Improving Multi-Center Generalizability of GAN-Based Fat Suppression using Federated Learning

TL;DR

The paper addresses poor cross-site generalizability of GAN-based synthesis of fat-suppressed knee MRIs from non-FS proton-density sequences. It proposes privacy-preserving federated learning using a FedGAN-enabled aggregation to train a global pix2pix model (U-Net generator with a $70x70$ PatchGAN discriminator) across UMB and FastMRI data. Results show that FL yields external generalizability with SSIM scores comparable to or higher than single-site baselines on external test sets (e.g., $SSIM=0.63 \pm 0.13$ for UMB and $0.58 \pm 0.12$ for FastMRI), demonstrating improved performance over lone-site models. The study suggests that privacy-preserving multi-center collaboration can advance GAN-based FS MRI synthesis toward clinical deployment, while noting limitations from small datasets and the use of a single FL strategy as directions for future work.

Abstract

Generative Adversarial Network (GAN)-based synthesis of fat suppressed (FS) MRIs from non-FS proton density sequences has the potential to accelerate acquisition of knee MRIs. However, GANs trained on single-site data have poor generalizability to external data. We show that federated learning can improve multi-center generalizability of GANs for synthesizing FS MRIs, while facilitating privacy-preserving multi-institutional collaborations.

Figures (2)

• Figure 1: Privacy-preserving multi-center GAN-based synthesis of FS sequences using FL.
• Figure 2: Examples from (a) UMB and (b) FastMRI test sets of ground-truth (cols. 1-2) with their corresponding synthetic FS sequences from the four models (cols. 3-6).