If you have had a scan in the last couple of years at a reasonably well-equipped hospital, there is a fair chance software looked at it before a radiologist did. Not instead of. Before.
That is the quiet reality behind a decade of loud headlines, and it is worth explaining plainly, because the gap between what these systems do and what people assume they do is wide in both directions. They are less impressive than "the computer diagnoses you" and more useful than "it's all hype."
What the model is actually doing
A medical imaging model is a pattern matcher trained by example. You show it a very large number of images along with a label for each — this one has a pneumothorax, this one does not — and it adjusts millions of internal parameters until it reproduces those labels. Then you give it an image it has never seen and it produces a number.
There is no reasoning in the sense a clinician would recognise. The model does not know what a lung is, does not know the patient has been coughing for a week, and has no concept of what a false negative costs. It has learned a statistical mapping from pixel arrangements to labels, and it applies that mapping to whatever you give it.
This is why these systems are simultaneously superhuman on some tasks and idiotic on others. Asked to find a subtle density change that appears in a consistent form across thousands of training examples, a model is tireless and reliable in a way a human at the end of a shift is not. Asked about something it has never seen, it does not say so — it produces a confident number anyway, because producing a number is all it does.
The five things these tools do in practice
Strip away the marketing and deployed imaging AI falls into a small number of categories.
Triage. The model reads studies as they arrive and pushes likely-urgent ones up the queue — a bleed on a head CT, a collapsed lung, a large clot. The radiologist reads everything eventually; the model just changes the order. This is the most common deployment and the safest, because a wrong flag costs a few seconds of attention.
Detection aids. The model marks candidate findings for the reader to consider. Nodules on chest CT, calcifications on mammography, fractures on radiographs. Useful, and the category where there is genuine concern about readers becoming dependent on the marks.
Measurement. Automatic, reproducible quantification: how big is this nodule, how much has it grown since the last scan, what is the volume of this brain structure. Boring and extremely valuable, because measuring by hand is slow and two people doing it get different answers.
Image reconstruction. Learned algorithms that build a diagnostic-quality image from less raw data, which means shorter MRI scans and lower-dose CT. Patients benefit from this without ever hearing the word AI, and it may be the highest-impact application in the entire field.
Reporting help. Drafting report text, populating structured fields, suggesting protocols. Newest, fastest-moving, and furthest from settled.
Notice that none of these is "the AI makes the diagnosis." Outside of a couple of narrow screening tasks, deployed imaging AI produces an input to a radiologist's decision, and the radiologist signs the report.
What the famous studies showed, and did not
The most cited result in this space is a large study of an AI system for mammography screening, published in Nature in 2020. It was developed and evaluated on UK and US screening datasets, and reported reductions in both false positives and false negatives relative to the historical reading of those cases, along with an independent reader study in which the system outperformed the participating radiologists.
Two caveats, both important and both routinely omitted. The study was retrospective — it re-read cases that had already happened, with the outcomes known. And the system was a research artefact, not a product anyone could buy or deploy. "An AI outperformed radiologists in a study" and "you can use it at your hospital" are very different statements, and the distance between them is years of regulatory and engineering work.
The systems you actually encounter in a hospital are commercial products that have been through a clearance process. Chest radiograph analysis software such as Lunit INSIGHT CXR is representative of the category: a defined intended use, regulatory clearances in specific markets, and integration into the reading workflow rather than a standalone verdict. Less headline-grabbing than the research systems, and much more relevant to what patients experience.
The state of the market, honestly
Over a thousand AI-enabled devices have now been authorised by the FDA, with roughly three-quarters of them in radiology. That number sounds like an established industry and partly misrepresents one.
Most of those authorisations came through a pathway based on demonstrating substantial equivalence to an existing cleared device. That is a legitimate regulatory route, but it is not the same as demonstrating improved patient outcomes, and very few imaging AI products have evidence at that level. The ones that do — mostly in breast screening, where randomised trial data now exists — are the exception rather than the norm.
The vendor landscape has also consolidated considerably. Several of the best-known independent AI imaging companies of the late 2010s have been acquired by larger imaging or health-data firms, and the equipment manufacturers increasingly ship these capabilities built into the scanner rather than as third-party software. For hospitals this is mostly good news: fewer integration projects, more accountability. For the field it means the era of the standalone algorithm startup is largely over.
What to make of it as a patient
A few things that are true and worth knowing.
A person is still reading your scan. If software was involved, it flagged something or measured something, and a radiologist assessed it.
The software is regulated as a medical device, with a defined intended use. It is not a general-purpose diagnostic system, and using it outside that intended use is not permitted.
Performance figures quoted for these systems come from validation studies on particular populations and particular equipment. Whether they hold at your hospital depends on how carefully your hospital validated and monitors them, which is a real and variable thing.
And if a report mentions AI-assisted analysis, it is entirely reasonable to ask whether the radiologist confirmed the finding. The answer should be yes.
The direction of travel
Imaging volume keeps growing faster than the number of people trained to interpret it. That is the pressure driving all of this, and it is not going away.
The realistic contribution is not a machine that diagnoses. It is a set of instruments that let a radiologist get through a heavier list without the quality falling — urgent cases surfaced earlier, measurements done automatically, images acquired faster and at lower dose, normal studies moved through more efficiently.
That is a meaningful change to how a specialty works. It is just a much less cinematic one than the coverage suggests, and the field would be better served by describing it accurately.
Tags
Taresh Sharan
support@sharaninitiatives.com