This paper explores the use of federated learning (FL) to improve the multi-center generalizability of Generative Adversarial Network (GAN)-based synthesis of fat-suppressed (FS) MRI sequences from non-fat-suppressed proton density (PD) sequences. The authors hypothesize that FL can facilitate privacy-preserving multi-institutional collaborations to collectively train a global model, addressing the poor generalizability of GANs trained on single-site data.
The authors used two datasets: an internal University of Maryland (UMB) dataset and the publicly available FastMRI dataset. They trained four models: a single-site model with UMB data, a single-site model with FastMRI data, a centrally aggregated model with combined UMB and FastMRI data, and a 2-client FL model with distributed UMB and FastMRI data.
The results showed that the FL model exhibited significantly higher performance on external data compared to the single-site models, despite the data heterogeneity between the two datasets. The single-site models had poor generalizability to external data, emphasizing the importance of training GANs with larger multi-institutional datasets.
The authors acknowledge the limitations of their preliminary work, including the sub-optimal performance of the synthetic MRIs due to the small training dataset size, and the use of only the FedGAN strategy for aggregating weights in FL. They plan to address these limitations in future work.
In conclusion, the authors' preliminary results suggest that FL can improve the generalizability of GANs for synthesizing FS knee MRIs in the real-world while preserving patient privacy, representing an exciting step towards synthetic MRIs becoming a clinical reality.
To Another Language
from source content
arxiv.org
Viktige innsikter hentet fra
by Pranav Kulka... klokken arxiv.org 04-12-2024
https://arxiv.org/pdf/2404.07374.pdfDypere Spørsmål