The spec sheet starts with one number: 14,000 patients. Every pelvic MRI in that cohort was labelled by expert radiologists. The model built on those labels now reads scans beside the radiologist, not instead of her. The company behind it is Matricis, a French startup spun out of two public research institutions. Its product is called EndomAI, and it targets one disease only.
The held detail sits inside that dataset. The founders did not download their training data from anywhere. No open, labelled pelvic MRI dataset existed for them to use. They had to build the dataset first, scan by scan, before any model work could start.
The direct answer: EndomAI is an AI copilot for pelvic MRI. It supports the radiologist during the read, flags the patterns linked to endometriosis, and produces a structured report. The radiologist keeps the final call on every scan. The system is in a first clinical pilot now, and it has published no results yet.
- The spec sheet comes first
- A disease that hides for years
- The dataset is the real product
- A second reader, not a replacement
- The pilot is the test that matters
- What a Canadian imaging group should ask
- The question in writing
The spec sheet comes first
Matricis describes EndomAI as a medical-grade copilot for women’s pelvic MRI reads. The training cohort holds 14,000 unique patients. The data comes from multiple centres across three continents. Expert pelvic imaging radiologists annotated the scans by hand.
The company is a spin-off of Inria and AP-HP, the Paris public hospital group. Raphaelle Taub, a physicist, co-founded the company and serves as chief executive. Elise Mekkaoui co-founded it with her and runs operations. The team pairs AI researchers with clinicians, including a head of radiology.
That mix matters for a buyer. The product was not built by a general imaging platform looking for a niche. It was built by a team that started from this one read. The narrowness is the design choice, and it shapes everything the pilot can prove.
A disease that hides for years
Endometriosis affects an estimated 10% of reproductive age women worldwide. That share works out to about 190 million people. The average time to diagnosis runs between 4 and 12 years, depending on the country. The seven years in this headline sits inside that published range.
The delay is not mainly a scanning problem. Symptoms overlap with other conditions, and early scans can look unremarkable. The lesions are small, scattered, and easy to read past on a busy list. A radiologist may see only a handful of confirmed cases in a month.
This is the gap the product aims at. A model trained on thousands of labelled cases carries pattern memory no single reader can hold. It does not get tired at the end of a long list. It applies the same checklist to the first scan and the fortieth.
The dataset is the real product
Most AI imaging stories start with the model. This one starts with the annotation work. Fourteen thousand patients is a large cohort for one condition. Each case needed expert labels before it could teach anything.
Taub has described the starting point plainly in interviews. There were zero open-source annotated MRI datasets of the female pelvis. None. So the company spent its early effort on data agreements, expert time, and labelling discipline.
That choice is also the moat. A rival can rent the same computing power this quarter. A rival cannot rent a decade of expert-labelled pelvic scans. The dataset is the asset a buyer should ask about first.
Canadian founders know this pattern well. Three Calgary AI startups store data in Canada. Their AI features can still send it abroad shows how data handling decides trust. In medical imaging, the labels matter even more than the storage address.
A second reader, not a replacement
EndomAI works inside the reading workflow. It maps the distribution and extent of lesions on the scan. It generates standardised, disease-specific reporting templates. It offers decision support and follow-up recommendations for the reading physician.
The radiologist still signs the report. In the pilot, every interpretation stays under the supervision of board-certified radiologists. The tool proposes; the physician disposes. That division is the only shape clinic adoption realistically takes.
The same shape shows up across Canadian deployments. An on-premises AI assistant keeps drawings in the building. Who reads the answer is still your problem makes the same point for engineering firms. The accountable human stays named, or the tool does not ship.
The pilot is the test that matters
Matricis announced its first United States clinical pilot in April 2026. The partner is SimonMed Imaging, one of the largest outpatient imaging providers in that country. The company’s site lists the first clinical pilots as starting May 2026. The initial phase covers about 200 patients with suspected or diagnosed endometriosis.
The pilot measures three things. It tracks diagnostic consistency between readers. It tracks lesion detection with and without the tool. It tracks radiologist confidence when reporting to referring physicians.
Those are the right measures, and they are not results. No outcome figures from this pilot are public yet. The company states that United States clearance is planned for the fourth quarter of 2026. Until the pilot reads out, every claim about accuracy is a plan, not a finding.
Buyers in adjacent fields apply the same discipline. Wealthsimple Penny asks before it pays. Your card limit sets the damage shows a Canadian product earning trust through limits. Ottawa’s $2 million AI governance challenge scores running the project above your code shows evaluators scoring delivery over demos. The pilot is Matricis’s version of that test.
What a Canadian imaging group should ask
Canadian patients wait years inside the same diagnostic range. Canadian radiology groups read the same kind of scans under the same time pressure. A second-reader model is the shape adoption would take here, too.
The questions are regulatory and practical, not technical. The tool would need Canadian medical device licensing before clinical use. The training data is European and American, so a buyer should ask about performance on Canadian scanners and protocols. The reporting templates would need to fit local referral habits.
None of those questions has a public answer today. That is normal for a company at pilot stage. What matters is that the company is testing in live clinics, with named partners, on published measures. Many imaging startups never get that far.
The question in writing
No hands-on test ran for this piece; the product is not available to test. So the close is the question a buyer should put in writing. Ask Matricis for the pilot’s lesion detection figures, split by reader experience, before any pricing talk. Ask what the tool missed, not only what it caught.
A startup that answers those two questions with data has earned a second meeting. One that answers with adjectives has told you what it is selling. Follow the pilot results as they publish, and judge this company on the readout, not the announcement.
