Pith. sign in

REVIEW 2 major objections 4 minor 35 references

Robustness of pathology foundation models should be read as a per-sample distribution of signed margins, not a pooled score; doing so exposes a confounder-dominated lower tail in every evaluated encoder and predicts downstream shortcut lear

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 02:13 UTC pith:H6ZER7ED

load-bearing objection A genuinely new sample-resolved robustness metric with a broad, honest evaluation, but the semantic reading of CRoMa<0 and the universal lower-tail claim need a null baseline before they can be taken at face value. the 2 major comments →

arxiv 2607.25497 v1 pith:H6ZER7ED submitted 2026-07-28 cs.CV cs.AI

Beyond Counts: A Distributional Robustness Margin For Pathology Foundation Models

Cl\'ement Grisi , Jeroen van der Laak , Geert Litjens This is my paper
classification cs.CV cs.AI
keywords pathology foundation modelsdistributional robustnessshortcut learningrepresentation geometrycross-confounder marginbatch effectsmodel selectionlower-tail robustness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that the standard count-based way of scoring robustness in pathology foundation models — pooling same-biology versus same-confounder neighbours within a fixed neighbourhood — hides exactly the information that matters: how far a sample sits from biological matches compared with confounder-driven distractors, and how much of the cohort sits on the wrong side. It introduces CRoMa, a signed per-sample margin between those two typed-neighbour distances, and treats the cohort-wide CRoMa distribution as the robustness readout. Across 20 tile-level and 4 slide-level encoders, the median margins rank models consistently, yet every encoder retains a confounder-dominated lower tail whose prevalence and severity vary independently of the median. The paper also shows that higher CRoMa predicts smaller shortcut-induced performance drops after supervised adaptation, framing model selection as a Pareto trade-off between typical robustness and worst-tail behaviour rather than a single score.

Core claim

For each sample, CRoMa_m(i) = (d^OS_m(i) - d^SO_m(i)) / (d^OS_m(i) + d^SO_m(i)), where d^SO is the mean cosine distance to the m=5 nearest neighbours sharing the biological label but from a different centre, and d^OS is the mean distance to the m=5 nearest neighbours sharing the centre but carrying a different biological label. Positive margins signal biology-dominant local geometry; negative margins signal confounder-dominant geometry. Because the margin is defined for every sample, it yields a full distribution over the cohort. The paper's empirical discovery is that no evaluated encoder is uniformly robust: even the most biology-dominant models have a negative lower tail, and the fraction

What carries the argument

The Cross-confounder Robustness Margin (CRoMa) is the central object: a per-sample signed ratio of mean distances to the nearest cross-confounder biological matches (SO) and nearest same-confounder biological distractors (OS), with m=5 neighbours per type. Being a ratio, it is invariant to global rescaling of embedding distances, which is what lets different models and datasets be compared on a common scale; geometrically it equals tan(theta - pi/4) in the typed-distance plane. Its work is to replace a pooled count with an empirical margin distribution, summarised by the median, F(0) (fraction of negative margins), and LTM10 (mean of the worst decile).

Load-bearing premise

The load-bearing premise is that every negative CRoMa margin indicates a sample whose local geometry is dominated by non-biological confounder structure; the paper gives no permutation or null baseline showing that a biology-only representation, or finite-sample nearest-neighbour fluctuation, would not produce comparable negative margins in a cohort of this size.

