Search

As one of the effective ways of ocular disease recognition, early fundus screening can help patients avoid unrecoverable blindness. Although deep learning is powerful for image-based ocular disease recognition, the performance mainly benefits from a large number of labeled data. For ocular disease, data collection and annotation in a single site usually take a lot of time. If multi-site data are obtained, there are two main issues: 1) the data privacy is easy to be leaked; 2) the domain gap among sites will influence the recognition performance. Inspired by the above, first, a Gaussian randomized mechanism is adopted in local sites, which are then engaged in a global model to preserve the data privacy of local sites and models. Second, to bridge the domain gap among different sites, a two-step domain adaptation method is introduced, which consists of a domain confusion module and a multi-expert learning strategy. Based on the above, a privacy-preserving federated learning framework with domain adaptation is constructed. In the experimental part, a multi-disease early fundus screening dataset, including a detailed ablation study and four experimental settings, is used to show the stepwise performance, which verifies the efficiency of our proposed framework.

Adopting deep learning in early fundus screening images benefits ocular disease recognition and helps patients avoid blindness in recent years. The robust data representation capability of deep learning is derived from numerous data and annotations. However, the fundus images collected from hospitals or institutes have privacy issues and obvious domain gaps, which greatly influence multi-site learning performance. In this work, a learning system with differential privacy and unsupervised domain regularizer is proposed for ocular disease recognition. First, a Laplace randomized mechanism is introduced to keep the privacy of local models and a global model is constructed via a weighted sum process. Second, an unsupervised domain regularizer, which converts the last fully-connected layer into two sub-layers and then adopts an MMD loss in the element-wise layers of source and target domains, is proposed for unsupervised domain adaptation. Numerous experiments, including four different settings, verify the performance in a multi-disease ocular dataset.