AI for Diagnostics and Pathology Labs
Laboratories are the most quantified environment in healthcare, which makes them unusually good ground for machine learning and unusually unforgiving about it, because everyone in the building already knows what a validation study looks like.
Labs already run validation studies, track precision and accuracy, and understand analytical performance in a way most healthcare organisations do not. That makes them excellent clients and it raises the bar considerably — a model with a hand-waving evaluation will not survive the first conversation with your quality manager, and it should not.
AI for diagnostics and pathology laboratories covers whole-slide image analysis, autoverification of routine results, quality control anomaly detection, workload routing and prioritisation, and result interpretation support — under IVDR where a product is placed on the market.
Digital pathology, and the variability that breaks it
Whole-slide image analysis works, and the thing that determines whether it keeps working is pre-analytical and technical variability that has nothing to do with the algorithm. A model trained on your slides, your scanner and your staining protocol will degrade when any of those changes, and all of them change.
| Source of variation | Effect on the model | Mitigation |
|---|---|---|
| Scanner make and model | Colour profile and resolution shift | Multi-scanner training data; scanner as a monitored covariate |
| Stain lot and protocol drift | Systematic colour shift over weeks | Stain normalisation; QC on colour statistics per batch |
| Section thickness and fixation | Texture and morphology change | Captured as metadata; performance reported by site |
| Laboratory of origin | Everything above, combined | Validation split by lab, never at random |
| Annotation variability | The label itself is noisy | Multi-reader annotation with adjudication and reported agreement |
Your pathologists disagree, and that is data
Inter-observer agreement on many pathology tasks is materially below what people assume, and a model trained on single-reader labels inherits that noise while appearing more certain than the underlying evidence supports. Measuring reader agreement first tells you the ceiling on achievable performance — and occasionally tells you the task as specified is not well-defined enough to model, which is a genuinely useful thing to learn in week three rather than month nine.
Autoverification, which is the highest-return work in most labs
Most routine results are normal, unremarkable and reviewed anyway. Extending autoverification beyond simple rule sets — using patient history, delta checks, instrument flags and analyte relationships — releases scientist time toward the results that need judgement. It is measurable, it is inside your existing quality framework, and it is unglamorous enough to be under-invested.
Design for asymmetric error
Releasing an abnormal result as normal is a clinical harm. Holding a normal result for review costs a few minutes. The threshold must reflect that asymmetry explicitly rather than optimising a balanced accuracy metric, and the choice belongs to your laboratory director rather than to whoever tunes the model.
Delta checks are where prediction adds most
A result that is within reference range but implausible given the patient's own history is the case rule-based autoverification handles worst and where a model that sees the trajectory adds real value.
Keep it inside your existing validation framework
Labs already know how to validate a method: precision, accuracy, reportable range, interference, comparison with the existing procedure. Treat the model as a method and validate it the same way. It is the fastest route to something your quality manager will sign, and it produces better evidence than most AI evaluations do.
Where the regulatory line falls
The critical distinction is whether you are using a system internally or placing it on the market. An in-house tool used within your own laboratory sits under laboratory-developed test rules and your accreditation body's expectations. A product sold to other laboratories is an in vitro diagnostic and falls under IVDR — and if it requires notified body assessment, it is also Annex I high-risk under the EU AI Act with obligations from 2 August 2028.
- In-house use. Your accreditation standard and laboratory director's validation requirements govern. Substantial but familiar, and the framework already exists.
- Placing on the market. IVDR conformity assessment plus, for most AI-enabled IVDs, AI Act Annex I obligations. A materially different programme — see AI for medical device manufacturers.
- Result interpretation support. Suggesting an interpretation to a pathologist is decision support; issuing it to a clinician is a diagnostic claim.
- Screening applications. Where a model determines what a human never sees, the evidence bar is higher, because a missed case is invisible by construction.
- Post-market performance monitoring is now an explicit expectation, and it requires telemetry designed in rather than added.
How an engagement runs
Validated the way a laboratory validates a method, because that framework is better than most AI evaluations.
Scope and reader agreement
The task, its regulatory position, and inter-observer agreement on the labels — which sets the ceiling on achievable performance.
Data assembly and variability audit
Scanners, stains, sites and protocols represented, with variability quantified before anything is trained.
