The Atlas is a WSI-first selection guide covering more than 50 reviewed repositories, cohorts,
challenges, derived benchmarks, and WSI-language resources. Every card links to the
official host and records the release, scientific use, data hierarchy, WSI relationship,
access route, terms, count scope, source cohorts, and the mistakes most likely to invalidate
an experiment.
It includes conventional stained whole slides; full-slide representations converted from
scanner files; documented WSI-derived regions, patches, and cell annotations; and reports,
outcomes, spatial assays, or molecular measurements that remain linked to pathology
slides. It intentionally excludes generic microscopy, radiology, textbook-figure VQA, and
web-image corpora that have no defensible WSI lineage.
Counts are release-scoped, not interchangeable “sample” totals. Hover a count to see
its scope, then confirm it against the official manifest before sizing an experiment. A
card without a count means that no stable, authoritative aggregate was available—not zero
data.
Showing 57 of 57 reviewed records.
Native pyramidal WSIChallenge access
2019 · Challenge
ACDC@LungHP
A lung-biopsy whole-slide benchmark with tissue-level carcinoma polygons for cancer detection and segmentation in clinical biopsy specimens.
Tissue segmentationDiagnosis & gradinglung
200patients150 training and 50 test biopsy samples
200slidescomplete challenge cohort
Release
ISBI 2019 challenge release
Units
patients, cases, slides, ROIs
Modalities
whole-slide image, tumour annotation
Formats
pyramidal TIFF WSI, XML polygons
Last verified
License / terms
Challenge terms; no standalone dataset license stated on the official pages
Use with care: The official training release has 150 annotated WSIs and the remaining 50 are test data.
Additional cautions
Some WSIs contain multiple biopsy fragments while only a selected sample is annotated; follow the organizer's exclusion guidance.
No standalone reuse license is stated on the reviewed official pages. Verify permission before redistribution.
Preserve the patient/sample split used by the challenge.
Use with care: The released pyramidal TIFFs were converted from 40× NDPI originals and retain 10× data at approximately 0.92 µm/px; they are not the scanner-native files.
Additional cautions
H&E and IHC slides are adjacent sections, not pixel-identical images. Registration must accommodate tissue deformation, missing tissue, and stain-specific appearance.
The official split contains 750 training, 100 validation, and 303 test patients; use the set column with case ID because an ID alone is not globally unique across split folders.
Challenge landmark targets and feedback rules differ from the openly downloadable images; consult the ACROBAT challenge when reproducing leaderboard results.
Use with care: The multi-scanner subset contains repeated digital scans of the same glass specimens. Group every scanner replica by physical specimen before splitting.
Additional cautions
Registered masks were transferred across scanner versions and can contain registration error.
Test labels remain hidden and the challenge protocol governs direct comparison.
Counts describe several distinct subsets; do not add scan counts to case counts.
Expert-labelled H&E breast whole-slide images and regions of interest for seven-class breast-lesion detection and carcinoma subtyping.
Diagnosis & gradingTissue segmentationbreast
189patientswhole-slide release
547slideswhole-slide release
4,539ROIscompanion labelled ROI set
Release
Database release accompanying the 2022 paper
Units
patients, slides, ROIs
Modalities
whole-slide image, region annotation
Formats
SVS, PNG, JSON
Last verified
License / terms
CC0 1.0
Use with care: Registration with complete identity and affiliation information is required before download.
Additional cautions
The companion set contains 4,539 labelled regions of interest; these are ROIs, not independent patients or raw WSIs, so they are not entered as patch counts.
ROIs inherit slide and patient identity. Preserve the official patient-aware splits and never randomly split ROIs.
Confirm the current site rules and required dataset-paper citation even though the publication describes the data as CC0.
Use with care: CAMELYON17 contributes 1,000 slides from 200 patients, with five slides per patient; CAMELYON16 is reused as lesion-level training data, so family totals must not be formed by naive addition.
Additional cautions
Keep the official patient grouping and train/test definitions; a slide-random split leaks patient and centre information.
CAMELYON16 is also the source of derived datasets such as PatchCamelyon, which are not independent external validation cohorts.
