Conversion of the Mayo LDCT Data to Synthetic Equivalent through the Diffusion Model for Training Denoising Networks with a Theoretically Perfect Privacy

Deep learning techniques are widely used in the medical imaging field; for example, low-dose CT denoising. However, all these methods usually require a large number of data samples, which are at risk of privacy leaking, expensive, and time-consuming. Because privacy and other concerns create challenges to data sharing, publicly available CT datasets are up to only a few thousand cases. Generating synthetic data provides a promising alternative to complement or replace training datasets without patient-specific information. Recently, diffusion models have gained popularity in the computer vision community with a solid theoretical foundation. In this paper, we employ latent diffusion models to generate synthetic images from a publicly available CT dataset-the Mayo Low-dose CT Challenge dataset. Then, an equivalent synthetic dataset was created. Furthermore, we use both the original Mayo CT dataset and the synthetic dataset to train the RED-CNN model respectively. The results show that the RED-CNN model achieved similar performance in the two cases, which suggests the feasibility of using synthetic data to conduct the low-dose CT research. Additionally, we use the latent diffusion model to augment the Mayo dataset. The results on the augmented dataset demonstrate an improved denoising performance.