Pith. sign in

REVIEW 5 major objections 5 minor 52 references

Subspecialty-Specific Foundation Model for Intelligent Gastrointestinal Pathology

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A GI-specialized foundation model, pretrained on 353 million multi-scale images and fine-tuned on selected tumor regions, tops general pathology models on 33 of 34 GI tasks and reaches 99.70% screening sensitivity.

desk verdict A serious GI-specific pathology foundation model with real engineering, currently undermined by an evaluation confound (supervised curation overlapping downstream tasks) and no reported decontamination check; needs major revision, not desk rejection. read the letter →

arxiv 2505.21928 v2 pith:4RVUUOX2 submitted 2025-05-28 eess.IV cs.AIcs.CVcs.LG

Lianghui Zhu , Xitong Ling , Minxi Ouyang , Xiaoping Liu , Tian Guan , Mingxi Fu , Zhiqiang Cheng , Fanglei Fu
show 19 more authors
Maomao Zeng Liming Liu Song Duan Qiang Huang Ying Xiao Jianming Li Shanming Lu Zhenghua Piao Mingxi Zhu Yibo Jin Shan Xu Qiming He Yizhi Wang Junru Cheng Xuanyu Wang Luxi Xie Houqiang Li Sufang Tian Yonghong He
This is my paper
classification eess.IVcs.AIcs.CVcs.LG
keywords gastrointestinalpathologyfoundationmodelself-supervisedlearningwhole-slideimagesearlycancerscreeningmultipleinstancemulti-scalepretrainingcomputational
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Digepath is a foundation model built specifically for gastrointestinal pathology rather than for all tissues. The paper reports that pretraining a vision transformer on 353 million multi-scale image patches from 210,043 H&E-stained GI slides, followed by a second stage in which a tumor-region classifier selects diagnostically important areas for further fine-tuning, yields a single encoder that beats general-purpose pathology foundation models on 33 of 34 GI tasks. The winning tasks span routine diagnosis, grading, segmentation, protein-expression prediction (PD-L1, HER2, P53), mutation and microsatellite-instability prediction, and survival prognosis. The same encoder powers an early cancer screening module that, in a nine-institution study, reached 99.70% sensitivity and 89.30% specificity for low-grade neoplasia, high-grade neoplasia, and malignancy. The authors argue this makes a single specialized encoder both more accurate and more deployable than pan-tissue models for routine GI pathology.

What carries the argument

The engine of the paper is a two-phase training loop. Phase I is self-supervised pretraining of a vision transformer on 353,478,334 multi-scale patches (2.5×, 5×, 10×, 20×) taken from 210,043 H&E GI slides. Phase II is a closed-loop data refinement: a tumor classifier trained on 471,443 pathologist-annotated regions scans the original slide pool, ranks regions by predicted tumor probability, and samples the highest-confidence tumor regions plus random non-tumor regions to build a balanced 83,206,828-patch fine-tuning set; the encoder is then fine-tuned on this set to become Digepath-V2. Downstream WSI tasks use attention-based multiple-instance learning over the encoder's patch embeddings, and segmentation tasks plug the encoder into a transformer-augmented U-Net. The confidence-weighted sampling formulas are what convert classifier output into a curated training curriculum.

What would settle it

If an independent audit shows that slides or patients used in Phase II ROI selection also appear in the test splits of the 34 benchmarks, or if general-purpose encoders given the same Phase II fine-tuning protocol match or beat Digepath, the 33-of-34 SOTA claim would be refuted; a clean test would be to hold out entire external institutions from both pretraining and fine-tuning and rerun the benchmarks.

Watch

Extended reading notes

Core claim

The central claim is that subspecialty pretraining, not just scale, is what makes a pathology foundation model effective. Digepath-V1 is first pretrained self-supervised on four magnification levels (2.5×, 5×, 10×, 20×) from 210,043 GI whole-slide images. Digepath-V2 then repeats training on a refined dataset: expert pathologists annotated 471,443 regions, a multiple-instance classifier learned from those annotations selects 1,305,328 tumor-rich regions from the same slide pool, and the encoder is fine-tuned on a balanced multi-scale patch set built from those regions. The authors argue this train-refine-repeat loop is what lets the model find sparsely distributed lesions in gigapixel slides, and they report state-of-the-art results on 33 of 34 downstream benchmarks, including 99.70% sensitivity in a nine-hospital early cancer screening validation. They also report that the second stage consistently improves on the first, by 3.81% average balanced accuracy on four diagnostic tasks and 4.18% on TNM staging.

Load-bearing premise

The entire comparison rests on treating Phase II, in which a classifier trained on GI tumor annotations curates the fine-tuning data, as a domain-general pretraining step rather than task-specific supervision that overlaps with the evaluation benchmarks.

