Pith. sign in

REVIEW 2 major objections 5 minor 77 references

One pathology foundation model can absorb the strengths of eight specialized teachers through staged distillation.

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 · grok-4.5

2026-07-13 02:22 UTC pith:RAQLMDDX

load-bearing objection Solid engineering paper that consolidates eight pathology teachers into one backbone and wins average rank on a 96-task suite; the novelty is staged modular distillation, not a new learning principle. the 2 major comments →

arxiv 2607.09526 v1 pith:RAQLMDDX submitted 2026-07-10 cs.CV cs.AI

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

Jiawen Li , Tian Guan , Huijuan Shi , Xitong Ling , Mingxi Fu , Anjia Han , Chao He , Yonghong He This is my paper
classification cs.CV cs.AI
keywords computational pathologyfoundation modelknowledge distillationvision-language modelwhole-slide imagemulti-stage traininghistopathologytransfer learning
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.

Pathology foundation models have been specialized by design: some capture local tissue morphology, some align images with diagnostic language, and some encode whole-slide context. That specialization fragments expertise across separate backbones, so different clinical tasks often need different pretrained models. ALICE is a single backbone trained by multi-stage agglomerative distillation that sequentially absorbs knowledge from eight existing teachers—three vision-only, three vision-language, and two slide-level models—into dedicated modules. Pretrained on nearly 25 million low-resolution tiles and over 150 thousand high-resolution regions, it is tested on 96 downstream tasks spanning region-of-interest analysis, multimodal retrieval and question answering, and whole-slide diagnosis, biomarker prediction, and survival. Across all three evaluation regimes it posts the best average rank among task-matched models. The paper argues that structured distillation can therefore replace a fragmented zoo of specialist models with one broadly transferable representation for computational pathology.

Core claim

Agglomerative distillation can consolidate complementary morphology-oriented, language-aligned, and whole-slide expertise from eight specialized pathology foundation models into a single modular backbone, and that unified model achieves the best average rank across vision-only, vision-language, and slide-level benchmarks among task-matched models.

What carries the argument

Multi-stage agglomerative distillation: three sequential stages that freeze earlier modules and align teacher-specific summary tokens (plus shift-equivariant patch tokens) with frozen teacher embeddings via a balanced angular distillation loss, so morphology, language, and slide-level knowledge accumulate without overwriting prior stages.

Load-bearing premise

That freezing earlier modules and using separate teacher-specific summary tokens lets complementary expert knowledge transfer cleanly without destructive interference or loss of the teachers’ specialized strengths.

What would settle it

A controlled ablation that trains only on the final slide-level stage, or collapses all eight teachers into one simultaneous distillation step, and then re-runs the full 96-task benchmark: if average ranks fall below the strongest individual teachers in each regime, the staged-consolidation claim fails.

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

If this is right

  • A single frozen backbone can replace separate vision-only, multimodal, and slide-level extractors for ROI classification, retrieval, segmentation, zero-shot and few-shot diagnosis, biomarker prediction, and survival analysis.
  • New specialized teachers can be absorbed later by adding a distillation stage rather than training an entirely new foundation model from scratch.
  • Clinical pipelines that currently switch models by task type can standardize on one representation while retaining or improving accuracy.
  • High-resolution WSI features distilled this way remain adaptable under both frozen linear probes and task-specific fine-tuning.

Where Pith is reading between the lines

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

  • If the modular design continues to scale, future pathology systems may treat foundation models as composable knowledge sources rather than monolithic end-points.
  • The same staged-distillation pattern could be applied outside pathology whenever complementary specialists exist at different spatial or semantic scales.
  • Prospective multi-site validation would be the next practical test of whether the consolidated representation survives real staining, scanner, and population shifts.

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 / 5 minor

Summary. ALICE is a unified pathology foundation model obtained by multi-stage agglomerative distillation of eight specialist teachers (UNI-2, Virchow-2, H-Opt-1; CONCH, MUSK, KEEP; TITAN, CARE) into a modular backbone with vision-only, multimodal, and slide-level transformers. Pretraining uses ~25M low-resolution TCGA tiles and ~156k high-resolution TCGA/HISTAI regions. The student is evaluated as a frozen or lightly adapted feature extractor across 21 scenarios, 96 tasks, and 48 sources spanning ROI classification/retrieval/segmentation/few-shot, vision–language zero-shot/retrieval/VQA/MI-Zero/PathPT/MICA, and WSI KNN/linear/biomarker/survival/fine-tuning. The central claim is that ALICE attains the best average rank among task-matched PFMs in all three evaluation settings (exceeding the second-best by 1.79, 6.39, and 3.04 percentage points).

