Logo image
Back
Federated learning for optic nerve axon segmentation in mice
Conference proceeding

Federated learning for optic nerve axon segmentation in mice

Durjoy Deb Dhruba, Adam Hedberg-Buenz, Ashelyn Mann, Michael G. Anderson and Mona Garvin
Vol.13925, pp.1392527-1392527-9
04/03/2026
DOI: 10.1117/12.3086498

View Online

Abstract

Optic nerve damage is a key pathological feature of glaucoma and is closely associated with vision loss in animal models. Experts often examine the axons of retinal ganglion cells (RGCs) in the optic nerve to assess the extent of this damage. Traditional manual approaches to measuring axon count and condition are time-consuming and subject to bias. Although centralized deep learning (DL)-based solutions have been explored, they are often limited by small datasets and data-sharing challenges across sites. To address this, we investigated the use of Federated Learning (FL) for decentralized training of DL models to segment RGC axons from images of histological cross sections of optic nerve. We implemented two FL setups using nnU-NetLite architectures across multiple clients (each representing a separate dataset). In the first setup, two clients were trained on structurally normal axons from different disease models, each presenting mild to moderate nerve damage. In the second setup, a third client was added, trained exclusively on severely damaged axons with a distinct segmentation objective. Only model weights were shared after each epoch without sharing data. The first two clients appeared to benefit from the global FL model, showing a trend toward higher Dice scores compared to their independently trained local models. When their datasets were combined for centralized training, performance was comparable to the federated setup, indicating that FL can offer similar segmentation quality without direct data sharing. In contrast, the third client trained on severely damaged axons showed reduced performance under global model sharing, highlighting the impact of task heterogeneity in FL. However, applying personalized local fine-tuning substantially restored performance for this client. These findings suggest that FL can be effective when clients share similar segmentation objectives, even if the underlying disease models differ. On the other hand, its performance may decline when segmentation tasks differ substantially between clients, motivating hybrid or personalized FL strategies.

Details

Metrics

1 Record Views
Logo image