Editorial extensions

If this is right

  • A single GI-specialized encoder can replace task-specific feature extractors across diagnosis, grading, segmentation, molecular prediction, and prognosis, removing the need to maintain multiple models in a clinical workflow.
  • The multi-scale pretraining makes predictions nearly resolution-invariant: STLC accuracy varies by only 0.91% across five input scales, so the same model can be used at the magnification a pathologist prefers.
  • Early GI cancer screening can run as an automated triage step with near-perfect sensitivity (99.70%) at acceptable specificity (89.30%), including detection of subtle lesions that are easy to miss.
  • Molecular biomarkers and prognosis can be read from routine H&E slides alone, with AUROC 83.27% for HER2, 88.41% for MSI, and a 71.82% concordance index for colon cancer survival, which could prioritize which cases need additional stains or sequencing.
  • The second-stage ROI refinement lifts performance consistently (average +3.81% balanced accuracy on four diagnostic tasks, +4.18% on TNM staging), so the loop itself, not just the dataset scale, is the source of the gain.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: an external-cohort benchmark, using slides from institutions that contributed none of the pretraining or Phase II data, would separate genuine transfer from in-distribution familiarity.
  • Beyond the paper: the same two-phase recipe could be re-run for other organ systems, but it requires organ-specific pathologist annotation at a scale of hundreds of thousands of regions; the transferable lesson may be the curation loop, not the specific weights.
  • Beyond the paper: the 89.30% specificity implies the screening module is best used as a triage prescreener that flags suspicious cases for full review, not as a standalone diagnosis; its net value would depend on the baseline miss rate of human screening in the deployment site.
  • Beyond the paper: reporting Phase II ablations across all 34 tasks, rather than a subset, would quantify how much of the overall SOTA margin comes from the ROI-mining stage versus the multi-scale pretraining.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper presents Digepath, a GI-pathology vision foundation model trained in two phases: first, DINOv2-based self-supervised pretraining on 353,478,334 multi-scale patches from 210,043 H&E-stained WSIs; second, a supervised ROI-mining phase in which a tumor classifier trained on 471,443 pathologist-annotated ROIs selects 1,305,328 tumor regions that are used to fine-tune the encoder, yielding Digepath-V2. The authors report state-of-the-art performance on 33 of 34 downstream GI tasks, including routine diagnosis, segmentation, molecular marker prediction, mutation prediction, and survival analysis, and describe a prospective nine-center early-cancer screening study achieving 99.70% sensitivity and 89.30% specificity. The manuscript also includes multi-resolution evaluations, few-shot and retrieval experiments, and a clinical agent pipeline.

Significance. If the evaluation is valid, Digepath would constitute a significant advance: a subspecialty-specific encoder with multi-scale pretraining that outperforms general pathology foundation models across a broad GI benchmark and a screening module with near-perfect sensitivity. The paper deserves credit for assembling a 34-task benchmark, incorporating public datasets, evaluating few-shot and retrieval properties, and attempting real-world multi-center deployment. However, the central comparison currently rests on two unresolved load-bearing assumptions: that the downstream test sets are disjoint from the pretraining corpus, and that comparing Digepath-V2, which received a supervised tumor-selection phase, with baselines that did not receive an equivalent phase is a fair foundation-model comparison. In addition, several dataset descriptions and training-data arithmetic are internally inconsistent, so the reported claims cannot yet be taken at face value.

