An Interpretable Chest X-ray Classification Framework Using Prototype Memory and Counterfactual Consistency.

Chest X-ray (CXR) interpretation requires recognition of subtle and heterogenous radiographic patterns, yet conventional deep learning models often rely on spurious image cues that lack anatomical or pathological relevance. This study presents CXR-NeXus, an interpretable chest X-ray classification framework designed to promote explicit visual reasoning and reliable decision behavior under weak supervision. The proposed method integrates prototype memory with counterfactual consistency to constrain model predictions to clinically meaningful pulmonary evidence. CXR-NeXus learns multiple class-specific prototypes that represent diverse radiographic phenotypes for COVID-19, pneumonia, tuberculosis, and normal cases. Each prediction is supported by similarity to representative prototype patterns, enabling transparent "this-looks-like-that" explanations at the image level. To further ensure causal dependence on pathological evidence, counterfactual chest X-rays are generated by Grad-CAM-guided lesion suppression constrained within coarse lung regions. The model is explicitly trained to reduce disease confidence when lesion evidence is removed, aligning prediction behavior with clinical intuition. In addition, an evidence-alignment regularization is introduced to penalize extra-pulmonary saliency and to localize attention changes induced by counterfactual perturbations. The overall learning objective jointly optimizes classification performance, margin-based prototype separation, counterfactual consistency, and anatomically plausible attention, without requiring pixel-level lesion annotations. Experiments on a four-class chest X-ray dataset demonstrate that the proposed framework improves macro-average F1 score, Receiver Operating Characteristic - Area Under the Curve (ROC-AUC), specificity, and probability calibration compared with strong baselines, while substantially reducing reliance on spurious visual cues. These results suggest that combining prototype-level semantic anchoring with counterfactual-driven visual reasoning provides a practical and interpretable solution for reliable chest X-ray analysis.