Computational Pathology Has a Prerequisite Nobody Talks About

Computational pathology has a prerequisite that radiology never had: the images have to exist as images.

Radiology went digital decades ago because the scanners produce digital output. There was no choice to make. Pathology did not, and mostly still has not. In most laboratories in the world, a pathologist puts a glass slide on a microscope stage and looks through the eyepieces. There is nothing for an algorithm to read.

This single fact explains more about the state of AI in pathology than anything about the algorithms, and it is consistently missing from discussions of the field. Before you can deploy a model, someone has to buy scanners, rebuild the laboratory workflow around them, and solve a storage problem that is genuinely large. That is a capital project measured in years, not a software purchase.

Why the images are hard

A whole slide image is not a big photograph. It is a different category of object.

Scanned at the 40x objective typically used for diagnostic work, a single slide produces an image on the order of 100,000 by 100,000 pixels. Uncompressed that is tens of gigabytes; stored as a compressed pyramidal format it is usually somewhere between half a gigabyte and a few gigabytes. A busy laboratory produces hundreds of slides a day. The annual storage requirement lands in the hundreds of terabytes, and pathology slides have retention obligations measured in years or decades.

For machine learning, the size creates an architectural problem. You cannot feed a 100,000-pixel-square image into a convolutional network. The standard approach is to tile the slide into thousands of small patches, process those, and aggregate. That works, but it introduces the defining difficulty of the field: the label applies to the slide, not to the tile.

A pathologist reports that a specimen contains carcinoma. They do not annotate which of the forty thousand tiles contain it, and asking them to would be prohibitively expensive at any useful scale. So the training signal is weak: somewhere in this bag of tiles there is evidence for this label, and the model has to work out where. Multiple instance learning and attention-based aggregation are the standard answers, and they work well enough to have produced the field's real results — but they need far more slides than a fully annotated approach would, and they are considerably harder to debug when they go wrong.

The other structural difficulty is stain variation. Haematoxylin and eosin staining is a chemical process, and its output depends on reagent batch, staining protocol, section thickness, fixation time and the scanner's colour response. Two laboratories staining consecutive sections of the same block produce visibly different images. A model trained at one institution frequently degrades at another for no reason more interesting than the pink being a different pink. Stain normalisation and colour augmentation are now routine countermeasures and they help substantially, but this remains one of the most reliable ways to make a pathology model fail in deployment.

What has actually been demonstrated

The landmark result is still the CAMELYON16 challenge, in which teams built algorithms to detect breast cancer metastases in sentinel lymph node whole slide images. The results were published in JAMA in 2017, and the comparison design is what makes them worth quoting carefully.

The best-performing algorithms achieved higher discrimination than a panel of eleven pathologists working under a simulated time constraint meant to mimic routine workflow, and performed comparably to a pathologist interpreting the same slides without any time limit.

Read that precisely. The headline "AI beats pathologists" comes from the time-constrained comparison, and the time constraint was doing real work — small metastases are easy to miss when you are scanning a lot of tissue quickly, which is exactly the failure mode an exhaustive automated search should address. Against a pathologist given unlimited time, the algorithms were comparable, not superior. Both findings are useful; they say different things.

On the regulatory side, the enabling events were structural rather than algorithmic. The FDA first permitted marketing of a whole slide imaging system for primary diagnosis in 2017, which is what made digital reading legally viable in the US at all. An AI product for prostate cancer detection in biopsy slides received De Novo authorisation in 2021 as an adjunct to the pathologist's review. The number of authorised AI pathology products remains small — pathology is a long way behind radiology on this, and the digitisation gap is most of the reason why.

The workforce argument, handled honestly

AI in pathology is often justified by workforce shortage. A study covered by the College of American Pathologists found the practising US pathologist workforce had been undercounted by nearly 40 percent in previous estimates, which is a useful corrective to some of the more alarming projections.

The pressure is real but it is not uniform. It is acute in subspecialties, acute in particular regions, and largely absent in well-staffed academic centres. Global distribution is where the shortage is genuinely severe, and that is also where the capital cost of scanning infrastructure is hardest to justify — a familiar and uncomfortable pattern.

What I would resist is the arithmetic that says a model reads slides faster, therefore a laboratory needs fewer pathologists. In every deployment I am aware of, the model is an adjunct: it pre-screens, flags regions, or quantifies something, and the pathologist still examines the case and signs the report. The time saved is real but it is a fraction of the case, not the whole of it, and some of it is given back to reviewing the model's flags.

Where computational pathology is genuinely strong

Three categories stand out, and they are not the ones that get the coverage.

Exhaustive search on large, sparse tissue. Finding a small focus of metastatic tumour in a lymph node, or a few malignant glands in a long prostate core, is a needle-in-haystack task where tirelessness beats expertise. This is the clearest win and the one CAMELYON16 demonstrated.

Quantification. Counting mitoses, scoring immunohistochemistry, estimating tumour-infiltrating lymphocytes, measuring proportions. These are tasks where human performance is known to be variable and where the answer feeds directly into treatment decisions. Replacing an eyeballed estimate with a reproducible count is a smaller-sounding contribution than diagnosis and a more defensible one.

Prediction from morphology. The genuinely novel direction: models that predict molecular or genomic features, or treatment response, from routine stained slides — information a pathologist cannot extract by eye at all. This is research-stage, it needs careful external validation, and it is the only place where these systems are doing something categorically beyond what a human does rather than doing the same thing faster.

What I would ask before buying one

Is the laboratory digitised, and if not, is the scanner and storage programme funded? Everything else is downstream of this.

Was the model validated on slides stained and scanned by this laboratory, with this protocol, on this scanner? Cross-site validation in pathology is not optional, because stain is a site-specific variable.

What exactly is the intended use — screening, adjunct, quantification — and does the reporting workflow match it? An adjunct that appears after the pathologist has reached a conclusion changes nothing.

How is performance monitored when the staining protocol changes, as it will?

And who signs the report? The answer should be a pathologist, and the governance should say so in writing.

The realistic position

Computational pathology is roughly where radiology AI was several years ago, held back less by modelling than by the fact that most of the world's pathology is still analogue. The technical results are real and, in the exhaustive-search and quantification categories, quite convincing. The molecular-prediction work is the most scientifically interesting thing happening in medical imaging right now.

None of it reaches a patient until laboratories are digitised, and that is a slide-scanner and data-centre problem rather than a machine learning one. It is not the part anybody wants to write about. It is the part that determines the timeline.