Label-wise reliability-aware classifier for robust chest X-ray multi-label classification

doi:10.1016/j.eswa.2026.131438

Label-wise reliability-aware classifier for robust chest X-ray multi-label classification

2026-02-03

0

PRE

AI

DOI：10.1016/j.eswa.2026.131438

Original

Original request for help

Save

Abstract

En 中文

Chest X-ray (CXR) multi-label classifiers are commonly trained with labels extracted from clinical reports, which are often incomplete and noisy. Under such label noise, we observe that performance degrades severely on tail classes (e.g., rare diseases), because these categories are under-represented and easily overwhelmed by corrupted annotations. As a result, existing methods can misidentify tail classes as noise and downweight their contribution to optimization during training. To address this issue, we propose LRC-CXR (Label-wise Reliability-aware Classifier for Chest X-ray), which calculates per-label reliability and selectively corrects noisy labels, preventing tail classes from being inadvertently under-trained. First, a Medical Description Bank provides lesion-aware textual prompts that guide the visual encoder toward diagnostically relevant patterns. Second, LRC-CXR models per-label reliability with a two-component Gaussian Mixture Model to distinguish clean, inseparable, and noisy labels. Third, only labels identified as noisy are refined via feature-space k-nearest-neighbor smoothing, while clean and inseparable labels are trained with stronger objectives through a hierarchical loss. Experiments on ChestX-ray14, CheXpert, and PadChest, including high-noise stress tests, show that LRC-CXR improves overall AUC/F1 and substantially boosts tail-class recall and robustness under label noise.

Keywords:

Label noise

Multi-label classification

Chest X-ray

Tail class

Reliability-aware modeling

Label-wise reliability-aware classifier for robust chest X-ray multi-label classification

Abstract

Journal

Organization