Pith. sign in

REVIEW 36 cited by

Virchow2: Scaling Self-Supervised Mixed Magnification Models in Pathology

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2408.00738 v3 pith:4YAKZJD5 submitted 2024-08-01 cs.CV

Eric Zimmermann , Eugene Vorontsov , Julian Viret , Adam Casson , Michal Zelechowski , George Shaikovski , Neil Tenenholtz , James Hall
show 6 more authors
David Klimstra Razik Yousfi Thomas Fuchs Nicolo Fusi Siqi Liu Kristen Severson
This is my paper · ORCID
classification cs.CV
keywords modelsdatascalemillionmodelparameterpathologyperformance
verification ladder T0 review T1 audit T2 compute T3 formal

Signed reviews

No signed human review yet.

0 comments
read the original abstract

Foundation models are rapidly being developed for computational pathology applications. However, it remains an open question which factors are most important for downstream performance with data scale and diversity, model size, and training algorithm all playing a role. In this work, we propose algorithmic modifications, tailored for pathology, and we present the result of scaling both data and model size, surpassing previous studies in both dimensions. We introduce three new models: 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, with diverse tissues, originating institutions, and stains. We achieve state of the art performance on 12 tile-level tasks, as compared to the top performing competing models. Our results suggest that data diversity and domain-specific methods can outperform models that only scale in the number of parameters, but, on average, performance benefits from the combination of domain-specific methods, data scale, and model scale.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 36 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. OpenAlex reports about 44 citations worldwide. Full citation record

  1. Self-supervision drives representational convergence in medical foundation models more than clinical supervision

    cs.CV 2026-07 conditional novelty 7.0 of 10

    Representational convergence among medical image encoders is modest, driven mainly by the self-supervised pretraining objective rather than clinical supervision or scale, yet still sufficient for cross-encoder and cro...

  2. A Clinically Validated Foundation Model for Comprehensive Lung Pathology Interpretation

    eess.IV 2026-05 unverdicted novelty 7.0 of 10

    PulmoFoundation achieves 92.3% average AUC on 32 lung pathology tasks in prospective validation and raises pathologist accuracy from 83.8% to 91.7% in a crossover RCT.

  3. One Model to Magnify Them All: Efficient Scale-Invariant Histopathology via Conditional Normalization and Continuous Magnification Training

    cs.CV 2026-08 conditional novelty 6.0 of 10

    Conditional Layer Normalization with continuous magnification training lets one CNN handle arbitrary pixel sizes in histopathology, matching or exceeding a five-model ensemble on PANDA.

  4. Harnessing Adversarial Distillation to Customise Debiased, Disease-Specific Pathology Foundation Models for Breast Cancer

    cs.CV 2026-08 conditional novelty 6.0 of 10

    SmartStu distills multiple teacher pathology models into compact breast-cancer encoders with an adversarial noise model and self-supervision, matching or improving external-cohort accuracy at over 30x smaller size.

  5. A Distributional Robustness Margin For Pathology Foundation Models

    cs.CV 2026-07 conditional novelty 6.0 of 10

    CRoMa scores each pathology image embedding by the margin between cross-site biological matches and same-site biological distractors, revealing distributional lower tails that pooled robustness scores hide.

  6. LaGuadia: Language-Guided Adaptive Distillation from Pathology Foundation Models

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Language-guided adaptive multi-teacher distillation yields an 87M pathology encoder that matches or exceeds GigaPath and UNI on WSI captioning, VQA, and MIL tasks.

  7. ALICE: Learning a General-Purpose Pathology Foundation Model from Vision, Vision-Language, and Slide-Level Experts

    cs.CV 2026-07 accept novelty 6.0 of 10

    Multi-stage agglomerative distillation consolidates eight vision, vision-language, and slide-level pathology teachers into one backbone that ranks first on average across 96 downstream tasks.

  8. Multi-Teacher Contrastive Distillation for Edge-Efficient Pathology Foundation Models

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Multi-teacher contrastive distillation (MuCoDi) compresses Virchow2/UNI2/H-Optimus-1 into edge encoders that reach 71.0% external AUROC (vs 71.8% best teacher) and up to 605× Raspberry Pi speedup.

  9. The Good, the Bad, and the Brittle: Benchmarking Robustness and Generalisation of Histopathology Foundation Models

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Mid-sized pathology foundation models match or beat billion-parameter ones on clinically realistic perturbations and distribution-shift tests, so scaling alone has largely saturated for robustness.

