Almost everything written about AI in pathology skips the part that actually determines whether any of it happens. Before a model can read a slide, the slide has to exist as a file. And converting a pathology department from glass to pixels is a capital project, an IT project, a workflow redesign and a regulatory exercise — none of which involve machine learning at all.
I have watched this sequence from the algorithm side, and the ordering is worth stating plainly: digitisation first, everything else after. Labs that skipped straight to evaluating AI vendors without solving the scanning and storage problem ended up with nothing deployed.
The 2017 Threshold
The formal starting point in the United States was the FDA's decision to allow marketing of the first whole-slide imaging system for primary diagnosis. Before that authorisation, a pathologist in the US could use digital images for consultation, education or research, but the diagnosis of record had to be made by looking through a microscope.
That single decision mattered more for the field than any algorithm published before or since. It made it economically rational for a laboratory to install scanners, because the digital image could now be the thing the pathologist actually signs out on rather than a convenience copy. Every computational pathology deployment since sits on that foundation.
The pandemic accelerated what followed. When pathologists needed to sign out cases from home, regulators relaxed constraints on remote digital reading, and a transition that had been projected over a decade compressed into a couple of years in many institutions. Some of that was temporary. A lot of it stuck, because once a department has discovered that its staff can work from anywhere, the recruiting and retention implications are hard to give back.
What Digitisation Actually Costs
The scanner is the visible expense and usually not the dominant one.
A whole-slide image at diagnostic resolution runs to one to several gigabytes compressed. A department producing hundreds of thousands of slides a year is therefore generating storage in the hundreds of terabytes annually, and pathology material has retention requirements measured in years or decades. The storage problem is not a one-time purchase; it is a permanently growing line item, and it has to be fast enough that a pathologist opening a case does not wait.
Then there is the network. Moving multi-gigabyte images between the scanner, the archive, the viewer and any analysis service requires infrastructure that most hospital networks were not built for. Then integration with the laboratory information system, so that images attach to the right accession. Then the viewer software, which has to be good enough that pathologists prefer it to the microscope — a high bar, because the microscope is an extremely refined instrument and the people using it have spent years developing motor habits around it.
And then validation. Professional guidance recommends that a laboratory validate its own digital pipeline against glass on a substantial set of routine cases, read by the same pathologists with a washout interval between the two reads, before using it for primary diagnosis. This is the right requirement and it is real work — the validation has to cover the specimen types the lab actually handles.
None of this appears in the AI conversation, and all of it happens first.
What You Get Before Any AI
It is worth being clear that digitisation pays for itself on grounds that have nothing to do with algorithms.
A digital slide can be in two places at once. Subspecialist consultation that used to mean packing glass into a courier envelope and waiting days becomes a shared link. For a general hospital without a resident haematopathologist or dermatopathologist, that changes what can be diagnosed locally and how fast.
Slides stop being consumable. Glass breaks, gets misfiled, and stains fade over years. A digital archive does none of those things, and it is searchable — which turns a decade of accumulated cases into a resource for teaching, quality assurance and research rather than a room full of trays nobody opens.
Pathologists can work remotely, which is now a recruitment argument in a specialty with real staffing pressure.
And measurement becomes reproducible. Counting mitoses, quantifying a stain's intensity, measuring tumour extent — these are tasks humans do inconsistently, and doing them computationally removes a source of variability even when the computation is not remotely intelligent.
Then the Algorithms
Once the images exist, the useful applications are narrower than the excitement suggests but genuinely valuable.
The most convincing evidence I know of is not about detection at all. It is about consistency on a judgement that pathologists are known to disagree on. A 2020 study in Modern Pathology examined whether AI assistance improved Gleason grading of prostate biopsies, and it did: agreement with an expert reference standard rose from 0.799 to 0.872, and on external validation data from 0.733 to 0.786. The largest improvements went to the pathologists who had started out performing below the standalone algorithm.
That is the shape of the realistic benefit. Not a machine that diagnoses, but a machine that pulls the distribution of human performance upward and narrows its spread — which for a patient whose treatment depends on whether the biopsy is called 3+4 or 4+3 is a substantive improvement.
Beyond that: triage of a worklist so that likely-abnormal cases are read first, quality assurance where a model re-reads signed-out cases and flags disagreements for review, quantification of biomarkers that would otherwise be eyeballed, and screening support in high-volume, low-yield settings.
Where It Falls Over
Laboratory-to-laboratory variation is the dominant technical failure mode. Stain intensity depends on reagents, protocol, section thickness and fixation; scanners differ in colour rendering and focus. A model trained at one institution can degrade at another in ways that no internal validation will reveal. Normalisation and aggressive colour augmentation help. External validation is the only thing that tells you whether they helped enough.
Rare entities are systematically underserved. These models learn from what is in the archive, and the archive reflects prevalence. Common cancers are well represented; the unusual lymphoma subtype, the rare soft tissue tumour and the odd infection are not. This is exactly inverted from clinical need — the common case is the one a general pathologist handles comfortably, and the rare case is where help would be worth most. Any honest deployment treats model output on unusual material as uninformative rather than reassuring.
Automation bias is the clinical failure mode. A pathologist who reviews a model-negative slide less carefully than they would have otherwise has traded one error source for another, and the new one is harder to detect because it is correlated across cases. Workflow design matters more than model accuracy here.
Context is missing, and pathology is unusually context-dependent. The diagnosis often depends on the clinical history, the radiology, the site the specimen came from, prior biopsies, and sometimes a phone call to the surgeon. A model sees pixels. It is contributing one input to a judgement, and the framing of its output should make that obvious rather than presenting a confident label.
The Access Argument, Honestly
The most compelling case for this technology is global. Pathologist density varies enormously between countries, and in parts of sub-Saharan Africa the ratio is on the order of one pathologist per million people or worse. Cancer treatment depends on a tissue diagnosis. Where there is no pathologist, there is either no diagnosis or a long, expensive wait.
Digital pathology plus telepathology addresses this directly, and AI triage could extend it further by prioritising which cases need a scarce human expert.
The honest complication is that the bottleneck is frequently not the algorithm. It is the scanner, which is expensive. It is the bandwidth, because uploading a two-gigabyte image over an unreliable connection is not a solved problem. It is the histology laboratory upstream — you cannot scan a slide that was never cut and stained well, and tissue processing quality is itself a scarce capability. And it is validation: a model trained on tissue from one part of the world, processed in one kind of laboratory, cannot be assumed to work on material prepared differently in a different population.
Deployments that have worked in low-resource settings have generally led with connectivity and human expert networks, with AI as an accelerant rather than the premise. I find that ordering persuasive, and it matches what the field study literature on deploying medical AI in low-resource clinics keeps finding: the technology is rarely the hard part.
Where This Lands
Pathology is following radiology's path about a decade behind, and the delay is almost entirely about digitisation rather than about algorithms. The modelling problems in pathology are, if anything, more tractable than in radiology — the images are enormous but the task structure is often cleaner.
What that means practically: the interesting question for a laboratory is not which AI vendor to pick. It is whether the scanning, storage, integration and validation work is done, because that work is the precondition for everything and takes years. Labs that did it are now in a position to adopt computational tools quickly as they mature. Labs that did not are not, regardless of how good the models get.
The microscope is not going away tomorrow, and the pathologist is not being replaced. What is changing is the substrate — from glass that exists in one place to files that exist everywhere — and everything computational follows from that change rather than causing it.
Tags
Taresh Sharan