PRLx-GAN

Repository for Synthetic Generation and Latent Projection Denoising of Rim Lesions in Multiple Sclerosis published in Synthetic Data at CVPR 2025.

Summary

Paramagnetic rim lesions (PRLs) are a rare but highly prognostic lesion subtype in multiple sclerosis, visible only on susceptibility ($\chi$) contrasts. This work presents a generative framework to:

Synthesize new rim lesion maps that address class imbalance in training data
Enable a novel denoising method to resolve radiologist disagreements on noisy labels, "ambiguous rim lesions".

Latent projection iterations to identify the denoised rim lesion from a "noisy" rim lesion.

Uncurated synthetic rim lesion susceptibilities can be found in png
Pretrained weights are located in net

Preliminaries

To download the pretrained weights, ensure Git Large File Service is installed

sudo apt-get install git-lfs
git lfs install

The main.sh script will skip retraining unless \your\QSM\data is replaced by a valid path

Installation

Clone the repository with

git clone https://github.com/agr78/PRLx-GAN.git

Navigate to the repository

cd PRLx-GAN

Run the setup script

source ./src/main.sh

Wait...then check the generated and denoised outputs

cd ./out

Publications

If this code is used, please cite the following:

Conference Paper: A. G. Roberts et al., "Synthetic Generation and Latent Projection Denoising of Rim Lesions in Multiple Sclerosis," Synthetic Data for Computer Vision at CVPR, 2025.

BibTex

@inproceedings{
roberts2025synthetic,
title={Synthetic Generation and Latent Projection Denoising of Rim Lesions in Multiple Sclerosis},
author={Alexandra Grace Roberts and Ha Manh Luu and Mert Sisman and Alexey V. Dimov and Ceren Tozlu and Ilhami Kovanlikaya and Susan Gauthier and Thanh D. Nguyen and Yi Wang},
booktitle={Synthetic Data for Computer Vision Workshop @ CVPR 2025},
year={2025},
url={https://openreview.net/forum?id=wFkiqB5spT}
}

Acknowledgements

This method relies on the StyleGAN2-ADA architecture developed by @tkarras.

Contact

Please direct questions to Alexandra Roberts at [email protected].

agr78
/

PRLx-GAN