Test labels and challenge-evaluation conventions differ between CAMELYON16 and CAMELYON17.
The current official Data page states CC0, while a legacy official download page has shown CC BY-NC-ND 4.0; record the exact host and terms presented when you obtain the files.
TCIA Data Usage Policy and collection-specific terms; linked molecular data may have separate policies
Use with care: CPTAC is a family of disease-specific TCIA collections; record collection names, versions, and download manifests rather than a mutable aggregate count.
Additional cautions
The former CPTAC pathology-download index was deprecated in April 2026; use the current TCIA collection browser and filter for CPTAC program collections.
Pathology images, radiology, proteomics, genomics, and clinical data may live in different NCI services and require careful case-identifier linkage.
Verify the data-usage policy and citation requirements on every selected TCIA collection page before download or redistribution.
A large native-slide neuropathology atlas spanning common and rare brain tumours, controls, clinical descriptors, and a deliberately broad diagnostic taxonomy.
Archives & pretrainingDiagnosis & gradingbraincentral nervous system
2,880patientstumour and control cohort reported by the data paper
EBRAINS Human Data Gateway agreement; verify the current non-commercial human-data terms
Provenance / overlap
Medical University of Vienna neuropathology archive, 1995–2019
Use with care: Access to the human slide data uses the EBRAINS controlled-data workflow and data-use agreement; the paper's open-access license is not automatically the dataset license.
Additional cautions
Diagnostic labels follow the 2016 revised WHO classification and predate the current molecularly integrated CNS tumour taxonomy.
Rare entities can contain very few cases, while strategically oversampled entities have many slides; report per-class patient counts.
Some patients contribute multiple slides. Split by patient and retain temporal/specimen context.
The cohort includes 47 non-tumour controls; define whether controls form a diagnostic class before evaluation.
A mixed whole-slide and tissue-microarray lymphoma cohort linking H&E and IHC morphology to clinical, cytogenetic, treatment, follow-up, and survival variables.
CC BY-NC 4.0 plus TCIA citation and data-usage requirements
Provenance / overlap
Stanford Hospital DLBCL cohort
Use with care: The headline 246 images mix 204 H&E WSIs from 149 cases with 42 multi-stain TMA scans covering 209 cases; they are not one homogeneous WSI count.
Additional cautions
Whole-slide, TMA, ROI, patch, and nucleus products are nested. Split by patient before selecting any derived product.
The release is non-commercial and requires TCIA attribution.
Survival analyses must document endpoint, censoring, missingness, treatment variables, and whether derived morphology features were selected using the evaluation cohort.
Spatial transcriptomics and matched H&E section images from HER2-positive breast cancers, with processed expression, coordinates, and limited expert morphology annotations.
Spatial omicsMolecular targets & outcomesbreast
8patientsHER2-positive breast-cancer patients
36slidesspatial-transcriptomics tissue sections
13,620spotsapproximately, across released sections
HER2ST source cohort; also represented in later aggregations such as HEST
Use with care: The histology products are brightfield section JPEGs rather than scanner-native pyramidal whole-slide files.
Additional cautions
No authoritative machine-learning split is supplied. Keep all sections and spots from one patient in the same partition.
Detailed pathologist morphology annotations cover only a subset of the sections.
The archive access instructions, including published passwords, are maintained by the authors; cite the Zenodo CC BY 4.0 data release rather than inferring rights from the code repository.
Audit sample identifiers before combining this release with HEST.
Use with care: The official page reports 510 cases in total, while some publications report 509 slides; preserve the release-specific discrepancy instead of silently changing it.
Additional cautions
Each MIRAX slide requires its matching DAT directory. Moving only the MRXS file produces an incomplete WSI.
Use the fixed challenge split and class prevalence when comparing with the leaderboard.
The no-derivatives, non-commercial license is restrictive; verify that any intended redistribution or transformed-data release is permitted.
A standardized collection of spatial-transcriptomics samples paired with aligned H&E whole-slide images, plus a benchmark for morphology-to-gene-expression prediction.
Use with care: The official repository reported 1,276 paired samples at verification time; HEST is actively versioned, so cite and pin the exact release.
