Fed-CoT: Co-teachers for Federated Semi-supervised MS Lesion Segmentation

Federated learning (FL) is an emerging technique for obtaining a global model while ensuring the data privacy of each client, which is particularly significant in protecting the patients’ privacy when conducting medical image analysis. However, previous FL methods for medical images typically assume a fully supervised setting where each client’s data is fully annotated, disregarding the fact that obtaining such extensive annotations may present significant obstacles due to the need for specialized expertise and the associated overhead costs. In this work, we focus on lesion segmentation for brain MRI images and propose a federated semi-supervised framework to address this problem. Formally, we introduce a Federated Co-Teachers algorithm (Fed-CoT) that extends the prevalent Mean Teacher algorithm into the federated learning framework, and demonstrate its effectiveness. Particularly, in Fed-CoT, two teacher models, namely sync-teacher and async-teacher, which capitalize on different weight updating schemes are leveraged to provide informative consistency regularization and to avoid overfitting to the noise of targets generated by a single teacher model. Our experimental results validate the merits of our proposed method and suggest that the federated learning model can benefit from extra data even without annotations. This approach relaxes the requirement for client participation in federated learning, making it easier to deploy in real applications.