Table 4.
AUCs of models with and without proposed augmentation in detecting demographic attributes in CXR.
| Aug | Race |
Age |
Sex | |||||
|---|---|---|---|---|---|---|---|---|
| Asian | Black | White | 0–40 | 40–60 | 60–80 | 80- | ||
| w/o | 0.937 [0.931–0.943] | 0.954 [0.951–0.956] | 0.950 [0.947–0.952] | 0.965 [0.963–0.967] | 0.834 [0.830–0.838] | 0.795 [0.791–0.799] | 0.899 [0.896–0.903] | 0.992 [0.992–0.993] |
| w/ | 0.712 [0.700–0.724] | 0.826 [0.820–0.831] | 0.821 [0.816–0.825] | 0.843 [0.837–0.850] | 0.678 [0.673–0.683] | 0.581 [0.576–0.587] | 0.818 [0.813–0.823] | 0.986 [0.985–0.987] |
Lower values indicate a weaker ability to recognise race, age, or sex based on CXR images. The model trained using augmented images is hard to recognise demographic attributes from CXR. Aug.: Augmentation. The minimum values are highlighted in bold.