Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

DepViT-CAD: Deployable Vision Transformer-Based Cancer Diagnosis in Histopathology

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read DepViT-CAD claims a multi-attention vision transformer trained on 1008 whole-slide images can diagnose 10 cancer types plus non-tumor tissue, reaching 94.11% sensitivity on TCGA slides and 92% on routine clinical cases.

desk verdict Useful empirical CAD paper undermined by internal tile-count contradictions that break the organ-shortcut argument; conditional accept after major revision. read the letter →

arxiv 2507.10250 v1 pith:GABHWN4Z submitted 2025-07-14 eess.IV cs.AIcs.CVcs.LG

Ashkan Shakarami , Lorenzo Nicole , Rocco Cappellesso , Angelo Paolo Dei Tos , Stefano Ghidoni This is my paper
classification eess.IVcs.AIcs.CVcs.LG
keywords cancerdiagnosishistopathologywhole-slideimagesvisiontransformermulti-classclassificationclinicaldeploymentgliomasubtypesdigitalpathology
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

DepViT-CAD is an AI diagnostic system that aims to classify a whole-slide histopathology image into one of 11 categories: 10 cancer types, including fine-grained subtypes, plus non-tumor tissue. The paper claims that its MAViT architecture, trained on 1008 slides, reaches diagnostic sensitivity of 94.11% on 275 held-out TCGA slides and 92% on 50 routine biopsy and surgical cases. These numbers are presented as evidence that the model learns tumor-specific morphology rather than organ identity, which would make it a deployable second opinion for pathologists facing hard subtype distinctions such as glioblastoma versus astrocytoma and lung adenocarcinoma versus squamous cell carcinoma.

What carries the argument

The load-bearing component is MAViT, a hybrid vision transformer built on EfficientNet-B3, whose Vision Transformer Module applies linear self-attention to a 32x32 feature map and whose Dual Fusion Strategy aligns shallow, intermediate, and deep features by resizing before merging them. A D-Graph visualization sorts predictions by biopsy versus surgical modality and by correctness, while final slide-level diagnosis comes from majority voting over patch-level predictions. This combination is what lets the model integrate local textures with global context while keeping inference fast enough for deployment.

What would settle it

A concrete test would be to build a held-out set in which each organ contributes both tumor and non-tumor slides, then measure per-class accuracy when tissue of origin is fully removed from the training data; if accuracy for held-out organs drops to chance, the tumor-specificity reading is wrong. A multi-site external set with 500 or more slides that brings sensitivity well below the reported 92% would similarly falsify the deployment claim.

Watch

Extended reading notes

Core claim

DepViT-CAD is an end-to-end computer-aided diagnosis framework that classifies H&E-stained whole-slide images into 11 categories using a patch-based vision transformer. The central discovery is that MAViT, an EfficientNet-B3 backbone augmented with a Multi-Attention Vision Transformer module and a Dual Fusion Strategy, reaches 98.12% tile-level accuracy, 94.11% patient-level sensitivity, and 94.37% F1-score on the TCGA test set, and 92% sensitivity on a real-world clinical cohort containing both biopsy and surgical specimens. The authors interpret the high accuracy as pathology-relevant rather than organ-identity classification, pointing to the model's ability to label tumor and non-tumor regions within a single slide and to the training choice of 100 randomly sampled patches per patient. Misclassifications are mostly confined to histologically similar glioma subtypes, which the paper reads as evidence that the model is tackling genuine diagnostic difficulty rather than exploiting organ cues.

Load-bearing premise

The load-bearing premise is that 100 randomly sampled 512x512 patches per slide, with a mix of tumor and non-tumor tiles, are enough to prevent the model from latching onto organ-level shortcuts that are present in TCGA.

Editorial extensions

If this is right

  • If the central claim is correct, a pathologist can submit an unannotated whole slide or a targeted ROI and receive a subtype-level diagnostic vote without manual annotation.
  • The system can label tumor and non-tumor regions in the same slide, meaning it localizes disease rather than simply naming the organ of origin.
  • The hard clinical distinctions, such as glioma subtypes and lung adenocarcinoma versus squamous cell carcinoma, are handled with errors mostly confined to morphologically overlapping gliomas.
  • The modular design allows deployment to swap in lighter CNNs when speed matters or larger transformers when accuracy is the priority.
  • The authors identify soft-label aggregation as a future improvement that could reduce ambiguity when two classes receive similar confidence scores.

