Pathology is the strangest case in clinical AI: it has produced some of the most striking research results in the entire field and one of its smallest cleared footprints. Algorithms have matched expert readers on cancer-detection tasks in published studies for years, yet the number of FDA-authorized pathology tools, and the number of laboratories running them, remains modest. This guide reads both sides plainly — what has actually been authorized, and what the strongest studies measured — with every figure tied to a primary source. As of July 2026.
The digital prerequisite
Radiology has been natively digital for decades; pathology has not. A pathologist's raw material is a glass slide under a microscope, and before an algorithm can read it, the slide has to become a whole-slide image — a gigapixel scan of the tissue. That digitization step is the precondition for everything below, and it is the main reason pathology trails radiology in deployment: the scanners, storage, and workflow have had to be built out laboratory by laboratory. Where radiology AI could ride existing digital infrastructure, computational pathology had to lay its own track first.
What is FDA-authorized
On the FDA's AI-Enabled Medical Device List 1, pathology is present but thin. The landmark is 2021, when the FDA granted its first De Novo authorization for an AI system in digital pathology — a tool that flags regions of a prostate biopsy image suspicious for cancer for a pathologist to review 1. The design is assistive by construction: the software points, the pathologist decides and signs out the case.
The evidence behind that assistive model is quantified. In a study of AI-assisted reading of prostate core-needle-biopsy whole-slide images, pathologists' average sensitivity for cancer rose from 74% unassisted to 90% with AI assistance, with no statistically significant change in specificity (97%), and readers spent less time per slide 2. That is the shape of the value proposition in cleared pathology AI today: a second read that catches more of what a tired human eye misses, while a pathologist stays in the loop as the decision-maker.
The research spine
The published literature runs far deeper than the cleared list, and three studies anchor it.
The first established that computational pathology could work at scale without hand-drawn annotations. A weakly supervised deep-learning system trained on 44,732 whole-slide images from 15,187 patients — using only the reported diagnoses as labels — reached AUC above 0.98 for prostate cancer, basal cell carcinoma, and breast-cancer metastases in lymph nodes, and its authors showed it could let pathologists exclude 65-75% of slides while retaining 100% sensitivity 3. The practical promise there is triage: safely setting aside clearly benign slides so human attention concentrates where it counts.
The second is the CAMELYON16 lymph-node evaluation, which pitted algorithms against pathologists on the same 129 whole-slide images. The top algorithm reached an AUC of 0.994, while a panel of 11 pathologists reading under a routine-workflow time constraint averaged an AUC of 0.810; a single pathologist reading without a time limit reached 0.966 4. The nuance is the whole point: the algorithm did not beat unhurried expert reading, but it matched or exceeded what pathologists achieved under realistic time pressure — which is the condition most sign-outs actually happen in.
The third is the PANDA challenge, the largest test of automated Gleason grading. 1,290 developers built algorithms on 10,616 digitized prostate biopsies, and on external validation the best reached quadratically weighted kappa of 0.862 in a US cohort and 0.868 in a European cohort against expert uropathologists 5 — agreement that held across different populations, laboratories, and reference standards, which is exactly the generalization that usually fails.
| Study | Scale | What it measured | Headline result | Source |
|---|---|---|---|---|
| Weakly supervised system | 44,732 slides / 15,187 patients | Detection across 3 cancers | AUC > 0.98; exclude 65-75% of slides at 100% sensitivity | 3 |
| CAMELYON16 | 129 slides, 11 pathologists | Lymph-node metastasis | Top algorithm AUC 0.994 vs panel 0.810 (time-limited) | 4 |
| PANDA challenge | 10,616 biopsies, 1,290 teams | Gleason grading | Kappa 0.862 (US) / 0.868 (EU) vs uropathologists | 5 |
Read these as task-specific results rather than a general verdict. Each measures one cancer type on curated cohorts, and the strongest — CAMELYON16 and PANDA — are challenge datasets, which reward performance on a fixed benchmark rather than on the messy stream of a working laboratory. The discipline for weighing any such number is in our guide on how to read an AI validation study.
Why the cleared footprint stays small
If the research is this strong, why is the authorized list this thin? Three reasons, each visible in the primary record.
Digitization lag. As above, pathology is still building the whole-slide infrastructure that AI needs, so fewer sites can run these tools even where they exist.
Radiology sets the pace. The FDA record is lopsided by specialty. Of the 168 machine-learning devices authorized in 2024, radiology took 74.4%, with cardiovascular and neurology next; pathology is not among the leading panels 7. The clearance pipeline is dominated by imaging.
