Enhancing Fairness in Skin Lesion Classification for Medical Diagnosis Using Prune Learning

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models enhance skin lesion classification accuracy for medical diagnoses.
• Potential biases in models related to skin color can negatively impact diagnostic outcomes.
• Ensuring fairness is complex due to skin tone variability and computational demands.

Purpose of the Study:

• To propose a fairness algorithm for skin lesion classification that addresses bias across diverse skin tones.
• To improve diagnostic fairness and model efficiency in AI-powered dermatology tools.

Main Methods:

• Developed a fairness algorithm calculating feature map skewness in VGG and ViT networks.
• Reduced skin tone-related channels to focus on lesion areas.
• Applied the method to VGG11 and ViT-B16 models.

Main Results:

• Achieved 15-20% improvement in fairness metrics on average.
• Maintained accuracy and F1-score within 0.01 of the baseline.
• Reduced model size by 16% (VGG11) and memory footprint (ViT-B16).

Conclusions:

• The proposed algorithm effectively mitigates bias in skin lesion classification across different skin tones.
• The method enhances fairness and efficiency without requiring skin tone labels at inference.
• This approach offers a practical solution for equitable AI in dermatology.