Causal Disentanglement via DenseNet Generative Models for Explainable AI in Chest X-Ray Diagnosis

Abstract

Deep learning models for medical image diagnosis often operate as black boxes, providing limited insight into the reasoning behind their predictions. This opacity hinders clinical trust and adoption, particularly in high-stakes applications such as chest X-ray interpretation. We propose a novel framework that integrates a structurally causal conditional variational autoencoder (SC-CVAE) with a DenseNet-121 backbone to transform opaque feature embeddings into interpretable, intervenable latent representations. The DenseNet-121 extracts hierarchical feature maps from chest X-rays, which are aggregated into a composite tensor and fed into the SC-CVAE. The SC-CVAE encodes these features into a set of latent factors that correspond to clinically meaningful concepts, such as pathology shape, intensity, anatomical location, and scanner artifacts. A domain-derived causal graph governs the relationships among these factors, and a mutual information regularizer enforces disentanglement by penalizing statistical dependence between causally unrelated variables. The decoder reconstructs the feature tensor from the latent factors, and an adversarial training mechanism ensures that concept edits remain localized and medically plausible. For explanation generation, we perform interventions on specific latent factors—for example, setting the shape factor to a counterfactual value—and decode the modified representation to produce a new feature tensor. The difference in downstream classification probabilities between the original and counterfactual features quantifies the concept-level contribution of each factor to the diagnostic decision. The entire system is trained end-to-end with a combined objective that includes reconstruction, mutual information, adversarial, and classification losses. Our approach yields, for each input X-ray, pathology probabilities, a Grad-CAM heatmap for spatial localization, and a concept-level contribution vector. This multi-faceted output enables clinicians to understand not only where the model focuses but also why specific visual features drive the diagnostic outcome. The proposed method thus advances explainable AI in medical imaging by bridging the gap between high-performance deep learning and clinically meaningful interpretability.