LTCXNet: Tackling Long-Tailed Multi-label Classification and Racial Bias in Chest X-Ray Analysis

Chin-Wei Huang, Chi-Yu Chen, Mu-Yi Shen, Kuan-Chang Shih, Shih-Chih Lin, Po-Chih Kuo

Abstract: Chest X-ray (CXR) classification faces challenges from long-tailed, multi-label data distributions and demographic biases in medical AI systems. To address these, we present LTCXNet – a framework combining ConvNeXt, ML-Decoder, and multi-branch learning – evaluated on Pruned MIMIC-CXR-LT dataset curated for long-tail scenarios. The model achieves large performance gains especially in rare classes, with 79% and 48% improvements in detecting Pneumoperitoneum and Pneumomediastinum respectively. We introduce “mAUCr” fairness metric to quantify racial group performance disparities, demonstrating LTCXNet’s superior fairness in tail class subgroups compared to existing long-tail methods. This work advances medical imaging analysis by addressing both class imbalance and demographic bias through novel architectural integration and evaluation metrics. Our code is available on code.

External IDs:dblp:conf/miccai/HuangCSSLK25