major comments (5)
  1. [Methods: Pretraining dataset; Dataset of downstream tasks] The manuscript does not report any decontamination check between the 210,043-WSI GI pretraining corpus collected from ZN, LZ, and FJ and the downstream test sets, many of which come from the same hospital pool (ZN, LZ, FJ, SZT). Because Phase II mines ROIs from that same slide pool, the 33/34 SOTA claims and the 99.70% screening sensitivity could reflect memorization of previously seen tissue rather than transfer. Please provide slide-level and patient-level overlap analysis (for example, duplicate-image hashing, identifier exclusion, or slide-level train/test separation) and report performance separately on the subset that is provably disjoint from pretraining.
  2. [Methods: Dual-phase ROI-based optimization framework] Phase II is not a self-supervised pretraining step in the standard sense: a tumor classifier trained on 471,443 pathologist-annotated ROIs selects 1,305,328 tumor regions on which V2 is fine-tuned. The early-screening positive class in STMSCR, BOWSCR, and ESO-2cls (LIN, HIN, malignancy) is essentially the same semantic category as the tumor class used for curation. Since UNI, Gigapath, and Conch-V1.5 were not given an analogous curation phase, the comparisons in Figs. 2c and 3a are not like-for-like foundation-model comparisons. Please report Digepath-V1 as the primary comparison when claiming benefits of pretraining, or provide an equivalent supervised data-selection phase to the baselines.
  3. [Results: 1. Dual-phase pretraining] The stage-II arithmetic is internally inconsistent. With 2,048 × 2,048 ROIs and 256 × 256 patches, one ROI yields 64 patches, not 24; 1,305,328 × 24 = 31,327,872 is unexplained. In addition, if an equal number of non-tumor patches is added, the total should be 62,655,744, not the reported 83,206,828. Please provide the exact tiling and sampling rule and correct the totals, as this number is used to characterize the fine-tuning dataset.
  4. [Methods: Dataset of downstream tasks (CRC-100K, ESO-4cls, LHN-3cls)] Several dataset descriptions contradict each other and affect the reported task counts. Results §3.2 calls CRC-100K a 10-class task, while Methods lists nine tissue categories; Results §3.3 calls ESO-4cls a four-class task, while Methods states five classes but lists only four; Results §3.7 reports 384 slides for LHN-3cls, while Methods reports 462 WSIs. These inconsistencies must be reconciled so that the 34-task benchmark and each experimental protocol are unambiguous.
  5. [Results: 5.1 Early cancer screening; Methods: Prospective multi-center study] The 'nine independent medical institutions' claim is overstated. Three of the nine sites (ZN, LZ, FJ) contributed the pretraining data, and SZT is a source of several downstream datasets, so the 99.70% sensitivity is not fully external to model development. Please report sensitivity, specificity, and accuracy separately for the five centers that were not involved in pretraining or downstream development (CG, SX, NC, NB, LG) and phrase the abstract claim accordingly.
minor comments (5)
  1. [Data and code availability] The Methods state that all referenced code repositories are publicly accessible, but Code availability says the Digepath code and weights 'can be accessed later' and private images are available only on request; please clarify the exact release timeline and access conditions.
  2. [Figure 4d and §4.3] The mutation names 'TNN' and 'NARAS' in Figure 4d and the text are typos for TTN and NRAS; please correct them throughout.
  3. [Methods: Early esophageal cancer screening datasets] The ESO-4cls description says 'five classes' but lists only four classes; the text also contains 'an self-built' instead of 'a self-built' in several dataset descriptions.
  4. [Section 1 and Equation (9)] The phrase '31,327,872 million patches' contains an extra 'million', and Equation (9) uses 'Possion' instead of Poisson; please fix these typographical errors.
  5. [General] The model name is rendered inconsistently as 'Digepath' and 'DigPath' across the abstract, main text, and figures; please unify the naming convention.

Circularity Check

1 steps flagged · score 6.0 of 10

Phase II supervised tumor curation is reused as the early-screening positive label, so the headline 99.70% sensitivity reduces in part to the same tumor/non-tumor signal used to fine-tune the encoder.

  1. fitted input called prediction [Methods, 'Dual-phase ROI-based optimization framework' and 'Dynamic ROI selection strategy' (Eqs. 7–9); Methods, 'Early gastric cancer screening'; Results §5.1]
    "The annotated ROI was then used to train a tumor classifier, which processed the original dataset to identify 1,305,328 tumor regions, subdivided into 31,327,872 patches with size of 256 × 256. An equal number of non-tumor patches were randomly sampled to create a multi-scale dataset including 83,206,828 patches for fine-tuning, yielding the enhanced Digepath-V2 model."

    Phase II is presented as continued pretraining, but it is supervised: a binary tumor classifier is fitted to 471,443 pathologist-annotated ROIs, and its predicted tumor confidence is used to select the ROIs for fine-tuning (N1 = ceil(12 × ptumor/0.7)). The early-cancer screening evaluation then defines positive cases as 'LIN, HIN, or confirmed malignant tumors'—the same neoplastic/tumor distinction the Phase II classifier was trained to detect, and from the same hospital slide pool used for many downstream tests. The 99.70% screening sensitivity therefore largely measures a model whose encoder was explicitly tuned to separate tumor from non-tumor on overlapping data, rather than testing self-supervised transfer.

full rationale

Digepath's Phase I is a standard DINOv2 self-supervised pretraining step and is not circular by itself. The circularity enters in Phase II, which is described as continued pretraining but is in fact a supervised tumor-detection loop: a tumor classifier is fit to 471,443 pathologist-annotated ROIs, its confidence scores select the tumor regions used to fine-tune the encoder, and the downstream early-cancer screening module then defines positivity as LIN, HIN, or malignancy—the same lesion classes the Phase II classifier was built to detect. The reported 99.70% sensitivity is therefore partly a re-measurement of the Phase II curation signal, and comparisons to UNI, Gigapath, and Conch-V1.5 are asymmetric because those baselines did not receive an equivalent supervised Phase II. This makes the headline screening 'prediction' partially circular by construction. It does not make the whole paper circular: external benchmarks (CRC-100K, UNITOPATHO, CAMEL, CRAG, and TCGA survival/mutation tasks) are independent of the Phase II curation pool and show only modest gains, so those results retain independent content. No load-bearing self-citation chain or imported uniqueness theorem is present. A separate risk of slide overlap between pretraining and in-house test sets is a contamination concern, not a circularity, and is not scored here. Score 6 reflects partial circularity confined to the Phase-II-dependent screening and in-house SOTA claims.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