Additional cautions
Hugging Face login, contact-information sharing, and acceptance of the dataset conditions are required.
A spatial-transcriptomics sample is not necessarily equivalent to a unique patient or case; preserve source-dataset and sample-level provenance.
The full release is larger than 2 TB; use the official query/download tooling for selected samples instead of mirroring the collection.
Upstream source datasets can carry additional attribution or usage requirements.
A fully region-annotated gastric endoscopic-submucosal-dissection WSI cohort spanning inflammation, metaplasia, dysplasia, and early cancer histotypes.
Tissue segmentationDiagnosis & gradingstomach
44patientspublic HiESD cohort
104slidespublic HiESD cohort
308specimenscontinuous tissue strips on public slides
Release
Figshare Version 1
Units
patients, specimens, slides, ROIs
Modalities
whole-slide image, region annotation, quality-control mask
Formats
SVS, XML annotations, PNG masks and thumbnails, CSV mappings
Last verified
License / terms
CC BY 4.0
Use with care: The two external validation cohorts—155 slides from 20 patients—are private and are not part of the public release.
Additional cautions
The authors explicitly describe the annotations as region-level rather than precise individual-gland boundaries.
Several tissue strips can share one slide and several slides can share one patient; split at patient level.
Use the supplied QC masks and inspect scanner/site effects before comparing with the private external results.
Official subset table snapshot verified 2026-07-27
Units
cases, slides, reports
Modalities
whole-slide image, report, clinical label
Formats
TIFF, JSON
Last verified
License / terms
CC BY-NC 4.0
Use with care: Hugging Face login, contact-information sharing, and acceptance of the dataset conditions are required before accessing the hosted files.
Additional cautions
Counts have changed across the paper, website, and hosted releases; the values here follow the official GitHub subset table verified on 2026-07-27 and must be pinned to a Hugging Face revision for experiments.
Do not confuse the CC BY-NC HISTAI release with HistAI's separately licensed commercial CellDX cohorts.
The associated reports and instruction derivatives may contain label, translation, or slide-report alignment noise; audit a task-specific sample before training.
Case counts are not asserted to equal unique patients.
An independent 2026 data-quality audit should be reviewed alongside the original release before high-stakes use.
A multilingual instruction-tuning resource generated from HISTAI slide-report pairs for whole-slide pathology description, reasoning, dialogue, and visual question answering.
WSI-languagebreastcolon
24,259casesreleased instruction rows
Release
Hugging Face release accompanying MIDL 2026
Units
cases, slides, reports
Modalities
WSI-linked instructions, report, conversation
Formats
JSONL, Parquet
Last verified
License / terms
CC BY-NC 4.0
Provenance / overlap
HISTAI
Use with care: The instruction repository links to HISTAI filenames but does not redistribute the raw whole-slide images; obtain those separately under the HISTAI terms.
Additional cautions
The 24,259 rows contain nested synthetic conversational attributes and more than 1.1 million instruction-response pairs; row count is not instruction-pair count.
Instructions, translations, reasoning traces, and treatment-oriented answers are synthetic and require expert review before evaluation or clinical research.
The source reports and generated records inherit HISTAI slide-report alignment and label-quality limitations.
Use case-level splitting and check for overlap with any model pretraining that used HISTAI.
A colorectal-biopsy cohort providing native MIRAX WSIs, diagnostic labels, and a large explicitly derived patch product for screening-oriented modeling.
A multi-domain mitotic-figure benchmark of H&E regions from human and canine tumours, spanning scanners, laboratories, species, and seven tumour types.
Cells, mitosis & TILsQC & domain shiftbreastlung
503specimensone source specimen per ROI
503ROIsreleased two-square-millimetre regions
26,288cellsmitoses plus hard-negative cells
Release
Figshare collection and Scientific Data 2023 release
Units
specimens, ROIs, cells
Modalities
region image, cell annotation
Formats
TIFF, JSON, SQLite, CSV
Last verified
License / terms
CC0 1.0 for the Figshare data items; MIT for repository code
Provenance / overlap
Multiple human and canine institutional cohorts
Use with care: MIDOG++ publishes one 2-square-millimetre region of interest from each of 503 source specimens, not the raw whole-slide images.
