pith. machine review for the scientific record. sign in

arxiv: 2604.23799 · v1 · submitted 2026-04-26 · 💻 cs.CV

Recognition: unknown

VitaminP: cross-modal learning enables whole-cell segmentation from routine histology

Elizve N. Barrientos Toro, Karina B. Pinao Gonzales, Patient Mosaic Team, Paul Acosta, Pingjun Chen, Xiaoxi Pan, Yasin Shokrollahi, Yinyin Yuan

Pith reviewed 2026-05-08 06:40 UTC · model grok-4.3

classification 💻 cs.CV
keywords whole-cell segmentationcross-modal learningH&E stainingmultiplex immunofluorescencedigital pathologycancer segmentationboundary transferhistology analysis
0
0 comments X

The pith

Cross-modal learning from paired H&E-mIF data enables whole-cell segmentation on routine histology images.

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

VitaminP trains on paired H&E and multiplex immunofluorescence images so that molecular boundary cues from mIF can be transferred to standard H&E stains. This overcomes the weak cytoplasmic contrast that normally restricts H&E analysis to nuclei alone. The resulting model was trained on more than seven million cells across 34 cancer types and outperforms prior segmentation methods on both public and unseen rare-cancer datasets. An open-source inference platform accompanies the method to support wider use in pathology and spatial omics.

Core claim

By learning from paired H&E-mIF data, VitaminP transfers molecular boundary information from mIF to overcome cytoplasmic contrast in H&E, establishing cross-modal supervision as a general strategy for recovering missing biological structure.

What carries the argument

Cross-modal supervision that learns to predict whole-cell boundaries by treating mIF-derived labels as ground truth for corresponding H&E images.

If this is right

  • Whole-cell morphology becomes measurable on any standard H&E slide without requiring multiplex immunofluorescence.
  • Spatial analyses in precision pathology can move beyond nuclear-only readouts to full cell shapes and neighborhoods.
  • The approach generalizes across dozens of cancer types, including rare ones not represented in training.
  • Open-source inference tools lower the barrier for labs to adopt whole-cell segmentation at scale.

Where Pith is reading between the lines

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

  • Routine clinical H&E archives could be re-analyzed for cell-level features that were previously inaccessible.
  • Similar cross-modal pairing might recover other missing signals in low-contrast imaging modalities.
  • Integration with spatial transcriptomics could tighten correspondence between cell boundaries and molecular profiles.

Load-bearing premise

Paired H&E-mIF images supply reliable, aligned ground-truth boundaries that the model can learn and apply to new H&E images without major domain shift or annotation errors.

What would settle it

Manually annotate whole-cell boundaries on a fresh H&E dataset from an unseen cancer type and test whether VitaminP accuracy drops below the best single-modality H&E methods.

Figures

Figures reproduced from arXiv: 2604.23799 by Elizve N. Barrientos Toro, Karina B. Pinao Gonzales, Patient Mosaic Team, Paul Acosta, Pingjun Chen, Xiaoxi Pan, Yasin Shokrollahi, Yinyin Yuan.

Figure 2
Figure 2. Figure 2: Performance benchmark of VitaminP and comparison methods on nuclear and whole-cell segmentation. view at source ↗
Figure 5
Figure 5. Figure 5: VitaminPScope platform for interactive whole-slide image analysis and quantitative pathology outputs. a, view at source ↗
read the original abstract

Accurate whole-cell and nuclear segmentation is essential for precision pathology and spatial omics, yet routine hematoxylin and eosin (H&E) staining provides limited cytoplasmic contrast, restricting analyses to nuclei. Multiplex immunofluorescence (mIF) facilitates precise whole-cell delineation but remains constrained by cost and accessibility. We introduce VitaminP, a cross-modal learning framework enabling whole cell segmentation from H&E images. By learning from paired H&E-mIF data, VitaminP transfers molecular boundary information from mIF to overcome cytoplasmic contrast in H&E, establishing cross-modal supervision as a general strategy for recovering missing biological structure. We train VitaminP on 14 public datasets covering 34 cancer types and over 7 million instances, integrating publicly available labels with extensive annotations generated in this study, forming one of the largest resources for segmentation. VitaminP outperforms four state-of-the-art methods and generalizes to unseen datasets, including an in-house dataset spanning 24 rare cancer types. We further developed VitaminPScope, an open-source platform providing an interface for scalable inference and enabling broad adoption.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces VitaminP, a cross-modal supervised learning framework that trains on paired H&E and multiplex immunofluorescence (mIF) images to enable accurate whole-cell segmentation directly from routine H&E stains. It claims to assemble one of the largest segmentation resources by combining 14 public datasets (34 cancer types, >7 million instances) with new annotations, outperforming four state-of-the-art methods, and generalizing to held-out public datasets plus an in-house collection spanning 24 rare cancer types. An open-source inference platform (VitaminPScope) is also released.

