Pith. sign in

REVIEW 2 major objections 5 minor 28 references

Multi-teacher contrastive distillation packs large pathology foundation models into tiny edge encoders that nearly match teacher accuracy while running hundreds of times faster on ordinary hardware.

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

2026-07-11 06:12 UTC pith:O6I2L2PS

load-bearing objection Solid applied distillation paper: near-teacher CPTAC AUROC at orders-of-magnitude smaller size, with real Raspberry Pi numbers; main caveat is the teacher-solvable task filter, not the reported gaps. the 2 major comments →

arxiv 2607.05533 v1 pith:O6I2L2PS submitted 2026-07-06 cs.CV

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

Tim Lenz , Maurice Heide , Marco Gustav , Nic G. Reitsam , Jakob Nikolas Kather This is my paper
classification cs.CV
keywords Computational pathologyFoundation modelsContrastive distillationKnowledge distillationEdge deploymentMobileOneRepViTWhole-slide images
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.

Large pathology foundation models improve whole-slide analysis but are too big and slow for local use in pathology departments. This paper proposes MuCoDi, which trains compact MobileOne and RepViT student encoders by contrastive distillation from three frozen teacher models, using cached tile embeddings as keys instead of regressing teacher features one by one. Students are pretrained on 14.3 million tiles from only 11.8 thousand TCGA slides, then frozen and evaluated with slide-level heads on 23 clinical classification tasks with external validation. RepViT students stay within about 0.8 percentage points of the best teacher external AUROC while cutting size and compute by orders of magnitude; sub-million-parameter MobileOne students still reach roughly 66.5–66.9% external AUROC with up to 605-fold single-tile speedup on a Raspberry Pi 5. The claim is that PFM-quality pathology representations can run on departmental workstations or microscope-attached edge devices, keeping sensitive images local and lowering the barrier to routine AI assistance.

Core claim

MuCoDi distills complementary frozen tile embeddings from Virchow2, UNI2, and H-Optimus-1 into edge-oriented MobileOne and RepViT students via a multi-teacher contrastive objective. On 23 external clinical tasks, MuCoEdge-R2.3 and R1.5 reach 71.0% AUROC (within 0.8 points of the best teacher at 71.8%), with R2.3 leading external F1; sub-million MobileOne students retain 66.5–66.9% AUROC while delivering up to 605× single-tile Raspberry Pi speedup over Virchow2.

What carries the argument

MuCoDi: a MoCo-v3-style contrastive distillation loss in which cached multi-teacher PFM embeddings replace momentum-encoder keys, so a shared lightweight student backbone and teacher-specific projection heads pull same-tile student queries toward teacher positives against in-batch negatives.

Load-bearing premise

The reported near-teacher retention only covers clinical tasks that the large teachers already solve reasonably well on internal data, so the edge trade-off is proven mainly for morphology signals the teachers already capture.

What would settle it

Measure MuCoEdge students on clinically important endpoints where all three teachers score below about 0.65 internal AUROC: if students then fall far below usable accuracy while remaining fast, the claim that PFM-quality representations have been moved to the edge would not hold for those endpoints.

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

If this is right

  • Pathology departments can run near-teacher tile encoders on ordinary workstations and low-cost edge devices instead of dedicated GPUs or cloud pipelines.
  • Student pretraining needs only about 11.8K unlabeled whole-slide images, far fewer than the teachers, lowering the data barrier for compact model development.
  • A practical accuracy–latency spectrum exists: larger RepViT students for higher accuracy and sub-million MobileOne students for strict latency budgets.
  • Local inference keeps sensitive whole-slide images inside the department, reducing data movement and privacy risk.
  • Compact encoders support future microscope- or workstation-integrated assistance for screening, triage, and biomarker prescreening.

Where Pith is reading between the lines

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

  • If the multi-teacher contrastive recipe generalizes, departments could re-distill newer foundation models into the same edge backbones without repeating huge-scale teacher pretraining.
  • Because the benchmark filters for teacher-solvable endpoints, claims about clinical readiness should be re-checked on tasks teachers still miss before deployment decisions.
  • Pairing these students with adaptive patch selection could further cut whole-slide latency for real-time triage on the same hardware.
  • End-to-end workflow studies on Raspberry-Pi-class devices would test whether tile speedups translate into diagnostic time savings at the bench.

