Medical Image Generation based on Latent Diffusion Models

The application of deep learning networks in med-ical image analysis has become increasingly mature. However, the availability of a large amount of medical image data for training deep learning networks is hindered by various factors such as workload of doctors, different devices, and ethical privacy policies. These factors can limit the full potential of deep learning network models. In recent years, generative models, particularly diffusion models, have made significant progress in synthesizing realistic images in various domains. Nevertheless, there is currently limited research on the application of Latent Diffusion Models (LDMs) in medical image generation, especially when generating different types of medical images using a unified process. In our research, we explore the use of LDMs to generate synthetic images from various medical image datasets. We train the LDMs on the BreastMRI, Hand, HeadCT, and CXR images extracted from the MedNIST dataset. By employing a consistent LDMs architecture, we aim to generate images from different medical device sources. Additionally, we assess the similarity between the generated images and the real images to evaluate the performance of our approach.