Causally Constrained and Anatomically Grounded Interpretability in Deep Learning for Chest Radiograph Diagnosis

Abstract

Although deep learning systems have proved to have high potential in classifying chest radiographs, their low level of transparency still limits their widespread clinical use on a regular basis. Most current models produce visual explanations that are generated post-prediction, and frequently point out regions that do not have anatomical or diagnostic significance. This paper introduces a learning approach that focuses on interpretability where the anatomical structure and causal relevance are directly factored into network optimization. The given model limits attribution response in areas of the lungs, inhibits the background spuriously, and assesses prediction dependence by means of controlled counterfactual masking. Moreover, there are intermediate feature channels that are also analyzed to reveal patterns that are related to clinically significant radiographic features.The method was tested with NIH ChestX-ray14 dataset to detect pneumonia. The final model had a total classification accuracy of 96.8% and an AUC of 0.983. In addition to predictive performance, there was a significant improvement in the quality of the explanation: the overlap of attribution maps of lung-region showed up to 88% by non-relevant activation was low (7%), and the saliency in perturbation was over 30% better than gradient-based methods. These findings suggest that incorporating structural interpretability constraints at training can lead to improving diagnostic transparency and preserving high performance, providing a more trustworthy basis to use in clinical imaging settings.