Federated Training of EVA-I2I: Multi-Domain and Multi-Region Translation in Decentralised Scenarios

  • chair:Medical Imaging for Modeling and Simulation
  • type:Master thesis
  • tutor:

    M.Sc. Ciro Benito Raggio

  • person in charge:

    B.Sc. Paul Bidner

  • Motivation
    Federated Learning (FL) enables the collaborative training of machine learning models across multiple institutions without sharing raw data, thereby addressing privacy concerns that are critical in healthcare scenarios. Recent advances in image-to-image (I2I) translation, such as EVA-I2I [1], have shown remarkable performance in multi-domain and multi-region translation tasks, leveraging prompt-based control to generalise even to unseen domains (zero-shot translation). However, EVA-I2I has so far been trained in a centralised setting, which limits its applicability in real-world clinical environments where data are distributed, heterogeneous, and privacy-sensitive.

    Darstellung eines föderierten Lernsystems für Gehirn-MRT-Bildübersetzungen mit EVA I2I Modell.

    How can EVA-I2I be trained in a federated setting while preserving its ability to perform multi-domain translation and generalise to unseen domains?