Additional cautions
The cohort mixes human and canine specimens, seven tumour types, laboratories, and scanners; preserve the supplied cross-validation/domain metadata and report species-specific results.
The release contains labels for 11,937 mitotic figures and 14,351 hard-negative imposter cells established through pathologist consensus.
The Figshare data items are CC0, while the GitHub implementation is MIT; cite the Scientific Data paper even though CC0 does not require attribution.
A multi-organ nuclei-segmentation benchmark of H&E image regions sampled from TCGA slides, with expert nuclear-boundary annotations.
Cells, mitosis & TILsmulti-organ
44ROIs30 training and 14 test image regions
21,623nucleichallenge paper annotations
Release
MoNuSeg 2018 challenge release
Units
slides, ROIs, nuclei
Modalities
region image, nucleus annotation
Formats
TIFF, XML
Last verified
License / terms
CC BY-NC-SA 4.0
Provenance / overlap
TCGA
Use with care: MoNuSeg contains 1000-by-1000 image regions derived from TCGA WSIs, not the source slides; it can overlap a TCGA experiment even when filenames differ.
Additional cautions
The public benchmark has 30 training images and 14 test images. Preserve the official split and do not tune on the released test annotations.
The historical challenge prohibited external training data; distinguish strict challenge reproduction from a new study with a separately declared protocol.
The non-commercial, share-alike license applies to the challenge data; follow its attribution and derivative-distribution conditions.
A widely used nine-class colorectal tissue benchmark with color-normalized and non-normalized WSI-derived patches plus a patient-independent validation cohort.
A breast nucleus detection, classification, and segmentation resource with corrected, uncorrected, and multi-rater annotation products from TCGA slides.
Overlapping small- and large-field-of-view H&E patch pairs from TCGA, with cell point labels and tissue segmentation annotations for contextual cell detection.
paired region image, cell annotation, tissue annotation
Formats
PNG, JSON
Last verified
License / terms
CC BY-NC 4.0 plus the Lunit OCELOT Terms and Conditions
Provenance / overlap
TCGA
Use with care: OCELOT v1.0.1 distributes 667 paired fields of view derived from 304 TCGA WSIs; the raw source WSIs are not included and TCGA overlap must be audited.
Additional cautions
Use Zenodo version 1.0.1: it excludes test cases 586, 589, 609, and 615 because the maintainers found under-annotation.
The current Zenodo file is openly downloadable, but use remains non-commercial and subject to the linked Lunit terms.
Keep each small cell-detection field paired with its corresponding large tissue-context field and relative coordinates.
H&E liver whole-slide images from hepatocellular-carcinoma cases for whole-tumour segmentation and viable-tumour-burden estimation.
Tissue segmentationDiagnosis & gradingliver
100patientscomplete challenge cohort
100casescomplete challenge cohort
100slidescomplete challenge cohort
Release
PAIP 2019 challenge release
Units
patients, cases, slides
Modalities
whole-slide image, tumour annotation
Formats
SVS, XML, TIFF
Last verified
License / terms
CC BY-NC 4.0; access remains subject to the PAIP Data Use and Confidentiality Agreement
Use with care: Access requires a Grand Challenge account, challenge membership, and an approved signed DUA; download credentials and the gated slide files must not be redistributed.
Additional cautions
The official split is 50 training, 10 validation, and 40 test WSIs; validation and test annotations are withheld for challenge evaluation.
For training data, the TIF binary masks—not the accompanying XML polygons—are declared to be the segmentation ground truth.
The release selects one slide from each patient; keep the official case identifiers and splits when comparing with challenge results.
CC BY-NC 4.0; access remains subject to the PAIP Data Use and Confidentiality Agreement
Use with care: Access requires a Grand Challenge account, joining PAIP 2020, and an approved DUA; the organizers provide the gated link and credentials by email.
Additional cautions
The official split is 47 training, 31 validation, and 40 test WSIs; validation and test ground truth remains reserved by the challenge committee.
MSI-high classification is the primary task, but tumour-area segmentation was mandatory for a valid challenge submission.