Reading between the lines

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

  • A direct test of the tumor-specificity claim would be to train on only one patch per slide or to scramble patch order; if accuracy collapses on held-out organs, the 100-patch sampling may still carry organ cues.
  • Because the external cohort is 50 cases from a single institution, the 92% figure is only a lower-bound estimate of generalizability; multi-site external validation with hundreds of slides would tighten it.
  • The confusion between glioblastoma and oligodendroglioma could be probed with attention maps to see whether the model relies on the same necrosis and microvascular-proliferation features that pathologists use.
  • An extension that distinguishes tumor grade or molecular subtype within each histotype would test whether the learned features are truly pathology-relevant or merely coarse tissue stereotypes.
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

4 major / 6 minor

Summary. The paper presents DepViT-CAD, an end-to-end pipeline for multi-class cancer diagnosis from H&E whole-slide images, centered on a new architecture, MAViT, that combines an EfficientNet-B3 backbone with a Vision Transformer Module (VTM) and a Dual Fusion Strategy (DFS). The model is trained on a TCGA-derived dataset spanning 11 diagnostic categories and is evaluated at tile and patient levels on a held-out TCGA test set and on 50 routine clinical cases, reporting diagnostic sensitivities of 94.11% and 92%, respectively. The manuscript also includes ablations, comparisons with retrained baselines, and a visualization tool, D-Graph.

Significance. If the reported results are reproducible, this is a useful deployable multi-class cancer diagnosis system. The evaluation has clear strengths: the test split is patient-level, the baselines are retrained under the same conditions, an external routine clinical cohort is included, and the reporting follows TRIPOD. However, the central interpretative claim—that the model learns tumor-specific morphology rather than organ-identity shortcuts—is currently under-supported because of a direct contradiction between the stated 100-tiles-per-patient sampling and the reported tile counts. This is fixable with additional reporting and a controlled experiment, and it does not invalidate the engineering contribution, but it must be resolved before the generalization claim can be accepted.

major comments (4)
  1. [§3.1.1, Table 1, §5] The manuscript contradicts itself about the number of training tiles. Section 3.1.1 states that 'a representative subset of 100 tiles per patient was randomly sampled,' and Section 5 repeats that training uses 100 randomly selected patches per WSI. With 869 training patients (Table 1), this gives 86,900 training tiles, or at most 128,300 if the count is per WSI for all 1,283 slides. The paper instead reports 358,589 training tiles in Section 5 and 512,271 total tiles in Table 1. The reported per-tile training time of 0.0365 seconds is consistent with the 358,589-tile count, not with the 100-tile-per-patient claim. Since the 100-tile sampling assumption is the stated basis for the claim that the model learns tumor-specific features rather than organ-level architecture, this discrepancy must be resolved by reporting the actual sampling procedure and tile counts, and by providing a controlled experiment (for example, an organ-label probe or a test set that breaks the organ-tumor correlation) that supports the tumor-specificity interpretation.
  2. [Abstract, §4.1.1, Table 1] The abstract says MAViT was trained on expert-annotated patches from 1008 WSIs, whereas Section 4.1.1 and Table 1 report 1283 WSIs, and Table 1 sums to 1283. This is a consequential reporting inconsistency because the dataset size is part of the central claim. Please correct the number throughout and clarify whether 1008 refers to a filtered subset or is simply an error.
  3. [§5, Figures 10 and 12] The claim that the model relies on tumor-specific rather than organ-specific features is currently supported only by qualitative visual examples and a 50-case, single-center external set. Given the unresolved tile-count discrepancy, the evidence is insufficient to rule out an organ-identity shortcut. Please add a quantitative test, such as evaluating on a cohort with mixed organ chips or reporting per-class sensitivity stratified by organ, and also report confidence intervals for the external 92% sensitivity.
  4. [§3.2.2] The VTM subsection describes a 'Multi-Head Linear Attention mechanism inspired by the Linformer architecture' but gives the standard softmax attention formula, Omat = Softmax(QK^T / sqrt(dk)) V. Linformer achieves linear complexity by projecting the sequence dimension of K and V, which is not shown and is not implemented by the displayed equation. Please either correct the description or provide the actual linear-attention equations, since this is part of the claimed architectural novelty.
minor comments (6)
  1. [§3.1.3] The text refers to 'HistoCancer-CAD' in several places, while the paper title and rest of the manuscript use 'DepViT-CAD'; please unify the name.
  2. [Table 5, §4.2] Table 5 reports metrics that match the tile-level numbers in Section 4.2 (Sensitivity 88.60%, F1 89.54%), but the surrounding text discusses diagnostic accuracy in patient-level terms; please state explicitly in the caption whether these are tile-level or patient-level results.
  3. [§5] The number 358,589 training tiles should be reconciled with Table 1's 'Total number of tiles: 512271' and with the 100-tiles-per-patient description; if augmentation is included, please say so explicitly.
  4. [§3.3.3, Eqs. (3.9) and (3.10)] The D-Graph notation reuses the symbol \y_i for the vertical coordinate in Eq. (3.10) after defining it as the predicted class label in Eq. (3.9); please use a different symbol for the vertical coordinate.
  5. [Data Availability] The Data Availability section states that code and the WSI list 'will be' made available; for a reproducibility claim, please provide a current repository link or a statement of availability during review.
  6. [Table 2] The caption of Table 2 is ambiguous about what 'S1' through 'S5' denote and how the colored cells map to correct biopsy versus correct surgical cases; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported sensitivities are genuine out-of-sample evaluations, not re-statements of training inputs.