Significance. If the central claims hold, the work would be significant for precision pathology and spatial omics: routine H&E slides are ubiquitous while mIF is costly and low-throughput; a reliable transfer of molecular boundary information could remove the cytoplasmic-contrast bottleneck and enable scalable whole-cell analyses. The scale of the training corpus and the explicit generalization test on rare cancers are notable strengths. The open-source platform further lowers the barrier to adoption.

major comments (3)
  1. [§3 and §4.1] §3 (Data curation) and §4.1 (Experimental setup): The manuscript states that mIF boundaries provide the supervisory signal for H&E images, yet provides no quantitative assessment of H&E-mIF registration error (e.g., landmark-based Dice or Hausdorff distance) or inter-annotator agreement on the integrated labels across the 14 datasets. Without these metrics, it is impossible to rule out that reported gains arise from label noise rather than genuine cross-modal transfer.
  2. [§4.2 and Table 2] §4.2 (Results) and Table 2: The claim that VitaminP outperforms four SOTA methods is presented without an ablation that isolates the contribution of mIF supervision (e.g., training the identical architecture on H&E-only labels of equal volume). Consequently, it remains unclear whether performance differences are attributable to the cross-modal strategy or simply to the unprecedented training-set size.
  3. [§5] §5 (Generalization experiments): Generalization to the in-house set of 24 rare cancer types is asserted, but the paper does not report per-cancer-type performance breakdowns, domain-shift statistics (e.g., stain normalization metrics), or failure-case analysis. This information is required to substantiate that the learned mapping is staining-invariant rather than dataset-specific.
minor comments (2)
  1. [Abstract and §1] The abstract and §1 refer to “over 7 million instances” without clarifying whether this counts cells, patches, or slides; a precise definition would aid reproducibility.
  2. [Figure 3] Figure 3 (qualitative results) would benefit from side-by-side error maps or zoomed insets highlighting cytoplasmic boundary recovery on challenging H&E regions.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the work's potential impact. We address each major comment in detail below, indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [§3 and §4.1] §3 (Data curation) and §4.1 (Experimental setup): The manuscript states that mIF boundaries provide the supervisory signal for H&E images, yet provides no quantitative assessment of H&E-mIF registration error (e.g., landmark-based Dice or Hausdorff distance) or inter-annotator agreement on the integrated labels across the 14 datasets. Without these metrics, it is impossible to rule out that reported gains arise from label noise rather than genuine cross-modal transfer.

    Authors: We appreciate the referee's emphasis on data quality validation. The paired H&E-mIF images originate from public datasets where registration was performed by the original contributors using established methods; however, we agree that explicit quantification is valuable. In the revised manuscript, we will add quantitative metrics in §3, including average landmark-based Dice and Hausdorff distances computed on a sampled subset of pairs. For inter-annotator agreement on the newly generated annotations integrated across datasets, we will report agreement statistics (e.g., Dice overlap on a held-out annotation subset). These additions will help confirm that performance gains stem from cross-modal transfer rather than label artifacts. revision: yes

  2. Referee: [§4.2 and Table 2] §4.2 (Results) and Table 2: The claim that VitaminP outperforms four SOTA methods is presented without an ablation that isolates the contribution of mIF supervision (e.g., training the identical architecture on H&E-only labels of equal volume). Consequently, it remains unclear whether performance differences are attributable to the cross-modal strategy or simply to the unprecedented training-set size.

    Authors: This is a fair point regarding attribution. The SOTA baselines were evaluated using their original training protocols on smaller public corpora, while VitaminP leverages the scale and mIF-derived boundaries. To better isolate effects, the revision will include an ablation using our architecture trained on H&E-only labels from available subsets (where cytoplasmic annotations exist in the public data) and compare against the full mIF-supervised model. We will also expand the discussion in §4.2 to note that equivalent-volume H&E-only labels at this scale are not readily available without new annotation efforts, which underscores the practical advantage of the cross-modal approach. This will clarify the relative contributions. revision: partial

  3. Referee: [§5] §5 (Generalization experiments): Generalization to the in-house set of 24 rare cancer types is asserted, but the paper does not report per-cancer-type performance breakdowns, domain-shift statistics (e.g., stain normalization metrics), or failure-case analysis. This information is required to substantiate that the learned mapping is staining-invariant rather than dataset-specific.

