A CycleGAN Model to Synthesize Missing and Unpaired MRI Sequences for Under-Represented Multiple Sclerosis Lesions
Abstract
A CycleGAN model for synthesizing MRI sequences for Multiple Sclerosis (MS) lesions is provided. In particular, MS lesions are strongly under-represented and have shapes, styles, and positions, which could make them confusing with healthy tissues. We provide a fine-tuning procedure to allow MS lesion representation and segmentation also from imaging sequences by which MS lesions are almost canceled out. We have demonstrated that the synthesized images are usable by an automatic strategy to segment MS lesions with good performance, thus making it possible to use the proposed model either as a tool for data augmentation, for training automatic strategies, or for improving the performance of automatic tools when real data, from some imaging sequence, is missing or unavailable.
Cite
Text
Amato et al. "A CycleGAN Model to Synthesize Missing and Unpaired MRI Sequences for Under-Represented Multiple Sclerosis Lesions." European Conference on Computer Vision Workshops, 2024. doi:10.1007/978-3-031-91907-7_16Markdown
[Amato et al. "A CycleGAN Model to Synthesize Missing and Unpaired MRI Sequences for Under-Represented Multiple Sclerosis Lesions." European Conference on Computer Vision Workshops, 2024.](https://mlanthology.org/eccvw/2024/amato2024eccvw-cyclegan/) doi:10.1007/978-3-031-91907-7_16BibTeX
@inproceedings{amato2024eccvw-cyclegan,
title = {{A CycleGAN Model to Synthesize Missing and Unpaired MRI Sequences for Under-Represented Multiple Sclerosis Lesions}},
author = {Amato, Flavio D' and Cipriani, Alessia and Di Matteo, Alessandro and Lozzi, Daniele and Mattei, Enrico and Polsinelli, Matteo and Placidi, Giuseppe},
booktitle = {European Conference on Computer Vision Workshops},
year = {2024},
pages = {270-283},
doi = {10.1007/978-3-031-91907-7_16},
url = {https://mlanthology.org/eccvw/2024/amato2024eccvw-cyclegan/}
}