What would settle it

Reshuffle the confounder labels (or replace the encoder features with a matched representation that preserves biology but removes site signal) and recompute the CRoMa distribution. If a comparable fraction of samples still has CRoMa<0 — for instance, 10-30% lower-tail mass for the best encoders — then negative margins are not diagnostic of confounder dominance, and the universal LTM10<0 finding loses its force. A complementary check: train a model on confounder-shuffled pairs and see whether its lower-tail mass collapses toward zero.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Model rankings by median CRoMa are broadly stable across three tile-level benchmarks and a slide-level benchmark, suggesting the margin captures a reproducible property of an encoder rather than a dataset quirk.
  • No encoder is uniformly robust: every tile-level encoder has a confounder-dominated lower tail (LTM10 < 0), so pooled or median robustness scores systematically overstate worst-case deployment behaviour.
  • Models with very similar median margins can differ sharply in failure prevalence and severity, so model selection is naturally a Pareto problem balancing typical robustness against tail severity.
  • CRoMa computed from frozen embeddings tracks the shortcut-induced performance drop of biased linear probes (pooled in-domain Spearman around 0.84), offering an upstream proxy for transferability without repeated supervised training.
  • Because CRoMa assigns every sample a score, low-margin samples can be stratified by centre or biological subclass to locate where transferability is most at risk.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An implication the authors leave implicit is that CRoMa lower-tail summaries could serve as a cheap screening rule: run the margin on a small, deliberately structured multi-centre cohort and use F(0)/LTM10 to decide which models warrant expensive external validation.
  • The same per-sample margin could be used during pretraining as a signal for early stopping or data weighting to actively discourage confounder dominance, though the paper does not attempt this.
  • The universal negative-tail finding invites a null-model check: finite-sample nearest-neighbour statistics may generate negative margins even without confounder structure, so a shuffled-confounder or biology-only baseline would sharpen the semantic claim that CRoMa<0 means confounder dominance.
  • The distributional treatment generalises: other geometry metrics (local intrinsic dimension, nearest-neighbour entropy, reach) could be summarised per-sample in the same way, yielding tail-aware robustness profiles for other representation-learning domains.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes CRoMa, a sample-level representation-robustness metric for pathology foundation models. CRoMa compares, for each tile or slide, the mean cosine distance to the m nearest cross-confounder biological matches (SO neighbors) with the mean distance to the m nearest same-confounder biological distractors (OS neighbors), producing a signed margin in (-1,1). The authors argue that this margin is biology-dominant when positive and confounder-dominant when negative, and that aggregating per-sample margins into a median, F(0), and LTM10 yields a distributional robustness profile: typical robustness, failure prevalence, and lower-tail severity. They evaluate 20 frozen tile-level encoders on three PathoROB benchmarks and 4 slide-level encoders on a new PCaBiop benchmark, with a natural-image DINOv2-B control. They report that fixed-k RI and distance-weighted MaRI are nearly equivalent and are limited by low, model-dependent support; that CRoMa rankings are consistent across benchmarks; that every tile encoder has LTM10<0; and that median CRoMa correlates with downstream shortcut-induced performance drops (APD).

Significance. If the central interpretation holds, the paper makes a useful contribution: it identifies a concrete structural flaw in fixed-k robustness metrics (model-dependent sample support) and proposes a simple, scale-free, per-sample measure with a clear geometric reading. The study is empirically broad — 20 tile encoders across three benchmarks plus a slide-level panel — and the paper is unusually transparent about pretraining–evaluation overlap, operating-point choices, and sensitivity sweeps. The authors also state that implementations and evaluation manifests are released, which strengthens reproducibility. The major caveat is that the semantic claim 'CRoMa<0 means confounder dominance' is not calibrated against any null distribution; the universal LTM10<0 result may be partially or wholly a finite-sample/pool-size artifact. The paper also does not demonstrate that its distributional summaries (F(0), LTM10) add predictive value for downstream shortcut susceptibility beyond the median or beyond existing pooled scores. These are fixable with additional analyses, but they are load-bearing for the paper's headline interpretation.

major comments (2)
  1. [§2.4, §3.3; Eqs. (8)–(13)] The central semantic claim is that CRoMa<0 indicates confounder dominance. CRoMa_m(i) is the sign of a difference between two means of m nearest-neighbor distances, and that sign is not calibrated: it depends on the sizes of the SO and OS candidate pools. In Tolkach-ESCA each anchor has ~1,000 SO candidates but ~2,500 OS candidates (Table 1, Fig. 2); in PCaBiop-ISUP the pools are 250 vs 1,250, as acknowledged in Supp. A.12. A larger OS pool shrinks the expected nearest-OS distance by order-statistic effects alone, biasing CRoMa negative even for a representation with no true confounder signal. The paper uses balanced quartets only for the headline PCaBiop-ISUP CRoMa, but still reports occurrence-level F(0) and LTM10 over the imbalanced pools; Tolkach-ESCA uses the imbalanced pools directly. No permutation, random-feature, or matched-size-pool null is provided anywhere. Consequently, the
  2. [§3.5, Table 11] The claim that CRoMa 'anticipates downstream shortcut susceptibility' is supported by positive Spearman correlations, but the evidence is weaker than the text implies. Table 11 shows pooled APD_ID correlations of 0.84 for CRoMa versus 0.89 for both RI and MaRI; APD_OOD is 0.78 for CRoMa versus 0.76/0.74 for RI/MaRI. Moreover, the APD analysis uses only the median CRoMa, not F(0) or LTM10. Thus the downstream experiment validates the median margin, not the distributional profile that is the paper's main contribution. To support the distributional claim, please test whether F(0) or LTM10 adds predictive value after conditioning on median CRoMa (or on RI/MaRI), e.g., via partial correlations or a small regression model. Without this, the text overstates what the APD analysis demonstrates.
minor comments (4)
  1. [§2.7] The text says CRoMa is 'defined on the full evaluation cohort by construction', but it is defined only when at least m SO and m OS candidates exist; this holds for the chosen benchmarks at m=5, but it is a benchmark-design condition rather than a property of the metric. Please state this explicitly.
  2. [Supp. Table 10] The column labelled 'Boost' is not a ratio of CRoMa values but a between-cohort ratio of the typed-distance odds r = d^OS_m/d^SO_m. Please rename or define this clearly in the caption and text to avoid confusion with CRoMa itself.
  3. [Throughout] There are several typos and small formatting errors: 'Howver' (§3.4), 'distinc' (§3.5/Table 6 caption), 'completenes' (Supp. A.11), 'resuling' (Fig. 1 caption), 'MarRI' in Table 11 caption, and a duplicated 'and and' in the Discussion. Please copyedit carefully.
  4. [Tables 3–5] The pooled tables are dense and would benefit from a note explaining that conf bacc and Δ are diagnostic columns, not robustness scores; the bold 'best value' marking is confusing when applied to diagnostic columns.

