Preliminary Results on Improved Synthetic Image Generation for Melanoma Skin Cancer.

Abstract

Advances in computer vision have shown interesting results in synthetic image generation. Diffusion models have shown promising outputs while generating realistic images from textual inputs like stable diffusion and Imagen. However, their use in high-quality medical images is limited. Synthetic medical images may play an essential role in privacy-preserved artificial intelligence. In addition, these are also useful for augmenting small datasets. For this reason, this study proposed a stable diffusion-based algorithm for 3-dimensional (3D) skin cancer image generation. The target class in the proposed study is melanoma skin cancer. A lightweight low-rank adaptation technique (LoRA) has been used for training. The proposed approach can efficiently generate 3D images of size 512*512*3. The results have been compared with existing studies to validate the efficacy of the proposed study.