Geração de imagens de raios-X do tórax utilizando Latent Diffusion Models

Abstract

Medical imaging research is a challenging task due to, but not limited to, the scarcity of medical images. Usually, real data is used to carry out simulations. However, several factors limit the ability to use medical exam data. These include the need to protect sensitive and confidential patient information, the cost and accessibility of medical equipment, as well as ethical and legal issues related to data sharing. To mitigate these challenges, synthetic data generation offers a promising alternative, allowing to complement training datasets and conduct large-scale medical imaging research. Recently, Diffusion Models have gained prominence in the computer vision community by producing photorealistic synthetic images. In this context, this work proposes to explore the use of a specific type of Diffusion Models, the Latent Diffusion Models (LDM), in generating synthetic medical images from high-resolution thoracic images. To do this, we performed a fine-tuning on the Latent Diffusion Foundation Model, using images from the National Library of Medicine. As a contribution, we generated a set of high-resolution, excellent-quality healthy chest x-ray images. Additionally, we provide these generated images alongside the trained models for use in academic research or for training other models. As such, our approach offers the potential to overcome the data barrier in the medical domain.