Abstract

In this study, we propose a novel approach to multi-label classification of chest X-ray (CXR) images that prioritizes clinical interpretability while maintaining the efficiency of a streamlined, single-model, single-run training pipeline. Using the CheXpert dataset and VisualCheXbert-derived labels, we introduce hierarchical label groupings to reflect clinically meaningful relationships among diseases. To implement this, we developed a custom hierarchical binary cross-entropy (BCE) loss function that enforces label dependencies through fixed or data-driven penalty mechanisms. This approach aims to preserve diagnostic accuracy while producing structured and clinically coherent predictions. The official codebase supporting this work is available at https://github.com/the-mercury/CIHMLC.git.