Table 2 IoU and fairness scores for hip segmentation across different protected attributes, including race, sex, and age.
From: Fair AI-powered orthopedic image segmentation: addressing bias and promoting equitable healthcare
U-Net backbone | Model | White/Caucasian | Black/AA* | SER | SD | |
|---|---|---|---|---|---|---|
(a) Racial group IoU scores & fairness scores | ||||||
ResNet18 | Baseline | 0.876 | 0.867 | 1.070 | 0.004 | |
Balanced | 0.857 | 0.853 | 1.027 | 0.002 | ||
Stratified | 0.864 | 0.857 | 1.054 | 0.004 | ||
Group | 0.854 | 0.851 | 1.019 | 0.001 | ||
EfficientNet-B0 | Baseline | 0.869 | 0.862 | 1.051 | 0.003 | |
Balanced | 0.842 | 0.844 | 1.013 | 0.001 | ||
Stratified | 0.876 | 0.868 | 1.064 | 0.004 | ||
Group | 0.861 | 0.851 | 1.076 | 0.005 | ||
U-Net Backbone | Model | Male | Female | SER | SD | |
|---|---|---|---|---|---|---|
(b) Sex group IoU & fairness scores | ||||||
ResNet18 | Baseline | 0.874 | 0.871 | 1.023 | 0.001 | |
Balanced | 0.855 | 0.854 | 1.010 | 0.001 | ||
Stratified | 0.873 | 0.867 | 1.047 | 0.003 | ||
Group | 0.851 | 0.848 | 1.020 | 0.001 | ||
EfficientNet-B0 | Baseline | 0.869 | 0.864 | 1.033 | 0.002 | |
Balanced | 0.851 | 0.828 | 1.155 | 0.012 | ||
Stratified | 0.870 | 0.864 | 1.046 | 0.003 | ||
Group | 0.843 | 0.858 | 1.101 | 0.007 | ||
(c) Age group IoU & fairness scores | ||||||
|---|---|---|---|---|---|---|
U-Net Backbone | Model | Age 50 or Lower | Age 51–64 | Age 65–79 | SER | SD |
ResNet18 | Baseline | 0.868 | 0.874 | 0.871 | 1.044 | 0.002 |
Balanced | 0.855 | 0.833 | 0.834 | 1.147 | 0.010 | |
Stratified | 0.867 | 0.868 | 0.862 | 1.050 | 0.003 | |
Group | 0.768 | 0.852 | 0.852 | 1.574 | 0.040 | |
EfficientNet-B0 | Baseline | 0.866 | 0.868 | 0.864 | 1.025 | 0.001 |
Balanced | 0.833 | 0.817 | 0.814 | 1.115 | 0.008 | |
Stratified | 0.869 | 0.872 | 0.868 | 1.033 | 0.002 | |
Group | 0.743 | 0.862 | 0.843 | 1.854 | 0.052 | |