Significance. If the ranking results hold, the paper provides strong empirical evidence that structured multi-stage distillation can consolidate morphology-oriented, language-aligned, and whole-slide expertise into a single deployable backbone, reducing the need to maintain separate specialist PFMs. Strengths include the unusually broad 96-task / 48-source benchmark with means, SDs, p-values and average ranks; public code and weights; and an explicit modular architecture with teacher-specific summary tokens and balanced angular / shift-equivariant patch losses (Eqs. 1–2). The work is of clear practical interest to computational pathology and is a natural next step beyond GPFM-style unified distillation.

major comments (2)
  1. Methods (multi-stage agglomerative distillation) and Results: the load-bearing premise is that freezing earlier modules and using teacher-specific summary tokens transfers complementary expertise without catastrophic interference. The three-setting average-rank claim is already a strong empirical test of that premise, but the manuscript still lacks a controlled ablation that isolates (i) sequential freezing vs joint multi-teacher training and (ii) teacher-specific summary tokens vs a single shared summary. Without these, it remains unclear how much of the gain is due to the staged architecture versus simply having access to eight strong teachers and large pretraining data. A compact ablation table on a representative subset of ROI / multimodal / WSI tasks would make the central architectural claim fully load-bearing rather than correlational.
  2. Discussion and Data availability: pretraining draws heavily on TCGA-12K / TCGA-UT-8K and HISTAI, while many downstream WSI tasks also use TCGA-derived or overlapping public cohorts (CPTAC, EBRAINS, etc.). The paper does not quantify possible teacher–student or pretrain–downstream slide-level leakage (e.g., whether any teacher was itself trained on the same TCGA slides later used for evaluation). Even a short overlap audit or a held-out non-TCGA-only subset analysis would strengthen the claim that the ranking gains are not partly explained by shared data exposure.
minor comments (5)
  1. Figure 1 and Extended Data Figure 1: the modular three-transformer schematic is clear, but the exact mapping from student summary tokens to each teacher (and which heads are active at inference for multimodal tasks) could be stated more explicitly in the caption or Methods.
  2. Results text reports average-rank improvements of 1.79 / 6.39 / 3.04 percentage points; clarify whether these are rank differences or absolute performance (BAcc/AUC) differences, and ensure consistency with the radar plots and Extended Data tables.
  3. Several Extended Data tables (e.g., segmentation, survival) show ALICE second-best on individual tasks while still winning average rank; a short sentence acknowledging this pattern would help readers interpret the aggregate claim.
  4. Typos and wording: Abstract/Introduction “unifed” → “unified”; occasional figure-panel lettering inconsistencies (e.g., Figure 3e/e); ensure all 48 data sources listed in Methods appear in the Data availability section with stable URLs.
  5. Hyperparameter tables (Extended Data Tables 1–3) are thorough; a one-line note on total GPU-hours or wall-clock would aid reproducibility comparisons with GPFM and the teachers.

Circularity Check

0 steps flagged

No significant circularity: empirical ranks on external downstream tasks are not forced by the distillation objectives or by self-citation.

full rationale

ALICE is trained by multi-stage agglomerative distillation that aligns student summary and patch tokens to frozen teacher embeddings via angular and L2 losses (Methods, Eqs. 1–2 and stage descriptions). The load-bearing claim is that the resulting unified backbone attains the best average rank among task-matched PFMs across 21 scenarios / 96 tasks on 48 data sources. Those ranks are computed from independent evaluation protocols (linear probe, KNN, ProtoNet, PMT, zero-shot, MI-Zero, PathPT, MICA, Coxph, fine-tuning, etc.) on held-out public and institutional cohorts; they are not algebraic rearrangements of the distillation losses, nor are they obtained by fitting a parameter to a subset and then “predicting” a closely related quantity. Teacher models are independently released external checkpoints; self-citations (authors’ prior datasets such as SYSBM/SYSFL+, or GPFM) supply data or a baseline but do not define the ranking metrics or forbid alternatives. The architecture (teacher-specific summary tokens, frozen earlier modules) is an explicit design choice inspired by RADIO, not a uniqueness theorem imported from the authors. Consequently the central empirical result is self-contained against external benchmarks and exhibits no circular reduction of the claimed derivation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claim rests on standard self-supervised and distillation practice plus the empirical premise that multi-stage freezing plus teacher-specific tokens can consolidate heterogeneous pathology expertise. Hyperparameters are conventional; no new physical constants or unfalsifiable entities are introduced.

