DermNet: integrative CNN-ViT architecture for bias mitigation in dermatological diagnostics using advanced...

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin disease identification algorithms often exhibit bias against darker skin tones.
• Accurate dermatological diagnosis is crucial for effective treatment and patient outcomes.
• Existing methods struggle to segment skin lesions accurately across diverse skin colors.

Purpose of the Study:

• To develop an AI-driven method for unbiased skin disease identification in individuals with darker skin.
• To improve the accuracy and fairness of automated dermatological diagnostic tools.
• To address the limitations of current segmentation techniques in diverse skin tones.

Main Methods:

• A lesion-only, zero-shot, unsupervised approach was employed for initial analysis.
• The processed data was fed into a hybrid classifier, Dermnet, combining Vision Transformer and Convolutional Neural Network architectures.
• A novel segmentation technique, trained without labeled data, achieved high Intersection over Union (IOU) scores.

Main Results:

• The Dermnet classifier achieved a robust validation accuracy of approximately 81%.
• The unsupervised segmentation method attained an IOU of 90% across all skin colors.
• The approach effectively minimized the influence of skin tone on disease classification.

Conclusions:

• The proposed method significantly reduces bias in AI-based skin disease identification for people of color.
• The advanced segmentation technique ensures accurate lesion identification irrespective of skin pigmentation.
• This work paves the way for more equitable and accurate dermatological AI tools.