MORE | Management of Real-time Energy Data - 2nd Position in Federated Tumour Segmentation Challenge
FeTS is one of the largest Federated Learning initiatives (https://www.med.upenn.edu/cbica/fets/), where organisations with access to large databases of MRI scans seek to learn a model collaboratively. There are many challenges for learning a common model across such large databases; for instance, the datasets across the organisations are obtained under varying circumstances resulting in widely different distributions and sizes. Moreover, even the models trained in controlled settings may fail when tested in the wild due to drifts or distributional changes.
To this end, FeTS 2022 at MICCAI presented an open challenge to devise algorithms that can learn useful models in real-life settings with data distributed across 23 different institutions. IBM Research Europe (Ireland) participated in this challenge and devised a novel algorithm for robust federated learning across large datasets with non-iid distributions and secured 2nd rank in the competition. Their algorithm uses carefully designed optimisation methods that help arrive at effective models under computational constraints.
Team Members - Ambrish Rawat, Giulio Zizzo, Swanand Kadhe, Jonathan Epperlein, Stefano Braghin
(Ambrish and Giulio were supported by MORE)
FeTS event coordinators - Ujjwal Baid, Spyridon Bakas
FeTS 2022 Competition : http://miccai2022.fets.ai
Coverage in MICCAI Digest - https://www.rsipvision.com/MICCAI2022-Tuesday/26
Read the paper, which will be part of the MICCAI proceedings.
24 Oct 2022
