Virchow: Classical & Computational Pathology

• Virchow is a polyvalent term denoting both classical anatomical features (e.g., Virchow–Robin spaces) and advanced digital pathology models.
• The digital pathology models, including Virchow and Virchow2, use vision transformers pre-trained on millions of whole-slide images to achieve high performance in cancer detection.
• Domain-specific training techniques such as LoRA adaptation and mixed-magnification approaches enhance tasks like mitosis classification, tissue segmentation, and slide-level retrieval.

Virchow is a polyvalent term in contemporary biomedical literature. In classical medical usage, it survives most visibly in the eponym “Virchow–Robin spaces,” the perivascular spaces that surround small penetrating arteries and arterioles as they pass from the subarachnoid space into the brain parenchyma (Midya et al., 2015). In computational pathology, “Virchow” designates a lineage of pathology-specific vision foundation models—most prominently Virchow and Virchow2—trained with DINOv2-style self-supervision on very large whole-slide image corpora and reused as backbones for cancer detection, biomarker prediction, mitosis classification, nuclei analysis, retrieval, and slide-level multimodal modeling (Vorontsov et al., 2023, Zimmermann et al., 2024).

1. Historical and eponymic scope

The medical eponym “Virchow–Robin space” refers to Rudolf Virchow and Charles-Philippe Robin, who described these perivascular spaces in the 19th century (Midya et al., 2015). In the neuroradiologic literature considered here, these spaces are defined anatomically as perivascular channels surrounding small penetrating arteries and arterioles as they enter the brain parenchyma from the subarachnoid space, particularly along vessels in the basal ganglia, midbrain, and deep gray matter (Midya et al., 2015).

In current computational pathology literature, the same name is reused for a family of foundation models. The term therefore functions in two non-equivalent but historically linked domains: an anatomical-neuropathologic one centered on vascular–parenchymal interfaces, and a machine-learning one centered on pathology-specific visual representation learning (Midya et al., 2015, Vorontsov et al., 2023). A plausible implication is that the term now indexes both a classical pathology heritage and a contemporary technical ecosystem.

2. Virchow as a million-slide digital pathology foundation model

The original Virchow model is a self-supervised vision transformer for digital pathology with 632 million parameters and a ViT-H/14 backbone, trained using DINOv2 on 1.5 million H&E whole-slide images from ~120,000 patients across 17 tissue groups at Memorial Sloan Kettering Cancer Center (Vorontsov et al., 2023). Training tiles are extracted at 20× magnification (0.5 $\mu$m/pixel) as 224×224 patches, after tissue foreground detection on downsampled slides; tiles are retained when at least 25% of the tile area contains tissue (Vorontsov et al., 2023).

Virchow is trained at tile level rather than whole-slide level. At inference, the “Virchow embedding” is a 2560-dimensional vector formed by concatenating the output class token and the mean of the patch tokens:

$$\mathbf{z} = \left[\, \mathbf{c} \;\|\; \frac{1}{K}\sum_{i=1}^{K}\mathbf{p}_i \,\right].$$

Here, the class token and the mean patch-token vector are each 1280-dimensional (Vorontsov et al., 2023, Shaikovski et al., 2024).

Empirically, the original study reports a 0.949 overall specimen-level AUC across 17 different cancer types and 0.937 AUC on 7 rare cancer types (Vorontsov et al., 2023). The same work also reports state-of-the-art tile-level and biomarker-prediction performance, including the highest AUROC among the compared models for ColonMSI, BladderFGFR, and LungEGFR (Vorontsov et al., 2023). These results established Virchow as a general-purpose pathology backbone rather than a task-specific classifier.

3. Virchow2 and scaling in mixed-magnification pathology models

Virchow2 extends the original framework by scaling both data and pathology-specific training design. The Virchow2 paper introduces Virchow2, a 632 million parameter vision transformer, Virchow2G, a 1.9 billion parameter vision transformer, and Virchow2G Mini, a 22 million parameter distillation of Virchow2G, each trained with 3.1 million histopathology whole slide images (Zimmermann et al., 2024). The training corpus is explicitly described as diverse in tissues, originating institutions, and stains, with mixed magnifications including 5×, 10×, 20×, and 40× (Zimmermann et al., 2024).

