Rethinking fairness in medical imaging: Maximizing group-specific performance with application to skin disease...

Area of Science:

• Medical Image Computing
• Artificial Intelligence in Healthcare
• Clinical Diagnostics

Background:

• Current medical image computing models face a fairness-accuracy trade-off, where improving fairness for some groups can decrease accuracy for others.
• This compromise is problematic in clinical settings due to the high stakes of diagnostic errors.

Purpose of the Study:

• To reframe fairness as maximizing diagnostic accuracy for each patient group.
• To introduce a novel data reweighting algorithm, SPARE, to achieve group-specific accuracy without the traditional trade-off.

Main Methods:

• Developed SPARE, a data reweighting algorithm evaluating training samples by utility and group similarity.
• Trained group-specific models leveraging additional computational resources.
• Applied SPARE to skin disease datasets.

Main Results:

• SPARE significantly improved group-specific diagnostic accuracy.
• The algorithm maintained comparable fairness metrics.
• Demonstrated avoidance of the traditional fairness-accuracy trade-off.

Conclusions:

• SPARE offers a practical paradigm for improving clinical reliability in medical image analysis.
• The approach enhances fairness by optimizing accuracy for each patient group individually.
• This method holds promise for more equitable and reliable AI in healthcare.