Medical Imaging AI Raises New Privacy Risks: What You Need to Know

Medical Imaging AI Raises New Privacy Risks: What You Need to Know

Artificial intelligence is reshaping how doctors interpret X‑rays, MRIs and CT scans. Algorithms can spot tumors, fractures and early signs of disease faster than the human eye in many cases. That promise, however, comes with a cost few patients consider: your medical images may be feeding these systems in ways that expose sensitive health data to risks you never agreed to. Recent findings from the Radiological Society of North America (RSNA) have drawn attention to privacy vulnerabilities that are still poorly understood by the public.

What happened

The RSNA report, presented in May 2026, examined how AI models used in medical imaging interact with patient data. Among the key issues: some AI systems can inadvertently “memorize” parts of the training data, including entire patient images. When a model is later queried or attacked, it may reconstruct those images, revealing identifiable features—faces, tattoos, unique anatomy—that were supposed to be stripped away. Researchers also flagged that many AI tools rely on cloud‑based processing, where images are sent to external servers for analysis, adding an extra link in the data chain that can be harder to secure.

Another concern is re‑identification. Even after images are de‑identified according to standard rules (removing names, dates, and ID numbers), the images themselves contain unique patterns that can be matched back to an individual using other databases. The RSNA report noted that current de‑identification methods, originally designed for structured data like spreadsheets, often fail when applied to complex image files.

Why it matters

For most patients, an MRI or a mammogram is a one‑time, transactional event. You get the scan, you get the results, and you assume the image stays inside your health system’s secure archive. The reality is more complicated. AI‑powered imaging tools increasingly operate as software‑as‑a‑service, meaning your hospital or clinic uploads images to a vendor’s cloud server. That vendor may then use those images—sometimes broadly, sometimes without explicit patient consent—to improve their algorithms. In a 2025 survey cited in the RSNA report, fewer than 15% of patients were aware that their medical images might be used for AI development, and most wanted to be asked first.

The potential harms go beyond embarrassment. Medical images can reveal pregnancy, genetic markers, substance abuse, or surgical history. If a dataset is breached—and healthcare data breaches have been steadily rising—those images become public. Re‑identification attacks can link them to individuals, opening the door to discrimination by insurers, employers, or landlords. And because AI models can be retained and reused for years, the risk persists long after the original scan is taken.

What readers can do

You don’t need to be a security expert to reduce your exposure. Start by asking your provider a few concrete questions before any imaging exam:

• “Will my images be shared with any third parties for AI development or research?”
• “Are my images stored locally or sent to a cloud service? If so, where, and how long are they kept?”
• “Can I opt out of having my data used for training or algorithm improvement without affecting my care?”

Many hospitals have a standard consent form that covers data use for “quality improvement” or “education.” Read it carefully. If the language is vague, ask for clarification or request an addendum that limits use of your images to your own diagnosis. Some institutions allow you to sign a partial consent that permits AI use only in fully anonymized form.

Additionally, if you are participating in a research study that involves imaging, ask specifically how the researchers plan to de‑identify the images and what protections they have in place against re‑identification. The National Institutes of Health and other bodies recommend that studies describe these safeguards in their consent forms, but enforcement varies.

Sources

The findings described here draw primarily from the RSNA report “Medical Imaging AI Opens a Pandora’s Box of Privacy‑Related Risks,” published in May 2026. That report synthesizes research from multiple academic medical centers, patient advocacy surveys, and earlier literature on privacy attacks in machine learning. It is publicly available through RSNA’s website (rsna.org). Additional context comes from published analyses of healthcare data breaches by the U.S. Department of Health and Human Services’ Office for Civil Rights, which tracks breaches affecting 500 or more individuals. While HIPAA covers many provider‑side protections, it does not fully address the secondary reuse of images by third‑party AI vendors—a gap that lawmakers and professional societies are just beginning to examine.