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. The paper introduces MuCoDi, a multi-teacher contrastive distillation framework that adapts MoCo v3 so that cached embeddings from frozen Virchow2, UNI2, and H-Optimus-1 replace momentum-encoder keys. Lightweight MobileOne and RepViT students (MuCoEdge family) are pretrained on 14.3M tiles from 11.8K TCGA WSIs and evaluated as frozen tile encoders with STAMP transformer MIL heads on 23 binary clinical endpoints, using patient-level 5-fold TCGA training and held-out CPTAC external validation. The central empirical claim is that RepViT students retain near-teacher external performance (MuCoEdge-R2.3/R1.5 at 71.0% AUROC vs Virchow2 71.8%; R2.3 best external F1 51.8%) at orders-of-magnitude smaller size, while sub-million MobileOne variants retain 66.5–66.9% AUROC with up to 605× single-tile Raspberry Pi 5 speedup over Virchow2.

Significance. If the reported retention and latency numbers hold under the stated protocol, the work is a concrete step toward local, departmental deployment of PFM-quality tile representations without GPU clusters. Strengths include multi-teacher contrastive (not pure regression) distillation into true edge CNN/RepViT backbones, external CPTAC validation with three metrics and CI half-widths, per-organ breakdowns, parameter/GFLOP tables, and timed single-tile Raspberry Pi measurements without quantization. Pretraining on only 11.8K WSIs while matching or exceeding LiteFM on several external metrics is practically relevant. The contribution is primarily empirical systems/engineering rather than a new theoretical principle, but the efficiency–accuracy spectrum is useful for the field.

major comments (2)
  1. Section 3.2 task selection requires that all three teachers achieve ≥0.65 internal TCGA AUROC on every endpoint before CPTAC evaluation. This conditions the benchmark on morphology signals already captured by the teachers, so the near-teacher retention claim (Tables 2–3, abstract) is retention of already-solvable signals rather than discovery of new ones. The paper should either (a) report the full unfiltered task pool and how many endpoints were excluded, or (b) add a sensitivity analysis on endpoints below the threshold, and state the scope limitation explicitly in the abstract/conclusion.
  2. The multi-teacher contrastive design is presented as central (Eqs. 1–2, Fig. 1), yet the manuscript lacks load-bearing ablations: single-teacher vs multi-teacher, contrastive InfoNCE vs feature regression, and with/without teacher-specific projection heads. Without these, it is unclear whether the reported 71.0% external AUROC and edge trade-offs require the full MuCoDi objective or would be obtained by simpler distillation into the same backbones. A minimal ablation table on a subset of models/tasks is needed to support the methodological claim.
minor comments (5)
  1. Table 2 CI half-widths are large relative to the 0.8 pp teacher–student AUROC gap; a brief note on whether differences are within uncertainty would help interpretation.
  2. Edge benchmark (Section 3.3) is single-tile, batch-1, float32, single-threaded on Raspberry Pi 5. Clarify that WSI-level throughput (thousands of tiles) and any memory constraints are not measured.
  3. LiteFM is noted to include CPTAC in pretraining while MuCoEdge does not; make this comparison caveat more prominent near Table 2 / Fig. 2.
  4. Code is anonymous; ensure public release of student weights, teacher-cache recipe, and exact task lists for reproducibility after acceptance.
  5. Minor typos and spacing artifacts (e.g., “Multi-T eacher”, “a nd”, “Uni ted Kingdom”) should be cleaned in production.

Circularity Check

0 steps flagged

No circularity: empirical multi-teacher distillation with held-out CPTAC evaluation; performance and speedups are measured, not forced by definition or self-citation.

full rationale

MuCoDi is an empirical engineering paper. Students are trained with a standard InfoNCE/MoCo-v3-style contrastive objective against cached frozen embeddings from external teachers (Virchow2, UNI2, H-Optimus-1); the loss (Eqs. 1–2) does not encode the downstream AUROC/F1 targets. Downstream protocol freezes the student, trains MIL heads only on TCGA (patient-stratified 5-fold CV via STAMP), and deploys unchanged to matched CPTAC—so external metrics are not fitted inputs renamed as predictions. Edge latencies are measured single-tile float32 timings on a fixed Raspberry Pi 5 setup. Self-citations (COBRA, STAMP, prior benchmarking) supply related multi-PFM context and evaluation tooling but do not supply a uniqueness theorem, ansatz, or numerical target that forces the reported near-teacher gaps or 327–605× speedups. The Section 3.2 filter (endpoints where all teachers already reach ≥0.65 internal AUROC) scopes the benchmark to teacher-solvable morphology signals; that is a selection bias concern, not a circular reduction of the reported numbers to their inputs. No self-definitional loop, no fitted-parameter-as-prediction, and no load-bearing self-citation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

