Computational Pathology: What CAMELYON16 Really Showed

There is a detail in the CAMELYON16 results that almost never gets quoted, and it changes the meaning of the whole study.

The challenge asked algorithms to detect breast cancer metastases in sentinel lymph node sections. The results were published in JAMA in 2017, and the headline was that the best algorithms exceeded the performance of a panel of eleven pathologists. That is what got reported.

The detail: the pathologist panel was working under a simulated time constraint, roughly mimicking the pace of a real workday. One pathologist was also allowed to read the same slides without any time limit. That pathologist scored far above the time-constrained panel — close to the best algorithms.

So the honest reading is not that the algorithm was more capable than a pathologist. It is that the algorithm did not get tired, did not have to triage its attention across a full worklist, and did not miss small metastases because it had ninety more slides to get through. The comparison was never really about diagnostic skill. It was about sustained attention over a gigapixel image, which is a task human visual systems are genuinely poorly suited to and machines are genuinely well suited to.

Once you see that framing, most of computational pathology makes more sense.

The Scale Problem

A whole-slide image at diagnostic resolution is on the order of a hundred thousand pixels on a side — a gigapixel image, one to several gigabytes compressed. A pathologist navigating it at high magnification is sampling. They move systematically, they know where to look based on architecture visible at low power, and they are very good at this. But a single cluster of malignant cells occupying a tiny fraction of the tissue can be missed, and the literature on inter-observer variability reflects that.

This scale is also why pathology AI is architecturally different from radiology AI. You cannot feed a gigapixel image to a network. The standard approach tiles the slide into thousands of small patches, encodes each one, and aggregates — usually with multiple instance learning, because the available label is at the slide level ("this biopsy contains cancer") rather than at the pixel level. Nobody has time to annotate every malignant cell on ten thousand slides.

That weak supervision is both the enabling trick and a persistent source of trouble. The model learns to associate patch-level features with a slide-level outcome, and it can pick up correlations that have nothing to do with the biology — the tissue block came from a different lab, was cut thicker, was stained on a different day.

The Distribution Shift Problem Is Worse Here

In radiology, a model can learn the scanner. In pathology, a model can learn the laboratory, and the effect is larger.

Haematoxylin and eosin staining varies with reagent batch, staining protocol, section thickness, fixation time, and how long the block has been sitting around. Two labs staining serial sections of the same tissue produce visibly different images. Add scanner differences on top and the variation between institutions is substantial.

The consequence is that a model trained on one lab's slides can degrade sharply on another's, and the degradation can be invisible without deliberate testing. Stain normalisation and heavy colour augmentation during training help considerably. They do not eliminate the problem, and any pathology AI claim without external validation on independent laboratories should be treated as provisional.

Where the Evidence Is Solid

The strongest evidence in this field, to my mind, is not about AI replacing a read. It is about AI improving a notoriously inconsistent human judgement.

Gleason grading of prostate biopsies is the canonical example of a subjective, high-stakes, poorly reproducible task. The difference between a 3+4 and a 4+3 changes management, and pathologists disagree about it constantly.

A 2020 study in Modern Pathology tested whether AI assistance improved pathologists' grading. It did. Agreement with an expert reference standard rose from 0.799 to 0.872 on the main assessment, and from 0.733 to 0.786 on external validation data. The pathologists who gained the most were those who had initially performed below the standalone algorithm. Participants rated the visual overlays highlighting tumour patterns as the most useful part.

I like this study because of what it measures. Not "can the machine beat the human", but "does the combination do better than the human alone, and for whom". The answer — yes, and most for those who needed it most — is both more useful and more plausible than a superiority claim.

What Is Actually Cleared

Regulatory reality in pathology is considerably behind radiology, for a straightforward reason: the underlying digitisation came later. You cannot deploy computational pathology in a lab that still reads glass under a microscope, and most labs did until recently.

The landmark is the FDA's 2021 De Novo authorisation of an AI system for prostate biopsy assessment — the first artificial intelligence authorised for digital pathology in the United States. Read the indication carefully, because it is narrower than the coverage suggested: the system operates as an adjunct that flags tissue areas with the highest likelihood of harbouring cancer, for a pathologist to review. It does not make the diagnosis. The pathologist does, as before.

In Europe, a larger number of products carry CE marking, and the in-vitro diagnostic regulation that came into force has substantially tightened what that requires. The practical situation in most of the world is that computational pathology is used as a second read, a quality assurance step, or a triage mechanism — not as a primary diagnostic authority.

The Genuinely Interesting Frontier

Predicting molecular characteristics directly from routine H&E morphology is the part of this field I find most scientifically interesting, and it deserves careful framing.

Certain genetic alterations produce subtle morphological changes that models can learn to recognise — microsatellite instability in colorectal cancer being the most studied example. The models work better than chance and considerably better than pathologists attempting the same judgement by eye. Reported performance is moderate rather than excellent, and varies by target and by cohort.

The realistic use is triage, not replacement of molecular testing. If a model can identify which cases are most likely to be positive, a lab can prioritise confirmatory testing, which matters when the test is expensive, slow, or consumes scarce tissue. Any claim that H&E-based prediction substitutes for sequencing is wrong, and every serious group working on this says so themselves.

Foundation models are the other current frontier — large encoders trained with self-supervision on enormous archives of unlabelled slides, then adapted to specific tasks with far less labelled data. Several have been released and the transfer learning results are genuinely good. Whether they solve the cross-institution generalisation problem or merely move it is, as far as I can tell, still open.

What Digitisation Actually Buys

It is easy to lose the unglamorous benefits in the AI conversation, and for many laboratories they are the larger part of the value.

A digital slide can be in two places at once. Second opinions that once required physically shipping glass across the country happen in minutes. Subspecialist expertise becomes reachable from a general hospital. Pathologists can work remotely, which changed staffing possibilities permanently. Archives become searchable. Teaching sets become infinite instead of consumable.

None of that requires a neural network. All of it requires the scanners, the storage and the workflow software that AI deployment also depends on, which is why digitisation is the prerequisite step and why adoption is slower than the algorithm literature implies.

The Realistic Picture

Computational pathology works best in the same shape as everywhere else in medical AI: as a second reader that catches what sustained attention misses, as a consistency aid on subjective grading, and as a triage mechanism for prioritising cases and tests.

The failure modes worth watching are automation bias — a pathologist who stops looking as carefully at slides the model called negative — and the cross-laboratory generalisation gap, which is the most likely cause of a system that validated well and then underperformed after deployment.

What I would not expect soon is autonomous diagnosis. Pathology diagnosis integrates the slide with clinical history, prior specimens, radiology, molecular results and, frequently, a conversation with the surgeon about what was actually taken and from where. The pixel analysis is one input to that. It is the input most amenable to automation, which is exactly why it is being automated first, and it is not the whole of the reasoning.

The microscope is not being replaced by a machine that thinks. It is being replaced by a scanner, a screen, and a very good attention mechanism — and the pathologist is still the one who decides what it means.