Research & publications / Medical image segmentation
Uncertainty-Aware Pneumothorax Segmentation and Selective Prediction in Chest Radiographs
Zenodo ¡ 2026
Uncertainty-aware pneumothorax segmentation in chest radiographs, examining calibration, selective prediction, and the reliability of model uncertainty.
Preprint
Overview
This study examines pneumothorax segmentation and uncertainty estimation in chest radiographs. The author manuscript presents a comparative evaluation of segmentation, calibration, and selective prediction.
Research question
Do uncertainty estimates reliably identify segmentation errors and support selective prediction in pneumothorax chest radiographs?
Methods
A ResNet34 U-Net is evaluated with Monte Carlo Dropout and deep ensembles. Analyses include segmentation quality, calibration, error discrimination, risk-coverage behaviour, and projection subgroups.
Data & evaluation setup
SIIM-ACR pneumothorax dataset, with patient-isolated evaluation described in the manuscript. Dataset files and trained model weights are not included in the Zenodo deposit.
Results
The manuscript reports limitations in checkpoint-selection stability and a fixed operating threshold. Ensemble uncertainty rankings can perform worse than random, underscoring the need to evaluate uncertainty rather than assume it is reliable.
Limitations
This is a preprint and does not establish clinical readiness. Checkpoint-selection variance, threshold reliability, and the generalization of uncertainty rankings require further validation.
Collaborators & acknowledgements
The study uses the SIIM-ACR dataset and established U-Net, ResNet, Monte Carlo Dropout, and ensemble methods; original sources retain their authorship and terms.