HealthcareMachine LearningOn-PremiseNVIDIA DGX BasePODNVIDIA DGX A100NVIDIA Base CommandNVIDIA AI EnterpriseMONAINVIDIA FLAREFlywheelPure Storage FlashBlade

Accelerating the Radiological Workflow With AI at University of Wisconsin–Madison

University of Wisconsin–Madison Department of Radiology

The UW-Madison Department of Radiology uses NVIDIA DGX BasePOD with the MONAI imaging framework (integrated into Flywheel's healthcare data platform) and NVIDIA FLARE for federated learning to curate imaging datasets and rapidly iterate on AI models for tasks like pediatric bone age assessment and opportunistic screening. Ten thousand abdominal CT cases that previously took six to eight months to process manually can now be processed in a day; over one million images can be processed in under a day. The tools are being deployed via containers to a 21-site global clinical trial.

Overview

The UW-Madison Department of Radiology uses NVIDIA DGX BasePOD with the MONAI imaging framework (integrated into Flywheel's healthcare data platform) and NVIDIA FLARE for federated learning to curate imaging datasets and rapidly iterate on AI models for tasks like pediatric bone age assessment and opportunistic screening. Ten thousand abdominal CT cases that previously took six to eight months to process manually can now be processed in a day; over one million images can be processed in under a day. The tools are being deployed via containers to a 21-site global clinical trial.

The challenge

UW–Madison's Department of Radiology wanted to use AI to speed up tedious tasks in radiologic interpretation, such as pediatric bone age assessment, and to enable opportunistic screening, but faced limited and imbalanced data, data drawn from disparate sources (multiple vendors, PACS, EMR, radiology dictation software), irreproducibility of studies, and subjectivity of interpretation without ground-truth analysis.

The solution

Dr. John Garrett's team uses the MONAI imaging framework integrated into Flywheel's healthcare data management platform to preprocess data, including de-identification and labeling, curating and normalizing data from multiple systems and hospitals. The university, in collaboration with other hospitals, securely trains AI models for medical imaging, annotation and classification using NVIDIA FLARE (Federated Learning Application Runtime Environment) on NVIDIA DGX BasePOD (with NVIDIA DGX A100 for training), NVIDIA Base Command, NVIDIA AI Enterprise software, and Pure Storage FlashBlade.

Machine LearningComputer Vision

Reported business value

Ten thousand abdominal CT cases that previously took six to eight months to process manually can now be processed in a day, and over one million images can be processed in under a day. This directly resulted in published papers on fully automated deep learning tools for CT-based osteoporosis assessment, CT-based liver volume segmentation, and abdominal CT-based markers. Using AI, results such as automated bone-age analysis can be returned before a radiologist even picks up the images. UW-Madison is starting a clinical trial deploying these tools via containers to 21 sites around the world, where after gaining access to a container, sites have processed their first 100 cases within an hour.

Related entries

Other healthcare entries in the register.

All entries
HealthcareGenerative AIPublic Cloud

Novo Nordisk builds AI drug discovery platform on Azure with Microsoft Research

Novo Nordisk partnered with Microsoft Research to build an AI platform on Azure AI and data stacks spanning regulatory affairs, early research, drug discovery and trial design, using Azure OpenAI Service, Azure Cosmos DB and Azure Kubernetes Service, with Power BI and Power Apps for collaboration. The platform includes a copilot for researchers, shared reasoning-chain templates, and governance/auditing of how data and models are used. The teams published early results on predictive AI models for cardiovascular disease risk detection, including an algorithm that Novo Nordisk says predicts patients' cardiovascular risk better than the best clinical standards, drawing on more than 100 years of insulin research data.

96/100HighNovo NordiskPrimary source
HealthcareNatural Language ProcessingPublic Cloud

CDPHP modernizes infrastructure and improves medical data extraction with AWS AI/ML

CDPHP, a not-for-profit health plan serving 400,000 members in Upstate New York, used AWS services including Amazon Comprehend Medical, Amazon Textract, and Amazon SageMaker to automate its data processing pipeline for unstructured medical records and health data. The organization processed over seven million records during initial migration and now processes 3,000 electronic health records weekly. CDPHP achieved a 60% improvement in overall efficiency and reduced HEDIS report generation from 4-5 days (three data scientists) to two reports produced daily.

100/100HighCDPHP (Capital District Physicians' Health Plan Inc.)Primary source
HealthcareGenerative AIUnknown

Mayo Clinic deploys NVIDIA DGX SuperPOD to accelerate pathology foundation models

Mayo Clinic deployed an NVIDIA DGX SuperPOD with NVIDIA DGX B200 systems to support foundation model development for pathomics, drug discovery and precision medicine. In partnership with Aignostics, Mayo Clinic built the Atlas pathology foundation model, trained on more than 1.2 million histopathology whole-slide images. The new infrastructure is reducing four weeks of pathology slide analysis work to one week.

92/100HighMayo ClinicPrimary source
HealthcareGenerative AIPublic Cloud

Myriad Genetics speeds document processing with AWS GenAI Intelligent Document Processing Accelerator

Myriad Genetics partnered with the AWS Generative AI Innovation Center to replace an Amazon Textract/Comprehend pipeline with Amazon Bedrock foundation models (Nova Pro for classification, Nova Premier for extraction) using the open-source GenAI IDP Accelerator. Document classification accuracy rose from 94% to 98%, classification cost per page fell 77% (3.1 cents to 0.7 cents), and classification time fell 80% (8.5 minutes to 1.5 minutes per document). Automated key information extraction reached 90% accuracy matching the manual baseline, with a projected $132K in annual savings and 300 hours saved monthly across 9,000 prior authorizations in the Women's Health unit alone.

96/100HighMyriad GeneticsPrimary source

Was this helpful?

Your feedback helps us improve our use case database