free parameters (4)
  • peak learning rate / cosine schedule
    Chosen as 3e-4 with warmup and decay; standard but free choice that affects convergence of each distillation stage.
  • DAMP multiplicative noise strength
    Set to 0.05 / 0.02 across stages for robustness; hand-tuned hyperparameter.
  • number of summary / register tokens and projection heads
    Architectural free choices (3 summary tokens in vision stage, 2 in slide stage) that define how teacher knowledge is absorbed.
  • patch-crop size, grid, and loss weights (e.g., TITAN patch loss 0.5)
    Hand-selected to balance global and local distillation; not derived from first principles.
axioms (4)
  • domain assumption Frozen teacher embeddings constitute useful, complementary supervision that a student can absorb without access to the teachers' original training data or objectives.
    Core premise of agglomerative distillation (Methods, multi-stage section); standard in knowledge-distillation literature but not proven for pathology scale/modality mix.
  • ad hoc to paper Freezing earlier modules after each stage preserves previously acquired capabilities while later modules specialize.
    Explicit design choice of ALICE; if false, catastrophic forgetting would invalidate the unified-backbone claim.
  • standard math Standard self-supervised and contrastive losses (angular distillation, shift-equivariant patch matching, ALiBi attention) transfer morphological and semantic structure.
    Inherited from DINOv2, RADIO, and transformer literature; used without re-derivation.
  • domain assumption Public TCGA / HISTAI patches and the 48 downstream sources are sufficiently representative for ranking claims.
    Stated data sources; acknowledged limitation that prospective multi-institutional variation remains untested.
invented entities (2)
  • ALICE modular three-transformer backbone (vision-only + multimodal adaptor + slide-level ALiBi aggregator) independent evidence
    purpose: Provide dedicated slots that sequentially receive morphology, language, and whole-slide expertise while sharing a common visual encoder.
    New architectural assembly; independent evidence is the released weights and the empirical rank improvements, not an external physical prediction.
  • Teacher-specific summary tokens with balanced angular distillation loss no independent evidence
    purpose: Allow the student to absorb each teacher's global representation without forcing a single shared target space.
    Design device introduced for multi-teacher pathology distillation; validated only by downstream metrics inside the paper.

pith-pipeline@v1.1.0-grok45 · 68164 in / 2855 out tokens · 51968 ms · 2026-07-13T02:22:54.886580+00:00 · methodology

0 comments
read the original abstract

Foundation models are reshaping computational pathology, yet their capabilities remain shaped by pretraining objectives, data sources, and spatial scales, fragmenting complementary expertise across separate backbones. Here we present ALICE, a unified foundation model trained through multi-stage agglomerative distillation that sequentially distills eight vision-only, vision-language, and slide-level teacher models into dedicated modules of a single backbone. ALICE is pretrained on 24,985,184 tile-level pathology images and 155,604 high-resolution images, and evaluated across 21 task scenarios, 96 downstream tasks, and 48 data sources, spanning region-of-interest tissue analysis, vision-language multimodal evaluation, and whole-slide clinical assessment. In all three evaluation settings, ALICE achieved the best average rank among task-matched pathology foundation models. These results demonstrate that agglomerative distillation can consolidate complementary capabilities from specialized models into a unified backbone for broad computational pathology applications. The model is available at https://github.com/WonderLandxD/ALICE.

Figures

Figures reproduced from arXiv: 2607.09526 by Anjia Han, Chao He, Huijuan Shi, Jiawen Li, Mingxi Fu, Tian Guan, Xitong Ling, Yonghong He.