Circularity Check

0 steps flagged

No load-bearing circularity: CRoMa is an independently defined margin, validated against external APD probes; the only self-citation is minor and non-load-bearing.

full rationale

The central construction, CRoMa, is defined directly from typed-neighbor distances (Eq. 8-10) and is not fitted to APD, confounder accuracy, or any downstream target. The averaging radius m=5 is fixed a priori and shown to be stable across a sweep (Table 8), so the median CRoMa rankings are not self-fulfilling. The reported association between CRoMa and APD is an empirical correlation between a frozen-representation geometry metric and an independently trained linear-probe protocol, so it is not a fitted input renamed as a prediction. The near-equivalence of RI and MaRI is an empirical result, not a consequence of defining MaRI in terms of RI. The universal LTM10<0 finding is an empirical distributional fact: the metric's definition does not guarantee that the worst-decile mean is negative, so reporting it as evidence of a confounder-dominated lower tail is an interpretation, not a tautology. The paper itself flags a pool-size asymmetry in PCaBiop-ISUP (Section A.12) and mitigates it for the headline CRoMa via quartets, while leaving occurrence-level F(0) and LTM10 on imbalanced pools; this is a calibration/validity concern about the zero boundary, not a circular derivation. The only self-citation is reference [13] for the Prost40M model provenance in Table 2, which is not load-bearing for any robustness claim. The derivation is therefore self-contained; the noted limitations affect external validity, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central CRoMa metric has only one hand-set hyperparameter (m=5), with sensitivity evidence. The competing RI/MaRI comparisons rely on fitted or selected k and tau. The main unexamined modelling assumption is the semantic interpretation of a negative margin, which is added to support the lower-tail claims.

free parameters (4)
  • m (CRoMa averaging radius) = 5
    Fixed a priori; sensitivity sweep m=1..20 shows rank invariance (Supplementary A.3).
  • MaRI temperature tau = per-model median typed SO/OS distance
    Set per model to the median distance among typed neighbours; used only for MaRI, not CRoMa (Section 2.3).
  • RI/MaRI neighbourhood size k = per-model k*; dataset medians 11, 71, 61
    PathoROB protocol; model-dependent operating point for RI/MaRI (Section 2.2, 2.7).
  • LTM tail fraction alpha = 0.10
    Chosen by hand; sensitivity to alpha=0.05/0.20 provided (Supplementary A.2).
axioms (4)
  • domain assumption The SO/OS four-way typed-neighbour contrast is a valid operationalization of biology-vs-confounder dominance.
    Adopted from de Jong et al. and PathoROB in Section 2.1; CRoMa inherits this choice.
  • ad hoc to paper A negative CRoMa margin is interpretable as confounder dominance rather than finite-sample noise.
    Core semantic claim behind F(0) and LTM10 in Sections 2.4 and 3.3; no null model is supplied.
  • domain assumption Class labels shared across centres are biological matches, and centre is the only non-biological confounder.
    Benchmarks in Section 2.5 pair coarse class labels with acquisition site, e.g. tumour/normal across two centres.
  • standard math Cosine distance on L2-normalized embeddings is a valid similarity geometry for pathology representations.
    Standard practice, stated in Section 2.1; not derived in the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 32024 in / 12543 out tokens · 130166 ms · 2026-08-01T02:13:55.444681+00:00 · methodology

