Semi-Supervised Federated Learning for Equine Sesamoiditis Grading

Deep Learning (DL) shows promise for automated equine sesamoiditis grading, potentially improving equine healthcare. However, DL model training typically requires large amounts of labeled data, which demands centralized data collection and expert annotation, raising privacy and resource concerns. In view of this, we propose a Semi-Supervised Federated Learning framework for collaborative grading of equine sesamoiditis (SSFLgrading), designed for the scenario where only the server has labeled data while clients possess unlabeled data. In this framework, the server trains a DL model using private labeled data and generates pseudo labels for client data. Clients then train the model using these pseudo labels. On the server side, we formulate the grading problem as a regression task and introduce a designed ordinal loss to help the model better capture the ordinal nature of grading levels. On the client side, pseudo labels are often noisy due to data heterogeneity between server and client datasets. To mitigate this, we introduce a dual-selection strategy, which selects pseudo-labeled data using a proposed adaptive acceptance threshold-based pseudo labeling approach, and selectively updates specific components of the model. We evaluate the proposed framework on an internally collected equine sesamoiditis dataset and further validate its versatility on a public diabetic retinopathy grading dataset. The experimental results demonstrate the effectiveness of our framework in semi-supervised federated grading tasks.