-based models trained on the CheXpert dataset for 14 thoracic findings, together with a domain-aware TB classifier trained using a multi-source dataset comprising TBX11K, Shenzhen, and Montgomery collections. The TB subsystem incorporates source-balanced sampling, targeted augmentation policies, and threshold-calibrated inference to improve robustness under domain shift and unseen acquisition conditions. Unlike prior approaches that treat pathology classification and TB screening as independent tasks, the proposed framework combines general pathology context, disease- specific optimization, clinically aligned threshold calibration, and deployment-oriented interpretability within a single workflow. Experimental results show that the TB classifier achieves an AUC of 0.9545, including strong zero-shot generalization on the unseen Montgomery dataset (AUC 0.8522). For multi-label thoracic pathology classification, the DenseNet and ViT models achieve mean AUCs of 0.8180 and 0.8140, respectively, while weighted ensembling improves performance to 0.8223. The framework further provides Grad-CAM-based visual interpretability, probabilistic confidence estimation, and structured report generation to support clinical review workflows. These findings demonstrate that modular architecture design, domain-aware training, and clinically calibrated decision strategies can improve the robustness and practical applicability of AI-assisted CXR interpretation systems.