How is AI used in pathology?

AI performs foundational tissue-analysis tasks a spatial pathology workflow depends on, including segmenting individual cells within whole-slide images, classifying cell types, and extracting spatial features such as density and proximity. Increasingly, large pathology foundation models pretrained on millions of images are then adapted to specific clinical tasks, including detecting rare cancers where disease-specific training data is scarce.

What are pathology foundation models?

Large deep learning models pretrained on massive collections of whole-slide pathology images, then adapted to specific downstream tasks such as biomarker prediction or cancer detection. Virchow, one such model, was trained on 1.5 million whole-slide images and achieved 0.95 specimen-level AUROC for pan-cancer detection across nine common and seven rare cancers, matching specialized models on some rare variants despite using less task-specific training data.

How does deep learning analyze tissue images?

By segmenting individual cells within a whole-slide image, classifying each cell type, and extracting spatial features such as neighborhood composition and cell-to-cell proximity. Some models, such as TITAN, incorporate spatial reasoning directly into their architecture, grouping tissue regions by physical proximity during training rather than only computing spatial features from outputs afterward.

AI in spatial pathology | Drug Discovery News

Highlights in this article

Why spatial pathology needs AI
Cell classification and segmentation
Spatial feature extraction
Pathology foundation models
Validation and trust

Applying AI in pathology to spatial tissue data has moved from a research convenience to a practical necessity, driven by scale a human reviewer simply cannot match: a single whole-slide image can contain hundreds of thousands of individual cells, and a spatial biology dataset built from many such slides multiplies that count further. Deep learning is the layer that turns those raw pixel grids into the cell classifications, spatial features, and predictions a clinical or research question actually needs.


Key takeaways Cell foundation models showed promising overall zero-shot performance but consistently failed on challenging patches with low contrast, overlapping nuclei, or atypical morphologies in one large kidney pathology evaluation. A follow-up study specifically re-tested those same previously identified hard cases with newer models and found genuine improvement, a more honest measure of progress than reporting only fresh aggregate accuracy. Virchow, a large pathology foundation model, achieved 0.95 specimen-level AUROC for pan-cancer detection across nine common and seven rare cancers, matching specialized models on rare variants using less training data. TITAN, a multimodal whole-slide foundation model, groups tissue regions by spatial proximity during training, engineering spatial reasoning directly into the model architecture rather than adding it afterward. Validating an AI pathology model against previously identified failure cases, not just fresh data, is what actually demonstrates trustworthy progress.

Key takeaways

Cell foundation models showed promising overall zero-shot performance but consistently failed on challenging patches with low contrast, overlapping nuclei, or atypical morphologies in one large kidney pathology evaluation.
A follow-up study specifically re-tested those same previously identified hard cases with newer models and found genuine improvement, a more honest measure of progress than reporting only fresh aggregate accuracy.
Virchow, a large pathology foundation model, achieved 0.95 specimen-level AUROC for pan-cancer detection across nine common and seven rare cancers, matching specialized models on rare variants using less training data.
TITAN, a multimodal whole-slide foundation model, groups tissue regions by spatial proximity during training, engineering spatial reasoning directly into the model architecture rather than adding it afterward.
Validating an AI pathology model against previously identified failure cases, not just fresh data, is what actually demonstrates trustworthy progress.

Why spatial pathology needs AI

The scale argument is straightforward, but worth stating precisely rather than gesturing at generally. A single whole-slide image, digitized at diagnostic resolution, can span more than 100,000 pixels in each spatial dimension and contain hundreds of thousands of individual cells. A spatial pathology dataset assembled from a meaningful cohort of patients multiplies that cell count into the millions. No manual review process, however expert, processes data at that scale exhaustively, which is precisely why AI has moved from a research convenience into infrastructure a spatial pathology workflow now depends on structurally.

That necessity applies specifically to three foundational tasks this guide addresses in turn: identifying individual cells within a tissue image, extracting meaningful spatial features from the resulting cell map, and building models general enough to transfer across the wide diversity of tissue types, stains and diseases a real pathology workflow encounters.

Cell classification and segmentation

Before any spatial feature can be computed, cells have to be located and delineated within the image, a task called segmentation, and then assigned to a type, classification. Both tasks are foundational, since every downstream spatial analysis depends on getting them right, and both remain genuinely difficult on the tissue images that matter most clinically.


The honest measure of progress is what happens to the cases a model used to get wrong A study evaluated three widely used cell foundation models, Cellpose, StarDist, and CellViT, on a large-scale kidney pathology dataset of 8,789 high-resolution image patches sampled from 2,542 whole-slide images. Despite promising zero-shot performance overall, these models consistently failed on challenging image patches specifically, those with low contrast, overlapping nuclei, or atypical morphologies, exactly the cases where getting segmentation right matters most because they are the hardest to interpret by eye as well. A follow-up study took the more rigorous approach: rather than testing newer models on a fresh, potentially easier dataset, it specifically re-evaluated the same 2,091 curated difficult samples the earlier study had already identified as hard cases, using four newer 2025 models, three CellViT++ variants and Cellpose-SAM. The newer models demonstrated genuine improvement in generalization and segmentation quality on exactly those previously failed cases. That design, returning to a documented set of known failures rather than reporting only aggregate accuracy on new data, is a considerably more honest and more clinically meaningful way to demonstrate progress.

