# Navigating real-world challenges: A case study on federated learning in computational pathology

## TL;DR

This paper explores the real-world challenges of using federated learning in computational pathology, highlighting practical issues and solutions for collaborative model training across institutions.

## Contribution

The study provides a transparent case study of federated learning in computational pathology, revealing practical challenges and solutions in a real-world clinical setting.

## Key findings

- The FL model performed best across all clients' test sets but did not outperform local models on individual client test sets.
- System and data heterogeneity caused long experiment durations, which were mitigated by optimizing local client epochs.
- Infrastructure design was limited by network restrictions, resolved by deploying the server on AWS within a semi-public network.

## Abstract

Federated learning (FL) allows institutions to collaboratively train deep learning models while maintaining data privacy, a critical aspect in fields like computational pathology (CPATH). However, existing studies focus on performance improvement in simulated environments and overlook practical aspects of FL. In this study, we address this need by transparently sharing the challenges encountered in the real-world application of FL for a clinical CPATH use case. We set up a FL framework consisting of three clients and a central server to jointly train deep learning models for digital immune phenotyping in metastatic melanoma, utilizing the NVIDIA Federated Learning Application Runtime Environment (NVIDIA FLARE) across four separate networks from institutes in four countries. Our findings reveal several key challenges: First, the FL model performs the best across all clients' test sets but does not outperform all local models on their own client test set. Second, long experiment duration due to system and data heterogeneity limited experiment frequency, alleviated by optimizing local client epochs. Third, infrastructure design was hindered by hospital and corporate network restrictions, necessitating an open port for the server, which we resolved by deploying the server on an Amazon Web Services infrastructure within a semi-public network. Lastly, effective experiment management required IT expertise and strong familiarity with NVIDIA FLARE to enable orchestration, code management, parameter configuration, and logging.

Our findings provide a practical perspective on implementing FL for CPATH, advocating for greater transparency in future research and the development of best practices and guidelines for implementing FL in real-world healthcare settings.