How Bias Gets Into Medical Imaging AI – and What’s Being Done About It

If you’ve had an X-ray, a mammogram, or a CT scan in the last few years, there’s a fair chance an algorithm looked at it before or alongside a radiologist. AI tools now flag lung nodules, prioritize urgent brain scans, and screen for diabetic eye disease. They’re fast, they don’t get tired, and in published studies they often perform impressively. But there’s a question that rarely makes the headlines: impressive for whom?

An AI model learns medicine from the images it was trained on. If those images came mostly from one country, one type of hospital, or one group of patients, the model learns that group’s version of medicine and quietly struggles with everyone else. I’ve worked around medical imaging datasets long enough to see how easily this happens, and why the teams that take it seriously invest in diverse, ethically sourced medical imaging data rather than whatever scans happen to be easiest to download.

Here’s a plain-language look at what AI bias in healthcare actually is, how it gets into medical imaging, who it affects, and what’s being done to fix it.

What Is AI Bias in Healthcare?

AI bias in healthcare means an algorithm works noticeably better for some groups of patients than for others. It isn’t usually a matter of anyone programming prejudice into the software. It’s a matter of math: a model gets very good at patterns it has seen many times and much less reliable on patterns it has rarely seen.

Think of a doctor who trained for ten years in a single hospital in one city. They’d be excellent with the patients who usually came through those doors. Put them in a clinic on another continent, with different body types, different common diseases, and different equipment, and they’d need time to adjust. Human doctors know to adjust. An AI model doesn’t know what it doesn’t know. It gives its answer with the same confidence either way.

How Does Bias Get Into Medical Imaging AI?

In my experience, bias almost never comes from a single dramatic mistake. It creeps in through several ordinary decisions, each of which seemed reasonable at the time:

• Who gets scanned. Training data comes from people who had access to imaging in the first place. Groups with less access to care are underrepresented before the project even starts.
• Where the data comes from. A widely cited 2020 analysis in JAMA found that most U.S. studies training image-based diagnostic AI drew their patient data from just three states: California, Massachusetts, and New York. Globally, the imbalance between high-income and low-income countries is even larger.
• Which machines took the images. Scanners from different manufacturers, older equipment, and different hospital settings produce images that look subtly different. A model trained on one hospital’s high-end scanners can stumble on another’s older machines.
• Who labeled the images. Every training image needs a correct answer attached, usually written by doctors. If those labels come from automated text mining of reports, or from specialists less familiar with how a condition appears in certain patients, the errors get baked into the model.
• How the model was tested. If the test set comes from the same narrow source as the training set, the model will look excellent on paper while hiding its weak spots.

Real Examples: When Medical AI Got It Wrong

This isn’t hypothetical. Two studies did a lot to bring the problem into the open.

In 2021, researchers led by Laleh Seyyed-Kalantari published a study in Nature Medicine examining AI models trained on three large public chest X-ray datasets. The models more often labeled sick patients as healthy when those patients were female, Black, Hispanic, or from lower-income backgrounds, and the gap was widest for people who fell into more than one of those groups. Being wrongly told your X-ray is normal can mean a delayed diagnosis. The study sparked real scientific debate about how to measure this kind of bias, which is a healthy sign; nobody serious argues the problem should be ignored.

Dermatology offers an even clearer picture. A Stanford-led team reported in Science Advances in 2022 that leading skin-cancer AI models performed noticeably worse on darker skin tones and on uncommon diseases. They built a new, deliberately diverse image set to test this. Two details stood out. First, the dermatologists who typically label training images also performed worse on darker skin. Second, when the researchers retrained the models using diverse images, the performance gap between light and dark skin closed.

That second finding is the hopeful part of this story. Bias in medical imaging AI isn’t a permanent flaw of the technology. In many cases, it’s a data problem with a data solution.

Why Diverse Medical Imaging Data Matters So Much

Researchers have tried clever algorithmic fixes for bias, and some help at the margins. But the dermatology study found that state-of-the-art training methods couldn’t correct the gap on their own. What closed it was showing the model the patients it had been missing.

“Diverse” is broader than most people assume. A truly representative imaging dataset accounts for:

• Patient demographics: age, sex, ethnicity, skin tone, and body size.
• Geography and disease patterns: a chest X-ray model trained where tuberculosis is rare may misread scans from regions where it’s common.
• Clinical settings: large academic hospitals, small community clinics, and everything in between.
• Equipment: multiple scanner manufacturers, models, and imaging protocols.
• Disease stages: not just textbook cases, but early, subtle, and unusual presentations.

