Scalable Multi-domain Federated Learning with Quality-Aware Prototype Learning for Collaborative Label Transfer in Diabetic Retinopathy Diagnosis

📘 Overview

This repository provides the full implementation of a three-stage federated learning framework for medical image analysis, integrating:

• Federated Self-Supervised Pretraining (FSSL) — MAE-based pretraining across hospitals
• Adapter-based Fine-Tuning — Parameter-efficient personalization with only 12% trainable parameters
• Collaborative Label Transfer (CLT) — Quality-Aware FedProto (QA-FedProto) for unlabeled and late-joining institutions

The framework is designed for parameter-efficient, label-efficient, and domain-generalized diabetic retinopathy diagnosis across heterogeneous clinical institutions.