Health data infrastructure · Real-world imaging
Segmed
A Palo Alto company that collects medical scans from hospitals under revenue sharing deals, strips the identifiers, and sells the result to AI developers and drug companies. It says the library holds about 150 million imaging studies. In January 2026 an outside statistician certified the de-identification.
The paper came before the company
In February 2020, while its founders were sitting in a Y Combinator batch, the journal Radiology published a review titled “Preparing Medical Imaging Data for Machine Learning.” Ten authors signed it. Four of them, Martin Willemink, Wojciech Adam Koszek, Cailin Hardell and Jie Wu, are the four people later listed as Segmed founders.
The complaint in that paper reads now like a business plan. Medical AI was starving, the authors wrote, because “most research groups and industry have limited data access based on small sample sizes from small geographic areas.”
Six years later the company built on that sentence sells access to scans from more than 2,800 sites, and the argument about who may do that is not settled.
Why the pictures are stuck
Radiology images do not live in the electronic health record. They sit in a separate picture archive, in files that carry far more than the picture. A DICOM file records the scanner model, the acquisition settings, dates, and often the patient name and medical record number written straight into the pixels at the top of the image.
Cleaning that is three jobs: scrub the text fields, run optical character recognition across the image and mask what it finds, then normalize the result so a search can tell a chest CT from an abdominal one when the hospital label got it wrong. Willemink described the practical obstacle in a 2024 interview: the data is siloed, “each using different IT systems that don’t integrate easily.” The bottleneck after that is commercial. Segmed has to sign the hospital, and in the same interview he named that as the second constraint on growth.
A Dutch radiologist who never finished residency
Willemink went to medical school in the Netherlands, added a master’s in clinical epidemiology and a PhD in medical imaging, and started a radiology residency. A grant took him to Stanford for what was meant to be one year. He stayed three as a faculty instructor and never went back to finish.
Koszek came the other way: a FreeBSD developer as a teenager in Poland, then Xilinx, then a startup Salesforce bought in 2017. He became chief technology officer. On his own site he writes that Segmed “started during Stanford Ignite program, and was a part of YC W20 batch, Alchemist Batch 21, StartX winter 2020.” He has since left to found another company.
Willemink ran Segmed through the Series A. In May 2025 the board brought in David Gascoigne, a statistician by training who had run commercial analytics at IQVIA and led Almaden Genomics, and moved Willemink to chief scientific officer. Jie Wu is now chief data officer.
Where the scans actually come from
No patient deals with Segmed. Health systems do. The company’s page for providers describes a revenue sharing model: the hospital has been storing the images for years, Segmed de-identifies and structures them, and researchers buy the result. Beacon Health System in Indiana signed in April 2024. Advocate Health, a 67 hospital system, went further and co-led the Series A five months later.
That puts the consent question one step upstream, at the provider. Segmed says it sources data only through authorized healthcare partners. Its own scientists were more exact in print. Writing in PLOS Digital Health in October 2025, Segmed staff and outside co-authors noted that opt in and opt out protocols exist but that their availability “varies depending on local laws and health provider standards.” They also wrote that identifiers are replaced with codes “so that the data can be re-identified or tracked back to the source institutions if necessary.” It is the most candid document the company has produced, and it carries a competing interests statement declaring that two of its authors are employees and shareholders of Segmed.
