Best medical imaging AI: radiology and pathology

A radiology AI vendor and a digital pathology AI vendor describe themselves as being in different businesses, and their buyers sit in different parts of the building with different budgets. Underneath, they sell the same thing: a model that looks at an image and returns a finding a clinician will act on. The AI Health Index tracks them as two categories because that is how a buyer encounters them, and grades them as one population here because every question that decides the purchase is shared.

Is it authorised, and for what intended use. Was it validated on anyone resembling my patients, on equipment resembling mine. Does the result reach the person reading, inside the software they already read in. Where do the images go, and what was the model trained on. Who is accountable for a finding it missed. Those five questions do not change when the image is a slide rather than a scan.

The practical consequence is that the two departments have more to learn from each other's diligence than either has from a general healthcare AI list. Pathology arrived at whole slide imaging later than radiology arrived at digital acquisition, and its buyers are re running an evaluation radiology has already made most of the mistakes in.

Most of healthcare AI is not a regulated medical device. Documentation tools, administrative automation and most decision support reach the market without a submission, on the reasoning that a clinician reviews the output. Imaging is the exception, because a model that detects a finding in a study is doing the thing a device does, and the AI Health Index distribution shows it plainly.

47 of the 82 vendors here hold a top grade on FDA and Regulatory Status. Across all 554 vendors the index grades, in every category, 87 hold that grade. So the majority of regulatory standing in the whole index is concentrated in this one lane, held by a small fraction of the vendors in it.

Read this table as a map of where the lane's real strengths are rather than as a scoreboard. Regulatory status and published evidence are genuinely concentrated here. Nothing else is, and two of the rows below them are concentrated only because the index wide count is itself tiny, which is a statement about the whole market rather than a compliment to this part of it.

This is the single most consequential misreading in imaging AI procurement, and it is easy to make because the regulatory language sounds like a performance finding. A premarket notification records that a device was reviewed and found substantially equivalent to a predicate for a stated intended use and population. It is a permission to market. It is not a conclusion that the product improves detection, shortens turnaround or changes what happens to a patient, and it says nothing about performance in your department.

The AI Health Index grades the two separately for exactly that reason. 47 vendors here hold a top grade on FDA and Regulatory Status and 33 hold one on Clinical and Operational Evidence. Only 26 hold both, which means 21 of the cleared vendors have regulatory standing and no published evidence that the product does anything.

That is not an accusation. Clearance and publication answer to different audiences on different timelines, and a recently authorised product may have evidence in progress. What it does establish is that a procurement process which stops at the clearance letter has verified the weaker of the two available facts, and has usually verified it because it was the one the vendor volunteered.

Ask for the summary document behind the authorisation and read the intended use statement in it, which is public. Then ask separately what has been published, by whom, on how many sites, and whether the configuration studied is the configuration being sold.

Every model in this lane learned from a particular distribution of images: certain scanners, certain protocols, certain reconstruction settings, or in pathology certain stains, scanners and slide preparation practices. Change the acquisition and you have changed the input, sometimes enough to matter and usually without any visible signal that it has happened. This is the lane's characteristic failure and it does not announce itself, because the model returns a confident finding either way.

17 of 82 vendors hold a top grade on Setting and Specialty Coverage, the axis that asks for a clear published statement of the settings, specialties and populations a product is validated for. Combine that with regulatory standing and published evidence and the list falls to 7, which is the narrowest defensible bar this page publishes.

The question to put to a vendor is not whether the model generalises, which every vendor answers yes to. It is which scanner manufacturers and protocols appear in the validation set, what the performance was on the ones that are not yours, and what the deployment does when it encounters an acquisition it was not validated on. A vendor that runs a site specific calibration before go live is telling you something true about the problem, and that is a good sign rather than a caveat.

Population level accuracy is the wrong unit for this lane. The output is a finding about a named patient, so a model that performs differently across groups does not produce a slightly different average, it produces missed findings concentrated in the group it performs worst on. Imaging has well documented versions of this, from skin tone in dermatological imaging to sex and body habitus in cardiac and chest work to demographic composition in the source datasets many of these products learned from.

2 of 82 vendors hold a top grade on AI Governance and Bias Disclosure, the axis requiring a published bias or fairness evaluation carrying its methodology and the population it was run on. Across the whole index only 11 vendors hold it in any category. Pair it with regulatory standing and the number is 2, because an authorisation does not require subgroup reporting and so the two almost never travel together.

The recourse side is worse and it is industry wide rather than particular to imaging. 0 vendors in this lane hold a top grade on AI Liability and Recourse, and across all 554 vendors the AI Health Index grades, exactly 1 does. Where this lane differs is the shape of the distribution: 32 of 82 sit one band below the top and only 22 sit at the bottom, better than most of the index, because a regulated device carries a labelled intended use and a limitations statement whether or not the vendor wanted to publish one. Regulation has done here what disclosure did not.

Imaging carries a data question the rest of healthcare AI does not. The payload is large, it is identifiable in ways text is not, and in pathology a single whole slide image can be enormous, which makes the deployment topology a real constraint rather than a preference. 7 of 82 vendors hold a top grade on Deployment Model and Data Residency, the axis covering where the product runs and where the data rests.

There is a second question underneath it that is specific to this lane. The training data for an imaging model is patient imagery, usually acquired from health systems under agreements the buyer never sees. 4 vendors hold a top grade on Model Supply Chain Disclosure while 29 sit at the bottom of that axis, so provenance is the scarcest disclosure here just as it is elsewhere in the index, but the thing being withheld is different. Elsewhere the question is which foundation model provider receives the data. Here it is also whose patients the model learned from, and under what consent.

And the finding has to arrive somewhere useful. 12 of 82 vendors hold a top grade on EHR and Interoperability Depth. In imaging that means the reading and reporting systems as much as the record, so establish specifically what the integration writes and where: an annotation on the study, a structured result into the report, a worklist reprioritisation, or a separate application somebody has to remember to open. The last of those is how a well performing product quietly fails to change anything.

The AI Health Index publishes no composite score, so this page does not rank. It reports who clears a named bar. Pick the failure your department cannot absorb, and take the list under it. The cuts run from widest to narrowest and reading them in order is the point: the first list is where most imaging procurement stops, and every list under it is shorter.

A vendor absent from a cut has not failed a product test. It has not published what the cut requires, on the date shown. That distinction is the whole basis of the grading framework, and it means a shortlist here is a list of vendors you can verify rather than a list of the best ones.

Every vendor carries a grade on all 15 axes before it is published at all, so this table compares the same questions answered for every vendor rather than the results of uneven research. Counts are of vendors, as of August 31, 2026.

The strong end is real and worth stating: 59 of 82 are genuinely AI products rather than platforms with a feature, which is the highest concentration of any lane the index grades, and 47 carry regulatory standing. Imaging AI is the part of this market that has most thoroughly done the hard institutional work.

The thin end is uncomfortable in a specific way. 2 vendor of 82 publishes a pricing basis, against 32 across the whole index, making this the most commercially closed lane the AI Health Index grades. 2 publish subgroup performance. 0 publish recourse. So the field that has taken regulation most seriously has taken commercial and population transparency least seriously, and those are not the same kind of rigour.

The middle is where a buyer should spend time. 13 of 82 hold a top grade on Autonomy and Oversight Model, and most of the lane sits one band below, which usually means the oversight model exists and is described rather than published in a form a reviewer can check.

Every vendor in this lane graded by the AI Health Index, each one on all 15 axes with a source and a date on every judgement. Inclusion is not purchasable and no vendor pays for placement or review.