The historical challenge disallowed external training data except ImageNet-pretrained models; state explicitly whether a new experiment reproduces or departs from that protocol.
CC BY-NC 4.0; access remains subject to the PAIP Data Use and Confidentiality Agreement
Use with care: Access requires a Grand Challenge account, joining PAIP 2021, and an approved DUA; the gated link and credentials are delivered after review.
Additional cautions
The official split is 150 training, 30 validation, and 60 test WSIs, balanced across the three organ groups; only training annotations are provided.
Test data can include cases without tumour, so the evaluation population is not identical to a tumour-only training cohort.
The site alternates between “pancreas” and “pancreatobiliary tract” wording; retain the official case metadata rather than collapsing that group into an inferred organ label.
The historical challenge prohibited external training data except ImageNet-pretrained models; document any modern protocol that changes this rule.
A multi-centre development collection of H&E prostate-biopsy whole-slide images for prostate-cancer detection and Gleason/ISUP grade prediction.
Diagnosis & gradingQC & domain shiftprostate
10,616slidespublic model-development set
Release
PANDA public model-development release
Units
patients, cases, slides
Modalities
whole-slide image, grade label, region annotation
Formats
TIFF, CSV
Last verified
License / terms
CC BY-NC-SA 4.0
Provenance / overlap
Radboud University Medical Center; Karolinska Institutet
Use with care: The 10,616-slide count describes the public model-development set, not the blinded tuning or external-validation cohorts reported in the paper.
Additional cautions
Kaggle account access and acceptance of the competition data rules are required.
Patient identifiers suitable for leakage-safe grouping are not uniformly exposed; do not assume that a random slide split is a patient split.
Respect the non-commercial ShareAlike license and cite the PANDA challenge paper.
A pan-cancer collection of H&E image patches with instance masks and five clinically relevant nucleus classes for joint nuclei segmentation and classification.
CC BY-NC-SA 4.0; research use only and commercial use prohibited by the Warwick usage rules
Provenance / overlap
TCGA; GTEx and other source material documented by the authors
Use with care: PanNuke distributes pre-extracted image patches and nucleus masks, not the source whole-slide images; do not describe it as a downloadable WSI cohort.
Additional cautions
The current Warwick release is three fold archives. Preserve those folds for published comparisons and record each archive checksum and array shape after download.
The 7,904-patch and 205,343-nucleus totals describe the three-fold release; verify array shapes and archive checksums because downstream summaries sometimes report different aggregates.
Audit provenance before combining PanNuke with TCGA or other nuclei benchmarks because the patches were assembled from multiple source collections.
A multi-scanner prostate-biopsy WSI resource from an underrepresented Middle Eastern population with three independent pathologist grades per physical slide.
Diagnosis & gradingQC & domain shiftprostate
185patientscomplete release
339slidesunique physical glass slides
1,017imagesthree scanner files per physical slide
Release
BioImage Archive S-BIAD2323
Units
patients, slides, images
Modalities
whole-slide image, Gleason grade, scanner replica
Formats
SVS, NDPI, tabular grades and metadata
Last verified
License / terms
CC BY 4.0
Provenance / overlap
Erbil prostate-biopsy cohort
Use with care: The 1,017 files are three scans of 339 physical slides. A scan-random split would create direct cross-scanner leakage.
Additional cautions
Split by patient and physical glass-slide identity, then use scanner replicas only within the assigned fold.
The release preserves three independent pathologist assessments rather than forcing a single consensus label; define the target before training.
This 2026 release should be pinned by BioImage Archive accession and file manifest.
A TCGA-linked whole-slide caption resource that derives pathology descriptions from diagnostic reports and exposes slide identifiers for WSI-language representation learning.
CSV metadata, text caption, NPY feature, linked SVS at GDC
Last verified
License / terms
No standalone data license; TCGA/GDC terms govern slides and reports, while repository code has separate terms
Provenance / overlap
TCGA
Use with care: The resource links captions and derived features to TCGA material; it does not redistribute the native whole-slide files.
Additional cautions
The authors describe roughly ten thousand WSI-text pairs, but the repository does not expose one immutable release manifest with a stable aggregate count. Recompute counts from the exact revision used.
