Federated Learning boosts the performance of Brain Tumour Segmentation models

• Brain Imaging & Neuro Epidemiology
• Brain Imaging & Neuro Epidemiology, DoPH
• Brain Imaging & Neuro Epidemiology, DoCR

Federated Learning (FL) is quickly establishing itself as a trustworthy technique for collaboratively training shared models while keeping data decentralized and private, a requirement when dealing with medical data.

At the LIH, Dr Olivier Keunen of the BraINE group and his collaborators Dr Ann-Christin Hau and Prof Simone Niclou have long been involved in the Federated Tumour Segmentation (FeTS) initiative, initially led by Dr Spyridon Bakas at the University of Pennsylvania. The study, involving 71 sites across 6 continents, showed the feasibility and benefits of training models with FL in the context of rare brain tumour diseases. The results, published in Nature Communications, have received considerable attention to date.

We are happy to report a second milestone achievement of the consortium, with the publication of the results of the FeTS challenge in Nature Communications again. The challenge benchmarked FL and segmentation algorithms across 32 sites, discussing techniques such as adaptive weight aggregation and client sampling to improve efficiency. The study also reports that while average generalization was good, worst-case performance highlighted data-specific failures, emphasizing the need for multi-site validation in healthcare AI.

“At BraINE, we are proud to be involved in initiatives that are shaping the future of medical AI, addressing some of its key data accessibility challenges,” said Dr Olivier Keunen, head of the BraINE group.

Recently, FeTS coordinated with Duke University, Indiana University and the RANO group, achieving another milestone in the training of models for the challenging task of segmenting brain tumours post therapy. These exciting results will be presented at the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) in South Korea later this year.