Digepath, Digepath-V1/V2, and DigeTools are software artifacts, not newly postulated physical entities. No new forces, particles, dimensions, or conserved quantities are introduced. The ledger instead records hand-set data-sampling hyperparameters and background modeling assumptions.

free parameters (4)
  • ROI sampling constant in N1 (12/0.7) = 12, 0.7
    Equation (7) N1 = ceil(12 x ptumor / 0.7) sets the number of high-confidence tumor ROIs per positive WSI. The constants are chosen by the authors to build a balanced 1:1 tumor:non-tumor dataset and are not derived from external constraints.
  • ROI sampling constant in N2 (4) = 4
    Equation (8) N2 = ceil(4 x (1 - ptumor)) sets the number of non-tumor ROIs sampled from tumor-positive slides; the arbitrary constant affects the stage-II training distribution.
  • ROI sampling constant in N3 (Poisson lambda=8) = lambda=8
    Equation (9) N3 = Poisson(lambda=8) sets the number of ROIs sampled from tumor-negative WSIs; arbitrary and affects stage-II data composition.
  • ROI attention threshold tau = not specified
    The DigeTools WSI Process extracts ROIs with attention scores above threshold tau; the value is not given and determines screening module outputs.
assumptions (5)
  • domain assumption DINOv2 self-supervised pretraining on large unlabeled patch corpora yields transferable representations for downstream medical imaging tasks.
    The entire two-stage design assumes that the SSL representations learned in Phase I transfer to the 24 diagnostic and 8 molecular or prognostic tasks; no proof is given, only empirical benchmarks.
  • ad hoc to paper Multi-magnification pretraining at 2.5x, 5x, 10x, and 20x captures clinically useful domain features that single-magnification pretraining misses.
    This is the paper's key architectural premise, motivated by clinical practice but not independently validated outside the reported benchmarks.
  • ad hoc to paper A tumor/non-tumor classifier trained on 471,443 pathologist ROIs identifies diagnostically valuable regions reliably enough to improve pretraining.
    The phase-II closed loop depends on this; if the classifier propagates annotation noise, V2's gains are not attributable to ROI mining.
  • domain assumption H&E morphology contains sufficient signal to predict PD-L1, HER2, P53, MSI, MUC16/TTN/NRAS mutations, and survival.
    Molecular profiling and prognosis sections rely on this assumption; performances are often modest (AUROC 60 to 88 percent), so the assumption is only weakly supported.
  • domain assumption The WHO-based labels of LIN, HIN, malignancy, and benign in the nine-center screening study are correct and consistently applied across institutions.
    The 99.70 percent sensitivity claim is measured against this ground truth; the text itself describes consensus reclassification of some cases, so label stability is a live concern.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Subspecialty-Specific Foundation Model for Intelligent Gastrointestinal Pathology." pith.science (2026). https://pith.science/paper/4RVUUOX2

@misc{pith2026250521928,
  author       = {Pith},
  title        = {Pith review of: Subspecialty-Specific Foundation Model for Intelligent Gastrointestinal Pathology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4RVUUOX2}},
  note         = {Machine review of arXiv:2505.21928}
}
read the original abstract

Gastrointestinal (GI) diseases represent a clinically significant burden, necessitating precise diagnostic approaches to optimize patient outcomes. Conventional histopathological diagnosis suffers from limited reproducibility and diagnostic variability. To overcome these limitations, we develop Digepath, a specialized foundation model for GI pathology. Our framework introduces a dual-phase iterative optimization strategy combining pretraining with fine-screening, specifically designed to address the detection of sparsely distributed lesion areas in whole-slide images. Digepath is pretrained on over 353 million multi-scale images from 210,043 H&E-stained slides of GI diseases. It attains state-of-the-art performance on 33 out of 34 tasks related to GI pathology, including pathological diagnosis, protein expression status prediction, gene mutation prediction, and prognosis evaluation. We further translate the intelligent screening module for early GI cancer and achieve near-perfect 99.70% sensitivity across nine independent medical institutions. This work not only advances AI-driven precision pathology for GI diseases but also bridge critical gaps in histopathological practice.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 45 canonical work pages