Figure 1
Figure 1. Figure 1: Overview of ALICE training and evaluation. a-c, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Vision-only feature transfer performance of ALICE. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Multimodal feature transfer performance of ALICE. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Slide-level feature transfer performance of ALICE. [PITH_FULL_IMAGE:figures/full_fig_p011_4.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

77 extracted references · 7 linked inside Pith

  1. [1]

    H.et al.Artificial intelligence for digital and computational pathology.Nature Reviews Bioengineering1, 930–949 (2023)

    Song, A. H.et al.Artificial intelligence for digital and computational pathology.Nature Reviews Bioengineering1, 930–949 (2023)

    2023

  2. [2]

    Zheng, Y .et al.Lazyslide: accessible and interoperable whole-slide image analysis.Nature Methods1–4 (2026)

    2026

  3. [3]

    Y .et al.Data-efficient and weakly supervised computational pathology on whole-slide images.Nature Biomedical Engineering5, 555–570 (2021)

    Lu, M. Y .et al.Data-efficient and weakly supervised computational pathology on whole-slide images.Nature Biomedical Engineering5, 555–570 (2021)

    2021

  4. [4]

    Jiang, R.et al.A transformer-based weakly supervised computational pathology method for clinical-grade diagnosis and molecular marker discovery of gliomas.Nature Machine Intelli- gence6, 876–891 (2024)

    2024

  5. [5]

    Campanella, G.et al.Clinical-grade computational pathology using weakly supervised deep learning on whole slide images.Nature Medicine25, 1301–1309 (2019)

    2019

  6. [6]

    Tsai, P.-C.et al.Histopathology images predict multi-omics aberrations and prognoses in colorectal cancer patients.Nature Communications14, 2102 (2023)

    2023

  7. [7]

    Zhu, L.et al.An accurate prediction of the origin for bone metastatic cancer using deep learning on digital pathological images.EBioMedicine87(2023)

    2023

  8. [8]

    J.et al.Towards a general-purpose foundation model for computational pathology

    Chen, R. J.et al.Towards a general-purpose foundation model for computational pathology. Nature Medicine30, 850–862 (2024)

    2024

  9. [9]

    Y .et al.A visual-language foundation model for computational pathology.Nature Medicine30, 863–874 (2024)

    Lu, M. Y .et al.A visual-language foundation model for computational pathology.Nature Medicine30, 863–874 (2024)

    2024

  10. [10]

    Yan, F.et al.Pathorchestra: A comprehensive foundation model for computational pathology with over 100 diverse clinical-grade tasks.npj Digital Medicine8, 695 (2025)

    2025

  11. [11]

    Nature630, 181–188 (2024)

    Xu, H.et al.A whole-slide foundation model for digital pathology from real-world data. Nature630, 181–188 (2024)

    2024

  12. [12]

    Neidlinger, P.et al.Benchmarking foundation models as feature extractors for weakly super- vised computational pathology.Nature Biomedical Engineering1–11 (2025)

    2025

  13. [13]

    Tizhoosh, H. R. Rethinking foundation models in pathology.Nature Biomedical Engineering 1–6 (2026)

    2026

  14. [14]

    V orontsov, E.et al.A foundation model for clinical-grade computational pathology and rare cancers detection.Nature Medicine30, 2924–2935 (2024)

    2024

  15. [15]

    Xiang, J.et al.A vision–language foundation model for precision oncology.Nature638, 769–778 (2025). 30

    2025

  16. [16]

    Campanella, G.et al.A clinical benchmark of public self-supervised pathology foundation models.Nature Communications16, 3640 (2025)

    2025

  17. [17]

    Ding, T.et al.A multimodal whole-slide foundation model for pathology.Nature Medicine 1–13 (2025)

    2025

  18. [18]

    InProceedings of the Computer Vision and Pattern Recognition Conference, 22487–22497 (2025)

    Heinrich, G.et al.Radiov2.5: Improved baselines for agglomerative vision foundation models. InProceedings of the Computer Vision and Pattern Recognition Conference, 22487–22497 (2025)

    2025

  19. [19]

    Zhu, C.et al.Efficient universal perception encoder.arXiv preprint arXiv:2603.22387(2026)

    arXiv 2026

  20. [20]

    Ma, J.et al.A generalizable pathology foundation model using a unified knowledge distilla- tion pretraining framework.Nature Biomedical Engineering10, 545–564 (2026)

    2026

  21. [21]

    Zimmermann, E.et al.Virchow2: Scaling self-supervised mixed magnification models in pathology.arXiv preprint arXiv:2408.00738(2024)

    Pith/arXiv arXiv 2024

  22. [22]

    Scalbert, M.et al.Abstract lb174: H-optimus-1: A foundation model for computational histopathology.Cancer Research86, LB174–LB174 (2026)

    2026

  23. [23]