Evgenia Mocanu, CEO and Founder of medDARE, who has spent more than 20 years in AI and technology, describes the gap she sees most often:

“When teams ask for a diverse dataset, they usually mean demographics, and that matters. But a model can fail just as badly because every scan came from one manufacturer’s machines or one hospital’s protocols. Real diversity means collecting images across countries, equipment, and clinical settings, and doing it ethically, with proper consent, anonymization, and compliance in every jurisdiction. It’s slower than downloading a public dataset. But it’s the only way a model built in one part of the world can be trusted in another.”

– Evgenia Mocanu, CEO & Founder, medDARE

Doesn’t a Scan Just Show Anatomy? Why Would Race or Sex Matter?

It’s a fair question. A lung is a lung, right? Not quite. Bone density, breast tissue density, body fat distribution, and how common certain diseases are all vary across populations. These differences affect how a condition looks on an image.

There’s also a more surprising finding. Research published in The Lancet Digital Health in 2022 showed that AI models could predict a patient’s self-reported race from chest X-rays and other scans with high accuracy, even though trained radiologists can’t. Nobody fully understands what the models are picking up on. That’s exactly why testing performance separately for different patient groups matters: a model may be using information we can’t see.

What Happens When Biased AI Reaches Patients?

The consequences depend on how the tool is used. A biased triage tool might push some patients further down the reading list, delaying care. A biased screening tool might miss early cancers in some groups while over-flagging others, leading to unnecessary follow-up tests and anxiety.

On a global scale, the stakes are higher still. AI is often promoted as a way to bring specialist-level screening to regions with few radiologists or dermatologists, which is a genuinely exciting goal. But if those tools were trained almost entirely on patients from wealthy countries, the communities with the most to gain may be the ones the tools serve worst.

How Are Researchers and Regulators Addressing AI Bias?

The good news is that the field has moved from “is this a problem?” to “how do we fix it?” The World Health Organization’s guidance on the ethics and governance of AI for health lists inclusiveness and equity among its six core principles, calling for AI that works across age, gender, income, and other characteristics. In Europe, the EU AI Act classifies most medical AI as high-risk and requires developers to examine their training data for possible biases and to use datasets that are relevant and representative. In the United States, the FDA increasingly expects developers to report how their devices perform across patient subgroups.

Inside development teams, the practical fixes look like this:

• Sourcing data from multiple countries and clinical settings, not just the most convenient hospital partner.
• Using expert, physician-reviewed labels instead of relying solely on automated text extraction from reports.
• Testing performance separately for different patient groups before release, not just reporting one overall accuracy number.
• Monitoring tools after deployment, because a hospital’s patients and equipment can differ from anything in the original test.
• Being transparent about where the training data came from and who it does and doesn’t represent.

What Can Patients Do About AI Bias in Healthcare?

You can’t audit an algorithm yourself, but you’re not powerless either. It’s reasonable to ask your provider whether AI was involved in reading your scan and whether a physician reviewed the result. If your symptoms don’t match a “normal” result, say so and ask about a second read. Your own knowledge of your body is information no algorithm has.

It also helps to keep this in perspective. Used well, medical imaging AI catches things humans miss and speeds up care for many patients. The goal isn’t to fear the technology. It’s to make sure it’s built on data that reflects everyone who will depend on it.

Frequently Asked Questions

What is an example of AI bias in healthcare?

One well-known example comes from chest X-ray AI: studies found that models trained on large public datasets more often missed disease in female, Black, Hispanic, and lower-income patients. In dermatology, several skin-cancer AI models have performed worse on darker skin tones because their training images showed mostly lighter skin.

Can AI in medical imaging be completely unbiased?

Probably not completely, just as human doctors aren’t. But bias can be measured and significantly reduced. Diverse training data, expert labeling, and testing results across patient groups have been shown to narrow performance gaps considerably.

Why can’t developers just use more data to fix AI bias?

More data only helps if it’s the right data. Adding another million scans from the same hospitals and the same population makes a model more confident about that population, not better for anyone else. What matters is variety in patients, locations, and equipment, along with accurate labels.

Is my medical data used to train AI?

It may be, depending on where you live and the policies of your healthcare provider. Reputable AI projects use anonymized data, with names and identifying details removed, and follow privacy laws such as HIPAA in the United States and GDPR in Europe. You can ask your provider about their data-sharing policies.

Should I trust an AI-assisted diagnosis?

In most settings, AI supports a physician rather than replacing one, and the final report is reviewed by a doctor. AI-assisted reads can be very accurate, but it’s always reasonable to ask questions, request a second opinion, or mention if a result doesn’t match how you feel.

Which areas of medicine are most affected by AI bias?

Radiology and dermatology get the most attention because they are image-heavy fields where AI is already widely used. Radiology accounts for the majority of AI-enabled medical devices authorized by the FDA, so the quality and diversity of imaging data matters enormously there.