  1. [1]

    Global burden of five major types of gastrointestinalcancer[J].PrzGastroenterol.2024;19(3):236–254

    Singh A, Ferlay J, Siegel RL, et al. Global burden of five major types of gastrointestinalcancer[J].PrzGastroenterol.2024;19(3):236–254

    work page 2024

  2. [2]

    Concomitant radiosurgical and targeted oncological treatment improves the outcome of patients with brain metastases from gastrointestinalcancer.RadiatOncol.2023Dec9;18(1):197

    Khalaveh F, Cho A, Shaltout A, et al. Concomitant radiosurgical and targeted oncological treatment improves the outcome of patients with brain metastases from gastrointestinalcancer.RadiatOncol.2023Dec9;18(1):197. [3]SungH,SiegelRL,TorreLA,et al.Globalpatternsinexcessbodyweightandthe associatedcancerburden[J].CACancerJClin.2019;69(2):88–112. [4]BrayF,Lave...

    work page 2019

  3. [3]

    Multiscale pretraining enables robust representation

    Routine clinicopathological diagnosis Aligned with the WHO Classification of Digestive System Tumours, this study established a comprehensive validation framework spanning three anatomical regions (esophageal, gastric, and intestinal) across 24 clinical tasks, incorporating classification and segmentation at both ROI and WSI-level. Digepath demonstrated s...

    work page

  4. [4]

    Molecular profiling and prognostic prediction outcomes Tumor molecular profiling guides therapeutic decision-making and is indispensable for precision oncology. We present a comprehensive evaluation of Digepath’s ability toinfermolecularprofilingdirectlyfromhistopathologyimages,whichareprediction of therapeutic protein target expression, determination of ...

    work page

  5. [5]

    Clinical translation and implementation 5.1 Early cancer screening Pathologists face the critical yet challenging task of detecting rare early-stage malignancies among vast numbers of gastrointestinal biopsy specimens, which is a time-intensive process. To transform this paradigm, we implemented an AI-powered early gastric cancer screening module based on...

    work page

  6. [6]

    In few-shot learning, the choice of 'way' has a significant impact on task difficulty and model performance

    Few- shot learning WeassessedDigepath’slabelefficiencyonthedatasetsofSTLCandUNITOPATHO using the non-parametric SimpleShot framework—a robust baseline widely adopted in the few-shot classification literature[52]. In few-shot learning, the choice of 'way' has a significant impact on task difficulty and model performance. Typically, increasing the number of...

    work page

  7. [7]

    The ProtoNet first convert all training images into embedding vectors, then performs mean-poolingonembeddingsofthesamecategorytoobtainprototyperepresentations

    ROI retrieval ROI retrieval is implemented through Prototypical Network[53] (ProtoNet). The ProtoNet first convert all training images into embedding vectors, then performs mean-poolingonembeddingsofthesamecategorytoobtainprototyperepresentations. We benchmarked histopathological image retrieval across two ROI-level tasks, with eachtestsampleclassifiedbas...

    work page

  8. [8]

    Global Cancer Observatory: Cancer Today (Version1.0).InternationalAgencyforResearchonCancer; 2024.AccessedFebruary 1,2024

    Ferlay J, Ervik M, Lam F, et al. Global Cancer Observatory: Cancer Today (Version1.0).InternationalAgencyforResearchonCancer; 2024.AccessedFebruary 1,2024

    work page 2024

Show all 52 references
  1. [9]

    Practical Considerations in Diagnosing and Managing Early-OnsetGICancers[J].JClinOncol.2022Aug20;40(24):2662–2680

    Lumish MA, Cercek A. Practical Considerations in Diagnosing and Managing Early-OnsetGICancers[J].JClinOncol.2022Aug20;40(24):2662–2680

  2. [10]

    Diagnosis to dissection: AI's role in earlydetection andsurgical intervention for gastric cancer[J].JRobot Surg

    Pattilachan TM, Christodoulou M, Ross S. Diagnosis to dissection: AI's role in earlydetection andsurgical intervention for gastric cancer[J].JRobot Surg. 2024 Jun 20;18(1):259

    2024

  3. [11]

    Levin B, Lieberman D A, McFarland B, et al. Screening and surveillance for the early detection of colorectal cancer and adenomatous polyps, 2008: a joint guideline from the American Cancer Society, the US Multi-Society Task Force on Colorectal Cancer, and the American College ...