    Zhou, X.et al.Knowledge-enhanced pretraining for vision-language pathology foundation model on cancer diagnosis.Cancer Cell44, 777–791 (2026)

    2026

  24. [24]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21078–21088 (2026)

    Zhang, D.et al.Care: A molecular-guided foundation model with adaptive region modeling for whole slide image analysis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21078–21088 (2026)

    2026

  25. [25]

    & Welling, M

    Ilse, M., Tomczak, J. & Welling, M. Attention-based deep multiple instance learning. In International conference on machine learning, 2127–2136 (PMLR, 2018)

    2018

  26. [26]

    N., Falcone, G

    Acosta, J. N., Falcone, G. J., Rajpurkar, P. & Topol, E. J. Multimodal biomedical ai.Nature medicine28, 1773–1784 (2022)

    2022

  27. [27]

    Y .et al.Visual language pretrained multiple instance zero-shot transfer for histopathol- ogy images

    Lu, M. Y .et al.Visual language pretrained multiple instance zero-shot transfer for histopathol- ogy images. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 19764–19775 (2023)

    2023

  28. [28]

    He, D.et al.Boosting pathology foundation models via few-shot prompt-tuning for rare cancer subtyping.Nature Communications(2026)

    2026

  29. [29]

    Nature Communications16, 7163 (2025)

    D ¨orrich, M.et al.A multimodal dataset for precision oncology in head and neck cancer. Nature Communications16, 7163 (2025)

    2025

  30. [30]

    Chowdhury, S.et al.Proteogenomic analysis of chemo-refractory high-grade serous ovarian cancer.Cell186, 3476–3498 (2023). 31

    2023

  31. [31]

    Li, Z.et al.Ai-enabled virtual spatial proteomics from histopathology for interpretable biomarker discovery in lung cancer.Nature Medicine1–14 (2026)

    2026

  32. [32]

    Nature634, 970–978 (2024)

    Wang, X.et al.A pathology foundation model for cancer diagnosis and prognosis prediction. Nature634, 970–978 (2024)

    2024

  33. [33]

    Nature Communications(2025)

    Xu, Y .et al.A multimodal knowledge-enhanced whole-slide pathology foundation model. Nature Communications(2025)

    2025

  34. [34]

    & Ivanova, E

    Nechaev, D., Pchelnikov, A. & Ivanova, E. Histai: an open-source, large-scale whole slide image dataset for computational pathology.arXiv preprint arXiv:2505.12120(2025)

    Pith/arXiv arXiv 2025

  35. [35]

    Nature Medicine30, 1174–1190 (2024)

    Vaidya, A.et al.Demographic bias in misdiagnosis by computational pathology models. Nature Medicine30, 1174–1190 (2024)

    2024

  36. [36]

    InMedical Imaging with Deep Learning, 544–569 (PMLR, 2026)

    Li, J.et al.Stainnet: Scaling self-supervised foundation models on immunohistochemistry and special stains for computational pathology. InMedical Imaging with Deep Learning, 544–569 (PMLR, 2026)

    2026

  37. [37]

    InInternational Conference on Medical Image Computing and Computer- Assisted Intervention, 573–583 (Springer, 2025)

    Karasikov, M.et al.Training state-of-the-art pathology foundation models with orders of mag- nitude less data. InInternational Conference on Medical Image Computing and Computer- Assisted Intervention, 573–583 (Springer, 2025)

    2025

  38. [38]

    Oquab, M.et al.Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193(2023)

    Pith/arXiv arXiv 2023

  39. [39]

    Press, O., Smith, N. A. & Lewis, M. Train short, test long: Attention with linear biases enables input length extrapolation.arXiv preprint arXiv:2108.12409(2021)

    Pith/arXiv arXiv 2021

  40. [40]

    & Shen, C

    Chu, X., Tian, Z., Zhang, B., Wang, X. & Shen, C. Conditional positional encodings for vision transformers.arXiv preprint arXiv:2102.10882(2021)

    Pith/arXiv arXiv 2021

  41. [41]

    & Bojanowski, P

    Darcet, T., Oquab, M., Mairal, J. & Bojanowski, P. Vision transformers need registers. In International conference on learning representations, vol. 2024, 2632–2652 (2024)

    2024

  42. [42]

    & Kaski, S

    Trinh, T., Heinonen, M., Acerbi, L. & Kaski, S. Improving robustness to corruptions with multiplicative weight perturbations.Advances in Neural Information Processing Systems37, 35492–35516 (2024)

