Intel partners with Penn Medicine to develop brain tumor classifier with federated learning

To coincide with Brain Tumor Awareness Month, Intel today announced the details of a National Institutes of Health-funded program that will leverage AI to identify brain tumors while preserving privacy. Together with the Perelman School of Medicine at the University of Pennsylvania (Penn Medicine), the company will coordinate a federation of 29 international medical centers in the U.S., Canada, U.K., Germany, Switzerland, and India to train AI models using federated learning, a technique that enables sensitive data — in this case, one of the world’s largest brain tumor corpora — to remain at hospitals’ and organizations’ datacenters.

Intel and Penn Medicine’s approach will ideally contribute to the development of a tumor detection model that doesn’t inadvertently expose medical records. There’s an urgent need — according to the American Brain Tumor Association, nearly 80,000 people (including 4,600 children) will be diagnosed with a brain tumor this year, and early intervention can vastly improve health outcomes.

Google researchers proposed federated learning in a technical paper published in 2017, and since then, it’s been cited more than 300 times by research scientists, according to Arxiv.org. Intel and Penn Medicine were among the first to lead research on federated learning in health care, demonstrating it could be used to attain over 99% of the accuracy of a model trained in a traditional, non-private fashion.

Quite simply, federated learning is a technique that trains an AI algorithm across decentralized devices or servers (i.e., nodes) holding data samples without exchanging those samples, enabling multiple parties to build a model without liberally sharing data. A central server might be used to orchestrate the steps of the algorithm and act as a reference clock, or the arrangement might be peer-to-peer (in which case no such server exists). Regardless, local models are trained on local data samples, and the weights are exchanged among the models at some frequency to generate a global model.

This new Intel- and Penn Medicine-led program will build on the initial research, which was presented in 2018 at the International Conference on Medical Image Computing and Computer Assisted Intervention. For its part, Intel says it will tap in-house software and hardware to implement federated learning “in a manner that provides additional privacy protection” to both the machine learning algorithm and the data set.

The subset of institutions expected to participate in the program’s first phase includes the Hospital of the University of Pennsylvania, Washington University in St. Louis, the University of Pittsburgh Medical Center, Vanderbilt University, Queen’s University, Technical University of Munich, University of Bern, King’s College London, and Tata Memorial Hospital. The model their corpora train will be tested against an expanded version of the International Brain Tumor Segmentation challenge data set, which consists of MRI scans focusing on the segmentation of heterogeneous tumors (namely a type called glioma that starts in the glial cells of the brain).

Apple and Google use federated learning to improve the quality of emoji, word, phrase, and music predictions, but the technique is increasingly being applied in the medical domain, where regulations like HIPAA require a level of data anonymization.

Federated learning powers a product from Owkin, a startup company backed by Google investment wing GV, designed to help medical professionals conduct tests and experiments to predict disease evolution and drug toxicity.

Beyond this, the American College of Radiology, Brazilian imaging center Diagnosticos da America, Partners HealthCare, Ohio State University, and Stanford Medicine collaborated to develop a federated learning model using more than 130,000 images from 33,000 mammography studies. Last fall, Nvidia and King’s College London worked together on a federated learning neural network for brain tumor segmentation. And in a paper published in March 2019, researchers at MIT CSAIL, Harvard University Medical School, and Tsinghua University’s Academy of Arts and Design detailed a federated learning model that could anticipate hospital stay and patient mortality.

More recently, Nvidia began working with collaborators to release COVID-19-related models trained with federated learning through the company’s Clara Imaging Software platform.

The Informatics Technology for Cancer Research program of the NIH’s National Cancer Institute furnished the three-year, $1.2 million grant for the Intel and Penn Medicine federation’s effort. Penn Medicine Center’s Dr. Spyridon Bakas will serve as the principal investigator.