Is Your Medical Image Data Safe? The Privacy Risks of AI in Radiology
When you go in for an X-ray, MRI, or CT scan, you expect the results to help your doctor make a diagnosis. You likely don’t expect your images to be used for anything else—especially not to train an artificial intelligence model. Yet that is exactly what is happening in many hospitals and imaging centers. AI tools are being developed to detect fractures, tumors, and other abnormalities faster than humans can. But the same data that powers these algorithms also introduces privacy risks that few patients are aware of.
The Radiological Society of North America (RSNA) has flagged this issue, warning that medical imaging AI “opens a Pandora’s box of privacy-related risks.” The statement came amid growing use of AI in radiology and concerns that patient data is not as safe as many assume.
What Happened
In May 2026, RSNA published an article directly titled “Medical Imaging AI Opens a Pandora’s Box of Privacy-Related Risks.” The piece described how AI models trained on large sets of medical images can inadvertently expose sensitive patient information. While the exact details of the article are behind a paywall, the key concern is that traditional methods of de-identification—removing names, dates, and ID numbers—may not be enough when AI is involved.
Separately, RSNA has also addressed economic hurdles to AI adoption, but the privacy angle has received less public attention. The core issue is not new: researchers have shown that AI systems can sometimes “memorize” individual patient data and that adversaries can infer whether a specific person’s images were used to train a model.
Why It Matters
For the average patient, the risk may sound abstract, but it has real consequences. Here are three ways your medical images could be exposed:
Re-identification from de-identified images. Even after removing obvious identifiers, scans often contain enough anatomical detail or metadata (such as age, body dimensions, or scanner serial numbers) to link them back to a specific person, especially if combined with other public records.
Model inversion attacks. An attacker with access to a trained AI model can sometimes reconstruct images that are eerily similar to the originals used in training—potentially revealing a patient’s face or unique features.
Membership inference attacks. By querying the model, an adversary can determine whether a particular patient’s data was part of the training set. This alone can leak sensitive information, such as confirming that someone has a specific disease.
The problem is compounded by gaps in regulation. HIPAA, the main U.S. health privacy law, covers identifiable health information, but de-identified data is not subject to the same protections. And HIPAA was written before AI models became widespread; it does not explicitly address risks like model inversion. Meanwhile, many consent forms used by hospitals include broad language allowing data to be used for “research” or “quality improvement,” which can include training commercial AI tools.
What Readers Can Do
You do not have to stop getting medical imaging. But you can take steps to protect your privacy:
Ask your provider about data use. Before a scan, ask: “Will my images be used to train any AI models? If so, can I opt out?” Some facilities offer an opt-out; many do not, but asking signals that patients care.
Read the consent form carefully. Look for clauses about data sharing, research, or third-party use. If the language is vague, ask for clarification. If you are uncomfortable, you have the right to refuse treatment (though in an emergency this may not be practical).
Check with the radiology department directly. The person scheduling your scan may not know. Call the imaging center ahead of time and ask how they handle patient data for AI training.
Stay informed about new regulations. Some states have passed additional privacy laws (like California’s CPRA) that may offer more protections. Federal updates to HIPAA or new AI-specific legislation could also change the landscape.
Consider donating images to research only through trusted registries. Some patients want to contribute to science. If that is your goal, use platforms like the NIH’s Medical Imaging and Data Resource Center (MIDRC), which have clear data governance policies.
Sources
“Medical Imaging AI Opens a Pandora’s Box of Privacy-Related Risks.” Radiological Society of North America, 20 May 2026. Link (Google News archive)
“Radiologists Urge Economic Realism in AI Adoption.” RSNA, 26 May 2026. Link
RSNA 2023 coverage on AI in radiology (referenced for context).
These risks are not reason to avoid medical scans—but they are reason to ask questions. As AI becomes a routine part of radiology, the balance between innovation and patient privacy will depend on transparent practices and informed patients.