Healthcare and Life Sciences

Accelerating the Radiological Workflow With AI

Objective

University of Wisconsin–Madison uses NVIDIA DGX BasePOD™ with MONAI and NVIDIA FLARE™ (Federated Learning Application Runtime Environment) to generate comprehensive datasets and rapidly iterate on AI models to bring AI into clinical tools more quickly.

Customer

University of Wisconsin–Madison

Use Case

Accelerated Computing Tools & Techniques

Products

NVIDIA AI Enterprise
NVIDIA DGX
NVIDIA Base Command

AI Shortens Reading Time, Improves Diagnostic Accuracy, and Enables Earlier Detection of Disease

The University of Wisconsin–Madison (UW–Madison) Department of Radiology has been developing innovative imaging technologies for decades, applying research to daily clinical practice to improve human health. Their researchers wanted to leverage AI to speed up tedious tasks in radiologic interpretation and gain new information from the interpretation. The assessment of pediatric bone age, for instance, is a manual process that requires radiologists to match X-rays with images in a standard atlas of bone development, which is based on data from large numbers of other kids of the same gender and age. Finding the right image can help them measure bone age in years and plan appropriate care for their patient.

UW–Madison also wanted to use AI to improve patient outcomes via opportunistic screening, but limited data and disparate data sources created challenges. Opportunistic screening happens when you take additional, indirect findings from imaging studies (such as CT scans), to find out more about a patient’s future health. For example, using a given patient’s combination of wide-ranging health test data, you can apply a variety of automated tools to assess bone mineral density, aortic calcifications, and different types of fat to predict long-term health. This would also help improve patient triaging, like flagging when a CT should get bumped up a priority list.

Facing Issues of Limited Data and Multiple Data Sources

"We realized that in order to integrate AI into our workflows, we need to have two things: a lot of data and the right computing infrastructure," says John Garrett, PhD, assistant professor and director of imaging informatics at the Department of Radiology. "In healthcare, you’re faced with limited data, and in addition, the data you have might not be representative of the population. Furthermore, data often comes from disparate data sources, be it multiple vendors or from different systems like PACS, EMR, or radiology dictation software. You’re also faced with irreproducibility of studies and subjectivity of interpretation due to no ground-truth analysis, as radiologists' interpretations are often not black and white.

“Beyond the concerns for limited, imbalanced data, as well as the need for infrastructure to handle large, complex data, it was equally important for us to have tools to make AI training easy, portable, and reproducible as we roll out our research into daily clinical practice.”

Products Used

  • NVIDIA DGX BasePOD for healthcare and life sciences  (NVIDIA DGX™ A100 for training)
  • NVIDIA Base Command™ (DGX system software)
  • NVIDIA AI Enterprise software suite
  • MONAI for AI SDKs 
  • NVIDIA FLARE for federated learning
  • Flywheel research platform for data management
  • Pure Storage FlashBlade

Tapping Into NVIDIA Hardware and Software

Dr. Garrett’s team leverages MONAI integrated into Flywheel’s healthcare data management solution to preprocess data, including de-identification and labeling. They’re able to quickly and easily curate and normalize data from multiple systems and hospitals to generate comprehensive and unbiased datasets.

The university, in collaboration with other hospitals, is securely training AI models for medical imaging, annotation, and classification using NVIDIA FLARE on NVIDIA DGX BasePOD.. DGX provides powerful computation infrastructure for rapid iteration, which enables the university to improve the accuracy of their models.

Faster Results to Improve Patient Outcomes

“Not only can we do things faster,  we can also scale the dataset up and do what was not possible before,” Garrett commented. “Using the MONAI imaging framework integrated into Flywheel, with training done on NVIDIA DGX BasePOD, we can apply our state-of-the-art research tools to every single abdominal CT we've ever performed at UW–Madison since 2004. Ten thousand cases alone used to take six to eight months just to get through, and we can now process them in a day.” This effort directly resulted in published papers on fully automated deep learning tools to improve CT-based osteoporosis assessment, CT-based liver volume segmentation, and the derivation of abdominal CT-based markers.

This has had a major impact on the timeliness of the radiologic interpretation. UW–Madison can more easily bring AI into their clinical tools and provide results instantly. For example, using AI, they can send images to the automated bone-age analyzer and receive results before the radiologist can even pick up the images. 

With AI tools and frameworks from NVIDIA AI Enterprise, the university also easily replicated their workflows to other clinics and institutions. "We were starting a big clinical trial where we're deploying those tools to 21 sites around the world. In the past, we would have to ship somebody a computer or instructions on how to get this up and running. Today, we can just send them the container, get them rights to use it, and then in an hour they've already processed the first 100 cases. It has enabled us to move to a faster way to access these containers and share software."

“Using the MONAI imaging framework integrated into Flywheel, with training done on NVIDIA DGX BasePOD, we can apply our state-of-the-art research tools to every single abdominal CT we’ve ever performed at UW–Madison since 2004. Ten thousand cases alone used to take six to eight months just to get through, and we can now process them in a day.”

John Garrett, PhD, Assistant Professor and Director of Imaging Informatics,
Department of Radiology

Challenges

 
  • Radiologic interpretation can be time consuming and tedious, but AI tools show great promise to improve workflows. However, limited data can hinder the development of AI models.

Results

 
  • Over one million images can be processed in less than a day.
  • Ten thousand cases can be processed in a day, compared to six to eight months previously.

About University of Wisconsin–Madison Department of Radiology

The UW Department of Radiology is dedicated to improving human health by developing innovative imaging technology through basic and translational research in collaboration with colleagues at UW–Madison and beyond. The department supports the Wisconsin Idea—improving peoples’ lives beyond the university’s classroom walls by collaborating with industry to translate new technology into daily clinical practice.

About Flywheel

Flywheel is a biomedical research data platform for imaging and associated data. Flywheel gives tools to securely discover, manage, curate, and compute large amounts of data, either within an organization or with collaborators across the world.

Ready to Get Started?

To learn more about NVIDIA solutions for healthcare and life sciences, contact us.