0 comments
read the original abstract

Pathology foundation models are approaching clinical deployment, yet remain vulnerable to systematic non-biological variation across centres. Differences in tissue preparation, staining and scanning are strongly encoded in their representations, enabling shortcut learning and weakening generalisation across cohorts and institutions. The Robustness Index (RI) quantifies whether local representation geometry is dominated by biology or by non-biological variation, but its count-based formulation discards distance information. We show that adding distance weights changes little because the deeper limitation lies in RI's pooled, fixed-neighbourhood design, which obscures sample-level heterogeneity and effectively evaluates only a model-dependent subset of samples. We introduce the Cross-confounder Robustness Margin (CRoMa), a sample-resolved measure that directly compares distances to cross-confounder biological matches and same-confounder biological distractors. CRoMa recasts robustness as a cohort-wide margin distribution rather than a single pooled score. We evaluated frozen representations from 20 tile-level encoders across three benchmarks and 4 slide-level encoders on a fourth. Rankings by median CRoMa were broadly consistent across datasets, while the underlying distributions revealed substantial within-model heterogeneity. Every tile encoder retained a confounder-dominated lower tail, whose prevalence and severity varied markedly across models. These distinct robustness profiles frame model selection as a Pareto trade-off between typical and lower-tail robustness. Higher CRoMa was also associated with smaller shortcut-induced performance drops after supervised adaptation. By turning representation geometry into a distributional robustness readout that anticipates downstream shortcut susceptibility, CRoMa provides a principled basis for robustness assessment and model selection.

Figures

Figures reproduced from arXiv: 2607.25497 by Cl\'ement Grisi, Geert Litjens, Jeroen van der Laak.

Figure 1
Figure 1. Figure 1: Count- and distance-based robustness metrics. SO neighbours share biology but differ in confounder, whereas OS neighbours differ in biology but share the confounder. a) RI counts typed neighbours within a fixed-k neighbourhood and gives both models the same score. MaRI weights the same evidence by distance (edge width denotes weight), resuling in different scores for each model. b) If the fixed neighbourho… view at source ↗
Figure 2
Figure 2. Figure 2: Each benchmark is balanced across biological class and confounder. Cells show evaluated samples per biological-class×confounder combination: 5,100 tiles for Camelyon (20,400 total), 360 for TCGA-4x4 (5,760), 500 for Tolkach-ESCA (9,000) and 250 slides for PCaBiop (1,000). Rows denote biological classes and columns acquisition centres. The paired TCGA-2 × 2 design is shown in Supplementary [PITH_FULL_IMAGE… view at source ↗
Figure 3
Figure 3. Figure 3: Median and lower-tail CRoMa on Camelyon. Median CRoMa is plotted against worst-decile mean LTM10 for the 20 pathology encoders evaluated in this study. Higher values are preferable on both axes. Ringed points form the upper-right Pareto frontier, while shaded points are dominated on both axes. † marks the Camelyon-exposed encoder (GPFM). 14 [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-sample CRoMa distributions on Camelyon. Ridgeline distributions for the 20 pathology encoders and the natural-image control (DINOv2-B), ordered by median CRoMa. The dashed line denotes CRoMa = 0, while shading denotes the confounder-dominant region (CRoMa < 0). 15 [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Median and lower-tail CRoMa ranks across tile benchmarks. Each of the 3 tile benchmarks ranks the 20 pathology encoders by median CRoMa (rank 1 = highest median) and by worst-decile mean LTM10 (rank 1 = mildest tail). The axes show the mean rank across benchmarks, with lower (=better) ranks plotted toward the upper-right corner. Ringed points form the Pareto frontier, while shaded points are dominated on b… view at source ↗
Figure 6
Figure 6. Figure 6: CRoMa against APD on Camelyon. Columns show in-domain evaluation (left, test centres held fixed) and out-of-domain evaluation (right, unseen centres). Top row: linear-probe accuracyon the balanced test set as the centre–biology correlation in the training set increases from V =0 to V =1. Dashed curves denote the 20 tile-level encoders and error bars are 95% confidence intervals over repeated seeding. Botto… view at source ↗
Figure 7
Figure 7. Figure 7: Biology-matched example samples across acquisition centres. Each benchmark block shows two biological classes (rows) across all acquisition centres (columns), so within a row biology is fixed and only the centre changes. Because PCaBiop is a slide-level benchmark, its block shows representative tiles rather than whole slides, which make the visual differences between its centres easier to see. The resultin… view at source ↗
Figure 8
Figure 8. Figure 8: CRoMa is the angular offset from the diagonal. In the (d SO m , dOS m ) plane, CRoMai = tan(θi − π/4): points above the diagonal have more distant OS distractors and positive margins, points below have negative margins. P and λP lie on one ray, so they share θi and therefore the same margin. A.6 Metric rank agreement within each benchmark Benchmark RI vs MaRI CRoMa vs RI CRoMa vs MaRI Camelyon 0.99 0.95 0.… view at source ↗
Figure 9
Figure 9. Figure 9: The typed-distance plane measured on PCaBiop. One point per slide for the 4 slide-level encoders (m = 5). Outlines enclose 75% of an encoder’s slides and filled markers are per-encoder medians. Axes are logarithmic because median typed-distance radii span a factor of 80, from 0.010 (Prov-GigaPath) to 0.80 (MOOZY). The log map turns a common rescaling into a shift along the diagonal, so position along the d… view at source ↗
Figure 10
Figure 10. Figure 10: Per-sample CRoMa distributions on TCGA-4x4. Ridgeline distributions for 20 pathology encoders and the natural-image control (DINOv2-B), ordered by pooled median. The dashed line denotes CRoMa = 0; shading denotes CRoMa < 0. Per-model F(0) and LTM10 are reported in [PITH_FULL_IMAGE:figures/full_fig_p034_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Per-sample CRoMa distributions on Tolkach-ESCA. Ridgeline distributions for 20 pathology encoders and the natural-image control (DINOv2-B), ordered by pooled median. The dashed line denotes CRoMa = 0; shading denotes CRoMa < 0. Per-model F(0) and LTM10 are reported in [PITH_FULL_IMAGE:figures/full_fig_p035_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: CRoMa and APD on TCGA-4x4. Columns show in-domain evaluation (left; test centres held fixed) and out-of-domain evaluation (right; unseen centres). Top row, balanced-test linear￾probe accuracy as the centre–biology correlation in the training set increases from V =0 to V =1; dashed curves denote the 20 tile-level encoders and error bars are 95% t-intervals over seeds. Bot￾tom row, pooled CRoMa(m=5) versus … view at source ↗
Figure 13
Figure 13. Figure 13: CRoMa against APD on Tolkach-ESCA. Columns show in-domain evaluation (left; test centres held fixed) and out-of-domain evaluation (right; unseen centres). Top row, balanced￾test linear-probe accuracy as the centre–biology correlation in the training set increases from V =0 to V =1; dashed curves denote the 20 tile-level encoders and error bars are 95% t-intervals over seeds. Bottom row, pooled CRoMa(m=5) … view at source ↗
Figure 14
Figure 14. Figure 14: Median and lower-tail CRoMa on TCGA-4x4. Pooled median CRoMa against worst￾decile mean LTM10 for the 20 pathology encoders. Higher values are preferable on both axes. Ringed points form the upper-right Pareto frontier, while shaded points are dominated on both axes. Per￾model F(0) and LTM10 are reported in [PITH_FULL_IMAGE:figures/full_fig_p039_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Median and lower-tail CRoMa on Tolkach-ESCA. Pooled median CRoMa against worst-decile mean LTM10 for the 20 pathology encoders. Higher values are preferable on both axes. Ringed points form the upper-right Pareto frontier, while shaded points are dominated on both axes. Per-model F(0) and LTM10 are reported in [PITH_FULL_IMAGE:figures/full_fig_p040_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: The paired TCGA protocol comprises 94 balanced 2 × 2 quartets. The grid shows one generic quartet: two biological classes by two medical centres, with 300 tile occurrences per cell. Class and centre identities vary across quartets, spanning 21 cancer types and 26 centres. Typed-neighbour counts are pooled across quartets before scoring, for 112,800 tile occurrences in total. The dataset-wide TCGA-4x4 desi… view at source ↗
Figure 17
Figure 17. Figure 17: CRoMa distributions on PCaBiop. Ridgeline distributions for the 4 slide-level en￾coders on PCaBiop, ordered by pooled median. The dashed line denotes CRoMa = 0; shading denotes CRoMa < 0. Per-model F(0) and LTM10 are reported in [PITH_FULL_IMAGE:figures/full_fig_p043_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: PCaBiop-ISUP is balanced across ISUP grade and contributing centre. Rows denote ISUP grade groups (grade 0 is benign, grades 1–5 increasingly aggressive prostate cancer) and columns the acquisition centre. Cells show the evaluated biopsies per ISUP grade×centre combination: 250 slides in each of the 6 × 2 cells, for 3,000 slides in total. 44 [PITH_FULL_IMAGE:figures/full_fig_p044_18.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 12 linked inside Pith