    2008

  4. [12]

    Wang Z, Liu Y, Niu X. Application of artificial intelligence for improving early detection and prediction of therapeutic outcomes for gastric cancer in the era of precision oncology[C]//Seminars in Cancer Biology. Academic Press, 2023, 93: 83–96

    2023

  5. [13]

    Opening the doors of precision medicine: novel tools to assess intestinal barrier in inflammatory bowel disease and colitis-associatedneoplasia[J].Gut,2024,73(10):1749–1762

    Iacucci M, Santacroce G, Majumder S, et al. Opening the doors of precision medicine: novel tools to assess intestinal barrier in inflammatory bowel disease and colitis-associatedneoplasia[J].Gut,2024,73(10):1749–1762

    2024

  6. [14]

    Artificial intelligence in digital pathology: a systematic review and meta-analysis of diagnostic test accuracy[J]

    McGenity C, Clarke E L, Jennings C, et al. Artificial intelligence in digital pathology: a systematic review and meta-analysis of diagnostic test accuracy[J]. npj DigitalMedicine,2024,7(1):114. [12]PasechnikovV,ChukovS,FedorovE,etal.Gastriccancer:prevention,screening andearlyd...

    2024

  7. [15]

    Sequential injection-electrocoagulation vs

    Ma ZY, Yang Z, Liu J, et al. Sequential injection-electrocoagulation vs. traditional electrocoagulation haemostasis during endoscopic submucosal dissection: arandomizedcontrolledtrial.SurgEndosc.2025May16

  8. [16]

    Updated evaluation of endoscopic submucosal dissection versus surgery for early gastric cancer: A systematic review and meta-analysis[J].InternationalJournalofSurgery,2020,73:28–41

    Liu Q, Ding L, Qiu X, et al. Updated evaluation of endoscopic submucosal dissection versus surgery for early gastric cancer: A systematic review and meta-analysis[J].InternationalJournalofSurgery,2020,73:28–41

    2020

  9. [17]

    Incidence of metachronous cancer after endoscopic submucosal dissection: a comparison between undifferentiated-type and differentiated-type early gastric cancer[J]

    Ishioka M,YoshioT, MiyamotoY,et al. Incidence of metachronous cancer after endoscopic submucosal dissection: a comparison between undifferentiated-type and differentiated-type early gastric cancer[J]. Gastrointestinal Endoscopy, 2021, 93(3): 557–564.e1

    2021

  10. [18]

    Endoscopic submucosal dissection forearlygastriccancer:alarge-scalefeasibilitystudy[J].Gut,2009,58(3):331–336

    Isomoto H, Shikuwa S, Yamaguchi N, et al. Endoscopic submucosal dissection forearlygastriccancer:alarge-scalefeasibilitystudy[J].Gut,2009,58(3):331–336

    2009

  11. [19]

    Surgical management of gastric cancer: a review[J].JAMAsurgery,2022,157(5):446–454

    Li G Z, Doherty G M, Wang J. Surgical management of gastric cancer: a review[J].JAMAsurgery,2022,157(5):446–454

    2022

  12. [20]

    Focus on gastric cancer[J]

    Ushijima T, Sasako M. Focus on gastric cancer[J]. Cancer cell, 2004, 5(2): 121–125

    2004

  13. [21]

    Staging and surgical approaches in gastric cancer:Asystematicreview[J].Cancertreatmentreviews,2018,63:104–115

    Coburn N, Cosby R, Klein L, et al. Staging and surgical approaches in gastric cancer:Asystematicreview[J].Cancertreatmentreviews,2018,63:104–115. [22]BurzC,PopV,SilaghiC,etal.Prognosisandtreatmentofgastriccancer: a2024 update[J].Cancers,2024,16(9):1708

    2018

  14. [23]

    Gastric cancer treatment: recent progress and future perspectives[J].Journalofhematology&oncology,2023,16(1):57

    Guan W L, He Y, Xu R H. Gastric cancer treatment: recent progress and future perspectives[J].Journalofhematology&oncology,2023,16(1):57. [24]AcsB, Rantalainen M,Hartman J.Artificial intelligence as thenext steptowards precisionpathology[J].Journalofinternalmedicine,2020,288(1):62–81

    2023

  15. [25]

    Hallmarks of artificial intelligence contributionstoprecisiononcology[J].NatureCancer,2025:1–15

    Chang T G, Park S, Schäffer A A, et al. Hallmarks of artificial intelligence contributionstoprecisiononcology[J].NatureCancer,2025:1–15

    2025

  16. [26]

    A comprehensive assessment of artificial intelligence applications for cancer diagnosis[J].Artificial Intelligence Review, 2024, 57(7):179

    Singh G, Kamalja A, Patil R, et al. A comprehensive assessment of artificial intelligence applications for cancer diagnosis[J].Artificial Intelligence Review, 2024, 57(7):179

    2024

  17. [27]

    A pathologist–AI collaboration framework for enhancing diagnostic accuracies and efficiencies[J]

