Medical Facilities Use Federated Learning and AI to Improve Cancer Detection.
Using NVIDIA-powered Federated learning, a panel of experts from leading research institutions and medical facilities in the United States is assessing the effectiveness of federated learning and AI-assisted annotation in training AI models for tumor segmentation.
What Is Federated Learning?
A method for creating more precise, broadly applicable AI models that are trained on data from several data sources without compromising data security or privacy is called Federated learning. It enables cooperation between several enterprises on the creation of an AI model without allowing sensitive data to ever leave their systems.
“The only feasible way to stay ahead is to use Federated learning to create and test models at numerous locations simultaneously. It is a really useful tool.
The team, comprising collaborators from various universities such as Case Western, Georgetown, Mayo Clinic, University of California, San Diego, University of Florida, and Vanderbilt University, utilized NVIDIA FLARE (NVFlare), an open-source framework featuring strong security features, sophisticated privacy protection methods, and an adaptable system architecture, to assist with their most recent project.
The committee was given four NVIDIA RTX A5000 GPUs via the NVIDIA Academic Grant Program, and they were dispersed throughout the collaborating research institutions so that they could configure their workstations for Federated learning. Further collaborations demonstrated NVFLare’s adaptability by using NVIDIA GPUs in on-premises servers and cloud environments.
Federated Learning AI
Remote Multi-Party Cooperation
Federated learning reduces the danger of jeopardizing data security or privacy while enabling the development and validation of more precise and broadly applicable AI models from a variety of data sources. It makes it possible to create AI models using a group of data sources without the data ever leaving the specific location.
Features
Algorithms Preserving Privacy
With the help of privacy-preserving techniques from NVIDIA FLARE, every modification to the global model is kept secret and the server is unable to reverse-engineer the weights that users input or find any training data.
Workflows for Training and Evaluation
Learning algorithms for FedAvg, FedOpt, and FedProx are among the integrated workflow paradigms that leverage local and decentralized data to maintain the relevance of models at the edge.
Wide-ranging Management Instruments
Management tools provide orchestration via an admin portal, safe provisioning via SSL certificates, and visualization of Federated learning experiments using TensorBoard.
Accommodates Well-Known ML/DL Frameworks
Federated learning may be integrated into your present workflow with the help of the SDK, which has an adaptable architecture and works with PyTorch, Tensorflow, and even Numpy.
Wide-ranging API
Researchers may create novel federated workflow techniques, creative learning, and privacy-preserving algorithms thanks to its comprehensive and open-source API.
Reusable Construction Pieces
NVIDIA FLARE offers a reusable building element and example walkthrough that make it simple to conduct Federated learning experiments.
Breaking Federated Learning’s Code
For the initiative, which focused on renal cell carcinoma, a kind of kidney cancer, data from around fifty medical imaging investigations were submitted by each of the six collaborating medical institutes. An initial global model transmits model parameters to client servers in a Federated learning architecture. These parameters are used by each server to configure a localized version of the model that has been trained using the company’s confidential data.
Subsequently, the global model receives updated parameters from each of the local models, which are combined to create a new global model. Until the model’s predictions no longer become better with each training round, the cycle is repeated. In order to optimize for training speed, accuracy, and the quantity of imaging studies needed to train the model to the requisite degree of precision, the team experimented with model topologies and hyperparameters.
NVIDIA MONAI-Assisted AI-Assisted Annotation
The model’s training set was manually labeled during the project’s first phase. The team’s next step is using NVIDIA MONAI for AI-assisted annotation to assess the performance of the model with training data segmented using AI vs conventional annotation techniques.
“Federated learning activities are most difficult when data is not homogeneous across places. Individuals just label their data differently, utilize various imaging equipment, and follow different processes, according to Garrett. “It’s aim to determine whether adding MONAI to the Federated learning model during its second training improves overall annotation accuracy.”
The group is making use of MONAI Label, an image-labeling tool that cuts down on the time and effort required to produce new datasets by allowing users to design unique AI annotation applications. Prior to being utilized for model training, the segmentations produced by AI will be verified and improved by experts. Flywheel, a top medical imaging data and AI platform that has included NVIDIA MONAI into its services, hosts the data for both the human and AI-assisted annotation stages.
NVIDIA FLARE
The open-source, flexible, and domain-neutral NVIDIA Federated Learning Application Runtime Environment (NVIDIA FLARE) SDK is designed for Federated learning. Platform developers may use it to provide a safe, private solution for a dispersed multi-party cooperation, and academics and data scientists can modify current ML/DL process to a federated paradigm.
Maintaining Privacy in Multi-Party Collaboration
Create and verify more precise and broadly applicable AI models from a variety of data sources while reducing the possibility that data security and privacy may be jeopardized by including privacy-preserving algorithms and workflow techniques.
Quicken Research on AI
enables data scientists and researchers to modify the current ML/DL process (PyTorch, RAPIDS, Nemo, TensorFlow) to fit into a Federated learning model.
Open-Source Structure
A general-purpose, cross-domain Federated learning SDK with the goal of establishing a data science, research, and developer community.