full rationale

The reported diagnostic sensitivities (94.11% on 275 held-out TCGA patients and 92% on 50 CliniR-ds clinical cases) are obtained from a model trained on a patient-level training split and evaluated on disjoint test and external cohorts, so they are genuine out-of-sample results rather than reconstructions of training inputs. MAViT's components (EfficientNet-B3 backbone, VTM with Linformer-style linear attention, and DFS multi-scale fusion) are defined and ablated in Section 4.7, and WSI-level diagnosis is produced by explicit majority voting over patch predictions via Equations 1-3; none of these definitions is expressed in terms of the final diagnostic labels. The self-citations (SlideTiler [33], TCNN [15], and a prior CAD abstract [16]) are used as tooling or literature context, not as evidence for the central performance claim, so no load-bearing premise is justified only by a same-author citation. The internal inconsistency between the stated 100-random-patches-per-patient sampling (Section 3.1.1) and the reported tile counts (Table 1: 512,271 total tiles; Section 5: 358,589 training tiles) is a data-fidelity concern that weakens the organ-vs-tumor feature interpretation, but it is not a circular derivation: the classifier still optimizes cross-entropy on labeled patches and is evaluated on unseen patients. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is imported from self-citations. Accordingly, no circular step is identified.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central performance claim rests on the correctness of the TCGA/clinical labels, on the representative power of 100-tile sampling, and on the transferability of a TCGA-trained model to routine clinical slides. The free parameters are standard DL hyperparameters, none of which are derived from first principles or justified with sensitivity analyses.

free parameters (5)
  • Tiles per patient (training sampling) = 100
    Section 3.1.1: 'a representative subset of 100 tiles per patient was randomly sampled'; this hand-set number defines the training distribution and the claimed mitigation of organ-shortcut learning.
  • Patch size = 512x512x3
    Section 3.1.1: tile size used for both training and inference; no resolution ablation is reported.
  • Learning rate and batch size = 1e-4, batch 8
    Section 4.1.3: SGD learning rate 1e-4 and batch size 8 (GPU-limited); no sensitivity analysis, though these affect the trained model's performance.
  • Backbone choice = EfficientNet-B3, ImageNet pretrained
    Section 3.2.1: selected for balance of accuracy and efficiency; the ablation only tests one backbone, so the central architecture depends on this choice.
  • VTM and DFS hyperparameters = Not fully specified
    Attention dimensions, number of heads, projection rank, and fusion alignment parameters are not reported in detail, so a reimplementation cannot fix them without author contact.
assumptions (3)
  • domain assumption TCGA diagnostic labels (molecular/clinical) are accurate ground truth for the 11 classes.
    The model is trained and evaluated against these labels; any label noise transfers directly into reported sensitivity/specificity. Used throughout Section 4.
  • domain assumption A random sample of 100 patches per patient preserves diagnostic signal while removing global organ context.
    Section 3.1.1 and Section 5 invoke this to claim the model learns tumor-specific features; it is not validated by an organ-holdout experiment.
  • domain assumption H&E patches from TCGA and the single-center CliniR-ds are representative of routine clinical histopathology, including biopsy specimens.
    Section 4.1 claims external validity from 50 cases at one hospital; the training set has limited biopsy representation (acknowledged in Section 6), so the transferability premise is load-bearing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DepViT-CAD: Deployable Vision Transformer-Based Cancer Diagnosis in Histopathology." pith.science (2026). https://pith.science/paper/GABHWN4Z

