Enhancing Skin Cancer Detection Using Synthetic Dermoscopic Images via Stable Diffusion

Skin cancer is a major medical problem. Its early diagnosis is essential for successful treatment outcomes and effective reduction in mortality rates. Disease diagnosis is mainly performed using image data. Optimizing a robust and accurate system to detect skin tumors needs complex deep learning (DL) models and huge datasets of high-quality data. Due to medical data privacy, the medical data are not sufficiently available. Data augmentation using generative artificial intelligence techniques such as generative adversarial network (GAN) and diffusion models (DM) is state-of-the-art for generating high-quality synthetic images to improve the classifier training process. To the best of our knowledge, the topic of skin cancer image generation has not been sufficiently explored. In this study, we used the technique Low Rank Adaptation (LoRA) to fine-tune the stable diffusion DL architecture with the HAM10000 dataset to generate synthetic images of skin cancer, which could help improve the accuracy of cancer detection classifiers. With Frchet inception distance of 41.968, results indicate that the generated images are of high quality and very similar to the real images than those in the literature studies. We tested the generated images to train and optimize a collection of cancer detection classifiers. All classifiers consistently perform better after being trained using a combination of real and synthetic images. In addition, the generated images have been tested by an expert dermatologist. Generated images are expected to improve the accuracy of cancer detection and help reduce the medical and economic challenges of this dangerous disease.