  1. [1]

    Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S

    Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolf- sson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Dur- mus, ...

    Pith/arXiv arXiv 2022

  2. [2]

    Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge.Nature Medicine, 28(1):154–163, 2022

    Wouter Bulten, Kimmo Kartasalo, Po-Hsuan Cameron Chen, et al. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge.Nature Medicine, 28(1):154–163, 2022

    2022

  3. [3]

    A clinical benchmark of public self-supervised pathology foundation models.Nature Communications, 16(1):3640, 2025

    Gabriele Campanella, Shengjia Chen, Manbir Singh, Ruchika Verma, Silke Muehlstedt, Jen- nifer Zeng, Aryeh Stock, Matt Croken, Brandon Veremis, Abdulkadir Elmas, et al. A clinical benchmark of public self-supervised pathology foundation models.Nature Communications, 16(1):3640, 2025

    2025

  4. [4]

    Emerging properties in self-supervised vision transformers, 2021

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv´ e J´ egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers, 2021. arXiv:2104.14294

    Pith/arXiv arXiv 2021

  5. [5]

    Towards a general-purpose foundation model for computational pathology.Nat

    Richard J Chen, Tong Ding, Ming Y Lu, Drew F K Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, Mane Williams, Lukas Oldenburg, Luca L Weishaupt, Judy J Wang, Anurag Vaidya, Long Phi Le, Georg Gerber, Sharifa Sahai, Walt Williams, and Faisal Mahmood. Towards a general-purpose foundation model for computat...