    Authors: We agree that granular reporting would strengthen the generalization claims. In the revised §5 and supplementary materials, we will provide per-cancer-type performance breakdowns (e.g., Dice and IoU) for the 24 rare cancer types in the in-house set. We will also report domain-shift statistics, including stain variation metrics (such as color histogram distances) pre- and post-normalization, and include a dedicated failure-case analysis with representative examples and discussion of potential causes (e.g., rare morphological variants). These additions will better demonstrate staining invariance. revision: yes

Circularity Check

0 steps flagged

No circularity in empirical cross-modal supervised learning pipeline

full rationale

The paper presents VitaminP as a standard supervised deep learning framework trained on paired H&E-mIF images to transfer boundary information for H&E-only segmentation. All reported results derive from empirical training on large-scale paired datasets (14 public + in-house) followed by evaluation on held-out sets, with no mathematical derivations, equations, or first-principles claims that reduce outputs to inputs by construction. No self-citations load-bear the central method, no fitted parameters are renamed as predictions, and no uniqueness theorems or ansatzes are imported to force the architecture. The approach is self-contained against external benchmarks as a conventional cross-modal transfer learning setup.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the empirical success of a neural network trained with cross-modal supervision; the key unverified premise is that paired H&E-mIF data supplies accurate transferable boundary labels.

axioms (1)
  • domain assumption Paired H&E and mIF images from the same tissue section share identical cellular structures, allowing mIF-derived boundaries to serve as reliable supervision for H&E images
    This assumption underpins the entire cross-modal transfer strategy described in the abstract.
invented entities (1)
  • VitaminP no independent evidence
    purpose: Cross-modal learning framework for whole-cell segmentation from H&E
    Newly introduced method whose performance is the central claim of the paper.

pith-pipeline@v0.9.0 · 5513 in / 1435 out tokens · 65346 ms · 2026-05-08T06:40:55.220012+00:00 · methodology

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

53 extracted references · 47 canonical work pages · 5 internal anchors

  1. [1]

    Langlotz, and Akshay S

    Amirreza Mahbod et al. “NuInsSeg: A fully annotated dataset for nuclei instance segmentation in H&E-stained histological images”. en. In:Scientific Data11.1 (Mar. 2024), p. 295.issn: 2052-4463. doi: 10.1038/s41597- 024- 03117- 2 .url: https://www.nature.com/articles/s41597- 024-03117-2(visited on 01/12/2026)

    work page doi:10.1038/s41597- 2024

  2. [2]

    Advancements in automated nuclei segmentation for histopathology using you only look once-driven approaches: A systematic review

    Shyam Sundar Debsarkar et al. “Advancements in automated nuclei segmentation for histopathology using you only look once-driven approaches: A systematic review”. en. In:Computers in Biology and Medicine190 (May 2025), p. 110072.issn: 00104825.doi: 10.1016/j.compbiomed.2025.110072. url: https://linkinghub.elsevier.com/retrieve/pii/S0010482525004238 (visite...

    work page doi:10.1016/j.compbiomed.2025.110072 2025

  3. [3]

    Lonini et al

    Clare McGenity et al. “Artificial intelligence in digital pathology: a systematic review and meta-analysis of diagnostic test accuracy”. en. In:npj Digital Medicine7.1 (May 2024), p. 114.issn: 2398-6352.doi: 10.1038/s41746- 024- 01106- 8.url: https://www.nature.com/articles/s41746- 024- 01106-8(visited on 01/12/2026)

    work page doi:10.1038/s41746- 2024

  4. [4]

    Multiplex Immunofluorescence and the Digital Image Analysis Workflow for Evaluation of the Tumor Immune Environment in Translational Research

    Frank Rojas et al. “Multiplex Immunofluorescence and the Digital Image Analysis Workflow for Evaluation of the Tumor Immune Environment in Translational Research”. In:Frontiers in Oncology 12 (June 2022), p. 889886.issn: 2234-943X.doi: 10.3389/fonc.2022.889886 .url: https:// www.frontiersin.org/articles/10.3389/fonc.2022.889886/full(visited on 01/12/2026)

    work page doi:10.3389/fonc.2022.889886 2022

  5. [5]

