Fig. 2.

The model performance and fairness gap for identifying Edema from CXR images in different race, age, and sex groups. Each row represents the performance (y-axis) and fairness gap (x-axis) of each evaluation metric. The 95% confidence intervals are calculated from 1000 bootstrap iterations. Each plot represents a different de-bias technique including the baseline model, the proposed augmentation, balanced, stratified, adversarial learning, DistMatchMMD, DistMatchMean, and FairALM. AUC: Area Under the ROC Curve; BCE: Binary Cross Entropy; ECE: Expected Calibration Error.