The honest measure of progress is what happens to the cases a model used to get wrong

A study evaluated three widely used cell foundation models, Cellpose, StarDist, and CellViT, on a large-scale kidney pathology dataset of 8,789 high-resolution image patches sampled from 2,542 whole-slide images. Despite promising zero-shot performance overall, these models consistently failed on challenging image patches specifically, those with low contrast, overlapping nuclei, or atypical morphologies, exactly the cases where getting segmentation right matters most because they are the hardest to interpret by eye as well.

A follow-up study took the more rigorous approach: rather than testing newer models on a fresh, potentially easier dataset, it specifically re-evaluated the same 2,091 curated difficult samples the earlier study had already identified as hard cases, using four newer 2025 models, three CellViT++ variants and Cellpose-SAM. The newer models demonstrated genuine improvement in generalization and segmentation quality on exactly those previously failed cases. That design, returning to a documented set of known failures rather than reporting only aggregate accuracy on new data, is a considerably more honest and more clinically meaningful way to demonstrate progress.

Spatial feature extraction

Once cells are segmented and classified, extracting spatial features, density, proximity, neighborhood composition, and related measures already covered across this cluster, requires the model to reason about position explicitly rather than treating each cell as an independent, unlocated data point.

TITAN, a recent multimodal whole-slide foundation model, illustrates this design principle concretely rather than abstractly. Built on patch-level features from an underlying model called CONCH v1.5, TITAN incorporates a specifically spatial preprocessing step during training: to improve how regions of interest are sampled, segmented tissue contours are grouped based on their spatial proximity within the slide, addressing cases where multiple tissue regions are interspersed with background areas, a common challenge in biopsy samples specifically. That is spatial reasoning engineered directly into the model’s architecture and training procedure, not a feature computed as an afterthought once patch-level predictions already exist.

Our colleagues at Technology Networks survey the broader open-source software layer this kind of feature extraction depends on in Analyzing Spatial Biology Data: Tools and Workflows, covering toolkits such as Squidpy, Scanpy, Seurat, and Giotto that implement much of the spatial statistical analysis this section describes at a conceptual level.

Pathology foundation models

The clearest quantified demonstration that large-scale pretraining transfers usefully to a genuinely hard clinical problem comes from Virchow, described in Nature Medicine as the largest foundation model for computational pathology at the time of its publication.

Trained on 1.5 million whole-slide images from more than 100,000 patients, Virchow was evaluated on biomarker prediction, cell identification, and pan-cancer detection. A pan-cancer detector built on Virchow achieved a specimen-level area under the receiver operating characteristic curve of 0.95 across nine common and seven rare cancers. Critically, with less training data than tissue-specific models require, the Virchow-based detector matched or outperformed specialized, production-grade clinical models on some rare cancer variants specifically, precisely the disease category where labeled training data is inherently scarcest and where a general-purpose foundation model’s broad pretraining offers the clearest practical advantage.

That result matters for spatial pathology specifically because rare disease variants are exactly where a spatial biomarker program is least likely to have a large, disease-specific training set of its own. A foundation model pretrained broadly, then adapted to a specific rare-variant task, is a genuinely different and more tractable starting point than training a model from scratch on a necessarily small, disease-specific dataset.

Validation and trust

The kidney cell-segmentation re-evaluation described above is this guide’s template for what validation should actually look like, and it generalizes as a principle beyond that specific study.

Three practices distinguish a genuinely trustworthy validation claim from a merely impressive-sounding one.

Document specific failure cases, not only aggregate accuracy. An overall accuracy figure can mask that a model fails consistently on a specific, identifiable category of hard case, exactly what the initial Cellpose, StarDist, and CellViT evaluation surfaced.
Re-test newer models against those same documented failures. Testing a new model version only on a fresh dataset cannot show whether it actually fixed a previously identified weakness; returning to the same hard cases can.
Report performance transfer to genuinely underrepresented situations. Virchow’s rare-cancer performance is a meaningful validation claim specifically because rare cancers are underrepresented in training data generally, which is a harder and more clinically relevant test than performance on the common cancers a model has seen extensively.

The broader question of how morphology and molecular spatial data are computationally bridged, a related but distinct challenge from the cell-level and whole-slide analysis covered here, is developed in Spatial Biology and Digital / Computational Pathology.

For the broader convergence argument and clinical infrastructure this spoke builds on, see Translating Spatial Biology into the Clinic: Digital Pathology and Beyond, and for where this fits within the full spatial biology pipeline, Spatial Biology in Drug Discovery: From Target Discovery to Translational Medicine.

This article was produced under Drug Discovery News’s AI editorial policies.