    Mohammad Yosofvand et al.Spatial Immunophenotyping from Whole-Slide Multiplexed Tissue Imaging Using Convolutional Neural Networks. en. Aug. 2024.doi: 10.1101/2024.08.16.608247 .url: http://biorxiv.org/lookup/doi/10.1101/2024.08.16.608247(visited on 01/12/2026)

    work page doi:10.1101/2024.08.16.608247 2024

  6. [6]

    Machine learning methods for histopathological image analysis: Updates in 2024

    Daisuke Komura et al. “Machine learning methods for histopathological image analysis: Updates in 2024”. en. In:Computational and Structural Biotechnology Journal27 (2025), pp. 383–400.issn: 20010370.doi: 10.1016/j.csbj.2024.12.033 .url: https://linkinghub.elsevier.com/ retrieve/pii/S2001037024004549(visited on 01/12/2026). 16

    work page doi:10.1016/j.csbj.2024.12.033 2024

  7. [7]

    Continuous Radon Monitoring during Seven Years of V olcanic Unrest at Campi Flegrei Caldera (Italy)

    Gustav M¨ uller-Franzes et al. “Medical slice transformer for improved diagnosis and explainability on 3D medical images with DINOv2”. en. In:Scientific Reports15.1 (July 2025), p. 23979.issn: 2045-2322. doi: 10.1038/s41598- 025- 09041- 8 .url: https://www.nature.com/articles/s41598- 025-09041-8(visited on 01/12/2026)

    work page doi:10.1038/s41598- 2025

  8. [8]

    Version Number:

    Maxime Oquab et al.DINOv2: Learning Robust Visual Features without Supervision. Version Number:

  9. [9]

    2023.doi: 10.48550/ARXIV.2304.07193 .url: https://arxiv.org/abs/2304.07193 (visited on 01/12/2026)

    work page internal anchor Pith review doi:10.48550/arxiv.2304.07193 2023

  10. [10]

    Vision Transformers for Computational Histopathology

    Hongming Xu et al. “Vision Transformers for Computational Histopathology”. In:IEEE Reviews in Biomedical Engineering17 (2024), pp. 63–79.issn: 1937-3333, 1941-1189.doi: 10 . 1109 / RBME.2023.3297604.url: https://ieeexplore.ieee.org/document/10190115/ (visited on 01/12/2026)

    work page arXiv 2024

  11. [11]

    Nature Medicine30(10), 2924–2935 (2024) https://doi.org/10.1038/s41591-024-03141-0

    Eugene Vorontsov et al. “A foundation model for clinical-grade computational pathology and rare cancers detection”. en. In:Nature Medicine30.10 (Oct. 2024), pp. 2924–2935.issn: 1078-8956, 1546-170X.doi: 10.1038/s41591-024-03141-0 .url: https://www.nature.com/articles/ s41591-024-03141-0(visited on 01/30/2026)

    work page doi:10.1038/s41591-024-03141-0 2024

  12. [12]

    Version Number: 1

    Qingchen Tang et al.Prototype-Based Image Prompting for Weakly Supervised Histopathological Image Segmentation. Version Number: 1. 2025.doi: 10.48550/ARXIV.2503.12068.url: https: //arxiv.org/abs/2503.12068(visited on 01/12/2026)

    work page doi:10.48550/arxiv.2503.12068.url: 2025

  13. [13]

    Advancing breast cancer diagnosis: token vision transformers for faster and accurate classification of histopathology images

    Mouhamed Laid Abimouloud et al. “Advancing breast cancer diagnosis: token vision transformers for faster and accurate classification of histopathology images”. en. In:Visual Computing for Industry, Biomedicine, and Art8.1 (Jan. 2025), p. 1.issn: 2524-4442.doi: 10.1186/s42492-024-00181-8 . url: https://vciba.springeropen.com/articles/10.1186/s42492-024-001...

    work page doi:10.1186/s42492-024-00181-8 2025

  14. [14]

    Version Number: 1

    Andrey Ignatov et al.Histopathological Image Classification with Cell Morphology Aware Deep Neural Networks. Version Number: 1. 2024.doi: 10.48550/ARXIV.2407.08625 .url: https: //arxiv.org/abs/2407.08625(visited on 01/12/2026)

    work page doi:10.48550/arxiv.2407.08625 2024

  15. [15]