    2024

  6. [6]

    de Jong, Eric Marcus, and Jonas Teuwen

    Edwin D. de Jong, Eric Marcus, and Jonas Teuwen. Current pathology foundation models are unrobust to medical center differences, 2025. arXiv:2501.18055

    Pith/arXiv arXiv 2025

  7. [7]

    Biased data, biased AI: deep networks predict the acquisition site of TCGA images.Diagn

    Taher Dehkharghanian, Azam Asilian Bidgoli, Abtin Riasatian, Pooria Mazaheri, Clinton J V Campbell, Liron Pantanowitz, H R Tizhoosh, and Shahryar Rahnamayan. Biased data, biased AI: deep networks predict the acquisition site of TCGA images.Diagn. Pathol., 18(1):67, May 2023

    2023

  8. [8]

    Wagner, Andrew H

    Tong Ding, Sophia J. Wagner, Andrew H. Song, Richard J. Chen, Ming Y. Lu, Andrew Zhang, Anurag J. Vaidya, Guillaume Jaume, Muhammad Shaban, Ahrong Kim, Drew F. K. Williamson, Harry Robertson, Bowen Chen, Cristina Almagro-P´ erez, Paul Doucet, Sharifa Sahai, Chengkuan Chen, Christina S. Chen, Daisuke Komura, Akihiro Kawabe, Mieko Ochi, Shinya Sato, Tomoyuk...

    2025

  9. [9]

    Distill- ing foundation models for robust and efficient models in digital pathology

    Alexandre Filiot, Nicolas Dop, Oussama Tchita, Auriane Riou, R´ emy Dubois, Thomas Peeters, Daria Valter, Marin Scalbert, Charlie Saillard, Genevi` eve Robin, and Antoine Olivier. Distill- ing foundation models for robust and efficient models in digital pathology. InMedical Image 24 Computing and Computer Assisted Intervention – MICCAI 2025, volume 15966 ...

    2025

  10. [10]

    Scaling self-supervised learning for histopathol- ogy with masked image modeling.medRxiv, 2023

    Alexandre Filiot, Ridouane Ghermi, Antoine Olivier, Paul Jacob, Lucas Fidon, Alice Mac Kain, Charlie Saillard, and Jean-Baptiste Schiratti. Scaling self-supervised learning for histopathol- ogy with masked image modeling.medRxiv, 2023

    2023

  11. [11]

    Phikon-v2, a large and public feature extractor for biomarker prediction, 2024

    Alexandre Filiot, Paul Jacob, Alice Mac Kain, and Charlie Saillard. Phikon-v2, a large and public feature extractor for biomarker prediction, 2024. arXiv:2409.09173

    Pith/arXiv arXiv 2024

  12. [12]

    Shortcut learning in deep neural networks.Nat

    Robert Geirhos, J¨ orn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks.Nat. Mach. Intell., 2(11):665–673, November 2020

    2020

  13. [13]

    Deep learning from routine histology improves risk stratification for biochemical re- currence in prostate cancer, 2026

    Cl´ ement Grisi, Khrystyna Faryna, Nefise Uysal, Vittorio Agosti, Enrico Munari, Sol` ene- Florence Kammerer-Jacquet, Paulo Guilherme de Oliveira Salles, Yuri Tolkach, Reinhard B¨ uttner, Sofiya Semko, Maksym Pikul, Axel Heidenreich, Jeroen van der Laak, and Geert Litjens. Deep learning from routine histology improves risk stratification for biochemical r...

    arXiv 2026

  14. [14]

    The impact of site-specific digital histology signatures on deep learning model accuracy and bias.Nat

    Frederick M Howard, James Dolezal, Sara Kochanny, Jefree Schulte, Heather Chen, Lara Heij, Dezheng Huo, Rita Nanda, Olufunmilayo I Olopade, Jakob N Kather, Nicole Cipriani, Robert L Grossman, and Alexander T Pearson. The impact of site-specific digital histology signatures on deep learning model accuracy and bias.Nat. Commun., 12(1):4423, July 2021

    2021

  15. [15]

    Self-Supervised Visual Feature Learning With Deep Neu- ral Networks: A Survey .IEEE Transactions on Pattern Analysis & Machine Intelligence, 43(11):4037–4058, November 2021

    Longlong Jing and Yingli Tian. Self-Supervised Visual Feature Learning With Deep Neu- ral Networks: A Survey .IEEE Transactions on Pattern Analysis & Machine Intelligence, 43(11):4037–4058, November 2021

