Skip to content

Federated Learning for Healthcare

Welcome to the Federated Learning tutorial that will be run in conjunction with multiple conferences!

Federated Learning (FL) is increasingly important in privacy sensitive domains, such as healthcare, where sharing of private/patient data is a barrier to building models that generalize well in the real world and minimize bias.

In this tutorial, we will be presenting the COmprehensive Federated Ecosystem (COFE), which comprises of the following components:

  1. The Generally Nuanced Deep Learning Framework (GaNDLF) - gandlf.org
  2. MedPerf - www.medperf.org
  3. OpenFL - github.com/securefederatedai/openfl
  4. Hugging Face Hub - huggingface.co

In 2021, COFE was used to conduct the largest to-date real world federation, with a network of 71 healthcare institutions around the world, the Federated Tumor Segmentation (FeTS) Initiative [ref]. Furthermore, leveraging the collaborators of this real-world FL initiative, the first ever FL challenge was conducted, which focused on the tumor segmentation task, called The FeTS 2021 challenge [ref], which was conducted again in 2022 [ref]. Taking into consideration the value and the interest of the community in this new paradigm for data private multi-institutional collaborations and building upon our experience, we organize this tutorial on FL for healthcare.

Upcoming Events

Past Events

Organizing Committee

  • Sarthak Pati, Indiana University & MLCommons.
  • Patrick Foley, Intel Corporation.
  • Hasan Kassem, MLCommons.
  • Alex Karargyris, MLCommons.
  • Spyridon Bakas, Indiana University & MLCommons.

Tutorial Description

The aim of this tutorial is to facilitate education on how to perform Federated Learning on both simulated and real-world studies. Tutorial structure focuses on specific clearly indicated parts for beginners and for more advanced attendees. Data scientists of different medical imaging communities (e.g., radiology, pathology) are considered during this tutorial on the opportunities and challenges in developing and using FL for training Al models across institutions using privacy preserving techniques. We plan on covering a spectrum of techniques, from software-based approaches that can be considered a method or a metric (e.g., differential privacy), to hardware-based trusted execution computing environments (TEEs).

The motivation for the tutorial is driven by the need to train and validate deep learning models across data silos, to create models that gain knowledge from diverse patient populations and hence generalize well, mitigate bias, and pave the way towards addressing health disparities.