Raidium and the EIC Accelerator: A Foundation Model for Radiology

Raidium, founded in Paris in 2022 by radiologist Paul Hérent and data scientist Pierre Manceron, both formerly of the AI biotech Owkin, builds a "co-pilot" for radiology: an AI foundation model that learns from unlabelled scans and extracts measurable biomarkers for radiologists and for clinical trials. Its EIC Accelerator project, FoundRa, "Foundation Model-Based Biomarker Platform for Radiology and Clinical Trials", receives an EU contribution of €2,499,875 and runs from November 2024 to October 2026, according to CORDIS. Raidium raised a €16 million seed round co-led by Newfund and Kurma Partners with support from the EIC, according to BiopharmaTrend.

The main result so far is Curia, trained with self-supervised learning on the entire CT and MRI output of the Centre d'Imagerie du Nord in Paris over several years: 150,000 exams, 130 terabytes and more than 200 million slices, according to Raidium. On CuriaBench, a 19-task benchmark Raidium built, the large model reached 98.4% accuracy in organ classification on CT, against 88.2% for BiomedCLIP and 85.0% for MedImageInsight. The EIC report adds a liver fibrosis biomarker, automated prediction of portal hypertension developed with AP-HP's Beaujon hospital, an oncology biomarker panel, and a cohort builder in the Raidium Viewer, according to CORDIS.

Where Raidium Stands Now

Raidium published Curia's base model weights on Hugging Face, according to its model page, and has built research tools for clinical cohorts. The EIC report does not name paying pharmaceutical customers or hospital deployments, and no CE marking of a clinical product has been announced. The French register lists a small workforce band, and Raidium's accounts are not in the public INPI dataset.

The Technology: Learning From Unlabelled Scans

Most radiology AI is trained on scans labelled by hand for one task, which is slow and costly, the EIC report says. Curia learns general representations from unlabelled images using a Vision Transformer and the DINOv2 method, in 86-million and 300-million-parameter versions, and can then be adapted to many tasks with little labelled data. Raidium says it generalises from CT to MRI far better than rival models.

What the EIC Award Promised, and What the Record Shows

The project runs to October 2026, so it is too early for a final judgement. The model and first biomarkers exist and look technically strong, though the headline benchmark was built by Raidium itself. The platform's use in clinical trials and radiology practice, the commercial half of the promise, is not yet documented.

1. Foundation modelDeliveredCuria released.
2. Biomarker libraryFirst markersLiver, oncology.
3. Clinical-trial customersNone namedNot reported.
4. Regulatory clearanceNot announcedNo CE mark.

The Money

The seed round counts the EIC among its backers; whether that means EIC Fund equity or only the grant is not specified. Raidium's burn and revenue are not public.

What Cannot Be Checked

Why Europe Should Care

Radiologists are too few for the growing number of scans, Raidium's CEO has argued. Raidium has released part of its European-trained radiology model openly. Its value depends on turning research results into products hospitals and drug developers pay for.

The Verdict

Raidium and the EIC have not been asked for comment for this article. It is based entirely on public records and published statements, linked throughout.