@misc{pith2026250710250,
  author       = {Pith},
  title        = {Pith review of: DepViT-CAD: Deployable Vision Transformer-Based Cancer Diagnosis in Histopathology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GABHWN4Z}},
  note         = {Machine review of arXiv:2507.10250}
}
read the original abstract

Accurate and timely cancer diagnosis from histopathological slides is vital for effective clinical decision-making. This paper introduces DepViT-CAD, a deployable AI system for multi-class cancer diagnosis in histopathology. At its core is MAViT, a novel Multi-Attention Vision Transformer designed to capture fine-grained morphological patterns across diverse tumor types. MAViT was trained on expert-annotated patches from 1008 whole-slide images, covering 11 diagnostic categories, including 10 major cancers and non-tumor tissue. DepViT-CAD was validated on two independent cohorts: 275 WSIs from The Cancer Genome Atlas and 50 routine clinical cases from pathology labs, achieving diagnostic sensitivities of 94.11% and 92%, respectively. By combining state-of-the-art transformer architecture with large-scale real-world validation, DepViT-CAD offers a robust and scalable approach for AI-assisted cancer diagnostics. To support transparency and reproducibility, software and code will be made publicly available at GitHub.

Figures

Figures reproduced from arXiv: 2507.10250 by the authors.

Figure 1
Figure 1. Cancer Diagnostic Pipeline via DepViT￾CAD 3.1 Pre-processing To ensure flexibility in both development and deploy￾ment, the pre-processing pipeline was divided into two phases: model training and clinical application. 3 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. MAViT’s overview (Linked to Figure 3) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. MAViT Architecture; integrating Multi-Attention Transformers, Multi-Layer Feature Fusion, and [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Architecture of the VTM [21] them with the intermediate features. The intermedi￾ate features undergo a 1 × 1 convolution for refine￾ment and are then resized to 18 × 18 to match the processed deep features. Finally, the shallow features are upsampled and concatenated w…
Figure 5
Figure 5. Figure 5: DFS integrates early and late fusion for [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: DepViT-CAD’s Decision-Making and Visualization scheme. The output includes tumor and non [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 8
Figure 8. Figure 8: Displays the model’s confidence intervals, [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 7
Figure 7. Figure 7: Radar Graph of Accuracy, Sensitivity, Specificity, and F1-Score, with Confusion Matrix Be￾low, for MAViT and DepViT-CAD on Patient-Level Test Set. the right and left of the vertical axis, respectively. Each circle represents a sample, with its size corre￾sponding to th…
Figure 9
Figure 9. Figure 9: D-Graph visualizes the Medical Predictive Model’s results (A), with upper quadrants representing [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Visual outputs of DepViT-CAD processing WSI (part a) and ROI (part c) from two clinical spec [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 7
Figure 7. Figure 7: While architectural innovation alone may [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 12
Figure 12. Figure 12: Example of tumor and non-tumor differentiation within the same ROI in gastric carcinoma. [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Stress-Aware Resilient Neural Training

    cs.LG 2025-07 unverdicted novelty 4.0 of 10

    The abstract claims a stress-triggered noise-injection optimizer improves generalization, but the provided manuscript body is an unrelated, garbled pure-mathematics paper.

