The advancement of machine learning techniques in critical domains such as medicine requires new approach to data distribution and privacy. Medical datasets are not easily available for research; therefore, we need new algorithmic and infrastructural solutions to overcome this limitation. One of them is Federated Learning (FL), a technique allowing for distributed training of machine learning models without actual data exchange between centres. This new method is gaining momentum in the fields of AI and medicine and more research should performed on its applications in tasks related to medical data processing.

In our talk, we discuss the recent developments in FL for medical applications and compare existing FL frameworks in the context of their usability and reliability. We also present our recent results of large-scale, distributed FL experiments related to chest X-ray (CXR) image segmentation and classification with multi-centre data. Various aspects regarding the impact of differential privacy, data and features distributions on the training process is discussed. Infrastructural details, namely usage of HPC and cloud, are presented. Finally, we present an approach combining segmentation and classification tasks in the FL manner.