The generalization problem is harder to clear. A whole-slide image varies with the scanner, the stain batch, and the laboratory's fixation protocol far more than a standardized radiology series does, so a model that excels on one site's slides can drift on another's. That makes robust external validation — the property regulators and buyers should demand — both more necessary and more difficult in pathology. And as with the rest of the field, none of the FDA-authorized AI devices yet use large language models 6: cleared pathology AI is quantitative image analysis rather than a chatbot reading slides. The Good Machine Learning Practice principles from the FDA, Health Canada, and the MHRA (2021) frame the development expectations these devices are held to 8.
The generalization problem, up close
The reason a stunning benchmark can disappoint in a working laboratory deserves its own paragraph, because it is the crux of pathology AI. A whole-slide image is shaped by the entire pre-analytic chain: the scanner make and its colour profile, the stain batch and its intensity, the thickness of the section, the fixation time, and the laboratory's own habits. A model that learned on one institution's slides can read another's as subtly out-of-distribution — the same tissue, rendered in a palette it was never trained on. Interoperability compounds the issue, because whole-slide formats and viewers differ across vendors, so even moving an image between systems can change what a model sees. This is why the external-validation results in CAMELYON16 and PANDA carry more weight than any single-site accuracy figure: PANDA's agreement held across US and European cohorts and multiple laboratories 5, which is the property that predicts travel. A laboratory cannot read a vendor's headline number and assume it will reproduce on its own scanners — the generalization question is the deployment question.
Questions before you deploy
Because cleared pathology tools are assistive, adopting one is a decision about a workflow, a signature, and a liability — a pathologist still owns the case. A handful of questions decides whether a tool helps or merely adds a click.
- What is the cleared claim, and what stays with the pathologist? Authorized pathology AI flags regions suspicious for cancer for review 1; it does not sign out. Confirm exactly what the tool asserts and what remains a human judgement.
- Was it validated on your scanners and stains? Given how much a whole-slide image varies by site 5, ask whether the tool was tested on the scanner models and staining protocols your laboratory actually runs — the external-validation question in its most literal pathology form.
- Where does it save time, and at what review cost? The value in the assistive studies is a second read that raises sensitivity 2, or a triage that safely sets aside clearly benign slides 3. Measure the review time it adds against the misses it prevents.
- How are its errors surfaced? A tool that flags confidently and fails silently is more dangerous than one that abstains. Ask how false positives and false negatives present to the reader, and how the human in the loop is expected to catch them.
- Who monitors drift, and on what cadence? Stain suppliers change, scanners are replaced, case mix shifts. Post-market monitoring with a named owner — the Good Machine Learning Practice expectation 8 — is part of the purchase, rather than an afterthought.
The pattern mirrors radiology's: the published evidence tells you a task is achievable, and a local, blinded evaluation on your own slides tells you whether this particular tool achieves it in your laboratory. In a field where the scanner is part of the experiment, borrowed benchmarks travel least well of all. The corollary is a budget line procurement often forgets — the cost of validating and re-validating on your own material as scanners and stains change over the tool's working life, which is where an inexpensive licence can quietly become an expensive commitment.
How to read this guide
Four cautions travel with everything above. First, research accuracy is not deployed accuracy: challenge-dataset and single-institution results routinely overstate how a model behaves on your scanners and your case mix. Second, cleared tools are assistive: the value on the label is a better-supported pathologist, not an autonomous diagnosis, and the pathologist remains accountable for sign-out. Third, the specialty split is real: pathology's thin authorized footprint reflects infrastructure and generalization hurdles rather than a ceiling on what the science can do. Fourth, status is perishable — the device list grows, new prospective studies land, and every figure here is dated to when it was pulled.
Because clearance status, laboratory accreditation, and liability turn on these distinctions, confirm the current regulatory status and evidence for any specific tool with your compliance, laboratory-quality, or regulatory counsel before acting; a clearance is not the same as validation quality on your own material. We revisit this guide on a 180-day cycle and whenever the FDA authorizes a new pathology device or a prospective trial reports. For the neighboring specialties, see our guides on AI in radiology and AI in cardiology, and for the field-wide counts, our tracker of FDA-cleared AI devices by year and specialty.
Sources and method
This guide draws on the FDA's AI-Enabled Medical Device List 1; the assistive prostate-detection study behind the first authorized pathology AI 2; three research anchors — a 44,732-slide weakly supervised system 3, the CAMELYON16 lymph-node contest 4, and the PANDA grading challenge 5; a taxonomy of 1,016 FDA authorizations 6; a 2024 authorization audit 7; and the tri-regulator Good Machine Learning Practice principles 8. Every figure is tied to the primary source cited beside it, and each perishable number carries its date. Statuses and counts are current as of July 2026.