    2024

  43. [43]

    Huo, X.et al.A comprehensive ai model development framework for consistent gleason grading.Communications Medicine4, 84 (2024)

    2024

  44. [44]

    Database2022, baac093 (2022)

    Brancati, N.et al.Bracs: A dataset for breast carcinoma subtyping in h&e histology images. Database2022, baac093 (2022)

    2022

  45. [45]

    A., Oliveira, L

    Spanhol, F. A., Oliveira, L. S., Petitjean, C. & Heutte, L. A dataset for breast cancer histopathological image classification.IEEE Transactions on Biomedical Engineering63, 1455–1462 (2015). 32

    2015

  46. [46]

    & Br ¨uck, O

    Brummer, O., P ¨ol¨onen, P., Mustjoki, S. & Br ¨uck, O. Computational textural mapping harmonises sampling variation and reveals multidimensional histopathological fingerprints. British Journal of Cancer129, 683–695 (2023)

    2023

  47. [47]

    & Jin, M

    Zhu, C., Chen, W., Peng, T., Wang, Y . & Jin, M. Hard sample aware noise robust learning for histopathology image classification.IEEE Transactions on Medical Imaging41, 881–894 (2021)

    2021

  48. [48]

    N., Halama, N

    Kather, J. N., Halama, N. & Marx, A. 100,000 histological images of human colorectal cancer and healthy tissue (2018). URLhttps://doi.org/10.5281/zenodo.1214456

    doi:10.5281/zenodo.1214456 2018

  49. [49]

    N.et al.Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer.Nature medicine25, 1054–1056 (2019)

    Kather, J. N.et al.Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer.Nature medicine25, 1054–1056 (2019)

    2019

  50. [50]

    Tolkach, Y .et al.Artificial intelligence for tumour tissue detection and histological regression grading in oesophageal adenocarcinomas: a retrospective algorithm development and valida- tion study.The Lancet Digital Health5, e265–e275 (2023)

    2023

  51. [51]

    Lou, S.et al.A large histological images dataset of gastric cancer with tumour microenviron- ment annotation for ai.Scientific Data12, 138 (2025)

    2025

  52. [52]

    Leavey, P.et al.Osteosarcoma data from ut southwestern/ut dallas for viable and necrotic tumor assessment [data set].Cancer Imaging Arch14(2019)

    2019

  53. [53]

    A.et al.Unitopatho, a labeled histopathological dataset for colorectal polyps classification and adenoma dysplasia grading

    Barbano, C. A.et al.Unitopatho, a labeled histopathological dataset for colorectal polyps classification and adenoma dysplasia grading. In2021 IEEE International Conference on Image Processing (ICIP), 76–80 (IEEE, 2021)

    2021

  54. [54]

    Graham, S.et al.Hover-net: Simultaneous segmentation and classification of nuclei in multi- tissue histology images.Medical Image Analysis58, 101563 (2019)

    2019

  55. [55]

    Graham, S.et al.Mild-net: Minimal information loss dilated network for gland instance segmentation in colon histology images.Medical Image Analysis52, 199–211 (2019)

    2019

  56. [56]

    Sirinukunwattana, K.et al.Gland segmentation in colon histology images: The glas challenge contest.Medical Image Analysis35, 489–502 (2017)

    2017

  57. [57]

    Sitnik, D.et al.A dataset and a methodology for intraoperative computer-aided diagnosis of a metastatic colon cancer in a liver.Biomedical Signal Processing and Control66, 102402 (2021)

    2021

  58. [58]

    Hu, W.et al.Ebhi: A new enteroscope biopsy histopathological h&e image dataset for image classification evaluation.Physica Medica107, 102534 (2023)

    2023

  59. [59]

    & Madabhushi, A

    Janowczyk, A. & Madabhushi, A. Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases.Journal of Pathology Informatics7, 29 (2016). 33

    2016

  60. [60]

    & Rajpoot, N

    Gamper, J. & Rajpoot, N. Multiple instance captioning: Learning representations from histopathology textbooks and articles. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition(2021)

    2021

  61. [61]

    InEuropean Conference on Computer Vision, 56–73 (Springer, 2024)

    Sun, Y .et al.Pathmmu: A massive multimodal expert-level benchmark for understanding and reasoning in pathology. InEuropean Conference on Computer Vision, 56–73 (Springer, 2024)

    2024

  62. [62]