    SHIFT: speedy histological-to-immunofluorescent translation of a tumor signature enabled by deep learning

    Erik A. Burlingame et al. “SHIFT: speedy histological-to-immunofluorescent translation of a tumor signature enabled by deep learning”. en. In:Scientific Reports10.1 (Oct. 2020), p. 17507.issn: 2045-2322.doi: 10.1038/s41598-020-74500-3 .url: https://www.nature.com/articles/ s41598-020-74500-3(visited on 01/12/2026)

    work page doi:10.1038/s41598-020-74500-3 2020

  16. [16]

    HEMIT: H&E to Multiplex-Immunohistochemistry Image Translation with Dual- Branch Pix2pix Generator

    Chang Bian et al. “HEMIT: H&E to Multiplex-Immunohistochemistry Image Translation with Dual- Branch Pix2pix Generator”. en. In:Medical Image Computing and Computer Assisted Intervention – MICCAI 2024 Workshops. Ed. by Anna Schroder et al. Vol. 15401. Series Title: Lecture Notes in Computer Science. Cham: Springer Nature Switzerland, 2025, pp. 184–197.isbn...

    work page doi:10.1007/978-3-031-84525-3_16 2024

  17. [17]

    Highly multiplexed immunofluorescence imaging of human tissues and tumors using t-CyCIF and conventional optical microscopes

    Jia-Ren Lin et al. “Highly multiplexed immunofluorescence imaging of human tissues and tumors using t-CyCIF and conventional optical microscopes”. en. In:eLife7 (July 2018), e31657.issn: 2050-084X. 17 doi: 10.7554/eLife.31657.url: https://elifesciences.org/articles/31657 (visited on 01/12/2026)

    work page doi:10.7554/elife.31657.url: 2018

  18. [18]

    Deep learning-inferred multiplex immunofluorescence for immunohisto- chemical image quantification

    Parmida Ghahremani et al. “Deep learning-inferred multiplex immunofluorescence for immunohisto- chemical image quantification”. en. In:Nature Machine Intelligence4.4 (Apr. 2022), pp. 401–412.issn: 2522-5839.doi: 10.1038/s42256-022-00471-x .url: https://www.nature.com/articles/ s42256-022-00471-x(visited on 01/12/2026)

    work page doi:10.1038/s42256-022-00471-x 2022

  19. [19]

    ROSIE: AI generation of multiplex immunofluorescence staining from histopathology images

    Eric Wu et al. “ROSIE: AI generation of multiplex immunofluorescence staining from histopathology images”. en. In:Nature Communications16.1 (Aug. 2025), p. 7633.issn: 2041-1723.doi: 10.1038/ s41467- 025- 62346- 0.url: https://www.nature.com/articles/s41467- 025- 62346- 0 (visited on 01/12/2026)

    2025

  20. [20]

    Version Number: 1

    Guillaume Balezo et al.MIPHEI-ViT: Multiplex Immunofluorescence Prediction from H&E Images using ViT Foundation Models. Version Number: 1. 2025.doi:10.48550/ARXIV.2505.10294. url:https://arxiv.org/abs/2505.10294(visited on 01/12/2026)

    work page doi:10.48550/arxiv.2505.10294 2025

  21. [21]

    CellSAM: a foundation model for cell segmentation

    Markus Marks et al. “CellSAM: a foundation model for cell segmentation”. en. In:Nature Methods22.12 (Dec. 2025), pp. 2585–2593.issn: 1548-7091, 1548-7105.doi: 10.1038/s41592-025-02879-w . url:https://www.nature.com/articles/s41592-025-02879-w(visited on 01/12/2026)

    work page doi:10.1038/s41592-025-02879-w 2025

  22. [22]

    Marius Pachitariu et al.Cellpose-SAM: superhuman generalization for cellular segmentation. en. May 2025.doi: 10.1101/2025.04.28.651001.url: http://biorxiv.org/lookup/doi/10.1101/ 2025.04.28.651001(visited on 01/14/2026)

    work page doi:10.1101/2025.04.28.651001.url: 2025

  23. [23]

