Well below half of oncology results hold up when retested. Prof. Francesco Sardanelli, of the Istituto Scientifico San Raffaele in Milan, Italy, opened his talk at ECR 2026 with that number, and added: "I don't think the situation is different in medical imaging."
Twelve percent of radiology articles share their primary data. Forty-seven percent are open access. Eighty-two percent of imaging journals have data sharing policies, but almost none make them mandatory. The field has spent a decade moving toward open publishing, but the data that would allow anyone to check, challenge, or rebuild the findings is still almost never there.
Open access is not reproducibility
An open-access paper is free to read. A reproducible result is one that someone outside the original team can verify - by accessing the data, running the code, and seeing whether the numbers hold. In radiology, that second scenario is rarely met in practice.
"There maybe should be no paper on AI or developments or discoveries without an independent validation. That's critical. That's mandatory," said Prof. Luis Marti-Bonmati, of La Fe University and Polytechnic Hospital in Valencia, Spain.Courtesy Claudia Tschabuschnig
Researchers spend months preparing, anonymizing, and curating data for a study. Once the paper is published, the data disappears, the code is not deposited and the result sits in the literature, gets cited, gets built upon, and cannot be checked, argued Luis Marti-Bonmati, MD, PhD, of La Fe University and Polytechnic Hospital in Valencia, Spain.
"There should maybe be no paper on AI developments or discoveries without independent validation," Marti-Bonmati said. "That's critical. That's mandatory."
His proposal: journals affiliated with the European Society of Radiology (ESR) should consider requiring authors to deposit research data into Cancer Imaging Europe (EUCAIM), the EU's pan-European cancer imaging infrastructure, at the time of publication, not as an option, but as a condition.
EUCAIM provides standardized anonymization verification, Findable, Accessible, Interoperable, and Reusable (FAIR)-compliant metadata, and controlled access for reuse. It already exists, whether ESR editors will enforce that is another matter.
Your anonymized scan is probably not anonymous
"Imaging data is next to non-anonymizable, and clinical information makes it even worse," said Tobias Penzkofer, of Charité Berlin, Germany.Courtesy Claudia Tschabuschnig
Prof. Dr. Med. Tobias Penzkofer, of Charité, Universitätsmedizin Berlin, gave three examples from real research practice.
A team strips the DICOM header of chest CT scans before sharing them externally. The metadata is clean, but the dose report, burned into the pixel data itself, still contains the patient's full name. This, Penzkofer said, happens frequently.
Standard head CT scans collected for biomarker research. Facial reconstruction from the imaging data works. The patient is identifiable.
Sample data from a newly installed scanner class published online. Combined with scanner model, unusual patient age, and acquisition time window, the potential patient population narrows to a handful of individuals.
A single patient record, combining lab panels, timestamps, ICD-10 codes, and imaging metadata, contains between 50 and 105 bits of effective information. Thirty-three bits are sufficient to identify every human on the planet.
The assumption behind most institutional data-sharing agreements, that removing the DICOM header is adequate de-identification, does not hold.
"Just ignoring it doesn't make it go away," Penzkofer said.
Robust deidentification requires quality control checks, pixel scrubbing, defacing for head imaging, and UID handling. Most research pipelines do not include all of these.
Who funds the research determines whether the data exists
Susanna Lee, MD, PhD, editor of Radiology: Advances and a radiologist at Harvard University in Boston, laid out the funder matrix. Public funders, national research councils, NIH, Horizon Europe, almost always require a data sharing plan upfront and enforce compliance by withholding future funding. 
"Industry is on the other end. They rarely need a data sharing plan, and almost always the data is going to be protected," said Susanna Lee, editor of Radiology: Advances and a radiologist at Harvard in Boston, U.S.Courtesy Claudia Tschabuschnig
Industry funders almost never require one. Data from industry-funded trials is protected until regulatory clearance. Embargoes are standard. Violation is enforced by contract. Industry funding for radiology research is soon to be a majority in the U.S. and growing across Europe.
The AI tools entering clinical practice in Europe are built on data that European researchers cannot inspect, reproduce, or challenge. Not because anyone is hiding anything, but because the funding structure makes sharing legally restricted.
The cost of sharing is not the article processing charge. Data managers, de-identification labor, IT infrastructure, and long-term storage together account for roughly 6% of grant awards at one U.S. institution -- and at Harvard, Lee noted, that figure is higher.
For researchers without grants, the situation is worse. Anna Colarieti, MD, PhD, of the University of Milan, addressed the question Sardanelli raised: how does a junior researcher doing independent work, outside industry funding, without a major grant, pay for open-access publication?
The big five publishers control more than half of a market she put at around 30 billion dollars. Article processing charges run into thousands of euros. "If we need money, we have to ask," she said.
A culture that does not publish failure
Journals reject reproducibility studies because they lack novelty. A paper confirming that a prior result holds is told there is nothing new here.
Tugba Akinci D'Antonoli, MD, of University Hospital Basel in Basel, Switzerland, pointed to registered reports as a partial answer: pre-register the methodology, receive peer review before the results are known, commit to publish regardless of outcome. It removes the incentive to chase positive findings. Whether the result is positive or negative stops mattering to publication.
Neither registered reports nor mandatory data deposit has been adopted at scale in radiology.
"We listed more problems than solutions," Sardanelli said at the close. "But these are the problems we have now in open research."
Our full ECR 2026 coverage can be found here.
