Federated learning: training machine learning algorithms together without sharing patient data

29 healthcare and research institutions from the Netherlands, the United States, Canada, the United Kingdom, Germany, Switzerland and India deployed federated learning for brain tumour identification. Federated learning is a form of machine learning where agencies collaborate on training algorithms without sharing patient data and compromising patient privacy to do so. The project is being conducted in collaboration with Perelman School of Medicine at the University of Pennsylvania (Penn Medicine) and Intel Labs.

In 2019, 1,338 people suffered a brain tumour, according to data from KWF Cancer Society. These were exclusively malignant brain tumours. 871 people died of a brain tumour in the Netherlands in 2018. The 5-year survival rate is 24%.

Federated learning could potentially improve brain tumour identification This distributed form of machine learning was introduced by Google in 2017. The technology allows collaboration on deep learning projects, without sharing patient data in the process. Based on the approach introduced by Google, Intel Labs and Penn Medicine developed a technology. A partnership of 29 internationals healthcare and research institutions is now using this technology to train AI models together. In the Netherlands, Erasmus MC in Rotterdam is participating in the project.

AI models are trained using federated learning to identify brain tumours. Penn Medicine and Intel Labs published a paper on federated learning in 2018. With this paper, the parties aimed to demonstrate that this technology can train a model with 99% accuracy. This with no impact on data subjects whose data is used to train the models. The new research with this technology uses software and hardware from Intel.

"Our scientific community generally shares the view that machine-learning training requires sufficient and varied data and much more than is available to each separate institution," said lead researcher Dr Spyridon Bakas of the University of Pennsylvania's Center for Biomedical Image Computing and Analytics (CBICA).

"We are coordinating an association of 29 collaborating international healthcare and research institutions to train state-of-the-art AI models for healthcare using privacy-respecting machine-learning technologies, including federated learning. This year, the nexus will begin developing algorithms that identify brain tumours from a greatly expanded version of the International Brain Tumour Segmentation (BraTS) challenge dataset. This collaboration will give medical researchers access to much larger amounts of healthcare data, while ensuring the security of that data."

"AI holds great promise for early detection of brain tumours, but more data than each medical centre has on its own is needed to reach its full potential. Using Intel software and hardware and support from some of the smartest people at Intel Labs, we are working with the University of Pennsylvania and a nexus of 29 medical centres to advance brain tumour identification and protect sensitive patient data," said Jason Martin, chief engineer at Intel Labs.

Penn Medicine will receive $1.2 million to conduct research on federated learning over the next three years. The funding has been made available by the Informatics Technology for Cancer Research programme (ITCR) of the National Cancer Institute (NCI) of the National Institutes of Health (NIH).

The paper on federated learning published by Penn Medicine and Intel Labs in January 2019 can be found here. The blog post in which Google introduced federated learning in 2017 is available here.

Author: Wouter Hoeffnagel