    Version Number: 1

    Alexander Kirillov et al.Segment Anything. Version Number: 1. 2023.doi: 10.48550/ARXIV.2304. 02643.url:https://arxiv.org/abs/2304.02643(visited on 01/17/2026)

    work page doi:10.48550/arxiv.2304 2023

  24. [24]

    Version Number: 1

    Fabian H¨orst et al.CellViT++: Energy-Efficient and Adaptive Cell Segmentation and Classification Using Foundation Models. Version Number: 1. 2025.doi: 10.48550/ARXIV.2501.05269 .url: https://arxiv.org/abs/2501.05269(visited on 01/14/2026)

    work page doi:10.48550/arxiv.2501.05269 2025

  25. [26]

    Predicting transcriptional outcomes of novel multigene perturba- tions with GEARS

    Noah F. Greenwald et al. “Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep learning”. en. In:Nature Biotechnology40.4 (Apr. 2022), pp. 555–565.issn: 1087-0156, 1546-1696.doi: 10.1038/s41587- 021- 01094- 0 .url: https: //www.nature.com/articles/s41587-021-01094-0(visited on 01/12/2026)

    work page doi:10.1038/s41587- 2022

  26. [27]

    Hover-Net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images.Medical Image Analysis, 58:101563, 2019

    Simon Graham et al. “Hover-Net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images”. en. In:Medical Image Analysis58 (Dec. 2019), p. 101563.issn: 13618415.doi: 10.1016/j.media.2019.101563.url: https://linkinghub.elsevier.com/retrieve/pii/ S1361841519301045(visited on 01/12/2026)

    work page doi:10.1016/j.media.2019.101563.url: 2019

  27. [28]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Alexey Dosovitskiy et al.An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. Version Number: 2. 2020.doi: 10.48550/ARXIV.2010.11929.url: https://arxiv.org/ abs/2010.11929(visited on 04/13/2026). 18

    work page Pith review doi:10.48550/arxiv.2010.11929.url: 2020

  28. [29]

    Version Number:

    Phillip Isola et al.Image-to-Image Translation with Conditional Adversarial Networks. Version Number:

  29. [30]

    2016.doi: 10.48550/ARXIV.1611.07004 .url: https://arxiv.org/abs/1611.07004 (visited on 02/22/2026)

    work page doi:10.48550/arxiv.1611.07004 2016

  30. [31]

    High-plex immunofluorescence imaging and traditional histology of the same tissue section for discovering image-based biomarkers

    Jia-Ren Lin et al. “High-plex immunofluorescence imaging and traditional histology of the same tissue section for discovering image-based biomarkers”. en. In:Nature Cancer4.7 (June 2023), pp. 1036–1052.issn: 2662-1347.doi: 10.1038/s43018- 023- 00576- 1 .url: https://www. nature.com/articles/s43018-023-00576-1(visited on 01/16/2026)

    work page doi:10.1038/s43018- 2023

  31. [32]

    10x Genomics.10x Genomics Datasets.url: https://www.10xgenomics.com/datasets (visited on 03/16/2026)

    2026

  32. [33]

    Gamper, N

    Jevgenij Gamper et al.PanNuke Dataset Extension, Insights and Baselines. Version Number: 7. 2020. doi: 10.48550/ARXIV.2003.10778 .url: https://arxiv.org/abs/2003.10778 (visited on 02/22/2026)

    work page doi:10.48550/arxiv.2003.10778 2020

  33. [34]

    Llvip: A visible-infrared paired dataset for low-light vision

    Simon Graham et al. “Lizard: A Large-Scale Dataset for Colonic Nuclear Instance Segmentation and Classification”. In:2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). Montreal, BC, Canada: IEEE, Oct. 2021, pp. 684–693.isbn: 978-1-66540-191-3.doi: 10. 1109/ICCVW54120.2021.00082 .url: https://ieeexplore.ieee.org/document/9607772/ ...

    work page arXiv 2021

  34. [35]

    A Multi-Organ Nucleus Segmentation Challenge

    Neeraj Kumar et al. “A Multi-Organ Nucleus Segmentation Challenge”. In:IEEE Transactions on Medical Imaging39.5 (May 2020), pp. 1380–1391.issn: 0278-0062, 1558-254X.doi: 10.1109/ TMI.2019.2947628 .url: https://ieeexplore.ieee.org/document/8880654/ (visited on 02/22/2026)

    work page arXiv 2020

  35. [36]

    Segmentation of nuclei in histopathology images by deep regression of the distance map.IEEE Transactions on Medical Imaging, 38(2):448–459, 2019

