Abstract
Automated chest X-ray classifiers are increasingly used for triage and decision support, but fixed-threshold inference can be unsafe
when models produce overconfident errors or uncertain predictions. We study uncertainty-aware selective prediction for multi-label frontal chest radiograph classification using frozen MedSigLIP image features and lightweight prediction heads. We compare four uncertainty signals: MC-dropout variance, MC-dropout entropy, ensemble variance, and ensemble predictive entropy, alongside random-referral and confidence-based baselines. Study-level uncertainty is computed from per-label uncertainty scores, and the most uncertain studies are referred to expert review while the remaining cases are handled automatically. On a five-label CheXpert frontal-view task, ensemble predictive entropy achieves the highest macro-F1 at the 20% referral operating point, improving macro-F1 from 0.9265 to 0.9387 while reducing false negatives from 972 to 661, an approximately 32.0% reduction, and false positives from 1572 to 1122 on the automated subset. These findings suggest that uncertainty estimation can provide an actionable referral signal for human-in-the-loop chest
X-ray classification.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_017.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=I0i7lPNdO0
BibTex
@InProceedings{AttNad_UncertaintyBased_MICCAISAT2026,
author = { Attia, Nada A. AND Al-Atabany, Walid AND Selim, Sahar},
title = { { Uncertainty-Based Referral for Selective Chest X-Ray Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}