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

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 <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.88,0.93$</tex>, and 0.85 and macro F1-scores of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.881,0.842$</tex>, and 0.793 with the lowest site fairness gaps (Site Fairness Index (SFI) <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\approx 0.08$</tex> for accuracy and 0.09 for macro F1; lower is better), a 10-fold CV test accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.874 \pm 0.076$</tex> across all hospitals.