Diffusion Models for Realistic CT Image Generation
Egileak: Maialen Stephens Txurio Andrés Marcos Pilar Castellote José M. Santabárbara Miguel Ángel González
Data: 14.06.2023
Abstract
Generative networks, such as GANs, have been applied to the medical image domain, where they have demonstrated their ability to synthesize realistic-looking images. However, training these models is hard, as they show multiple challenges like mode collapse or vanishing gradients. Recently, diffusion probabilistic models have revealed to outperform GANs in the context of natural images. However, only a few early works have attempted their use in medical images. Hence, in this work we aim at evaluating the potential of diffusion models to generate realistic CT images. We make use of the Frechet Inception Distance to quantitatively evaluate the trained model at different epochs, as a metric to provide information about the fidelity of the synthetic samples, displaying a clear convergence of the value in the advance of the training process. We quantitative and qualitatively show to successfully achieve coherent and precise CT images, compared to real images.
BIB_text
title = {Diffusion Models for Realistic CT Image Generation},
pages = {335-344},
keywds = {
CT synthesis; Deep learning; Diffusion probabilistic model; Image generation; Medical imaging
}
abstract = {
Generative networks, such as GANs, have been applied to the medical image domain, where they have demonstrated their ability to synthesize realistic-looking images. However, training these models is hard, as they show multiple challenges like mode collapse or vanishing gradients. Recently, diffusion probabilistic models have revealed to outperform GANs in the context of natural images. However, only a few early works have attempted their use in medical images. Hence, in this work we aim at evaluating the potential of diffusion models to generate realistic CT images. We make use of the Frechet Inception Distance to quantitatively evaluate the trained model at different epochs, as a metric to provide information about the fidelity of the synthetic samples, displaying a clear convergence of the value in the advance of the training process. We quantitative and qualitatively show to successfully achieve coherent and precise CT images, compared to real images.
}
isbn = {978-981993310-5},
date = {2023-06-14},
}
