Thesis of Matthis Manthe

Federated machine learning for healthcare applications based on medical imaging

Start date: 01/10/2021
End date (estimated): 01/10/2024

Advisor: Carole Lartizien
Coadvisor: Stefan Duffner


The aim of this PhD is to investigate methodological research in the field of Federated Learning (FL)
for medical image analysis, and more specifically for the design of diagnosis and prognosis models of
brain pathology based on multimodality imaging.
Medical diagnosis or prognosis models are designed to assist clinicians either by highlighting
abnormal regions in an image, predicting a diagnosis or patient outcome. Those models require a
large amount of data to perform well, particularly in the era of large-scale deep neural networks. One
option to increase the training population is to promote multi-center clinical studies, which allow
gathering small to medium size heterogeneous datasets located in different clinical centers. In this
context, federated learning is extremely appealing to counterbalance the need to access large patient
cohorts by the responsibility to maintain the privacy of individual participants.
Current FL approaches generally distribute copies of a machine learning algorithm to the sites or
devices where the data is kept (nodes), performing training iterations locally, and returning the
results of the computation (for example, updated neural network weights) to a central repository to
update the main algorithm. However, simple distributed training does not offer provable privacy
guarantees to satisfy technical safe standards and may reveal information about the underlying
patients. Moreover, training more complex models, such as deep neural networks, distributed over
many sites may considerably increase the number and volume of exchanged messages and traffic and
entail scalability and security issues. Therefore, more effective algorithms of distributed training of
these models need to be developed, and the structure of these models (e.g. neural network
architecture) needs to be adapted to this distributed scenario. Finally, the current trend in machine
learning for medical image analysis is to develop non-conventional deep architectures accounting for
the specificity of medical data, e.g. learning with weak, partial, uncertain annotations, heterogeneous
data (multi-center). The embedding of such complex architectures within the framework of FL is thus
another key challenge to address in this thesis.