Load-bearing content is empirical ML: standard contrastive KD assumptions, public cohort data, frozen teacher embeddings as supervision, and a curated task suite. Free parameters are ordinary training/architecture choices; axioms are domain assumptions about tile features and MIL transfer; invented entities are the MuCoDi procedure and MuCoEdge model family, which are engineering constructs with experimental handles rather than new physical entities.

free parameters (4)
  • contrastive temperature τ
    Set to 0.2 for InfoNCE; controls softness of teacher-key matching and thus student feature geometry.
  • student pretraining schedule (10 epochs, base lr 2e-2, weight decay 1e-6, global batch 2048)
    Hand-chosen optimization settings that determine how completely students fit the multi-teacher contrastive objective on 14.3M tiles.
  • teacher-specific projection head output dims (1280 / 1536)
    Chosen to match teacher embedding dimensionalities; affects alignment capacity per teacher.
  • task inclusion threshold (teacher AUROC ≥ 0.65 on internal TCGA)
    Explicit selection rule that defines which of the clinical endpoints enter the 23-task aggregate; changes the reported mean metrics.
axioms (5)
  • domain assumption Cached frozen embeddings from Virchow2, UNI2, and H-Optimus-1 are sufficient multi-view supervision for learning transferable pathology tile features via contrastive matching.
    Core MuCoDi premise in §3.1; students never see teacher gradients or raw teacher weights, only keys.
  • domain assumption Slide-level clinical labels can be predicted from bag-of-tile student features with a frozen encoder plus STAMP transformer MIL trained only on TCGA.
    Evaluation protocol §3.3; standard weakly supervised pathology assumption.
  • domain assumption TCGA→CPTAC external deployment with matched binary endpoints is a valid measure of clinical generalization for these morphology signals.
    Used throughout Results; cohort shift is real but still same-domain public cancer cohorts.
  • ad hoc to paper InfoNCE with in-batch cross-GPU negatives (MoCo v3-style, without momentum encoder) transfers relational teacher structure better for edge students than pure feature regression.
    Design choice stated in abstract/§3.1; limited comparative ablation in the reported tables.
  • standard math Standard math of L2-normalized cosine similarity and temperature-scaled softmax (InfoNCE) is valid for multi-teacher distillation loss aggregation.
    Equations (1)–(2) in §3.1.
invented entities (2)
  • MuCoDi pretraining framework independent evidence
    purpose: Replace MoCo v3 momentum keys with multi-PFM cached embeddings and train lightweight students with summed multi-teacher contrastive loss.
    Named method of the paper; engineering procedure, not a physical entity. Independent evidence is the reported student metrics and edge latency.
  • MuCoEdge student model family (RepViT and MobileOne variants) independent evidence
    purpose: Concrete compact encoders claimed to retain near-teacher pathology representations under edge constraints.
    New model checkpoints/results; falsifiable via re-training and external task evaluation.

pith-pipeline@v1.1.0-grok45 · 17339 in / 3661 out tokens · 35259 ms · 2026-07-11T06:12:39.679701+00:00 · methodology

0 comments
read the original abstract

Computational pathology foundation models (PFMs) have advanced whole-slide image analysis. However, their size and inference cost hinder local deployment in pathology departments. We propose MuCoDi, a pretraining framework that distills frozen tile embeddings from multiple PFMs into compact edge-oriented encoders. Instead of regressing individual teacher features, MuCoDi trains lightweight MobileOne and RepViT students with a contrastive distillation objective adapted from MoCo v3, where cached Virchow2, UNI2, and H-Optimus-1 embeddings replace momentum-encoder keys. We pretrain students on 14.3M TCGA tiles from only 11.8K WSIs and evaluate frozen encoders on 23 clinically curated downstream classification tasks. RepViT-based MuCoEdge students retain near-teacher performance while reducing model size by orders of magnitude: MuCoEdge-R2.3 and MuCoEdge-R1.5 reach 71.0% external AUROC, within 0.8 percentage points of the best teacher (Virchow2, 71.8%), while MuCoEdge-R2.3 obtains the best external F1 and the second-best AUPRC (51.8% and 53.3%). MuCoEdge-R1.0 reaches 70.9% AUROC with only 6.4M parameters and 1.12 GFLOPs. On a Raspberry Pi 5, sub-million-parameter MobileOne students achieve up to 605-fold single-tile speedup over Virchow2 while retaining 66.5% to 66.9% external AUROC, demonstrating that PFM-quality pathology representations can be moved toward practical edge deployment. Code is available at https://anonymous.4open.science/r/mucodi-6243.