    Huang Z, Yang E, Shen J, et al. A pathologist–AI collaboration framework for enhancing diagnostic accuracies and efficiencies[J]. Nature Biomedical Engineering, 2025,9(4):455–470

    2025

  18. [28]

    AI in digital pathology: automated histopathological analysis for cancergradingandprognosticoutcomeprediction[J].IntJComputApplTechnolRes, 2022,11(11):400–12

    Kumar A. AI in digital pathology: automated histopathological analysis for cancergradingandprognosticoutcomeprediction[J].IntJComputApplTechnolRes, 2022,11(11):400–12. [29]VanderLaakJ,LitjensG,CiompiF.Deeplearninginhistopathology: thepathto theclinic[J].Naturemedicine,2021,27(...

    2022

  19. [30]

    DosovitskiyA, Beyer L, KolesnikovA, et al.An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale[C]//International Conference on LearningRepresentations.2020

    2020

  20. [31]

    Foundation Models Defining a New Era in Vision:ASurveyandOutlook[J].IEEETransactionsonPatternAnalysisandMachine Intelligence,2025

    Awais M, Naseer M, Khan S, et al. Foundation Models Defining a New Era in Vision:ASurveyandOutlook[J].IEEETransactionsonPatternAnalysisandMachine Intelligence,2025

    2025

  21. [32]

    An empirical study of training self-supervised vision transformers.In:ProceedingsIEEE/CVFIntConfComputVis.2021:9640–9649

    Chen X, Xie S, He K. An empirical study of training self-supervised vision transformers.In:ProceedingsIEEE/CVFIntConfComputVis.2021:9640–9649

    2021

  22. [33]

    Foundation Model for Predicting Prognosis and Adjuvant Therapy Benefit From Digital Pathology in GI Cancers[J]

    Wang X, JiangY,Yang S, et al. Foundation Model for Predicting Prognosis and Adjuvant Therapy Benefit From Digital Pathology in GI Cancers[J]. Journal of ClinicalOncology,2025:JCO-24-01501

    2025

  23. [34]

    CellViT++: Energy-Efficient and Adaptive Cell Segmentation and Classification Using Foundation Models[J]

    Hörst F, Rempe M, Becker H, et al. CellViT++: Energy-Efficient and Adaptive Cell Segmentation and Classification Using Foundation Models[J]. arXiv preprint arXiv:2501.05269,2025

    2025 arXiv

  24. [35]

    A pathology foundation model for cancer diagnosisandprognosisprediction[J].Nature,2024,634(8035):970–978

    Wang X, Zhao J, Marostica E, et al. A pathology foundation model for cancer diagnosisandprognosisprediction[J].Nature,2024,634(8035):970–978

    2024

  25. [36]

    ChenRJ,DingT,Lu MY,et al.Towardsageneral-purpose foundationmodel for computationalpathology[J].NatMed.2024;30:850–862

    2024

  26. [37]

    A whole-slide foundation model for digital pathologyfromreal-worlddata[J].Nature.2024;630:181–188

    Xu H, Usuyama N, Bagga J, et al. A whole-slide foundation model for digital pathologyfromreal-worlddata[J].Nature.2024;630:181–188

    2024

  27. [38]

    Multimodal whole slide foundation model forpathology[J].arXivpreprintarXiv.2024:2411.19666

    Ding T, Wagner SJ, Song AH, et al. Multimodal whole slide foundation model forpathology[J].arXivpreprintarXiv.2024:2411.19666

    2024 arXiv

  28. [39]

    MSCL-Net: Unleashing thepower of multi-scale and cross-layer learning in pathology image classification[J]

    Sun K,ZhengY,Yang X, et al. MSCL-Net: Unleashing thepower of multi-scale and cross-layer learning in pathology image classification[J]. Biomedical Signal ProcessingandControl,2025,108:107880. [40]WuX,ChenZ,PengC,etal.MMSRNet:Pathologicalimagesuper-resolutionby multi-task and m...

    2025

  29. [41]

    Multi-Scale Dynamic Sparse Token Multi-Instance Learning for Pathology Image Classification[J]

    Lei D, Zhang Y, Wang H, et al. Multi-Scale Dynamic Sparse Token Multi-Instance Learning for Pathology Image Classification[J]. IEEE Journal of BiomedicalandHealthInformatics,2024

    2024

  30. [42]

    Clinically applicable histopathological diagnosis system for gastric cancer detection using deep learning[J]

    Song Z, Zou S, Zhou W, et al. Clinically applicable histopathological diagnosis system for gastric cancer detection using deep learning[J]. Nat Commun. 2020;11(1):4294

    2020

  31. [43]

    Reinforcement Learning Finetunes Small Subnetworks in Large Language Models[J]