The paper’s central technical claim is that downstream performance depends not only on data scale and model size but also on domain-specific algorithmic changes (Zimmermann et al., 2024). Two such changes are emphasized. First, the geometric augmentation pipeline replaces standard crop-and-resize with Extended-Context Translation (ECT), which samples sub-crops from larger source tiles without the same degree of morphology-distorting resizing (Zimmermann et al., 2024). Second, the KoLeo diversity regularizer used in DINOv2 is replaced by a kernel-density-estimation entropy term:

$$H_\textrm{KDE}(f) = - \frac{1}{n} \sum_{i=1}^n \log \sum_{j = 1}^n k(\mathbf{z}_i, \mathbf{z}_j),$$

with an unnormalized von Mises–Fisher kernel

$$k_\textrm{vMF}(\mathbf{x}, \mathbf{y}) = \exp(\kappa \mathbf{x}^\top\mathbf{y}), \qquad \kappa = 5.$$

This modification is motivated by the high within-batch similarity of pathology tiles and by training-stability concerns at scale (Zimmermann et al., 2024).

The study reports state-of-the-art performance on 12 tile-level tasks and argues that data diversity and domain-specific methods can outperform models that only scale in the number of parameters, while also stating that average performance benefits from the combination of domain-specific methods, data scale, and model scale (Zimmermann et al., 2024). In downstream benchmark papers, Virchow2 is consistently summarized as a ViT-H/14 pathology foundation model with 632 million parameters and DINOv2 pretraining on millions of histopathology WSIs (Ramchandani et al., 21 Sep 2025, Banerjee et al., 26 Jun 2025).

4. Adaptation to mitosis, nuclei, and renal cell segmentation

A major line of work uses Virchow and Virchow2 as frozen or mostly frozen backbones with parameter-efficient adaptation. In atypical mitotic figure classification for MIDOG 2025, Virchow, Virchow2, and UNI are compared under Low-Rank Adaptation (LoRA), with Virchow used as a frozen backbone plus LoRA layers and a linear classifier head (Ramchandani et al., 21 Sep 2025). The best configuration—Virchow with LoRA rank 8, 224×224 resizing, threshold = 0.6, and a three-fold ensemble—achieved balanced accuracy 0.8837 on the preliminary test set and ranked joint 9th on the challenge leaderboard (Ramchandani et al., 21 Sep 2025).

A separate cross-dataset benchmark on atypical versus normal mitosis likewise identifies the “Virchow-line” as especially effective under LoRA-based adaptation (Banerjee et al., 26 Jun 2025). In that study, Virchow2 (LoRA) is the best-performing model on AMi-Br with balanced accuracy 0.8135 ± 0.0145 and AUROC 0.9026 ± 0.0051, while Virchow (LoRA) achieves the best balanced accuracy on AtNorM-Br (0.7696 ± 0.0198) and AtNorM-MD (0.7705 ± 0.0287) (Banerjee et al., 26 Jun 2025). These studies converge on the same operational pattern: linear probing alone is modest, whereas LoRA-based adaptation materially improves performance and cross-domain generalization.

Virchow2 has also been repurposed as a tissue-segmentation context provider in nuclei analysis. The TAND framework uses a frozen Virchow-2 tissue segmentation branch whose semantic tissue probability maps modulate the nuclei classification stream via multi-scale Spatial-FiLM (Xu et al., 17 Nov 2025). On PUMA, TAND (with Virchow-2 & FiLM) achieves detection F1 0.734 ± 0.026 and classification macro-F1 0.398 ± 0.024, outperforming the reported baselines; the tissue branch itself yields tissue Dice 0.545 vs 0.274 for U-Net (Xu et al., 17 Nov 2025). Reported per-class gains are especially large for epithelium (+0.371 F1), endothelium (+0.298 F1), and stroma (+0.199 F1) (Xu et al., 17 Nov 2025).

