ST-FL: Style Transfer Preprocessing in Federated Learning for COVID-19 Segmentation

Antonios Georgiadis, Varun Babbar, Fran Silavong, Sean J. Moran, Robert OtterarXiv 2022 · DOI · arXiv

Introduction

Deep learning has demonstrated significant potential in identifying COVID-19 infections from chest CT scans. However, the development of a global AI diagnostic tool faces two primary challenges: data privacy and data heterogeneity.

Privacy regulations, such as GDPR and HIPAA, prevent hospitals from easily sharing patient data. Federated Learning (FL) addresses this by training a central model on local data without the data ever leaving the hospital. However, heterogeneity remains a significant hurdle; CT scanners from different manufacturers or locations produce images with varying brightness, contrast, and noise levels. This "garbage in, garbage out" problem often causes the performance of FL models to degrade when learning from inconsistent data styles simultaneously. This paper introduces ST-FL (Style Transfer Federated Learning), a framework designed to standardize medical images at the local level before they are used in the federated training process.

The ST-FL Framework

The core innovation of ST-FL is the introduction of a generative AI preprocessing step at each client (hospital) node. Instead of sending updates based on raw, inconsistent data, the system standardizes images using CycleGANs.

  • Target Style Standardization: A small, representative set of high-quality CT images (20–50 images) is shared across all clients to serve as a "gold standard" style.
  • Denoising CycleGAN: Each client trains a CycleGAN to map its local, noisy images into the target style, acting as a universal translator for image quality.
  • Dual-Channel Input: To ensure robustness, the segmentation model (a U-Net) utilizes a concatenated input of both the original image and the style-transferred image. This allows the model to retain original data details if the style transfer process fails to capture specific nuances.
  • Client-Specific vs. Universal Approaches: The researchers evaluated two variations: "Universal," where a single denoiser is shared, and "Client-Specific," where each hospital trains a denoiser tailored to its specific hardware noise.

Experimental Setup

The study utilized two primary publicly available COVID-19 CT segmentation datasets:

  1. Coronacases (Synthetic): Clean scans were artificially corrupted with noise patterns (inversion, Gaussian noise, contrast shifts) to simulate diverse hospital environments.
  2. MedSeg (Semi-Synthetic): This dataset features real-world imperfections and noisy labels, providing a rigorous test for the framework's practical applicability.

Results and Analysis

ST-FL demonstrated substantial quantitative improvements across all test scenarios. The framework improved lesion segmentation, as measured by Dice scores, by 5% to 40% depending on the noise type.

Dice Score Improvement across various noise patterns for synthetic and semi-synthetic datasets.

A key finding was that Client-Specific CycleGANs significantly outperformed the Universal approach. Since noise is often hardware-specific, tailored local cleaning proved more effective than a one-size-fits-all filter. Furthermore, the focus on style transfer (brightness, contrast, texture) rather than structural changes ensured that critical medical features, such as lesions, were preserved while standardizing the background.

Comparative Performance

In a 5-client setup using the semi-synthetic dataset, the Client-Specific ST-FL achieved a Dice score of 0.462, compared to 0.390 for standard Federated Learning (FedAvg). Notably, in some tests, ST-FL even outperformed traditional centralized models, suggesting that local denoising acts as a powerful form of data augmentation and regularization.

Detailed performance metrics (Dice and IOU) comparing ST-FL variants against Vanilla FedAvg and Centralised training.

Conclusion and Future Work

While ST-FL marks a significant advancement in privacy-preserving medical AI, limitations remain. The framework struggled with heavy Gaussian noise, which can be so destructive that information is lost during translation. Additionally, while the method corrects for distribution shifts (visual style), it does not yet address structural shifts (anatomical differences). Future research will focus on scaling the framework to unbalanced datasets and addressing structural variations in patient scans.

Summary updated

BibTeX
@misc{georgiadis2022st,
  author = {Antonios Georgiadis and Varun Babbar and Fran Silavong and Sean J. Moran and Robert Otter},
  title = {{ST-FL: Style Transfer Preprocessing in Federated Learning for COVID-19 Segmentation}},
  year = {2022},
  eprint = {2203.13680},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2203.13680},
}
Figure