Enhancing Skin Cancer Detection Across Skin Tones with Style Transfer Augmentation

Skin cancer, accounts for a significant number of cancer-related deaths, with early detection being critical for improved outcomes. However, existing artificial intelligence (AI) models for skin lesion detection often underperform for individuals with darker skin tones due to biases in dermatology datasets. To address this, we developed a method that incorporates U-Net-based segmentation to isolate skin lesions and arbitrary style transfer to augment the dataset with diverse skin tones. Using convolutional neural networks (CNNs), our system classifies seven different skin diseases from the HAM10000 dataset. Our approach achieved an accuracy of 0.89, improving detection performance across varied skin tones and lesion types.