Home / Current Issue / Paper 1722521
Robustness of Generative Models in Radiological Image Synthesis: A Comparative Analysis of StyleGAN3 and Denoising Diffusion under Perturbations
Subject area: Science,Engineering and Technology · Area of research: Radiological Image Synthesis
Abstract
Synthetic radiological images can augment scarce datasets, but only if they remain faithful when the imaging pipeline is perturbed. We compare StyleGAN3 against a denoising diffusion probabilistic model (DDPM) on chest X-ray and brain MRI synthesis under additive noise, motion blur, JPEG compression, and a challenging MRI-to-CT domain shift. Diffusion sampling proves markedly more robust, preserving structural similarity and downstream classifier accuracy where the GAN's outputs distort, at the cost of slower generation.
How to cite this paper
Elena Roskova, Tariq Hassan, Meera Iyer "Robustness of Generative Models in Radiological Image Synthesis: A Comparative Analysis of StyleGAN3 and Denoising Diffusion under Perturbations" Iconic Research And Engineering Journals Volume 6 Issue 6 2022 Page 498-502
Elena Roskova, Tariq Hassan, Meera Iyer "Robustness of Generative Models in Radiological Image Synthesis: A Comparative Analysis of StyleGAN3 and Denoising Diffusion under Perturbations" Iconic Research And Engineering Journals, vol. 6, no. 6, Dec. 2022
Elena Roskova, Tariq Hassan, Meera Iyer (2022). Robustness of Generative Models in Radiological Image Synthesis: A Comparative Analysis of StyleGAN3 and Denoising Diffusion under Perturbations. Iconic Research And Engineering Journals, 6(6).
Elena Roskova, Tariq Hassan, Meera Iyer "Robustness of Generative Models in Radiological Image Synthesis: A Comparative Analysis of StyleGAN3 and Denoising Diffusion under Perturbations" Iconic Research And Engineering Journals, vol. 6, no. 6, Dec. 2022.
@article{1722521,
author = {Elena Roskova, Tariq Hassan, Meera Iyer},
title = {Robustness of Generative Models in Radiological Image Synthesis: A Comparative Analysis of StyleGAN3 and Denoising Diffusion under Perturbations},
journal = {Iconic Research And Engineering Journals},
year = {2022},
volume = {6},
number = {6},
pages = {498-502},
issn = {2456-8880},
url = {https://www.irejournals.com/formatedpaper/1722521.pdf},
abstract = {Synthetic radiological images can augment scarce datasets, but only if they remain faithful when the imaging pipeline is perturbed. We compare StyleGAN3 against a denoising diffusion probabilistic model (DDPM) on chest X-ray and brain MRI synthesis under additive noise, motion blur, JPEG compression, and a challenging MRI-to-CT domain shift. Diffusion sampling proves markedly more robust, preserving structural similarity and downstream classifier accuracy where the GAN's outputs distort, at the cost of slower generation.},
month = {December},
}