    2021

  16. [16]

    Saarthak Kapse, Mehmet Ayg¨ un, Elijah Cole, Emma Lundberg, Le Song, and Eric P. Xing. GenBio-PathFM: a state-of-the-art foundation model for histopathology.bioRxiv, 2026

    2026

  17. [17]

    Training state-of-the-art pathology foundation models with orders of magnitude less data, 2025

    Mikhail Karasikov, Joost van Doorn, Nicolas K¨ anzig, Melis Erdal Cesur, Hugo Mark Horlings, Robert Berke, Fei Tang, and Sebastian Ot´ alora. Training state-of-the-art pathology foundation models with orders of magnitude less data, 2025. arXiv:2504.05186

    Pith/arXiv arXiv 2025

  18. [18]

    Hosseini

    Yousef Kotp, Vincent Quoc-Huy Trinh, Christopher Pal, and Mahdi S. Hosseini. MOOZY: A patient-first foundation model for computational pathology, 2026. arXiv:2603.27048

    Pith/arXiv arXiv 2026

  19. [19]

    de Jong, Julius Hense, Hannah Marienwald, Jonas Dippel, Philip Naumann, Eric Marcus, Lukas Ruff, Maximilian Alber, Jonas Teuwen, Frederick Klauschen, and Klaus-Robert M¨ uller

    Jonah K¨ omen, Edwin D. de Jong, Julius Hense, Hannah Marienwald, Jonas Dippel, Philip Naumann, Eric Marcus, Lukas Ruff, Maximilian Alber, Jonas Teuwen, Frederick Klauschen, and Klaus-Robert M¨ uller. Towards robust foundation models for digital pathology, 2025. arXiv:2507.17845

    Pith/arXiv arXiv 2025

  20. [20]

    Do histopathological foundation models eliminate batch effects? a comparative study, 2024

    Jonah K¨ omen, Hannah Marienwald, Jonas Dippel, and Julius Hense. Do histopathological foundation models eliminate batch effects? a comparative study, 2024. arXiv:2411.05489

    Pith/arXiv arXiv 2024

  21. [21]

    Lu, Bowen Chen, Drew F

    Ming Y. Lu, Bowen Chen, Drew F. K. Williamson, Richard J. Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, Anil V. Parwani, Andrew Zhang, and Faisal Mahmood. A visual-language foundation model for computational pathology.Nature Medicine, 30(3):863–874, 2024. 25

    2024

  22. [22]

    A generalizable pathology foundation model using a unified knowledge distillation pretraining framework.Nature Biomedical Engineering, 10(3):545–564, 2026

    Jiabo Ma, Zhengrui Guo, Fengtao Zhou, Yihui Wang, Yingxue Xu, Jinbang Li, Fang Yan, Yu Cai, Zhengjie Zhu, Cheng Jin, Yi Lin, Xinrui Jiang, Chenglong Zhao, Danyi Li, Anjia Han, Zhenhui Li, Ronald Cheong Kin Chan, Jiguang Wang, Peng Fei, Kwang-Ting Cheng, Shaoting Zhang, Li Liang, and Hao Chen. A generalizable pathology foundation model using a unified know...

    2026

  23. [23]

    A benchmarking crisis in biomedical machine learning.Nature Medicine, 31(4):1060, April 2025

    Faisal Mahmood. A benchmarking crisis in biomedical machine learning.Nature Medicine, 31(4):1060, April 2025

    2025

  24. [24]

    MahmoodLab/UNI2-h

    Mahmood Lab. MahmoodLab/UNI2-h. Hugging Face model repository,https://huggingf ace.co/MahmoodLab/UNI2-h, 2025

    2025

  25. [25]

    Hibou: a family of foundational vision transformers for pathology, 2024

    Dmitry Nechaev, Alexey Pchelnikov, and Ekaterina Ivanova. Hibou: a family of foundational vision transformers for pathology, 2024. arXiv:2406.05074

    Pith/arXiv arXiv 2024

  26. [26]

    Benchmarking foundation models as feature extractors for weakly supervised computational pathology.Nat

    Peter Neidlinger, Omar S M El Nahhas, Hannah Sophie Muti, Tim Lenz, Michael Hoffmeister, Hermann Brenner, Marko van Treeck, Rupert Langer, Bastian Dislich, Hans Michael Behrens, Christoph R¨ ocken, Sebastian Foersch, Daniel Truhn, Antonio Marra, Oliver Lester Saldanha, and Jakob Nikolas Kather. Benchmarking foundation models as feature extractors for weak...

    2026

  27. [27]

    DINOv2: Learning robust visual features without supervision, 2024

    Maxime Oquab, Timoth´ ee Darcet, Th´ eo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khali- dov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herv´ e Jegou, Julien Mairal, Pat...