Reference graph

Works this paper leans on

50 extracted references · 45 canonical work pages · cited by 1 Pith paper

  1. [1]

    J. Rosai. Why microscopy will remain a corner- stone of surgical pathology. Laboratory Investi- gation, 87(5):403–408, 2007

    work page 2007

  2. [2]

    Pisapia, V

    P. Pisapia, V. L’Imperio, F. Galuppini, E. Saj- jadi, A. Russo, B. Cerbelli, and U. Malapelle. The evolving landscape of anatomic pathol- ogy. Critical Reviews in Oncology/Hematology , 178:103776, 2022

    work page 2022

  3. [3]

    D. M. Metter, T. J. Colgan, S. T. Leung, C. F. Timmons, and J. Y. Park. Trends in the us and canadian pathologist workforces from 2007 to 2017. JAMA Network Open , 2(5):e194337– e194337, 2019

    work page 2007

  4. [4]

    M¨ arkl, L

    B. M¨ arkl, L. F¨ uzesi, R. Huss, S. Bauer, and T. Schaller. Number of pathologists in germany: Comparison with european countries, usa, and canada. Virchows Archiv, 478:335–341, 2021

    work page 2021

  5. [5]

    Betmouni

    S. Betmouni. Diagnostic digital pathol- ogy implementation: Learning from the digital health experience. Digital Health , 7:20552076211020240, 2021

    work page 2021

  6. [6]

    M. G. Hanna, O. Ardon, V. E. Reuter, S. J. Sir- intrapun, C. England, D. S. Klimstra, and M. R. Hameed. Integrating digital pathology into clin- ical practice. Modern Pathology, 35(2):152–164, 2022

    work page 2022

  7. [7]

    Maier, C

    A. Maier, C. Syben, T. Lasser, and C. Riess. A gentle introduction to deep learning in medical image processing. Zeitschrift f¨ ur Medizinische Physik, 29(2):86–101, 2019

    work page 2019

  8. [8]

    Echle, N

    A. Echle, N. T. Rindtorff, T. J. Brinker, T. Luedde, A. T. Pearson, and J. N. Kather. Deep learning in cancer pathology: A new gen- eration of clinical biomarkers. British Journal of Cancer, 124(4):686–696, 2021

    work page 2021

Show all 50 references
  1. [9]

    A. H. Song, G. Jaume, D. F. Williamson, M. Y. Lu, A. Vaidya, T. R. Miller, and F. Mahmood. Artificial intelligence for digital and computa- tional pathology. Nature Reviews Bioengineer- ing, 1(12):930–949, 2023

    2023

  2. [10]

    V. Baxi, R. Edwards, M. Montalto, and S. Saha. Digital pathology and artificial intelligence in translational medicine and clinical practice. Modern Pathology, 35(1):23–32, 2022

    2022

  3. [11]

    P. J. van Diest, R. N. Flach, C. van Dooijew- eert, S. Makineli, G. E. Breimer, N. Stathonikos, and M. Veta. Pros and cons of artificial intelli- gence implementation in diagnostic pathology. Histopathology, 2024

    2024

  4. [12]

    Shakarami

    A. Shakarami. Transformative ai for automating histopathology workflows. University of Padova, 2025

    2025

  5. [13]

    J. H. Oh, W. Choi, E. Ko, M. Kang, A. Tannen- baum, and J. O. Deasy. Pathcnn: interpretable convolutional neural networks for survival pre- diction and pathway analysis applied to glioblas- toma. Bioinformatics, 37(Supplement 1):i443– i450, 2021

    2021

  6. [14]

    Tan and Q

    M. Tan and Q. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learn- ing, pages 6105–6114. PMLR, 2019

    2019

  7. [15]

    Shakarami, L

    A. Shakarami, L. Nicol` e, M. Terreran, A. P. Dei Tos, and S. Ghidoni. Tcnn: A transformer convolutional neural network for artifact classi- fication in whole slide images. Biomedical Signal Processing and Control, 84:104812, 2023

    2023

  8. [16]

    Shakarami, L

    A. Shakarami, L. Nicole, and S. Ghidoni. Ai for advanced cancer diagnosis: a cad system em- powered by a novel vision transformer network for histopathology analysis. Virchows Archiv , 485(S1):S397–S398, 2024

    2024

  9. [17]

    Shakarami, H

    A. Shakarami, H. Tarrah, and A. Mahdavi- Hormat. A cad system for diagnosing 22 alzheimer’s disease using 2d slices and an im- proved alexnet-svm method. Optik, 212:164237, 2020

    2020

  10. [18]