The certificate
On January 6, 2026 Segmed announced three Expert Determinations, the harder of the two routes HIPAA allows for calling data de-identified. The expert was Brad Malin of Vanderbilt, the Accenture Professor of Biomedical Informatics, Biostatistics and Computer Science, an elected fellow of the National Academy of Medicine whose re-identification research the Federal Trade Commission has cited in the Federal Register. The determinations covered structured fields, imaging pixels and radiology text separately. The alternative route, Safe Harbor, is a checklist of eighteen identifiers to delete. Expert Determination requires somebody to put a name to a risk judgment.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| HIPAA Expert Determination | Announced Jan 6, 2026. Three determinations across structured data, imaging pixels and radiology text, by Brad Malin, PhD, of Vanderbilt. His credentials are independently verifiable; the determination reports are not public. | Named certifier |
| Security certifications | SOC 2 Type 2 and ISO 27001, described as maintained for several years and regularly audited. Certificates not published. | Company-stated |
| Library size | 100 million studies announced Aug 10, 2023, still 100 million in the Bayer agreement of Dec 2024, about 150 million from Feb 2026 on. The unit shifts between releases: studies, exams and datasets are used interchangeably for the same figure. | Units differ |
| FDA clearances enabled | The impact report says 35 or more cleared AI solutions. The homepage says 50 or more. A co-founder’s site says the platform enabled “hundreds of AI, healthcare, and life-science companies obtain FDA 510(k) clearance.” No list is published. | Three figures |
| Provider network | 2,000 or more sites in Jun 2024, more than 2,800 by Jul 2026. Two providers have ever been named: Beacon Health System and Advocate Health. | Company-stated |
| Open dataset | Released Mar 5, 2026 with iMerit and Advocate Health. Breast tomosynthesis from 558 patients, 271 malignant and 287 benign, biopsy confirmed, free to registered users. Reported composition: 96% White, 1% Black or African American, 1% Asian, 1% Multi-racial. | Public record |
| Peer-reviewed output | Seven PubMed-indexed papers with Segmed-affiliated authors, 2020 to 2026, in Radiology, PLOS Digital Health and JMIR Medical Informatics among others. Reviews and opinion pieces, not validation studies of the de-identification pipeline. | Public record |
| Re-identification audit | No published adversarial test of Segmed data by a third party found as of Sep 22, 2026. An Expert Determination is a risk opinion, not an attack study. | Not found |
| Regulatory and legal record | No FDA 510(k) or PMA record, registered trial, NIH award or NSF award under the Segmed name, and no privacy enforcement action or lawsuit found. Segmed sells data rather than devices, so the absence of device records is expected. | Public record |
Read plainly: the privacy machinery is real and now carries a named expert with a serious record, which is more than most companies in this trade offer. The numbers describing the asset are looser. Three counts of FDA clearances circulate on company surfaces at once, the unit behind the 150 million figure moves between studies, exams and datasets, and the only dataset anyone outside can inspect is 558 patients from a single health system and 96% White. None of that contradicts the privacy claims. It does mean the scale claims rest entirely on the company’s own count.
What to watch
- Whether any Expert Determination report, or a summary of its methods and residual risk, is published.
- Whether the 150 million figure is restated with a consistent unit and a date.
- Named provider partners beyond Beacon and Advocate, and whether the revenue share is ever disclosed.
- Diversity of the next open dataset. The company sells representativeness; the one public sample does not show it.
- The Japanese leg of the MAPI deal announced in April 2026, which would move data into a different consent regime.
In their words
“Another key challenge we face is expanding our network of healthcare providers to source more data, a process that’s often slow due to the complex approval process within health systems.”
Martin Willemink, then CEO, JM Search interview, 2024 · Interview
“As a result, de-identification considerably reduces, but does not eliminate, the possibility of re-identification.”
Lutz de Araujo, Wu, Harvey, Lungren, Graham, Leiner and Willemink, PLOS Digital Health, 2025 · Peer reviewed
“Although patients may wish to exert control over their data, true individual control is impractical and risks placing an undue burden on individuals to protect their data from misuse.”
Same paper, PLOS Digital Health, 2025 · Peer reviewed
“Based on my evaluation, it is my opinion that Segmed’s database meets the HIPAA standard for de-identified data.”
Brad Malin, PhD, Vanderbilt University Medical Center, January 2026 release · Certifier
“we are chopping them up and creating products and selling it back to them”
Marielle Gross, bioethicist, Johns Hopkins, on the patient data trade, to STAT, 2025 · Independent
“Their trust and partnership have been instrumental in enabling us to create a diverse data warehouse representing over 40 U.S. states and 15 countries.”
Garth Schulz, COO and CFO, on passing 100 million studies, 2023 · Company release
Related companies
Sources
- Public recordForm D, Segmed, Inc., CIK 0001953854
- Public recordPreparing Medical Imaging Data for Machine Learning
- Public recordMedical Imaging Data Calls for a Thoughtful and Collaborative Approach to Data Governance
- Public recordBradley Malin, PhD, faculty profile
- IndependentThe companies paying hospitals to hand over patient data to train AI
- IndependentSegmed, a startup that gathers medical imaging data for AI development, raises $10M
- IndependentSegmed platform passed 100M studies
- InterviewUnlocking the Potential of AI in Healthcare: Transforming Medical Imaging with AI
- CompanySegmed Secures $10.4 Million Series A Funding
- CompanySegmed Appoints David Gascoigne as CEO in Leadership Transition
- CompanySegmed Secures Independent HIPAA Expert Determination
- CompanyiMerit, Segmed, and Advocate Health Announce Free Annotated 3D Mammography Data Set
- CompanySegmed website: About, For Healthcare Providers, Newsroom, privacy blog
- FounderAbout me, personal website of Adam Koszek
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