Figures

Figures reproduced from arXiv: 2607.05533 by Jakob Nikolas Kather, Marco Gustav, Maurice Heide, Nic G. Reitsam, Tim Lenz.

Figure 1
Figure 1. Figure 1: Overview of MuCoDi pretraining. Frozen PFM teachers provide cached tile embeddings, while a lightweight student encoder and teacher-specific projection heads are trained with a multi-teacher contrastive distillation loss. is increasingly urgent: global cancer incidence was close to 20 million new cases in 2022 and is projected to reach 35 million by 2050 [1]. At the same time, histopathology workforces are… view at source ↗
Figure 2
Figure 2. Figure 2: Efficiency and edge deployment. (a) MuCoEdge models compared with LiteFM variants and teacher foundation models on CPTAC external validation. (b) Single-tile inference speedup over Virchow2 on Raspberry Pi 5. Efficiency and edge deployment. The efficiency gains are substantial (Tab. 1, Fig. 2a). Speedups are reported against Virchow2, the most efficient high￾performing teacher reference in our benchmark. M… 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

28 extracted references · 10 canonical work pages

  1. [1]

    L., Soerjomataram, I., Jemal, A.: Global cancer statistics 2022: GLOBOCAN estimat es of incidence and mortality worldwide for 36 cancers in 185 countries

    Bray, F., Laversanne, M., Sung, H., Ferlay, J., Siegel, R. L., Soerjomataram, I., Jemal, A.: Global cancer statistics 2022: GLOBOCAN estimat es of incidence and mortality worldwide for 36 cancers in 185 countries. CA: A Ca ncer Journal for Clinicians 74(3), 229–263 (2024). https://doi.org/10.3322/caac.2183 4

    work page doi:10.3322/caac.2183 2022

  2. [2]

    arXiv preprint arXiv :2602.14010 (2026)

    Cai, Y., Jin, C., Ma, J., et al.: A deployment-friendly fou ndational framework for efficient computational pathology. arXiv preprint arXiv :2602.14010 (2026). https://doi.org/10.48550/arXiv.2602.14010

    work page doi:10.48550/arxiv.2602.14010 2026

  3. [3]

    Nature Medicine 30, 850–862 (2024)

    Chen, R.J., Ding, T., Lu, M.Y., Williamson, D.F.K., Jaume , G., Chen, B., Zhang, A., Shao, D., Song, A.H., Shaban, M., et al.: Towards a genera l-purpose foun- dation model for computational pathology. Nature Medicine 30, 850–862 (2024). https://doi.org/10.1038/s41591-024-02857-3

    doi:10.1038/s41591-024-02857-3 2024

  4. [4]

    In: Proceedings of the IEEE/CVF Internation al Conference on Com- puter Vision (2021)

    Chen, X., Xie, S., He, K.: An empirical study of training se lf-supervised vision transformers. In: Proceedings of the IEEE/CVF Internation al Conference on Com- puter Vision (2021)

    2021

  5. [5]

    EBioMedicine 107, 105276 (2024)

    Choudhury, D., Dolezal, J.M., Dyer, E., Kochanny, S., Ram esh, S., Howard, F.M., Margalus, J.R., Schroeder, A., Schulte, J., Garassin o, M.C., Kather, J.N., Pearson, A.T.: Developing a low-cost, open-source, l ocally manufac- tured workstation and computational pipeline for automate d histopathol- ogy evaluation using deep learning. EBioMedicine 107, 1052...

    work page doi:10.1016/j.ebiom.2024.105276 2024

  6. [6]

    Nature Protocols 20, 293–316 (2025)

    El Nahhas, O.S.M., van Treeck, M., Wölflein, G., Unger, M., Ligero, M., Lenz, T., Wagner, S.J., Hewitt, K.J., et al.: From whole-slide image t o biomarker prediction: end-to-end weakly supervised deep learning in computation al pathology. Nature Protocols 20, 293–316 (2025). https://doi.org/10.1038/s41596-024-0 1047-2

    work page doi:10.1038/s41596-024-0 2025

  7. [7]

    Cancer Discovery 3(10), 1108–1112 (2013)

