Interpretable lung-constrained RegNetY-ViT framework for pulmonary tuberculosis classification in chest X-rays with radiological feature-guided neuro-symbolic reasoning.

Introduction: Tuberculosis (TB) remains difficult to diagnose from chest X-rays due to overlapping radiological patterns with pneumonia, fibrotic scars, and other chronic lung abnormalities. Manual interpretation is highly dependent on expert experience and is often time-consuming, subjective, and prone to variability, especially in cases involving subtle or mixed lesion presentations. Methods: This study proposes a RegNetY-ViT hybrid framework for chest X-ray analysis using the TBX11K dataset. RegNetY captures fine-grained local spatial features such as cavitary margins, consolidation, and fibronodular patterns, while ViT models global contextual relationships across the lung fields. Interpretability is embedded within the pipeline through multi-lesion Grad-CAM, enabling the localization of multiple abnormal regions. These attention maps are further refined into lesion masks using lung-constrained segmentation with adaptive thresholding, active contour refinement, and overlap validation. Radiological biomarkers, including upper-lobe infiltrates, cavitary changes, heterogeneous consolidation, and fibrotic distortion, are extracted and fed into a neuro-symbolic fuzzy inference system to translate imaging features into rule-based diagnostic support. Results: The proposed hybrid model outperformed the standalone RegNetY and ViT models, particularly in detecting Active TB. It achieved an overall accuracy of 95.8%, macro-average sensitivity of 85.1%, macro-average specificity of 98.6%, and macro-average AUC of 87.2%. Interpretability analysis showed that Grad-CAM effectively localized disease-relevant lung regions, segmentation produced high consistency across lesion areas, and the neuro-symbolic layer generated clinically interpretable diagnostic explanations aligned with radiological biomarkers. Discussion: The proposed RegNetY-ViT hybrid framework improves TB classification performance while enhancing interpretability through lesion localization and neuro-symbolic reasoning, thereby supporting more transparent and clinically meaningful decision-making in chest X-ray analysis.