    Shakarami, M

    A. Shakarami, M. B. Menhaj, and H. Tarrah. Di- agnosing covid-19 disease using an efficient cad system. Optik, 241:167199, 2021

    2021

  11. [19]

    Shakarami, M

    A. Shakarami, M. B. Menhaj, A. Mahdavi- Hormat, and H. Tarrah. A fast and yet efficient yolov3 for blood cell detection. Biomedical Sig- nal Processing and Control , 66:102495, 2021

    2021

  12. [20]

    Lu, Drew F

    Ming Y. Lu, Drew F. K. Williamson, Tzu-Ming Chen, and et al. Data-efficient and weakly supervised computational pathology on whole- slide images. Nature Biomedical Engineering , 5(6):555–570, 2021

    2021

  13. [21]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 , 2020

    2010 arXiv

  14. [22]

    Touvron, M

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. J´ egou. Training data-efficient image transformers & distillation through attention. In International Confer- ence on Machine Learning , pages 10347–10357. PMLR, 2021

    2021

  15. [23]

    Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo. Swin transformer: Hierar- chical vision transformer using shifted windows. Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition , 2021

    2021

  16. [24]

    Trans- mil: Transformer based correlated multiple in- stance learning for whole slide image classifica- tion

    Zhihao Shao, Heng Bian, Yicheng Chen, Yingy- ing Wang, Jun Zhang, and Xiangyang Ji. Trans- mil: Transformer based correlated multiple in- stance learning for whole slide image classifica- tion. In Advances in Neural Information Pro- cessing Systems (NeurIPS) , volume 34, pages...

    2021

  17. [25]

    R. J. Chen, C. Chen, Y. Li, T. Y. Chen, A. D. Trister, R. G. Krishnan, and F. Mahmood. Scal- ing vision transformers to gigapixel images via hierarchical self-supervised learning. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 16...

    2022

  18. [26]

    G. Li, G. Wu, G. Xu, C. Li, Z. Zhu, Y. Ye, and H. Zhang. Pathological image classification via embedded fusion mutual learning. Biomedical Signal Processing and Control , 79:104181, 2023

    2023

  19. [27]

    R. J. Chen, T. Ding, M. Y. Lu, D. F. Williamson, G. Jaume, B. Chen, and F. Mah- mood. A general-purpose self-supervised model for computational pathology. arXiv preprint arXiv:2308.15474, 2023

    2023 arXiv

  20. [28]

    M. Y. Lu, B. Chen, D. F. Williamson, R. J. Chen, I. Liang, T. Ding, and F. Mahmood. A visual-language foundation model for computa- tional pathology. Nature Medicine , 30(3):863– 874, 2024

    2024

  21. [29]

    Y. Xu, Y. Wang, F. Zhou, J. Ma, S. Yang, H. Lin, and H. Chen. A multimodal knowledge- enhanced whole-slide pathology foundation model. arXiv preprint arXiv:2407.15362 , 2024

    2024 arXiv

  22. [30]

    A simple frame- work for contrastive learning of visual represen- tations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple frame- work for contrastive learning of visual represen- tations. In International Conference on Machine Learning, pages 1597–1607. PMLR, 2020

    2020

  23. [31]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9729–9738, 2020

    2020

  24. [32]

    R. J. Chen, M. Y. Lu, J. Wang, D. F. Williamson, S. J. Rodig, N. I. Lindeman, and F. Mahmood. Pathomic fusion: an integrated framework for fusing histopathology and ge- nomic features for cancer diagnosis and prog- nosis. IEEE Transactions on Medical Imaging , 41(4):757–770, 2022. 23

    2022

  25. [33]

    Barcellona, L

    L. Barcellona, L. Nicol` e, R. Cappellesso, A. P. Dei Tos, and S. Ghidoni. Slidetiler: A dataset creator software for boosting deep learning on histological whole slide images. Journal of Pathology Informatics, 15:100356, 2024

    2024

  26. [34]

    S. Wang, B. Z. Li, M. Khabsa, H. Fang, and H. Ma. Linformer: Self-attention with linear complexity. arXiv preprint , arXiv:2006.04768, 2020

    2006 arXiv

  27. [35]

    Y. Y. Huang and W. T. Chu. Learnable context in multiple instance learning for whole slide im- age classification and segmentation. Journal of Imaging Informatics in Medicine , 1:1–15, 2024

    2024

  28. [36]

    H. D. Couture, J. S. Marron, C. M. Perou, M. A. Troester, and M. Niethammer. Multi- ple instance learning for heterogeneous images: Training a cnn for histopathology. In Medi- cal Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference, Gr...