    Ellis, M.J., Gillette, M., Carr, S.A., Paulovich, A.G., S mith, R.D., Rodland, K.K., Townsend, R.R., Kinsinger, C., Mesri, M., Rodriguez, H., Li ebler, D.C.: Con- necting genomic alterations to cancer biology with proteom ics: The NCI clinical proteomic tumor analysis consortium. Cancer Discovery 3(10), 1108–1112 (2013). https://doi.org/10.1158/2159-8290....

    work page doi:10.1158/2159-8290.cd-13-0219 2013

  8. [8]

    I n: Medical Im- age Computing and Computer Assisted Intervention – MICCAI 2 025

    Filiot, A., Dop, N., Tchita, O., Riou, A., Dubois, R., Peet ers, T., Valter, D., Scalbert, M., Saillard, C., Robin, G., Olivier, A.: Dist illing foundation models for robust and efficient models in digital pathology. I n: Medical Im- age Computing and Computer Assisted Intervention – MICCAI 2 025. Lec- ture Notes in Computer Science, vol. 15966, pp. 162–172...

    work page doi:10.1007/978-3-032-04981-0_16 2026

  9. [9]

    In: NIPS Deep Learning and Representation Learning Wo rkshop (2015)

    Hinton, G., Vinyals, O., Dean, J.: Distilling the knowled ge in a neural net- work. In: NIPS Deep Learning and Representation Learning Wo rkshop (2015). https://doi.org/10.48550/arXiv.1503.02531

    Pith/arXiv arXiv doi:10.48550/arxiv.1503.02531 2015

  10. [10]

    In: Proceedings of the IEEE/CVF International Conference on Co mputer Vision

    Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Ta n, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., Le, Q.V., Adam, H.: Searching fo r MobileNetV3. In: Proceedings of the IEEE/CVF International Conference on Co mputer Vision. pp. 1314–1324 (2019)

    2019

  11. [11]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and P attern Recogni- tion (2025)

    Lenz, T., Neidlinger, P., Ligero, M., Wölflein, G., van Tr eeck, M., Kather, J.N.: Un- supervised foundation model-agnostic slide-level repres entation learning. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and P attern Recogni- tion (2025). https://doi.org/10.48550/arXiv.2411.1362 3

    work page doi:10.48550/arxiv.2411.1362 2025

  12. [12]

    In : Advances in Neural Information Processing Systems

    Li, Y., Yuan, G., Wen, Y., Hu, E., Evangelidis, G., Tulyak ov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at MobileNet speed. In : Advances in Neural Information Processing Systems. vol. 35, pp. 12934–12949 ( 2022)

    2022

  13. [13]

    Nature Biomedic al Engineering (2025)

    Ma, J., Guo, Z., Zhou, F., Wang, Y., Xu, Y., Li, J., Yan, F., Cai, Y., Zhu, Z., Jin, C., et al.: A generalizable pathology foundation model using a unified knowl- edge distillation pretraining framework. Nature Biomedic al Engineering (2025). https://doi.org/10.1038/s41551-025-01488-4

    doi:10.1038/s41551-025-01488-4 2025

  14. [14]

    Huggi ng Face model card (2025), https://huggingface.co/MahmoodLab/UNI2-h

    Mahmood Lab: UNI2-h: Pathology foundation model. Huggi ng Face model card (2025), https://huggingface.co/MahmoodLab/UNI2-h

    2025

  15. [15]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Mallya, M., Mirabadi, A.K., Farahani, H., Bashashati, A .: HistoMILKD: A multiple instance learning based multi-teacher knowledge distilla tion framework for whole slide image classification. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3390–3400 (2026)

    2026

  16. [16]

    Nature Biomedical Engi neering 10, 1113–1123 (2026)

    Neidlinger, P., El Nahhas, O.S.M., Muti, H.S., Lenz, T., Hoffmeister, M., Brenner, H., et al.: Benchmarking foundation models as feature extra ctors for weakly su- pervised computational pathology. Nature Biomedical Engi neering 10, 1113–1123 (2026). https://doi.org/10.1038/s41551-025-01516-3

    doi:10.1038/s41551-025-01516-3 2026

  17. [17]

    https://doi.org/10.48550/arXiv.2502

    Neidlinger, P., Lenz, T., Foersch, S., Loeffler, C.M.L., C lusmann, J., Gustav, M., Shaktah, L.A., et al.: A deep learning framework for efficient pathology image analysis (2025). https://doi.org/10.48550/arXiv.2502. 13027

    doi:10.48550/arxiv.2502 2025

  18. [18]

    van den Oord, A., Li, Y., Vinyals, O.: Representation lea rning with contrastive predictive coding (2018)

