A feasibility study for federated learning in the context of classification of biomarkers from optical coherence tomography (OCT) images

Abstract

Recent advancement in machine learning and deep learning requires centralized data for training. Federated Learning (FL) is a machine learning approach that deals with collaboratively training a model while keeping the training data decentralized. The introduction of FL reduces the privacy risk (data sharing) and the cost of memory usage from the traditional centralized approach. We developed a scalable baseline FL framework based on PyTorch, incorporated in a Docker container. We distributed our training data equally to our client servers and deployed the Docker container to train our models. This research paper focuses on creating a baseline FL workflow for OCT biomarkers classification to lower the risk inherent to centralized medical data. The best prediction accuracy (macro average F1-score) obtained from the FL approach (72.5%) is closer when compared to our centralized approach (73.6%).