    Pith/arXiv arXiv 2024

  28. [28]

    H-optimus-0.https://github.com/bioptimus/releases/ tree/main/models/h-optimus/v0, 2024

    Charlie Saillard, Rodolphe Jenatton, Felipe Llinares-L´ opez, Zelda Mariet, David Cahan´ e, Eric Durand, and Jean-Philippe Vert. H-optimus-0.https://github.com/bioptimus/releases/ tree/main/models/h-optimus/v0, 2024

    2024

  29. [29]

    Mariet, and Rodolphe Jenatton

    Marin Scalbert, Charlie Saillard, Thomas Peeters, Liam Gonzalez, Dasha Valter, Fe- lipe Llinares-L´ opez, Zelda E. Mariet, and Rodolphe Jenatton. Abstract LB174: H- optimus-1: a foundation model for computational histopathology.Cancer Research, 86(8 Supplement):LB174, 2026. Model weights:https://huggingface.co/bioptimus/H -optimus-1

    2026

  30. [30]

    Kunz, Juan A

    George Shaikovski, Adam Casson, Kristen Severson, Eric Zimmermann, Yi Kan Wang, Jeremy D. Kunz, Juan A. Retamero, Gerard Oakley, David Klimstra, Christopher Kanan, Matthew Hanna, Michal Zelechowski, Julian Viret, Neil Tenenholtz, James Hall, Nicolo Fusi, Razik Yousfi, Peter Hamilton, William A. Moye, Eugene Vorontsov, Siqi Liu, and Thomas J. Fuchs. PRISM:...

    Pith/arXiv arXiv 2024

  31. [31]

    A foundation model for clinical-grade computational pathology and rare cancers detection.Nat

    Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, Ellen Yang, Philippe Mathieu, Alexander van Eck, Donghun Lee, Julian Viret, Eric Robert, Yi Kan Wang, Jeremy D Kunz, Matthew C H Lee, Jan H Bernhard, Ran A Godrich, Gerard Oakley, Ewan Millar, M...

    2024

  32. [32]

    Nirschl, Joel Neal, Maximilian Diehn, Sen Yang, and Rui- jiang Li

    Jinxi Xiang, Xiyue Wang, Xiaoming Zhang, Yinghua Xi, Feyisope Eweje, Yijiang Chen, Yuchen Li, Colin Bergstrom, Matthew Gopaulchan, Ted Kim, Kun-Hsing Yu, Sierra Willens, Francesca Maria Olguin, Jeffrey J. Nirschl, Joel Neal, Maximilian Diehn, Sen Yang, and Rui- jiang Li. A vision–language foundation model for precision oncology.Nature, 638(8051):769– 778, 2025

    2025

  33. [33]

    A whole-slide foundation model for digital pathology from real-world data.Nature, 630(8015):181–188, June 2024

    Hanwen Xu, Naoto Usuyama, Jaspreet Bagga, Sheng Zhang, Rajesh Rao, Tristan Naumann, Cliff Wong, Zelalem Gero, Javier Gonz´ alez, Yu Gu, Yanbo Xu, Mu Wei, Wenhui Wang, Shum- ing Ma, Furu Wei, Jianwei Yang, Chunyuan Li, Jianfeng Gao, Jaylen Rosemon, Tucker Bower, Soohee Lee, Roshanthi Weerasinghe, Bill J Wright, Ari Robicsek, Brian Piening, Carlo Bifulco, S...

    2024

  34. [34]

    A multi- modal knowledge-enhanced whole-slide pathology foundation model.Nature Communications, 16(1):11406, 2025

    Yingxue Xu, Yihui Wang, Fengtao Zhou, Jiabo Ma, Cheng Jin, Shu Yang, Jinbang Li, Zhengyu Zhang, Chenglong Zhao, Huajun Zhou, Zhenhui Li, Huangjing Lin, Xin Wang, Jiguang Wang, Anjia Han, Ronald Cheong Kin Chan, Li Liang, Xiuming Zhang, and Hao Chen. A multi- modal knowledge-enhanced whole-slide pathology foundation model.Nature Communications, 16(1):11406, 2025

    2025

  35. [35]

    Virchow2: Scaling self-supervised mixed magnification models in pathology, 2024

    Eric Zimmermann, Eugene Vorontsov, Julian Viret, Adam Casson, Michal Zelechowski, George Shaikovski, Neil Tenenholtz, James Hall, David Klimstra, Razik Yousfi, Thomas Fuchs, Nicolo Fusi, Siqi Liu, and Kristen Severson. Virchow2: Scaling self-supervised mixed magnification models in pathology, 2024. arXiv:2408.00738. 27 A Supplementary material A.1 Example...

    Pith/arXiv arXiv 2024