Multi-Site Privacy-Preserving Cataract Detection from Fundus Images with EVA-02 Vision Transformers Through Federated Deep Learning

View at publisher Summarise with VIRAT
Authors: Hasan Shahriar Mahmud Mir, Md Mahbubur Rahman, Faysal Ahmed
Cited by: 0
References: 13

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Cataract is one of the main causes of reversible blindness worldwide. Early screening is important to prevent avoidable vision loss. Deep-learning-based models are now widely used in medical imaging. We build the system in two stages. First, we focus on the backbone model selection. We only consider light and tiny CNN and transformer backbones to keep the method computationally efficient and accurate. We design three classification heads (MLP, gated, and attention-based) and combine them with seven candidate backbones. Each head is evaluated with three different activation functions (SiLU, GELU, and Mish). This gives 63 total configurations, all are trained with the same protocol. From these experiments, we select an EVA-02 tiny vision transformer as our base model because it gives the best accuracy and F1-score. In the second phase, we apply Federated learning (FL) using the selected EVA-02 model to preserve data privacy and build a federated cataract detection system. We simulate three clinical sites in our study, each representing a different hospital dataset. We compare six FL algorithms FedAvg, FedProx, FedBN, FedMA, MOON, and FedProc with the same backbone, optimizer and data. Among FL methods, MOON offers the best balance between overall accuracy and fairness across the three hospitals. It achieves per-hospital accuracies of$0.88,0.93$, and 0.85 and macro F1-scores of$0.881,0.842$, and 0.793 with the lowest site fairness gaps (Site Fairness Index (SFI)$\approx 0.08$for accuracy and 0.09 for macro F1; lower is better), a 10-fold CV test accuracy of$0.874 \pm 0.076$across all hospitals.

Fields of Study & Subjects

Retinal Imaging and Analysis Privacy-Preserving Technologies in Data Advanced Neural Network Applications

Publication details

Journal / source
International Conference on Advanced Research in Computing (ICARC)
Pages
1–6
Publication date
18 February 2026
Language
EN