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

Area of Science:

• Computational Pathology
• Artificial Intelligence
• Medical Informatics

Background:

• Federated learning (FL) offers a privacy-preserving approach for collaborative deep learning model training.
• Existing research often overlooks practical challenges in real-world federated learning applications, particularly in sensitive domains like computational pathology (CPATH).

Purpose of the Study:

• To transparently document the challenges and practical considerations encountered during the real-world implementation of federated learning for a clinical computational pathology use case.
• To provide insights into overcoming infrastructure, network, and operational hurdles in deploying FL for digital immune phenotyping in metastatic melanoma.

Main Methods:

• A federated learning framework was established using NVIDIA FLARE, involving three clients and a central server across four international institutions.
• Deep learning models were jointly trained for digital immune phenotyping in metastatic melanoma, with attention to system and data heterogeneity.
• Solutions were developed for network restrictions and experiment management, including AWS deployment and optimization of local client epochs.

Main Results:

• The federated learning model demonstrated strong performance across client test sets, though not universally superior to all local models.
• Experiment duration was a significant challenge due to heterogeneity, mitigated by optimizing local client epochs.
• Network restrictions required creative infrastructure solutions, such as deploying the server on AWS.
• Effective experiment management demanded significant IT expertise and familiarity with the NVIDIA FLARE platform.

Conclusions:

• Real-world federated learning implementation in CPATH presents unique, practical challenges beyond simulated environments.
• Transparency in sharing these challenges is crucial for advancing FL adoption in healthcare.
• Further development of best practices and guidelines is needed for successful FL deployment in clinical settings.