Uncertainty-Aware Pneumothorax Segmentation and Selective Prediction in Chest Radiographs

Overview

This manuscript investigates uncertainty-aware pneumothorax segmentation and selective prediction in chest radiographs using a ResNet34 U-Net, Monte Carlo Dropout, and deep ensembles on the SIIM-ACR dataset.

Evaluation and limitations

The study examines patient-isolated segmentation, calibration, error discrimination, risk-coverage behaviour, and projection subgroups. It reports limitations including checkpoint-selection variance, an unreliable fixed operating threshold, and ensemble uncertainty rankings that can perform worse than random. These findings motivate further validation rather than demonstrating readiness for clinical deployment.

Supporting materials

Code and research documentation

This deposit contains the author manuscript sourced from repository commit 023e3314dea2b247d10b0c82d3c786db43b7cd50. It does not include the SIIM-ACR dataset or model weights. No journal acceptance or peer-review status is asserted by this deposit.