    Peter Naylor et al. “Segmentation of Nuclei in Histopathology Images by Deep Regression of the Distance Map”. In:IEEE Transactions on Medical Imaging38.2 (Feb. 2019), pp. 448–459.issn: 0278-0062, 1558-254X.doi: 10.1109/TMI.2018.2865709 .url: https://ieeexplore.ieee. org/document/8438559/(visited on 01/21/2026)

    work page doi:10.1109/tmi.2018.2865709 2019

  36. [37]

    CryoNuSeg: A dataset for nuclei instance segmentation of cryosectioned H&E- stained histological images

    Amirreza Mahbod et al. “CryoNuSeg: A dataset for nuclei instance segmentation of cryosectioned H&E- stained histological images”. en. In:Computers in Biology and Medicine132 (May 2021), p. 104349. issn: 00104825.doi: 10 . 1016 / j . compbiomed . 2021 . 104349.url: https : / / linkinghub . elsevier.com/retrieve/pii/S0010482521001438(visited on 02/22/2026)

    2021

  37. [38]

    Structured crowdsourcing enables convolutional segmentation of histology images

    Mohamed Amgad et al. “Structured crowdsourcing enables convolutional segmentation of histology images”. en. In:Bioinformatics35.18 (Sept. 2019). Ed. by Robert Murphy, pp. 3461–3467.issn: 1367-4803, 1367-4811.doi: 10.1093/bioinformatics/btz083.url: https://academic.oup. com/bioinformatics/article/35/18/3461/5307750(visited on 02/22/2026)

    work page doi:10.1093/bioinformatics/btz083.url: 2019

  38. [39]

    MoNuSAC2020: A Multi-Organ Nuclei Segmentation and Classification Challenge

    Ruchika Verma et al. “MoNuSAC2020: A Multi-Organ Nuclei Segmentation and Classification Challenge”. In:IEEE Transactions on Medical Imaging40.12 (Dec. 2021), pp. 3413–3423.issn: 0278-0062, 1558-254X.doi: 10.1109/TMI.2021.3085712 .url: https://ieeexplore.ieee. org/document/9446924/(visited on 02/22/2026)

    work page doi:10.1109/tmi.2021.3085712 2021

  39. [40]

    A Dataset and a Technique for Generalized Nuclear Segmentation for Computational Pathology

    Neeraj Kumar et al. “A Dataset and a Technique for Generalized Nuclear Segmentation for Computational Pathology”. In:IEEE Transactions on Medical Imaging36.7 (July 2017), pp. 1550–1560.issn: 0278- 19 0062, 1558-254X.doi: 10.1109/TMI.2017.2677499 .url: https://ieeexplore.ieee.org/ document/7872382/(visited on 01/21/2026)

    work page doi:10.1109/tmi.2017.2677499 2017

  40. [41]

    Methods for Segmentation and Classification of Digital Microscopy Tissue Images

    Quoc Dang Vu et al. “Methods for Segmentation and Classification of Digital Microscopy Tissue Images”. In:Frontiers in Bioengineering and Biotechnology7 (Apr. 2019), p. 53.issn: 2296-4185. doi: 10.3389/fbioe.2019.00053.url: https://www.frontiersin.org/article/10.3389/ fbioe.2019.00053/full(visited on 02/22/2026)

    work page doi:10.3389/fbioe.2019.00053.url: 2019

  41. [43]

    A panoptic segmentation dataset and deep-learning approach for explainable scoring of tumor-infiltrating lymphocytes

    Shangke Liu et al. “A panoptic segmentation dataset and deep-learning approach for explainable scoring of tumor-infiltrating lymphocytes”. en. In:npj Breast Cancer10.1 (June 2024), p. 52.issn: 2374-4677. doi: 10.1038/s41523- 024- 00663- 1 .url: https://www.nature.com/articles/s41523- 024-00663-1(visited on 02/22/2026)

    work page doi:10.1038/s41523- 2024

  42. [44]

    Restaining-based annotation for cancer histology segmentation to overcome annotation-related limitations among pathologists

    Daisuke Komura et al. “Restaining-based annotation for cancer histology segmentation to overcome annotation-related limitations among pathologists”. en. In:Patterns4.2 (Feb. 2023), p. 100688.issn: 26663899.doi: 10.1016/j.patter.2023.100688.url: https://linkinghub.elsevier.com/ retrieve/pii/S2666389923000193(visited on 03/26/2026)

    work page doi:10.1016/j.patter.2023.100688.url: 2023

  43. [45]