    2018

  19. [19]

    Model release (202 4), https://github.com/bioptimus/releases/tree/main/models/h-optimus/v0 MuCoDi for Edge-Efficient Pathology Foundation Models 11

    Saillard, C., Jenatton, R., Llinares-Lopez, F., Mariet , Z., Cahane, D., Durand, E., Vert, J.P.: H-optimus-0. Model release (202 4), https://github.com/bioptimus/releases/tree/main/models/h-optimus/v0 MuCoDi for Edge-Efficient Pathology Foundation Models 11

  20. [20]

    In: Proceedings of the American Association for Cancer Researc h Annual Meeting 2026; Part 2 (Late-Breaking, Clinical Trial, and Invited Ab stracts)

    Scalbert, M., Saillard, C., Peeters, T., Gonzalez, L., V alter, D., Llinares-Lopez, F., et al.: H-optimus-1: A foundation model for computation al histopathology. In: Proceedings of the American Association for Cancer Researc h Annual Meeting 2026; Part 2 (Late-Breaking, Clinical Trial, and Invited Ab stracts). vol. 86, p. LB174. AACR (2026). https://doi....

    work page doi:10.1158/1538-7445 2026

  21. [21]

    In: International Conference on Learning Representa tions (2020), https://openreview.net/forum?id=SkgpBJrtvS

    Tian, Y., Krishnan, D., Isola, P.: Contrastive represen tation distilla- tion. In: International Conference on Learning Representa tions (2020), https://openreview.net/forum?id=SkgpBJrtvS

    2020

  22. [22]

    : MobileOne: An improved one millisecond mobile backbone

    Vasu, P.K.A., Gabriel, J., Zhu, J., Tuzel, O., Ranjan, A. : MobileOne: An improved one millisecond mobile backbone. In: Proceedings of the IEE E/CVF Conference on Computer Vision and Pattern Recognition. pp. 7907–7917 ( 2023)

    2023

  23. [23]

    Nature Medicine 30, 2924–2935 (2024)

    Vorontsov, E., Bozkurt, A., Casson, A., et al.: A foundat ion model for clinical- grade computational pathology and rare cancers detection. Nature Medicine 30, 2924–2935 (2024). https://doi.org/10.1038/s41591-024- 03141-0

    doi:10.1038/s41591-024- 2024

  24. [24]

    Diagnostic Pathology 19, 163 (2024)

    Walsh, E., Orsi, N.M.: The current troubled state of the g lobal pathol- ogy workforce: A concise review. Diagnostic Pathology 19, 163 (2024). https://doi.org/10.1186/s13000-024-01590-2

    work page doi:10.1186/s13000-024-01590-2 2024

  25. [25]

    In: Proceedings of the IEEE/CVF Confe rence on Computer Vision and Pattern Recognition

    Wang, A., Chen, H., Lin, Z., Han, J., Ding, G.: RepViT: Rev isiting mobile CNN from ViT perspective. In: Proceedings of the IEEE/CVF Confe rence on Computer Vision and Pattern Recognition. pp. 15909–15920 (2024)

    2024

  26. [26]

    R.M., Ozenberger, B.A., Ellrott, K., Shmulevich, I., et al.: The cancer genome atlas pan-cancer analysis project

    Weinstein, J.N., Collisson, E.A., Mills, G.B., Shaw, K. R.M., Ozenberger, B.A., Ellrott, K., Shmulevich, I., et al.: The cancer genome atlas pan-cancer analysis project. Nature Genetics 45, 1113–1120 (2013). https://doi.org/10.1038/ng.2764

    doi:10.1038/ng.2764 2013

  27. [27]

    Nature 630, 181–188 (2024)

    Xu, H., Usuyama, N., Bagga, J., Zhang, S., Rao, R., Nauman n, T., Wong, C., et al.: A whole-slide foundation model for digital pathology from r eal-world data. Nature 630, 181–188 (2024). https://doi.org/10.1038/s41586-024-0 7441-w

    work page doi:10.1038/s41586-024-0 2024

  28. [28]

    https://doi.org/10.48550/arXiv.2408.00738

    Zimmermann, E., Vorontsov, E., Viret, J., Casson, A., Ze lechowski, M., et al.: Virchow2: Scaling self-supervised mixed magnification mod els in pathology (2024). https://doi.org/10.48550/arXiv.2408.00738

    Pith/arXiv arXiv doi:10.48550/arxiv.2408.00738 2024