AI in Radiology: What It Can and Cannot Do Today, Honestly

AI in radiology today is a strong assistant on narrow tasks and a poor substitute for a radiologist on the whole job. It genuinely helps with flagging findings a tired eye might skip, sorting urgent scans to the front of the queue, and measuring more consistently than a hurried hand. What it cannot do is take in a patient's history, weigh an ambiguous shadow against the rest of the image, and decide what the picture means for this person. That second kind of work is still the radiologist's, and the evidence says it should stay there.

I write this as a physician-scientist who builds and evaluates clinical AI, not as a radiologist, and imaging is one place where the hype runs furthest ahead of the proof. A definition helps. An imaging AI tool is software trained to recognize patterns in pixels, good at the narrow task it was trained on and silent about everything it was not. Hold that next to any claim you hear, and most of the confusion clears up. This is general education, not medical advice, and questions about your own imaging belong with your own clinician.

What can AI actually do well in radiology?

It does best at three real things: flagging, triage, and measurement. Each is a narrow, repeatable visual task where consistency matters more than judgment, and that is where pattern recognition earns its keep.

Flagging means the software marks a region that may contain a finding so a human looks again. Used as a second set of eyes, this can catch the subtle thing that a long shift and a full worklist help a person miss. The value is not that the model knows more than the radiologist. It is that the model never tires and treats the last scan of the day like the first.

Triage means sorting scans by likely urgency so the worrying ones surface sooner. A tool that pushes a possible acute finding to the top of the queue does not make the diagnosis. It shortens the time before a qualified person sees the study, which on a busy service can matter a great deal.

Measurement means doing the quantitative, tedious work more consistently than a hurried hand. Tracking how a nodule has changed between two scans is arithmetic on pixels, and a machine that does it the same way every time removes a real source of noise.

The honest shape of all three is the same. There is a clear right answer, a human supervising, and the machine handling volume while the person keeps the meaning.

What can AI not do in radiology?

It cannot do the part of radiology that is reasoning rather than recognition. A radiologist does not simply detect a finding. They decide whether it fits the clinical story, what else it could be, and what should happen next. That work draws on the history, the prior scans, and a trained sense of how disease behaves. The model sees pixels and nothing else.

A model also fails quietly outside the narrow world it learned from. Train a tool on images from one kind of scanner and one type of patient, and it can stumble when the equipment or the population changes, often with no signal that it has. A wrong answer delivered with the same calm confidence as a right one is the dangerous failure mode. A tool built to find one thing will also miss the unrelated finding in the corner of the same image, the one a human eye catches by looking at the whole picture.

So the gap is not about raw accuracy on a benchmark. It is about what a benchmark leaves out: context, the unexpected, and the rare case that least resembles the training data.

Does AI replace the radiologist?

No, and the framing is the problem. The useful question is not whether software can match a radiologist on a test set. It is how a radiologist plus a tool performs on real patients, including the ones who do not fit the average. Posed that way, the strongest results come from the pair.

There is a quiet trap worth naming. A tool that is right almost all of the time trains its user to stop checking the rest, which is exactly when the rare wrong answer slips through. Keeping the radiologist genuinely in charge, with time and information to disagree, is a design requirement, not a courtesy.

The accountability point is simpler. Someone has to answer to the patient when a reading is wrong, and that someone has to be a person. A model can be accurate and still owe nobody an explanation.

How should you judge a radiology AI tool before trusting it?

Hold it to a standard stricter than a published accuracy number. A few questions separate a validated tool from a marketed one.

Was it tested where it will be used? A score earned on the developer's own curated images tells you the model fits data like its training set, not how it behaves on your scanners and your patients.

What is the tool actually for? A flagging aid, a triage sorter, and a measurement helper carry different risks, and a tool sold as one should not drift into another. The narrower the stated intended use, the easier it is to trust. Does the tool also know when to step back? A model that always has an opinion is more dangerous than one that recognizes a case outside what it was validated for.

Does the marketing match the evidence? A common pattern in health technology marketing is to describe a tool as more autonomous, more general, and more proven than its studies support. The cure is boring. Read what the tool was tested on, in whom, and against which human standard. That is what the rules behind software as a medical device keep asking, the framework I studied while earning training in medical device regulations: define the intended use, prove it where it will live, and name who is responsible.

Where this leaves us

The radiologist stays central not because of tradition but because the hardest part of reading a scan was never the detection. It was the judgment, the context, and the willingness to be answerable for what the picture means. Keep that in human hands, hold every tool to validation rather than its slide deck, and AI becomes what it should be: a way to make a good radiologist faster and a tired one harder to fool.