Build and method validation
Precision, accuracy, comparison with the existing procedure, and performance broken down by scanner, site and stain batch.
Parallel running
Alongside the existing process, with discordant cases reviewed individually rather than summarised.
Phased deployment and QC integration
Into the LIS workflow, with model performance monitored inside your existing quality control regime.
What you receive
Evidence in the form your laboratory director already trusts.
Reader agreement study
What the achievable ceiling is, before anyone spends money trying to exceed it.
Variability assessment
How much scanner, stain, site and protocol variation exists and what it does to performance.
Method validation package
Precision, accuracy, comparison and reportable range — in the format your accreditation body expects.
Model in the LIS workflow
Where the result is actually reviewed, not in a separate application.
Autoverification threshold analysis
Asymmetric error explicitly costed, with the decision left to your laboratory director.
QC integration
Model performance monitored inside the quality control regime you already run.
Is this the right starting point?
Worth being direct. There are situations in diagnostics and pathology where custom AI work is the wrong spend, and those are listed rather than buried.
Worth doing if
- High-volume routine work where autoverification could safely be extended.
- Digital pathology with slides across more than one scanner or site.
- Turnaround time constrained by review capacity rather than by analytical time.
- You want validation done to laboratory standards rather than to software standards.
- You are considering placing a tool on the market and need the IVDR position settled early.
Do something else if
- The instrument vendor already offers this validated — buy it.
- Slides are not digitised. Scanning comes first, and it is a substantial programme.
- Reader agreement on the task is poor and the task has not been redefined.
- The bottleneck is specimen transport or pre-analytics, which no model touches.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Does a model trained on our slides work on another lab's?
Usually worse, sometimes much worse. Scanner colour profiles, stain lots and protocols, section thickness and fixation all shift the image in ways the model has learned as signal. Multi-site training data and stain normalisation help, and validation split by laboratory rather than at random is what tells you the truth. Any performance figure from a single-site study should be treated as an upper bound.
Is our in-house tool regulated?
Used only within your own laboratory, it generally sits under laboratory-developed test rules and your accreditation body's validation expectations rather than under IVDR. Placed on the market for other laboratories, it is an in vitro diagnostic under IVDR — and where notified body assessment is required, also Annex I high-risk under the EU AI Act from 2 August 2028. The distinction is worth settling before you build, because it changes the programme entirely.
How much autoverification is safe?
That is your laboratory director's decision and the model's job is to inform it honestly. The design principle is asymmetric: releasing an abnormal result as normal is a clinical harm, holding a normal result costs minutes, and the threshold should reflect that rather than optimising a balanced metric. We provide the performance at each candidate threshold, with the miss rate for each abnormality class stated plainly, and the decision stays where accountability sits.
What if our pathologists disagree about the labels?
Then measure it first, because inter-observer agreement sets the ceiling on what any model can achieve and it is frequently lower than people expect. Multi-reader annotation with adjudication gives better labels and an honest uncertainty estimate. Occasionally the finding is that the task as specified is not well enough defined to model — which is a useful thing to discover in week three rather than in month nine.
Can AI replace a pathologist?
No, and the framing misses where the value is. What works is prioritising the worklist so urgent cases surface first, pre-screening to direct attention within a slide, quantifying what is tedious to count by eye, and flagging discordance for a second look. Each of those makes a pathologist faster and more consistent. Systems designed to remove the pathologist have a poor record and a worse regulatory path.
Related verticals
Organisations in diagnostics and pathology usually share data, buyers or regulators with these. All fourteen are listed on the Healthcare & Life Sciences page.
AI for Hospitals and Health Systems
Imaging triage, ambient documentation, patient flow forecasting and coding — with clinical safety governance built in rather than added.
Read more →AI for Medical Device Manufacturers
Model development under device regulation — PCCPs, Annex I readiness, and monitoring designed for the submission.
Read more →AI for Biotechnology
Experimental data infrastructure, screening analysis and phenotyping — starting with reproducibility, which is usually the actual problem.
Read more →Tell us what the problem looks like.
Thirty minutes, no charge, no deck. We will tell you whether this is an AI problem, a data problem, or a process problem — and we will say when the honest answer is to buy something rather than build it.