  10. MOOZY: A Patient-First Foundation Model for Computational Pathology

    cs.CV 2026-03 conditional novelty 6.0 of 10

    Patient-level pretraining with a case transformer and multi-task public supervision yields transferable WSI embeddings that beat larger slide-centric models on held-out pathology tasks.

  11. Towards Robust Foundation Models for Digital Pathology

    eess.IV 2025-07 conditional novelty 6.0 of 10

    PathoROB shows that all 20 evaluated pathology foundation models encode medical center information and that lower robustness correlates with larger downstream performance drops.

  12. Single GPU Task Adaptation of Pathology Foundation Models for Whole Slide Image Analysis

    cs.CV 2025-06 conditional novelty 6.0 of 10

    TAPFM adapts pathology foundation models on a single GPU for WSI mutation prediction, outperforming fixed-feature and end-to-end fine-tuning baselines.

  13. A Foundation Model for Spatial Proteomics

    cs.CV 2025-06 conditional novelty 6.0 of 10

    KRONOS, a self-supervised foundation model for spatial proteomics, outperforms existing vision models on cell phenotyping, retrieval, region classification, and label-efficient tasks across 11 cohorts.

  14. The Butterfly Effect in Pathology: Exploring Security in Pathology Foundation Models

    cs.CV 2025-05 conditional novelty 6.0 of 10

    A label-free attack that perturbs only 0.1% of patches in a whole-slide image can shift the model's global representation and substantially degrade downstream pathology task accuracy.

  15. PathBench: A comprehensive comparison benchmark for pathology foundation models towards precision oncology

    cs.CV 2025-05 reject novelty 6.0 of 10

    A large private multi-center benchmark of 19 pathology foundation models on 64 tasks finds Virchow2 and H-Optimus-1 best overall, with vision-language models lagging behind.

  16. Mind the Gap: Evaluating Patch Embeddings from General-Purpose and Histopathology Foundation Models for Cell Segmentation and Classification

    cs.CV 2025-02 conditional novelty 6.0 of 10

    General-purpose Swin Transformer and ConvNeXt encoders outperformed histopathology-specific ViT foundation models for cell instance segmentation and classification across PanNuke, CoNIC, and CytoDArk0.

  17. CPath-Omni: A Unified Multimodal Foundation Model for Patch and Whole Slide Image Analysis in Computational Pathology

    cs.CV 2024-12 conditional novelty 6.0 of 10

    A single 15B pathology model unifies patch-level and whole-slide tasks and reports state-of-the-art results on 39 of 42 benchmark datasets.

  18. Is Self-Supervision Enough? Benchmarking Foundation Models Against End-to-End Training for Mitotic Figure Classification

    cs.CV 2024-12 conditional novelty 6.0 of 10

    End-to-end trained ResNet50 outperforms linear probing of five foundation models for mitotic figure classification across data sizes and domains.

  19. Multimodal Whole Slide Foundation Model for Pathology

    eess.IV 2024-11 conditional novelty 6.0 of 10

    A three-stage pretrained vision-language whole-slide foundation model for pathology outperforms existing slide encoders on subtyping, molecular, survival, retrieval, and report generation tasks.

  20. Assessment of Conditional Diffusion Model for Synthetic Histopathology Image Generation

    cs.LG 2026-08 conditional novelty 5.0 of 10

    Pathology-pretrained feature metrics, especially a modified Inception Score, correlate better with downstream nuclei segmentation performance than ImageNet-based scores on synthetic histopathology images.

  21. Understanding Synergistic Interactions among Pathology Foundation Models via Adaptive Fusion

    cs.CV 2026-08 conditional novelty 5.0 of 10

    A lightweight gating network that fuses frozen pathology foundation models improves several downstream benchmarks, but one headline metric is contradicted by the paper's own table.

  22. Beyond Classification: Pathology Foundation Models as Detection Encoders for Mitotic Figures

    cs.CV 2026-07 accept novelty 5.0 of 10

    Frozen pathology FMs (especially H-optimus-0 and Virchow) support mitotic-figure detection competitively with end-to-end ResNet50 and transfer slightly better to TUPAC16.

