Mislabel Detection in Multi-Label Chest X-Rays via Prototype-Weighted Neighborhood Consistency in CoAtNet Embedding Space

Large-scale chest X-ray (CXR) datasets often rely on report-derived or weak labels, introducing missing and incorrect annotations that can degrade downstream models and limit trust. We study training-free mislabel detection in multi-label CXRs by scoring neighborhood label consistency in a fixed embedding space. Using the NIH Chest X-ray Kaggle sample (5606 CXRs), we extract intermediate CoAtNet features and obtain 64-dimensional embeddings with a frozen CoAtNet backbone and a lightweight refinement head. On top of these embeddings, we compare kNN consistency baselines with distance weighting and label-set similarity against LPV-DW-CS, clustered prototype voting weighted by distance and cluster support. We evaluate three synthetic label-noise regimes with review budgets matched to the corruption rate: random single-label (5% and 20%), boundary-noise (20% corruption within the lowest-density 20% subset), and disjoint-label replacement (20% within that subset). LPV-DW-CS yields the highest downstream macro-AUROC after filtering top-ranked samples (up to 0.8860), while kNN variants achieve higher Recall@budget at the same review rates (up to 99.44%). An image-only expert Likert review of top-ranked real samples finds substantial label-set inconsistencies (54.1% for LPV-DW-CS-280-A; 60.5% for KNN-DW-LSS), supporting neighborhood-consistency ranking as a practical, training-free tool for targeted dataset auditing.