AI Cancer Detection: What a Strong Study Must Show

What the headline number usually is

AI tools are being built to read mammograms, CT scans, pathology slides, and other health data. Most news stories about them report one number. Often it is the share of cancers the software found in a stored set of images.

That number is a place to start. It is not proof that the tool helps people.

This page explains what federal sources say. It is not medical advice and does not suggest a test or treatment.

What NCI says about AI in cancer research

NCI describes artificial intelligence as software that uses data to make predictions or create new content. These programs can find patterns in very large sets of data. NCI funds this work across cancer biology, screening, detection, diagnosis, drug discovery, and care delivery.

Some tools have already reached the clinic. NCI notes that the FDA has authorized software that helps pathologists mark areas of a prostate biopsy image that may hold cancer.

NCI's own wording is careful. It says these uses have potential when the work is done in an ethical and scientifically rigorous way.

The questions a validation study has to answer

NCI scientists and their partners have named the open questions plainly. Are these tools ready to leave the lab? Will they actually help patients? And will the benefit reach everyone, or only some people?

A strong study takes those questions in order.

• Was the model locked before testing began? A tool that keeps changing during a study has not really been tested.
• Was it tested somewhere else? How a tool performs at the hospitals that built it says little about other hospitals.
• Did the test group match the real users? Age, sex, race and ethnicity, scanner type, and disease mix all matter.
• Were wrong results counted? Cancers found is only half of the result.
• Was the whole process compared with usual care? The software is not the care. The pathway is.

Public reporting should break the results down by group. One overall score can hide a tool that works well for some people and poorly for others.

Finding more is not the same as helping more

NCI's screening overview is blunt about this. Screening tests have risks. False-positive results are possible. False-negative results are possible. And a screening test can find a cancer that would never have caused symptoms or harm. NCI calls that overdiagnosis, and the treatment that follows it overtreatment.

A false positive usually leads to more tests and procedures, and those carry risks of their own. A tool built to flag more findings can raise all of these at once. So a rise in detection is not automatically good news. Our guide to the benefits and harms of cancer screening walks through the trade-off.

Limits worth stating plainly

• High accuracy on stored images does not prove better care in daily practice.
• An AI score is not a diagnosis. Diagnosis still rests on biopsy and expert review.
• Results from one hospital or one group of patients may not carry over to another.
• FDA has an authorized-device list, but a place on that list is not evidence that a tool saves lives.

What AI does not replace

Clinical judgment, follow-up testing, quality checks, and clear talk with patients are all still needed when AI is part of the process. Nothing here changes current screening guidance, which depends on cancer type, age, and personal risk.

Questions worth asking about a detection study

• Was the system tested on future patients, or only on old records?
• Were people from different groups and care settings included?
• How many people were flagged who turned out not to have cancer?
• What happens after the software flags a finding, or misses one?
• Did anyone measure whether patients ended up better off?

How this article was prepared

An AI-assisted editorial system helped prepare this page. No named medical reviewer has reviewed it unless one is listed.

The National Cancer Information Foundation publishes Cancer Explained. This page is for learning. It is not medical advice and does not suggest a test or treatment.

See an error, old source, or unclear wording? Tell us.