    Mukherjee S, Yuan L, Hakkani-Tur D, et al. Reinforcement Learning Finetunes Small Subnetworks in Large Language Models[J]. arXiv preprint arXiv:2505.11711, 2025

    2025

  32. [44]

    EMORL: Ensemble Multi-Objective Reinforcement Learning for Efficient and Flexible LLM Fine-Tuning[J]

    Kong L, Yang C, Neufang S, et al. EMORL: Ensemble Multi-Objective Reinforcement Learning for Efficient and Flexible LLM Fine-Tuning[J]. arXiv preprintarXiv:2505.02579,2025

    2025 arXiv

  33. [45]

    Transfer learning with adaptive fine-tuning[J]

    Vrbančič G, Podgorelec V. Transfer learning with adaptive fine-tuning[J]. IEEE Access,2020,8:196197–196211

    2020

  34. [46]

    Fine tuning deep learning models for breasttumorclassification[J].ScientificReports,2024,14(1):10753

    HeikalA, El-GhamryA, Elmougy S, et al. Fine tuning deep learning models for breasttumorclassification[J].ScientificReports,2024,14(1):10753. [47]OquabM,DarcetT,MoutakanniT,etal.Dinov2:Learningrobustvisualfeatures withoutsupervision[J].arXivpreprintarXiv:2304.07193,2023

    2024 arXiv

  35. [48]

    Attention-based deep multiple instance learning[C]//Internationalconferenceonmachinelearning.PMLR,2018:2127–2136

    Ilse M, Tomczak J, Welling M. Attention-based deep multiple instance learning[C]//Internationalconferenceonmachinelearning.PMLR,2018:2127–2136

    2018

  36. [49]

    TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers[J]

    Chen J, Mei J, Li X, et al. TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers[J]. Medical Image Analysis,2024,97:103280

    2024

  37. [50]

    Clinical and pathological staging of gastric cancer: Current perspectives and implications[J]

    Zhu Z, Gong Y, Xu H. Clinical and pathological staging of gastric cancer: Current perspectives and implications[J]. European Journal of Surgical Oncology, 2020,46(10):e14–e19. [51]WeiserMR.AJCC8thedition:colorectalcancer[J].Annalsofsurgicaloncology, 2018,25:1454–1455

    2020

  38. [52]

    Simpleshot: Revisiting nearest-neighbor classification for few-shot learning[J]

    Wang Y, Chao W L, Weinberger K Q, et al. Simpleshot: Revisiting nearest-neighbor classification for few-shot learning[J]. arXiv preprint arXiv:1911.04623,2019

    1911 arXiv

  39. [53]

    tumor" ROI, we selected the top N₁ ROIs withthehighestclassificationconfidenceforthe

    Snell J, Swersky K, Zemel R. Prototypical networks for few-shot learning[J]. Advancesinneuralinformationprocessingsystems,2017,30. Methods Inrecentyears,foundationmodelshavedemonstratedremarkabletransfercapabilities innatural image analysis.ArchitecturessuchasVisionTransformer...

    2017

  40. [54]

    An empirical study of training self-supervised vision transformers[C]//Proceedings of the IEEE/CVF international conference on computer vision.2021:9640–9649

    Chen X, Xie S, He K. An empirical study of training self-supervised vision transformers[C]//Proceedings of the IEEE/CVF international conference on computer vision.2021:9640–9649

    2021

  41. [55]

    Data-efficient and weakly supervised computational pathology on whole-slide images[J]

    Lu M Y, Williamson D F K, Chen T Y, et al. Data-efficient and weakly supervised computational pathology on whole-slide images[J]. Nature biomedical engineering,2021,5(6):555–70

    2021

  42. [56]

    Transmil: Transformer based correlated multiple instance learning for whole slide image classification[J]

    Shao Z, Bian H, Chen Y, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification[J]. Advances in neural informationprocessingsystems,2021,34:2136–2147

    2021

  43. [57]

    Li B, Li Y, Eliceiri K W. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning[C]//ProceedingsoftheIEEE/CVFconferenceoncomputervisionandpattern recognition.2021:14318–14328. [58]Li J,ChenY,ChuH,et al. Dyn...

    2021

  44. [60]

    Coca: Contrastive captioners are image-text foundationmodels[J].arXivpreprintarXiv:2205.01917,2022

    Yu J, Wang Z, Vasudevan V, et al. Coca: Contrastive captioners are image-text foundationmodels[J].arXivpreprintarXiv:2205.01917,2022. Extended Data Figure 1 | ROI visualizations across models. a–c. Visualizationsof fivemodelsonSTLC. d–g.VisualizationsoffivemodelsonUNITOPATHO. ...

    2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.