    A.et al.Lung and colon cancer histopathological image dataset (lc25000)

    Borkowski, A. A.et al.Lung and colon cancer histopathological image dataset (lc25000). arXiv preprint arXiv:1912.12142(2019)

    Pith/arXiv arXiv 1912

  63. [63]

    B.et al.Viable and necrotic tumor assessment from whole slide images of osteosarcoma using machine-learning and deep-learning models.PloS One14, e0210706 (2019)

    Arunachalam, H. B.et al.Viable and necrotic tumor assessment from whole slide images of osteosarcoma using machine-learning and deep-learning models.PloS One14, e0210706 (2019)

    2019

  64. [64]

    A., Molina, R

    Silva-Rodr ´ıguez, J., Colomer, A., Sales, M. A., Molina, R. & Naranjo, V . Going deeper through the gleason scoring scale: An automatic end-to-end system for histology prostate grading and cribriform pattern detection.Computer Methods and Programs in Biomedicine 195, 105637 (2020)

    2020

  65. [65]

    Kriegsmann, K.et al.Deep learning for the detection of anatomical tissue structures and neoplasms of the skin on scanned histopathological tissue sections.Frontiers in Oncology12, 1022967 (2022)

    2022

  66. [66]

    Han, C.et al.Wsss4luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma.arXiv preprint arXiv:2204.06455(2022)

    Pith/arXiv arXiv 2022

  67. [67]

    Xu, F.et al.Predicting axillary lymph node metastasis in early breast cancer using deep learning on primary tumor biopsy slides.Frontiers in Oncology11, 759007 (2021)

    2021

  68. [68]

    Ling, X.et al.Comprehensive benchmark dataset for pathological lymph node metastasis in breast cancer sections.Scientific Data12, 1381 (2025)

    2025

  69. [69]

    Zhu, M.et al.Development and evaluation of a deep neural network for histologic classifi- cation of renal cell carcinoma on biopsy and surgical resection slides.Scientific Reports11, 7080 (2021)

    2021

  70. [70]

    Scientific Data9, 55 (2022)

    Roetzer-Pejrimovsky, T.et al.The digital brain tumour atlas, an open histopathology resource. Scientific Data9, 55 (2022)

    2022

  71. [71]

    ´A.et al.Huncrc: annotated pathological slides to enhance deep learning applications in colorectal cancer screening.Scientific Data9, 370 (2022)

    Pataki, B. ´A.et al.Huncrc: annotated pathological slides to enhance deep learning applications in colorectal cancer screening.Scientific Data9, 370 (2022)

    2022

  72. [72]

    Bulten, W.et al.Artificial intelligence for diagnosis and gleason grading of prostate cancer: the panda challenge.Nature medicine28, 154–163 (2022)

    2022

  73. [73]

    Li, J.et al.Diagnostic text-guided representation learning in hierarchical classification for pathological whole slide image.Medical Image Analysis103894 (2025). 34

    2025

  74. [74]

    H.et al.A partial label-based machine learning approach for cervical whole-slide image classification: the winning tissuenet solution

    Fick, R. H.et al.A partial label-based machine learning approach for cervical whole-slide image classification: the winning tissuenet solution. In2021 43rd Annual International Con- ference of the IEEE Engineering in Medicine & Biology Society (EMBC), 2127–2131 (IEEE, 2021)

    2021

  75. [75]

    Asadi-Aghbolaghi, M.et al.Machine learning-driven histotype diagnosis of ovarian carci- noma: insights from the ocean ai challenge.medRxiv2024–04 (2024)

    2024

  76. [76]

    H., Marshall, C., Harris-Birtill, D

    Myles, C., Um, I. H., Marshall, C., Harris-Birtill, D. & Harrison, D. J. Surgen: 1020 h&e-stained whole-slide images with survival and genetic markers.GigaScience14, giaf086 (2025)

    2025

  77. [77]

    ballooning drop

    Snell, J., Swersky, K. & Zemel, R. Prototypical networks for few-shot learning.Advances in Neural Information Processing Systems30(2017). 35 … … Transformer layer Transformer layer … … Transformer layer Transformer layer Patch tokens Register tokens Summary tokens Transformer layer × 32 𝑙𝑎𝑦𝑒𝑟𝑠 a. Vision-only transformer … … … … × 2 𝑙𝑎𝑦𝑒𝑟𝑠 Teacher-align to...

    arXiv 2017