When the AI stain invents tissue

I spend my working life looking at spectra, so I have a specific reflex about instruments: I want to know what happens when the instrument is wrong, and how I would find out. Most measurement disciplines are built around that second question. You run a blank, a standard, a known sample. The point of a reference is not that it is interesting. The point is that it is the thing that can contradict you.

Virtual staining is a genuinely impressive technology that quietly removes the reference. This report is about what a group at UCLA found when they went looking for the consequences.

What virtual staining is

Histopathology has worked the same way for over a century. You take tissue, fix it, embed it, slice it, and stain it. Haematoxylin and eosin, H&E, is the workhorse: haematoxylin turns nuclei blue-purple, eosin turns cytoplasm and extracellular matrix shades of pink. A pathologist looks at the result and makes a diagnosis.

Virtual staining replaces the chemistry with a neural network. You image the unstained tissue using its own autofluorescence, feed that into a deep generative model trained on paired autofluorescence and chemically-stained images, and the model outputs a picture that looks like an H&E slide.

The advantages are real, and I do not want to undersell them. No staining reagents. No 30 to 60 minutes on a slide stainer. Less stain-to-stain variability between labs. The tissue is preserved rather than consumed, so you can run other assays on it. You can generate several different virtual stains from one acquisition. For a field that is short on technicians and long on turnaround time, this is a serious proposition.

And here is the structural fact that drives everything else. In deployment, the whole point is that you do not make the chemically stained slide. That is the saving. Which means the reference image, the thing you would compare against to check the model, does not exist.

The paper states this plainly:

in the deployment phase of the VS model there would be no HS available, and therefore, supervised evaluation metrics based on ground-truth images cannot be used in the VS workflow

Read that as a measurement problem rather than a software problem and the shape of it becomes familiar. You have built a pipeline whose economic justification is the deletion of its own control.

Two kinds of wrong

The UCLA group, led by Aydogan Ozcan at the California NanoSystems Institute, split the failure modes into two categories. The distinction is the most useful idea in the paper.

The first they call technical or unrealistic hallucinations: blurring, tissue-fold artifacts, aberrantly stained regions. These look like the artifacts that pathologists already deal with on glass slides every day, and a trained expert spots them. They are annoying, not dangerous.

The second category is the problem. Realistic hallucinations are outputs where the model has replaced real tissue components with invented ones that look entirely plausible. In the authors' words:

These realistic hallucinations might mislead pathologists, deceiving them to diagnose features that do not appear in the actual tissue specimen, although looking realistic and believable from the perspective of tissue staining quality.

The paper spells out what that means clinically: hallucinated tumour cells appearing inside a benign section, hallucinated mitotic figures altering a tumour grade, hallucinated lymphocytes in a tumour microenvironment changing a predicted treatment response.

This is the inversion that makes generative models different from ordinary instruments. A conventional instrument that fails usually produces obvious garbage, and garbage is self-announcing. A generative model trained on an adversarial objective is explicitly optimized to produce output that a discriminator cannot distinguish from real. Plausibility is the training target. So when it fails, it does not fail toward noise. It fails toward convincing.

The result that should stop you

The team built an autonomous checker called AQuA, and I will come to how it works, but the number worth leading with is not AQuA's accuracy. It is what happened to the humans.

Three board-certified pathologists independently and blindly scored virtually stained images of human kidney and lung tissue, rating nuclear, cytoplasmic and extracellular stain quality from 1 to 4, without access to the chemically stained ground truth. Exactly the condition a deployed virtual-staining workflow creates.

On images from good models, the pathologists and the algorithm largely agreed. Fine.

On images from deliberately degraded models containing realistic hallucinations, the paper reports:

pathologists failed to flag these hallucinated images and could not identify the mismatch between VS images and their HS counterparts when the access to HS ground truth was unavailable

Board-certified specialists, looking at images in which nuclei were missing from glomeruli and renal tubules, signed off on them as adequately stained and ready to read. Not because they were careless. Because the images looked right, and there was nothing available to look right against.

The paper makes it concrete with a lung transplant rejection case. The chemically stained reference showed a prominent mononuclear cell infiltrate, indicating acute cellular rejection. One degraded virtual-staining model produced images with obvious artifacts, aberrant colour and blurred nuclear boundaries, which any reviewer would reject. A second degraded model produced images that looked good but had dropped many lymphocyte nuclei. The authors note this could mislead a pathologist into assessing the case as "non-rejection" rather than "rejection."

The clean-looking failure is the one that changes the diagnosis. The ugly one just gets rerun.

Why the usual metrics do not save you

This is the part I found most instructive, and it generalizes well beyond pathology.

The obvious response is to score the images with standard similarity metrics: mean squared error, Pearson correlation, peak signal-to-noise ratio. The researchers tried. For both kidney and lung, the distributions of good and hallucinated images under those metrics "present strong overlap and cannot be well separated with simple thresholding."

Sit with that. These are the supervised metrics, computed with the ground-truth chemical stain in hand, and they still could not reliably separate hallucinated images from good ones. Pixel-level agreement is simply not the axis on which this failure lives. A model can drop a population of lymphocyte nuclei, which is diagnostically decisive, while barely moving a whole-image error statistic. The error is semantic, and the metrics are structural.

Anyone who works with instruments will recognize this. It is the same reason a good residual does not certify a fit, and the same reason a strong correlation coefficient tells you nothing about whether you assigned the peak correctly. I wrote about a version of this in the XPS peak that probably is not sp3 carbon: a number can behave impeccably while the interpretation sitting on top of it is wrong.