    Incorporating the image formation process into deep learning improves network performance

    Clarence Yapp et al. “Highly multiplexed 3D profiling of cell states and immune niches in human tumors”. en. In:Nature Methods22.10 (Oct. 2025), pp. 2180–2193.issn: 1548-7091, 1548-7105.doi: 10.1038/s41592- 025- 02824- x.url: https://www.nature.com/articles/s41592- 025- 02824-x(visited on 04/16/2026)

    work page doi:10.1038/s41592- 2025

  44. [46]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    Olaf Ronneberger et al.U-Net: Convolutional Networks for Biomedical Image Segmentation. Version Number: 1. 2015.doi: 10.48550/ARXIV.1505.04597 .url: https://arxiv.org/abs/1505. 04597(visited on 02/27/2026)

    work page internal anchor Pith review doi:10.48550/arxiv.1505.04597 2015

  45. [47]

    Rethinking Atrous Convolution for Semantic Image Segmentation

    Liang-Chieh Chen et al.Rethinking Atrous Convolution for Semantic Image Segmentation. Version Number: 3. 2017.doi: 10.48550/ARXIV.1706.05587 .url: https://arxiv.org/abs/1706. 05587(visited on 02/27/2026)

    work page internal anchor Pith review doi:10.48550/arxiv.1706.05587 2017

  46. [48]

    Group normalization

    Yuxin Wu et al.Group Normalization. Version Number: 3. 2018.doi:10.48550/ARXIV.1803.08494. url:https://arxiv.org/abs/1803.08494(visited on 03/27/2026)

    work page doi:10.48550/arxiv.1803.08494 2018

  47. [49]

    Gaussian Error Linear Units (GELUs)

    Dan Hendrycks et al.Gaussian Error Linear Units (GELUs). Version Number: 5. 2016.doi:10.48550/ ARXIV.1606.08415.url:https://arxiv.org/abs/1606.08415(visited on 03/27/2026)

    work page internal anchor Pith review arXiv 2016

  48. [50]

    URLhttps://doi.org/10.48550/arXiv

    Jie Hu et al.Squeeze-and-Excitation Networks. Version Number: 4. 2017.doi: 10.48550/ARXIV. 1709.01507.url:https://arxiv.org/abs/1709.01507(visited on 02/27/2026)

    work page internal anchor Pith review doi:10.48550/arxiv 2017

  49. [51]

    An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution , publisher =

    Rosanne Liu et al.An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution. Version Number: 2. 2018.doi: 10.48550/ARXIV.1807.03247.url: https://arxiv.org/abs/ 1807.03247(visited on 02/27/2026)

    work page doi:10.48550/arxiv.1807.03247.url: 2018

  50. [52]

    Version Number: 1

    Fausto Milletari et al.V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. Version Number: 1. 2016.doi: 10 . 48550 / ARXIV . 1606 . 04797.url: https : //arxiv.org/abs/1606.04797(visited on 02/27/2026). 20

    work page arXiv 2016

  51. [53]

    https://doi.org/10.48550/arXiv.1708.02002

    Tsung-Yi Lin et al.Focal Loss for Dense Object Detection. Version Number: 2. 2017.doi: 10.48550/ ARXIV.1708.02002.url:https://arxiv.org/abs/1708.02002(visited on 02/27/2026)

    work page arXiv 2017

  52. [54]

    Version Number: 3

    Ilya Loshchilov et al.Decoupled Weight Decay Regularization. Version Number: 3. 2017.doi: 10 . 48550 / ARXIV . 1711 . 05101.url: https : / / arxiv . org / abs / 1711 . 05101(visited on 02/27/2026)

    2017

  53. [55]

    OpenSlide: A vendor-neutral software foundation for digital pathology

    Adam Goode et al. “OpenSlide: A vendor-neutral software foundation for digital pathology”. en. In: Journal of Pathology Informatics4.1 (Jan. 2013), p. 27.issn: 21533539.doi: 10.4103/2153-3539. 119005.url: https : / / linkinghub . elsevier . com / retrieve / pii / S2153353922006484 (visited on 02/27/2026). Methods Dataset construction To enable cross-modal...

    work page doi:10.4103/2153-3539 2013