In kidney pathology, Virchow appears as the backbone of CellViT++ [Virchow], one of several 2025 cell foundation models evaluated on 2,091 challenging renal patches in a human-in-the-loop rating framework (Wang et al., 1 Oct 2025). CellViT++ [Virchow] is the strongest standalone model, with 40.3% of predictions rated “Good” and 0.9% rated “Bad” (Wang et al., 1 Oct 2025). A fusion-based ensemble of four models further increases the “Good” rate to 62.2% and reduces “Bad” to 0.4% (Wang et al., 1 Oct 2025). This suggests that Virchow’s value extends beyond classification into nuclei instance segmentation and ensemble-based error reduction.

5. Slide-level modeling and zero-shot retrieval

Because Virchow is fundamentally a tile-level encoder, several studies address the slide-level aggregation problem explicitly. PRISM is a slide-level multimodal generative foundation model that builds on Virchow tile embeddings and clinical report text (Shaikovski et al., 2024). Virchow provides fixed 2560-dimensional tile embeddings, while PRISM uses a Perceiver slide encoder with 513 learned latents and 8 blocks to aggregate large tile sets into slide embeddings and latent features for language cross-attention (Shaikovski et al., 2024).

PRISM is trained jointly with a contrastive image–text objective and a report-generation objective, enabling several modes of use: report generation, zero-shot classification from text prompts, linear probing on slide embeddings, and fine-tuning for biomarker prediction (Shaikovski et al., 2024). The paper states that using text prompts, PRISM achieves zero-shot cancer detection and sub-typing performance approaching and surpassing that of a supervised aggregator model, that linear classifiers on PRISM slide embeddings surpass supervised aggregator models, and that an aggregator initialized with PRISM and trained on as little as 10% of the training data can outperform a supervised baseline that uses all of the data (Shaikovski et al., 2024). For example, on MSK DCIS vs IDC, the baseline supervised aggregator reaches 0.876 AUROC, whereas the PRISM linear probe reaches 0.939 (Shaikovski et al., 2024).

Zero-shot whole-slide retrieval provides a stricter test of whether Virchow’s patch embeddings transfer without any task-specific adaptation. In the Yottixel-based retrieval study, Virchow embeddings are used in a strictly zero-shot pipeline on TCGA, spanning 23 organs and 117 cancer subtypes (Alfasly et al., 2024). Yottixel–Virchow achieves Top-1 F1 = 41% ± 13%, Majority @ Top-3 F1 = 43% ± 15%, and Majority @ Top-5 F1 = 41% ± 13% (Alfasly et al., 2024). This clearly exceeds the older Yottixel–DenseNet baseline (28% ± 13%, 28% ± 13%, 27% ± 13%) but remains close to, and in some settings slightly below, other pathology foundation models such as UNI and GigaPath (Alfasly et al., 2024). A plausible implication is that strong patch embeddings do not by themselves solve WSI-level retrieval; the slide aggregation and similarity formulation remain decisive.

6. Representational geometry and robustness

Virchow2 has also been studied as an object of analysis rather than merely as a backbone. In a representational similarity analysis across six computational pathology foundation models, Virchow2 has the lowest average similarity to the other CPath models, with mean Spearman correlation 0.419 (Mishra et al., 18 Sep 2025). Its lowest similarity is with UNI2, even though both use a DINOv2-style self-distillation paradigm (Mishra et al., 18 Sep 2025). The same study finds strong slide-dependence but weak disease-dependence in Virchow2’s embeddings: slide specificity $\delta = 0.615$ with range [0.551, 0.667], and disease specificity $\delta = 0.120$ with range [0.058, 0.162], the latter being the lowest disease-specificity among the compared CPath models (Mishra et al., 18 Sep 2025).

