Deep hiearchical multi-label classification applied to chest X-ray abnormality taxonomies

Haomin Chen, Shun Miao, Daguang Xu, Gregory D. Hager, Adam P. Harrison

Abstract: Highlights•A two-stage hierarchical multi-label classification approach (HLUP-finetune) is outlined for chest X-rays.•A numerically stable method for calculation of cross-entropy loss of unconditional probabilities is derived.•In AUCs, especially when examining samples conditioned on >=1 abnormalities being present.•HLUP-finetune also demonstrates performance gains over baseline methods under incompletely labelled scenarios.