Federated training of segmentation models for radiotherapy treatment planning

A general problem in medical imaging is access to large datasets for training deep models. In this presentation Prof. Anders Eklund will demonstrate how they have trained segmentation models between Linköping, Lund and Umeå through federated learning. Each city exported local data from the oncology department, preprocessed the data in the same way and then setup a computer which joined the federation. The initial results show that the global model performs a bit better than the local models, even if there are differences in the local data.