Federated Deep Learning for Classifying Glaucomatous Optic Neuropathy from Optical Coherence Tomography Volumetric Scans: A Privacy-preserving Multi-national Study

Abstract

Purpose : We aim to develop privacy-preserving deep-learning (DL) models with federated learning (FL), a technique taking advantage of datasets across multiple “clients” (i.e., different end-users or centers) without centralizing or sharing data, to classify glaucomatous optic neuropathy (GON) from 3D optical coherence tomography (OCT) volumetric scans.

Methods : This is a multi-national study. We collected datasets of OCT scans from 7 eye centers. Each dataset was a “client” and split into training, tuning, and testing sets with a ratio of 7:1:2. We experimented with three kinds of DL architectures to develop 3D FL-Models. Figure 1 illustrates the FL process, which consists of a “central server” and 7 local clients. The central server maintains a “Global Model” and coordinates clients’ updates on their local models. To be specific, each client trained locally on its own training set and then updated the model parameters to the central server after one training epoch. The central server aggregated the updates from each local model to refine the Global Model and then redistributed the updated Global Model to all the clients. Subsequently, each client continued fine-tuning locally with its tuning set based on the updated Global Model. This process repeated back and forth until the Global Model converged which was then tested on each client’s testing set. For performance comparison, we used all data from 7 clients to develop Joint-Models which served as the upper bound and were tested on the same testing sets as FL-Models.

Results : We used 8,436 volumetric scans from 2,192 patients (Table 1). FL-Models developed by three architectures achieved the area under the receiver operating curve (AUROC) values with ranges of 0.784-0.993, 0.805-0.996, and 0.809-0.991 in 7 clients, respectively. Joint-Models achieved AUROC values of 0.775-0.997, 0.807-0.999, and 0.822-0.996, respectively. In each client, the FL-Models had accuracy, sensitivity, and specificity similar to corresponding Joint-Models (Table 2).

Conclusions : The 3D FL-Models showed performance non-inferior to Joint-Models and achieved good generalizability without sharing any patient data among multiple centers. Our results demonstrated that the FL technique can ensure data security and enhance the feasibility of implementing DL-based OCT analysis to identify GON in real-world clinics.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.