Nvidia Utilizes Federated Learning to Develop Medical Imaging AI

2019-10-14

Federated learning is a machine learning approach that eliminates the need for creating a single data lake for training models when using client-server architecture. It allows models to be trained on local devices and then aggregates insights from multiple machines into a central model. Researchers from NVIDIA and King's College London have utilized federated learning to train neural networks for tumor segmentation, a technique that enables data sharing between hospitals and researchers while protecting patient privacy. NVIDIA claims that this is a significant milestone in medical image analysis. Abdul Halabi, Director of Healthcare at NVIDIA, stated in a phone interview, "This innovation has two applications. One of them, which we released in August last year, is the best general model available. The other one, which we created earlier, is characterized by its robustness, good universality, and potential for scalability." Halabi concluded, "Most completed experiments have been based on synthetic or random datasets, but this experiment uses real hospital data. The biggest challenge with using the BRATS challenge is privacy." The potential impact of machine learning in healthcare is already evident today, with some computer vision systems demonstrating capabilities surpassing those of human radiologists. However, due to the extensive privacy requirements in healthcare, large datasets of hundreds of thousands of cases are not always available. This is why many researchers in this field choose to use synthetic datasets or specially curated datasets to address experimental challenges. The research conducted by NVIDIA and King's College London is an important step towards deploying secure federated learning, exploring the limitations of differential privacy to enhance the security of collaborative learning models. The study shows that without differential privacy, neural networks can still gain insights from exposed underlying data. Nicola Reike, Senior Researcher at NVIDIA, said, "By improving the way each participating node is modified, we enable the model to limit the granularity of information actually shared between institutions while storing updated data." The major contribution of this research is the ability to aggregate a small fraction of models, achieving the same accuracy and performance as models with a larger share. In recent medical imaging news, the ACR Data Science Institute announced that they will integrate NVIDIA's Clara AI toolkit into their platform this spring, which is a significant endorsement.

RECENT AI NEWS

RECENT AI TOOLS