Advanced image generation for cancer using diffusion models

Abstract

Deep neural networks have significantly advanced medical image analysis, yet their full potential is often limited by the relatively small dataset sizes. Generative modeling has stimulated attention for its potential applications in the synthesis of medical images. Recent advancements in diffusion models have exhibited a remarkable capacity for producing photorealistic images. Despite this promising development, the application of such models in the generation of medical images remains underexplored. In this study, we explored the potential of using diffusion models to generate medical images, with a particular emphasis on producing brain magnetic resonance imaging (MRI) scans, such as those depicting low-grade gliomas. Additionally, we examined the generation of contrast enhanced spectral mammography (CESM) images, as well as chest and lung X-ray images. Utilizing the Dreambooth platform, we trained stable diffusion models based on text prompts, class and instance images, subsequently prompting the trained models to produce medical images. The generation of medical imaging data presents a viable approach for preserving the anonymity of medical images, effectively reducing the likelihood of patient re-identification during the exchange of data for research. The findings of this study reveal that the application of diffusion models in generating images successfully captures attributes specific to oncology within imaging modalities. Consequently, this research establishes a framework that harnesses the power of artificial intelligence for the generation of cancer medical imagery.