How AQuA works

The solution is clever enough to describe. Since no chemical reference exists at deployment, they built one out of the model itself.

Alongside the forward virtual-staining model (autofluorescence to H&E), they train a reverse model (H&E back to autofluorescence). Then they iterate: stain, un-stain, re-stain, repeatedly. A well-behaved image survives this round trip roughly intact. An image containing a hallucination does not, because the invented content has no stable preimage, so the inconsistency compounds with each pass. The accumulated instability is the signal, and a classifier reads it. They add a majority-voting scheme across classifiers to firm up the decision.

The reported performance, on blind test sets from new patients: kidney, about 2,100 images, 99.8% accuracy and 99.8% sensitivity. Lung, about 2,400 images, 97.8% accuracy and 99.5% sensitivity. Agreement with the pathologists' consensus classification of 98.5%. On the hallucinated images from poor models, AQuA rejected 100%, including the realistic ones the pathologists passed.

It also generalized in ways that matter. Trained only on failures from early-stopped models, it correctly caught failures from an overfitted model, which is a different failure pattern it had never seen. Trained on kidney and tested on lung, and vice versa, it held a sensitivity of 98.9%.

Throughput is roughly 1 to 2 seconds per field of view, against an hour of expert examination. That is the practical argument.

What this does not establish

I want to be careful here, because "AI catches AI errors that fool doctors" is exactly the kind of sentence that travels further than it should.

The bad models were manufactured. The hallucinating models were made deliberately, mostly by taking early-stopped training checkpoints, plus a purposely overfitted one. That is a sound way to generate a labelled test set. It is not a measurement of how often a properly validated, clinically deployed virtual-staining model produces realistic hallucinations in practice. This paper establishes that the failure mode exists, that it is severe, and that experts cannot catch it unaided. It does not establish a rate. Anyone citing this as evidence that deployed virtual staining is frequently wrong is going beyond the data.

The expert panel was three people. Three board-certified pathologists is a reasonable panel for this kind of comparison, and consensus was required, but it is three people. "Superhuman" is the authors' word and it is doing a lot of work on a small denominator.

The watchdog is itself a trained model, with the same class of blind spot. The paper reports its one false negative honestly, and it is the most revealing detail in the whole study. AQuA accepted an image it should have rejected: a lung field composed almost entirely of red blood cells with scattered lymphocytes. The reason given is under-representation of red blood cells in the training data. So the hallucination detector failed on a distribution shift, which is precisely the failure category it was built to catch. Meanwhile the human pathologists caught that one, because they are familiar with procedure-related bleeding and know to disregard such regions.

That single disagreement is the honest summary of the whole field. The model and the human fail on different things. Neither is a general-purpose backstop for the other. A workflow that replaces the human with AQuA inherits AQuA's blind spots, and they are not the same blind spots.

Scope is two organs, one stain, one group. Human kidney and lung, virtual H&E, developed and evaluated by the same lab that built the staining models. The authors are explicit that extension to virtual immunohistochemistry and immunofluorescence is prospective: they write that "although we did not provide results on this capability in this manuscript," the framework might extend to stain-to-stain transformations. Independent replication on other tissues, other stains and other groups' models is the obvious next requirement.

And a general caution. This work is from June 2025 and I have not tracked whether AQuA has since been externally validated or moved toward regulatory review. Treat it as a well-executed demonstration of a real problem and a promising answer, not as deployed clinical infrastructure.

The signal

The transferable lesson is not about pathology. It is about a trade that is being made all over science right now, usually without being named.

Every time a generative model replaces a physical measurement step, the saving is the removal of the ground truth. That is not a side effect. That is the value proposition. Virtual staining is worth money precisely because you skip the stain. A simulated spectrum is worth having precisely because you skip the acquisition. A predicted structure is worth having precisely because you skip the crystallography.

So the question to ask of any such system is not "how accurate is it?" That number will be excellent, because it was measured in a validation setting where the ground truth still existed. The question is:

In deployment, what would tell me this is wrong?

If the honest answer is "an expert would notice," this paper is the reason to distrust that answer. The experts did not notice. They were not being lazy; they were being asked to detect a mismatch against a reference that had been removed for cost reasons, using an image that had been optimized to look correct.

Three practical habits fall out of it.

Ask what the model was optimized to fool. Adversarially-trained generators are not trying to be accurate, they are trying to be indistinguishable. Those objectives overlap most of the time and come apart exactly when it matters. Realism is not evidence.

Distrust aggregate similarity scores for semantic failures. MSE, correlation and PSNR could not separate hallucinated from clean images here even with the ground truth available. If someone validates a generative pipeline with a single global error metric, they have measured something, but not the thing that hurts.

Keep an audit fraction. The strongest practical answer is unglamorous and the paper's own framing points at it: keep chemically staining some percentage of cases, chosen at random, forever. Not for diagnosis, for surveillance. It reintroduces the control that the efficiency argument deleted, at a small fraction of the original cost, and it is the only mechanism that detects drift after the model was validated. This is the same instinct as my note in the baselines report: the boring comparator is what keeps the impressive system honest.

None of this is an argument against virtual staining. I think it will be genuinely useful and I would rather have it than not. It is an argument that removing a reference from a pipeline is a change to the pipeline's error-detection properties, not just to its cost, and that the two need to be priced together. The technology that saves you the control is the technology that most needs one.