A panel of experts from leading medical institutions and research organizations in the U.S. is leveraging NVIDIA’s federated learning technology to assess the effectiveness of AI-assisted annotation for training models that identify tumors.
Federated learning is a cutting-edge method that enables the creation of more precise and adaptable AI models by utilizing data from various sources while ensuring data security and privacy. This approach facilitates collaboration among institutions in developing AI models without transferring sensitive data off their servers.
“As privacy and data management challenges escalate, sharing data across sites and consolidating it in one location is becoming increasingly complex — and the pace of imaging AI advancements is outstripping the ability of research institutions to establish data-sharing agreements,” stated John Garrett, an associate professor of radiology at the University of Wisconsin–Madison. “Utilizing federated learning for simultaneous model development and testing at several sites is practically essential to keep pace. It is an invaluable tool.”
Garrett is involved with the Society for Imaging Informatics and Medicine (SIIM) Machine Learning Tools and Research Subcommittee, consisting of clinicians, researchers, and engineers dedicated to enhancing AI’s role in medical imaging. NVIDIA has partnered with this committee on federated learning initiatives since 2019.
“Federated learning methods promote heightened data privacy and security by adhering to regulations like GDPR and HIPAA,” explained committee chair Khaled Younis. “Moreover, these techniques lead to improved accuracy and generalization of models.”
To kick off their current 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), an open-source framework offering robust security features, advanced privacy safeguards, and a flexible architecture.
Supported by the NVIDIA Academic Grant Program, the committee was awarded four NVIDIA RTX A5000 GPUs, which were allocated to various research institutes to facilitate their federated learning setups. Additional collaborators made use of NVIDIA GPUs in both cloud and on-premises servers, showcasing the adaptability of NVFLare.
Decoding Federated Learning
Each of the six participating medical centers contributed data from approximately 50 medical imaging studies, primarily focusing on renal cell carcinoma, a type of kidney cancer.
“The fundamental principle of federated learning is that during training, we share the model rather than the data,” explained Yuankai Huo, an assistant professor of computer science and director of the Biomedical Data Representation and Learning Lab at Vanderbilt University.
In a federated learning framework, the initial global model shares parameters with client servers. Each server then configures a local version of the model, trained on the specific organization’s proprietary data. The updated parameters from all local models are sent back to the global model, where they are aggregated to refine the global model. This process continues until further training no longer enhances the model’s predictive accuracy.
The team explored various model architectures and tuning parameters to enhance training speed, accuracy, and the number of imaging studies needed to achieve the desired precision.
AI-Assisted Annotation Using NVIDIA MONAI
In the project’s initial phase, the model’s training data was labeled manually. In the subsequent phase, the team is incorporating NVIDIA MONAI for AI-assisted annotation to compare model performance with AI-segmented training data against traditional labeling methods.
“A major challenge faced in federated learning is the lack of uniformity in data across different sites. Institutions use various imaging technologies, adhere to differing protocols, and label their data in disparate ways,” noted Garrett. “Through retraining the federated learning model with the inclusion of MONAI, we hope to determine whether this enhances overall annotation accuracy.”
The team is employing MONAI Label, a tool for image labeling that allows users to create customized AI annotation applications, thereby simplifying and expediting the creation of new datasets. Experts will review and refine the AI-generated segmentations before these are implemented for model training.
Data from both the manual labeling and AI-assisted annotation stages is managed on Flywheel, a leading platform for medical imaging data and AI that has integrated NVIDIA MONAI into its services.
Upon the completion of the project, the team intends to release their methodology, annotated datasets, and pretrained models to further academic and professional pursuits.
“We aim not only to explore these tools,” Garrett stated, “but also to disseminate our findings so others can leverage these tools across the medical field.”
Apply for an NVIDIA Academic Grant
The NVIDIA Academic Grant Program promotes academic research by offering access to top-tier computing resources and technologies to researchers. Applications are currently open for full-time faculty at accredited academic institutions utilizing NVIDIA technology in projects related to simulation and modeling and generative AI and large language models.
Future application rounds will focus on initiatives in data science, graphics and vision, and edge AI, including federated learning.
.