After Macenko stain normalization, Virchow2’s slide specificity decreases to 0.547 and its disease specificity to 0.084 (Mishra et al., 18 Sep 2025). The authors report that Virchow2 then has the lowest slide-specificity among CPath models after normalization (Mishra et al., 18 Sep 2025). The paper also places Virchow2 within the more distributed representational regime of vision-only models, in contrast to the more compact representations of vision-LLMs (Mishra et al., 18 Sep 2025).

A separate robustness benchmark evaluates Virchow and Virchow-2 among twelve pathology foundation models using REET perturbations and Non-Redundant K-fold (NR-Kfold) validation (Yajnik et al., 5 Jul 2026). In that study, Virchow/Virchow2 and UNI2 are grouped in the 600–700M parameter range with strong transferability (Yajnik et al., 5 Jul 2026). Across perturbation categories, EXAONEPath, Virchow2, UNI2, and Phikon v2 consistently ranked highest by the Perturbation Performance Index (PPI), whereas larger billion-parameter models yielded no further gain and occasionally declined (Yajnik et al., 5 Jul 2026). Under NR-Kfold, large and mid-sized PFMs such as Virchow2, UNI2, and H-Optimus-0/1 sustain only modest accuracy loss with mean $\Delta \mathrm{AUC} \approx -0.07$ and fold-wise SD $\approx 0.09$–$0.12$ (Yajnik et al., 5 Jul 2026). The paper’s general conclusion is that mid-sized models such as UNI2/Virchow-2 achieve comparable or greater resilience than larger systems (Yajnik et al., 5 Jul 2026).

7. Virchow–Robin spaces in HIV-associated cryptococcosis

In neuroradiology, Virchow enters the literature primarily through Virchow–Robin spaces (VRS), which are anatomically the perivascular spaces surrounding penetrating arteries and arterioles as they pass from the subarachnoid space into brain parenchyma (Midya et al., 2015). In the HIV-associated cryptococcosis study, these spaces are central to recognition of central nervous system infection because Cryptococcus neoformans may expand the perivascular spaces within the gray matter, producing “soap-bubble lesions” or “gelatinous pseudocysts” (Midya et al., 2015).

The paper describes the characteristic MRI appearance of these enlarged spaces in cryptococcosis as relatively symmetrically in the basal ganglia bilaterally and in midbrain, hypointense or slightly hyperintense on T1-weighted images, hyperintense on T2-weighted images, non-enhancing on post-contrast T1-weighted images, and suppressing on FLAIR (Midya et al., 2015). In the study cohort, 2 patients with cryptococcosis showed characteristic imaging finding of enlarged Virchow Robin (perivascular) spaces (Midya et al., 2015). The lack of enhancement is attributed to preserved blood–brain barrier integrity and minimal inflammatory edema in profoundly immunosuppressed patients (Midya et al., 2015).

MR spectroscopy adds metabolic characterization. In the cryptococcosis cases with enlarged VRS, spectra showed elevated choline peaks in addition to reduced NAA, and no trehalose peaks were identified (Midya et al., 2015). The main spectroscopic metric is the NAA/Cr ratio,

$$\text{NAA/Cr} = \frac{[\text{N-acetyl aspartate}]}{[\text{creatine}]}.$$

Across HIV-positive cases overall, the mean NAA/Cr was approximately 1.62 (SD 0.42) versus 1.99 (SD 0.29) in controls, with significant reductions in the overall comparison and in both age strata ($$\mathbf{z} = \left[\, \mathbf{c} \;\|\; \frac{1}{K}\sum_{i=1}^{K}\mathbf{p}_i \,\right].$$0) (Midya et al., 2015). The same paper states that NAA/Cr ratio is reduced in HIV positive patients and is a marker for HIV infection of the brain even in the absence of imaging findings of HIV encephalopathy or when the patient is symptomatic due to neurological disease of other etiologies (Midya et al., 2015). Within this diagnostic framework, enlarged Virchow–Robin spaces supply the structural clue to cryptococcosis, while MR spectroscopy supplies metabolic evidence of inflammatory and neuronal involvement.