Bridging the Gap: AI-Driven Fusion of Skin Tones in Skin Cancer Datasets

Skin cancer is one of the most frequently diagnosed cancers in the world, and early diagnosis significantly improves outcomes for patients. While machine learning models are increasingly being used to detect skin cancers, these models generally rely on large and diverse image datasets that often lack sufficient representation of diverse skin tones. This work addresses the challenge of building a more inclusive skin cancer image dataset through generating synthetic images covering a wide range of skin tones. We experiment with three types of GANs: Conditional GAN (CGAN), Deep Convolutional GAN (DCGAN) with style transfer, and StyleGAN, augmenting a limited existing dataset. Concerning CGAN and DCGAN, style transfer failed to obtain resultant realistic and diverse skin cancer images whereas StyleGAN was efficient in the generation of high-quality images with much detail retention regarding lesions across all the skin tones. Experimental results indicate that StyleGAN attained the lowest Fréchet Inception Distance (FID) score with the highest Inception Score (IS). This means that this method attains better quality images, diversity, and clinical relevance. This research closes the data diversity gap in medical imaging to produce fairer and more accurate machine learning models for the diagnosis of skin cancers.