ARPA-H awards up to $42 million for HARBOR, a network to train medical AI on images that stay in hospitals
ARPA-H gave the University of Utah and Boston Children's Hospital a three-year contract worth up to $42 million to build HARBOR, a federated platform for training medical imaging AI without moving patient images out of hospitals, with FDA regulatory scientists as partners.
Docket, covering health regulation and AI in care
, 3 min read
$42MMaximum value of the three-year ARPA-H HARBOR contract
What to know
- The University of Utah and Boston Children's Hospital share a three-year ARPA-H contract worth up to $42 million, announced November 16.
- HARBOR is meant to let AI models learn from medical images held at many hospitals while the images stay under each hospital's control.
- The FDA's Office of Science and Engineering Laboratories is a partner; that role is research, not a regulatory clearance of anything built on the platform.
The usual complaint about medical imaging AI is not that the algorithms are weak. It is that the training data sits in hundreds of hospital archives, behind privacy rules and incompatible formats, so most models learn from a few large academic centers and then meet very different patients in the field. The Advanced Research Projects Agency for Health (ARPA-H) is now paying to try a different arrangement.
ARPA-H awarded the University of Utah and Boston Children's Hospital a three-year contract worth up to $42 million to build HARBOR, short for Health Asset Registry for Bioimaging, Orchestration & Reuse, according to a University of Utah announcement dated November 16. Healthcare IT News and DistilINFO reported the same amount and term.
What HARBOR is supposed to do
The design keeps images where they are. Hospitals host their own data, on public cloud or their own servers, in a secure federation, and analysis runs under each hospital's governance rather than by copying files into a central store, the university says. Its technical lead, Valerio Pascucci, director of Utah's Center for Extreme Data Management Analysis and Visualization, summed up the idea as "HARBOR flips the model."
The main technique is federated learning: an AI model is sent to each site's data, trained there, and only the model updates come back. The Utah team leads that work. The platform builds on federated storage from the National Science Data Fabric and on OpenViSUS data streaming, both named in the university release and by Healthcare IT News. HARBOR is also meant to help hospitals with the unglamorous steps that usually stall sharing: standardizing, cleaning and de-identifying images.
The stated targets are pediatric and rural medicine, two groups that are thin in most imaging datasets. Boston Children's work is led by P. Ellen Grant, a pediatric neuroradiologist. The team list includes researchers from Stanford, Indiana University, Children's National, Tulane and MLCommons, plus imaging companies PathPresenter and Instapath, and a new commercial venture led by Ouwen Huang that is meant to carry the platform forward. The release's headline calls the effort "Harbor Health Cloud"; DistilINFO calls the venture Harbor Health.
Where the FDA comes in
The contract sits in ARPA-H's Health Science Futures Office under program manager Ileana Hancu. The FDA's Office of Science and Engineering Laboratories, part of the device center, is a partner through Aldo Badano, director of its Division of Imaging, Diagnostics and Software Reliability. Both were also named in ARPA-H's first joint project with the FDA, a medical imaging data marketplace announced in March 2024 to support premarket review of AI devices.
Read that partnership carefully. OSEL is the FDA's regulatory science lab, and DistilINFO notes that taking part in HARBOR does not by itself establish FDA authorization of anything. A model trained on HARBOR would still need its own 510(k), De Novo or approval before it could be sold for clinical use, and its maker would still have to show its test data in a summary like those described in our guide to reading a 510(k) summary.
What is not known yet
The announcement names no participating hospitals beyond the two leads, no imaging volumes, and no milestones inside the three years. "Up to" $42 million is a ceiling, not a sum already paid. DistilINFO reported additional details, including a Massachusetts matching grant, that do not appear in the university release, so they are not used here.
Security will also matter. Medical imaging AI software is already a target: Amazon said this week that attackers' AI agents found and exploited a flaw in MONAI Label, an open-source imaging tool.
Why it matters
Imaging tools, such as CT triage software, make up much of the AI reaching hospitals, and the HARBOR team's premise is that too much of it learns from narrow data. If HARBOR works, developers and FDA reviewers could test models against children, rural patients and many scanner types without anyone moving patient images, which bears directly on whether a model that passed review keeps working after it ships, a question recent research on AI device updates has left open. If it does not, it will be another data platform that ended with the grant.
This article is general information about how technology works. It is not medical advice. Talk to a clinician about your own health.
Sources
- Unlocking the value of medical data while safeguarding patient privacy: University of Utah and Boston Children's Hospital Researchers Receive ARPA-H Award to Build Secure National Platform: Harbor Health Cloud, University of Utah Price College of Engineering, November 16, 2026
- ARPA-H awards $42M to Boston Children's and University of Utah, Healthcare IT News, November 2026
- Boston Children's Secures $42M Medical AI Funding, DistilINFO, November 18, 2026
- ARPA-H announces medical imaging data partnership with FDA, ARPA-H, March 12, 2024
AI law and policyDigital health and AI in careFDA and health regulation
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.