Researchers from NVIDIA and King’s College London have collaborated on a new federated learning system specifically designed for the interpretation of medical images.
Federated learning involves training a deep neural network with data from multiple healthcare providers, giving participants the ability to work together on a shared model without actually sharing patient data. The practice is viewed as a much more secure way to train neural networks.
“Each node trains its own local model and, periodically, submits it to a parameter server,” the team explained in its paper, as quoted in a prepared statement from NVIDIA. “The server accumulates and aggregates the individual contributions to yield a global model, which is then shared with all nodes.”
The researchers worked with a dataset that included MRI scans more than 250 patients with brain tumors. They found that their approach resulted in “a comparable segmentation performance” that kept the institutions’ data more secure.
“Federated learning has the potential of effectively aggregating knowledge across institutions learned locally from private data, thus further improving the accuracy, robustness, and generalization ability of the deep models,” according to the statement from NVIDIA.
Researchers who worked on this project will be sharing their findings at MICCAI 2019 in Shenzhen, China.