    2018

  29. [37]

    G. S. Collins, J. B. Reitsma, D. G. Altman, and K. G. Moons. Transparent reporting of a multi- variable prediction model for individual progno- sis or diagnosis (tripod): the tripod statement. Annals of Internal Medicine, 162(1):55–63, 2015

    2015

  30. [38]

    P. C. Austin, F. E. Harrell Jr, and D. van Klav- eren. Graphical calibration curves and the inte- grated calibration index (ici) for survival models. Statistics in Medicine , 39(21):2714–2742, 2020

    2020

  31. [39]

    Nagrani, S

    A. Nagrani, S. Yang, A. Arnab, A. Jansen, C. Schmid, and C. Sun. Attention bottlenecks for multimodal fusion. In Advances in Neural In- formation Processing Systems, volume 34, pages 14200–14213, 2021

    2021

  32. [40]

    Singh, R

    L. Singh, R. R. Janghel, and S. P. Sahu. A hybrid feature fusion strategy for early fusion and majority voting for late fusion towards melanocytic skin lesion detection. International Journal of Imaging Systems and Technology , 32(4):1231–1250, 2022

    2022

  33. [41]

    Maur ´ ıcio, I

    J. Maur ´ ıcio, I. Domingues, and J. Bernardino. Comparing vision transformers and convolu- tional neural networks for image classification: A literature review. Applied Sciences, 13(9):5521, 2023

    2023

  34. [42]

    Howard, M

    A. Howard, M. Sandler, G. Chu, L. C. Chen, B. Chen, M. Tan, and H. Adam. Search- ing for mobilenetv3. In Proceedings of the IEEE/CVF international conference on com- puter vision , pages 1314–1324, 2019

    2019

  35. [43]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep resid- ual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  36. [44]

    Visual language pretrained multiple instance zero-shot transfer for histopathology images

    Ming Y Lu, Bowen Chen, et al. Visual language pretrained multiple instance zero-shot transfer for histopathology images. In CVPR, 2023

    2023

  37. [45]

    A knowledge- enhanced pathology vision-language foundation model for cancer diagnosis

    Xiao Zhou, Luoyi Sun, et al. A knowledge- enhanced pathology vision-language foundation model for cancer diagnosis. arXiv preprint arXiv:2412.13126, 2024

    2024

  38. [46]

    Pan-cancer integrative histology-genomic analysis via mul- timodal deep learning

    Richard J Chen, Ming Y Lu, et al. Pan-cancer integrative histology-genomic analysis via mul- timodal deep learning. Cancer Cell , 40(8):865– 878, 2022

    2022

  39. [47]

    Lu, Tiffany Y

    Ming Y. Lu, Tiffany Y. Chen, Drew F. K. Williamson, Ming Y. Zhao, Mahmoud Shady, Jana Lipkova, and Faisal Mahmood. Ai-based pathology predicts origins for cancers of un- known primary. Nature, 594(7861):106–110, 2021

    2021

  40. [48]

    Diagnostic as- sessment of deep learning algorithms for detec- tion of lymph node metastases in women with breast cancer

    Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes van Diest, Bram van Ginneken, Nico Karssemeijer, Geert Litjens, et al. Diagnostic as- sessment of deep learning algorithms for detec- tion of lymph node metastases in women with breast cancer. In JAMA, volume 318, pages 2199–...

  41. [49]

    Detection of lymph node metastases in breast cancer from whole- slide histopathology images using deep learning: participation in the camelyon17 challenge

    Peter Bandi, Oscar Geessink, Qaiser F Man- son, Marije C Van Dijk, Markus Balkenhol, Meyke Hermsen, et al. Detection of lymph node metastases in breast cancer from whole- slide histopathology images using deep learning: participation in the camelyon17 challenge. In Medical Ima...

  42. [50]

    Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study

    Jakob Nikolas Kather, Jan Krisam, Piroon Charoentong, and et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS Medicine, 16(1):e1002730, 2019. 25

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.