  23. Robustifying pathology foundation models via fine-tuning

    cs.CV 2026-07 reject novelty 5.0 of 10

    A uniform fine-tuning step improves acquisition robustness and downstream performance across ten pathology foundation models, but the paper never discloses the fine-tuning recipe.

  24. HistoFID- Calibrating Frechet-distance evaluation across pathology foundation models

    eess.IV 2026-07 conditional novelty 5.0 of 10

    Raw Fréchet distances in pathology vary ~30-fold across encoders; dividing by each encoder's own within-cohort floor cuts cross-encoder variation by ~89% within and ~58% across cohorts.

  25. BRIGHT: A Collaborative Generalist-Specialist Foundation Model for Breast Pathology

    cs.CV 2026-03 conditional novelty 5.0 of 10

    Fine-tuning a generalist pathology model on 51,000 breast WSIs and concatenating its features with the original model yields top-1 performance on 21 of 24 internal breast-pathology tasks, but only 5 of 10 external tasks.

  26. Atlas 2 -- Foundation models for clinical deployment

    cs.CV 2026-01 conditional novelty 5.0 of 10

    Atlas 2 and its distilled variants set new average state-of-the-art results across 80 pathology benchmarks, with larger robustness margins over prior models.

  27. Pathology Foundation Models are Scanner Sensitive: Benchmark and Mitigation with Contrastive ScanGen Loss

    q-bio.QM 2025-07 conditional novelty 5.0 of 10

    Pathology foundation models produce scanner-dependent predictions, and the ScanGen contrastive loss reduces this scanner bias during fine-tuning for EGFR mutation prediction from whole slide images.

  28. Benchmarking histopathology foundation models in a multi-center dataset for skin cancer subtyping

    cs.CV 2025-06 conditional novelty 5.0 of 10

    A multi-center benchmark of eight histopathology foundation models on skin cancer subtyping finds that VIRCHOW-2 performs best, while a new silhouette-based metric captures each model's center-related feature bias.

  29. Accelerating Data Processing and Benchmarking of AI Models for Pathology

    cs.CV 2025-02 conditional novelty 5.0 of 10

    The authors introduce Trident, a WSI processing package, Patho-Bench, a benchmarking library, and 42 curated pathology tasks to standardize foundation model evaluation.

  30. Distilling foundation models for robust and efficient models in digital pathology

    cs.CV 2025-01 conditional novelty 5.0 of 10

    A distilled 86M-parameter pathology model reaches near state-of-the-art performance on EVA and HEST benchmarks and shows strong robustness to scanner and staining variation.

  31. Reusable specimen-level inference in computational pathology

    eess.IV 2025-01 conditional novelty 5.0 of 10

    SpinPath packages pretrained specimen-level pathology models with Python and browser inference tools and benchmarks them on an external breast cancer metastasis dataset.

  32. Atlas: A Novel Pathology Foundation Model by Mayo Clinic, Charit\'e, and Aignostics

    cs.CV 2025-01 conditional novelty 5.0 of 10

    Atlas, a 632M-parameter ViT pathology model trained on 1.2M multi-stain slides, achieves a 61.9 percent average on 21 public benchmarks, the best among seven leading foundation models.

  33. Are the Latent Representations of Foundation Models for Pathology Invariant to Rotation?

    eess.IV 2024-12 conditional novelty 5.0 of 10

    Rotated pathology image patches produce more similar latent representations in models that used rotation augmentation during self-supervised training.

  34. Leveraging Pathology Foundation Models for Panoptic Segmentation of Melanoma in H&E Images

    eess.IV 2025-07 conditional novelty 4.0 of 10

    A network combining a pathology foundation model (Virchow2) with an Efficient-UNet segments melanoma tissue types and won the PUMA challenge tissue segmentation task.

  35. A Survey on Computational Pathology Foundation Models: Datasets, Adaptation Strategies, and Evaluation Tasks

    cs.CV 2025-01 conditional novelty 4.0 of 10

    A systematic literature review of computational pathology foundation models that catalogs datasets, SSL adaptation strategies, and evaluation tasks into taxonomies.

  36. Towards Cellular-Scale Interpretability in Pathology Foundation Models for Biomarker Assessment

    cs.CV 2025-11 reject novelty 3.0 of 10

    JWTH achieves modest tissue-classification gains by adding attention pooling and stain augmentation to a DINOv3 backbone, but the biomarker claims in the abstract are unsupported by the experiments.

Pith tools