A team of experts from leading U.S. medical centers and research institutions is leveraging NVIDIA-powered federated learning to assess the impact of this technology and AI-assisted annotation in training AI models for tumor segmentation.

Federated learning is a method that allows the creation of more accurate and generalizable AI models using data from various sources without compromising data security or privacy. This approach enables multiple organizations to collaborate on developing an AI model without the need for sensitive data to leave their servers.

“With increasing privacy and data management constraints, it’s becoming more challenging to share data between sites and aggregate it in a central location — while imaging AI is advancing faster than research institutes can establish data-sharing agreements,” said John Garrett, an associate professor of radiology at the University of Wisconsin–Madison. “Adopting federated learning to simultaneously build and test models at multiple sites is practically the only way to keep pace. It’s an invaluable tool.”

Garrett is part of the Society for Imaging Informatics and Medicine (SIIM) Machine Learning Tools and Research Subcommittee, which includes clinicians, researchers, and engineers dedicated to advancing AI development and application in medical imaging. NVIDIA, a member of SIIM, has been collaborating with the committee on federated learning projects since 2019.

“Federated learning techniques enhance data privacy and security, complying with privacy regulations like GDPR and HIPAA,” said committee chair Khaled Younis. “Additionally, we observe improved model accuracy and generalization.”

For their latest project, the team — including contributors from Case Western, Georgetown University, the Mayo Clinic, the University of California, San Diego, the University of Florida, and Vanderbilt University — utilized NVIDIA FLARE (NVFlare). This open-source framework boasts robust security features, advanced privacy protection techniques, and a flexible system architecture.

Through the NVIDIA Academic Grant Program, the committee received four NVIDIA RTX A5000 GPUs, distributed across the participating research institutes to set up workstations for federated learning. Additional collaborators used NVIDIA GPUs in the cloud and on-premises servers, demonstrating NVFLare’s flexibility.

Cracking the Code for Federated Learning

Each of the six participating medical centers contributed data from approximately 50 medical imaging studies for the project, which focused on renal cell carcinoma, a type of kidney cancer.

“The concept with federated learning is to exchange the model during training rather than the data,” explained Yuankai Huo, assistant professor of computer science and director of the Biomedical Data Representation and Learning Lab at Vanderbilt University.

In a federated learning framework, an initial global model broadcasts model parameters to client servers. Each server uses these parameters to set up a local version of the model trained on the organization’s proprietary data. Updated parameters from each local model are then sent back to the global model, where they are aggregated to produce a new global model. This cycle continues until the model’s predictions no longer improve with each training round.

The group experimented with different model architectures and hyperparameters to optimize training speed, accuracy, and the number of imaging studies needed to achieve the desired level of precision.

AI-Assisted Annotation With NVIDIA MONAI

In the initial phase of the project, the training data for the model was labeled manually. For the next phase, the team is utilizing NVIDIA MONAI for AI-assisted annotation to compare how model performance varies with training data segmented using AI versus traditional annotation methods.

“The primary challenge with federated learning activities is that data from different sites is not highly uniform. Different imaging equipment, protocols, and labeling methods are used,” Garrett said. “By retraining the federated learning model with the addition of MONAI, we aim to determine if this improves overall annotation accuracy.”

The team is employing MONAI Label, an image-labeling tool that allows users to create custom AI annotation apps, reducing the time and effort required to develop new datasets. Experts will validate and refine the AI-generated segmentations before using them for model training.

Data for both the manual and AI-assisted annotation phases is hosted on Flywheel, a leading medical imaging data and AI platform that has integrated NVIDIA MONAI into its offerings.

Upon project completion, the team plans to publish their methodology, annotated datasets, and pretrained model to support future research.

“Our goal is not only to explore these tools,” Garrett said, “but also to publish our work so that others in the medical field can learn and utilize these tools.”

Apply for an NVIDIA Academic Grant

The NVIDIA Academic Grant Program supports academic research by providing world-class computing resources to researchers. Applications are now open for full-time faculty members at accredited academic institutions using NVIDIA technology to accelerate projects in simulation and modeling, generative AI and large language models.

Future application cycles will focus on projects in data science, graphics and vision, and edge AI, including federated learning.

For more information, visit the NVIDIA Blog.

For more Information, Refer to this article.