Federated Semi-Supervised Segmentation with Randomized Weight Perturbation

Medical image segmentation is challenging because of lack of annotated data and the reluctance of institutions sharing sensitive patient information. Recent advances in federated learning and semi-supervised learning enable training models with limited labels in a privacy-preserving manner. We propose FedWeP, a method for federated semi-supervised segmentation. FedWeP uses Randomized Weight Perturbation in which the server modulates model weights with Gaussian noise, and disseminates perturbed models to clients for semi-supervised training. Moreover, in order to facilitate consistency regularization at clients, models are perturbed in a controlled way so that the predictions from perturbed models do not deviate excessively, for which we propose similarity-based seed search. Importantly, our method requires only a single model to be exchanged between server and clients, unlike conventional semi-supervised methods using two models, halving the communication costs. We validate our approach using the COVID-19 infection dataset, and demonstrate that FedWeP outperforms the baseline using one model, and is comparable to those using two models, striking a balance between segmentation performance and communication costs.