AI in pathology gets plenty of attention, and it can be hard to tell working tools apart from conference demos. This guide is for pathologists, lab managers and hospital decision-makers who want a grounded view of what artificial intelligence does well in the laboratory today, where it still falls short, and how to adopt it without overpromising to clinicians or budget holders.
What AI in pathology actually covers
The phrase is an umbrella for several quite different technologies. Lumping them together leads to confused expectations, so it helps to separate them by the job they do.
- Image analysis: algorithms that detect, segment, classify or measure structures on digitized slides, such as tumor regions, nuclei or mitotic figures.
- Quantification support: scoring aids for immunohistochemistry and other stains where the task is counting or estimating proportions.
- Triage and prioritization: tools that flag slides or cases likely to contain significant findings so they can be reviewed earlier.
- Quality control: checks for blur, tissue folds, pen marks, bubbles or missing tissue before a slide reaches the pathologist.
- Language tools: speech recognition, report drafting, structuring free text and checking reports for internal consistency.
Most image-based tools depend on whole slide images, so they presuppose a digital workflow. If your lab is still deciding whether to scan, start with what digital pathology involves in practice before evaluating algorithms.
Where AI in pathology helps today
The strongest use cases share a pattern: the task is narrow, well defined, repetitive and easy to verify by eye.
Counting and measuring
Humans are slow and inconsistent at counting hundreds of cells or estimating percentages across a large area. Algorithms are good at exactly this kind of tedious quantification. Used as an aid, they can make scores more reproducible between pathologists and between days, provided the pathologist checks that the algorithm counted the right region.
Screening and triage
Some tools pre-screen slides and highlight areas of concern, or sort a worklist so that likely positive cases surface first. This does not shorten the time needed to read a slide carefully, but it can change the order of work so urgent findings are seen sooner.
A consistent second look
An algorithm does not get tired at the end of a long sign-out session. As a safety net that flags regions for a second glance, it can reduce the chance that a small focus is overlooked. The pathologist still decides what the flagged area means.
Report text
Language tools can transcribe dictation, propose standard phrasing, and catch internal contradictions, for example a diagnosis line that does not match the description or a missing required element. These are low-risk, high-frequency tasks where review is quick.
What it cannot do (yet)
The limits matter as much as the strengths, especially when a vendor presentation blurs them.
- It does not see the whole patient. A pathologist integrates the clinical history, imaging, gross findings, prior specimens and conversations with the clinician. Current image algorithms see pixels from one slide or one case.
- It is narrow. Most tools are trained for one task on one tissue type with one stain. A model that grades one tumor type tells you nothing about another.
- It can fail quietly on unfamiliar data. Performance may drop when scanners, staining protocols, tissue processing or patient populations differ from the training data, and the tool will not necessarily warn you.
- It struggles with rare entities. Uncommon diagnoses are, by definition, underrepresented in training sets.
- It cannot fix upstream problems. Poor fixation, thick sections or faded stains produce poor images, and poor images produce unreliable outputs.
- It cannot take responsibility. The report is signed by a pathologist, and that accountability does not transfer to software.
Regulation, validation and responsibility
In many jurisdictions, software intended to support diagnosis is regulated as a medical device, and requirements differ between countries and between intended uses. A regulatory clearance in one market says something about the evidence the manufacturer submitted. It does not prove the tool will perform the same way on your scanner, with your stains and your case mix.
For that reason, treat an AI tool like any other new method in your quality management system. Define its intended use, verify performance on your own material before routine use, document the result, train users, and monitor it over time. If your lab works under ISO 15189, the international standard for medical laboratory quality and competence, these steps fit naturally into how you already handle method changes.
Responsibility stays where it has always been. AI output is an input to the pathologist's judgment, recorded and reviewable, never a substitute for it.
How to start without overcommitting
A cautious, staged approach protects both patients and budgets.
- Pick one painful, measurable problem. A scoring task with poor reproducibility, a backlog in a specific case type, or slow report typing are better starting points than a general wish to use AI.
- Check your foundations. Image-based AI needs consistent slide quality, reliable scanning and a way to view results inside the normal workflow.
- Run a local evaluation. Use your own retrospective cases, compare output with signed-out diagnoses, and look closely at the failures, not only the averages.
- Pilot with a small group. Let a few pathologists use the tool in parallel with normal practice and collect their feedback on usability and trust.
- Decide, then monitor. If you adopt it, keep tracking agreement and errors, and recheck after any change to scanners, stains or software versions.
Common mistakes include buying a tool before the digital workflow is stable, accepting vendor accuracy figures without local testing, and introducing AI output in a separate application that pathologists have to switch to. Each of these turns a promising tool into shelfware.
Key takeaways
- AI in pathology is a family of narrow tools, not a single capability, so evaluate each one against a specific task.
- Today it is most useful for counting, triage, quality checks and report text, where output is easy to verify.
- It does not integrate clinical context, handles unfamiliar data poorly, and cannot take responsibility for a diagnosis.
- Local validation and ongoing monitoring under your quality system matter more than headline accuracy claims.
- Start with one measurable problem and a stable digital workflow.
If you are exploring where assistance fits in your lab, PATHOLOGY+ offers AI-assisted drafting and consistency checks of report descriptions that the pathologist reviews and signs. You can book a demo to see how it works in a real sign-out.