Captions are derived from diagnostic reports through OCR and language-model processing, so alignment, hallucination, and label-leakage audits are required.
Use patient-level TCGA identifiers for splitting and check overlap with every other TCGA-derived training or evaluation cohort.
A software repository license does not grant rights to the underlying TCGA slides or clinical text.
A prostatectomy epithelium-segmentation cohort with H&E whole-slide exports, IHC-derived epithelial masks, corrected masks, and a region-level test set.
SVS WSI, Sedeen XML annotations, XLSX clinical table
Last verified
License / terms
CC BY 3.0 plus TCIA citation and data-usage requirements
Provenance / overlap
Sunnybrook Health Sciences Centre post-neoadjuvant breast resections
Use with care: Post-treatment residual morphology is a distinct domain from untreated diagnostic breast cohorts; define the treatment time point explicitly.
Additional cautions
Multiple slides can share a patient. Split using anonymized patient IDs rather than slide filenames.
Training and test folders have different annotation-rater designs; do not treat every region or cell label as equivalently adjudicated.
Clinical-table rows correspond to slides and contain repeated patient information. Aggregate endpoints and missingness at patient level.
A multi-institution, multi-organ whole-slide and diagnostic-report challenge for histopathology report generation, with fixed training and hidden evaluation partitions.
WSI-languageDiagnosis & gradingbladderbreast
10,494casesWSI-report pairs across all challenge partitions
8,494casestraining WSI-report pairs
1,000casestest phase 1 pairs
1,000casestest phase 2 pairs
Release
REG2025 challenge release
Units
cases, slides, reports
Modalities
whole-slide image, diagnostic report
Formats
TIFF at 20×, JSON report
Last verified
License / terms
Challenge-specific terms; no standalone redistribution license stated on the reviewed data page
Provenance / overlap
Five contributing institutions across five countries
Use with care: This is a historical challenge release with fixed competition rules and hidden evaluation labels; benchmark claims must name the phase and rule set.
Additional cautions
The distributed 20× TIFF files are standardized challenge representations rather than guaranteed scanner-native source WSIs.
The organizer posted file corrections and image-decoding guidance during the challenge. Preserve the final manifest and audit TIFF photometric metadata.
No standalone dataset license was visible on the reviewed official data page; verify reuse and redistribution rights with the organizers.
Keep institution and case boundaries intact, and audit continuity or overlap before combining REG2025 with later REG cohorts.
CC BY-NC-SA 4.0 plus active REG2026 challenge rules and publication restrictions
Provenance / overlap
Contributing institutions in Germany, India, Japan, Korea, and Türkiye
Use with care: REG2026 is active and mutable: counts, corrections, rules, and deadlines can change. Record the download date and final challenge manifest.
Additional cautions
The challenge rules prohibit participant papers or preprints using the data until the official challenge meta-paper or proceedings become public; verify the current rule before dissemination.
The organizers have announced revised slides and reasoning annotations. Apply all official corrections and retain checksums.
The distributed 20× TIFF files are standardized challenge representations rather than guaranteed scanner-native source WSIs.
Treat the reports and reasoning traces as research annotations, not unquestioned clinical ground truth, and audit continuity with earlier REG cohorts.
A prostate grading resource with global Gleason labels and region-level grade annotations distributed as slide-derived image products.
Diagnosis & gradingTissue segmentationprostate
Release
Mendeley Data Version 2
Units
slides, patches, ROIs
Modalities
slide-derived image, Gleason annotation
Formats
JPG image products, annotation tables
Last verified
License / terms
CC BY 4.0
Provenance / overlap
Hospital Clínico Universitario de Valencia prostate-biopsy material
Use with care: Despite the dataset title, the downloadable archive is commonly consumed as extracted JPG image products rather than scanner-native pyramidal WSIs.
Additional cautions
Published SICAPv2 slide, patient, and patch totals vary by preprocessing convention; inspect Version 2 and report the exact manifest rather than copying an aggregate count.
All patches and grade regions from one source patient or slide must remain in the same split.
Do not compare patch-level results directly with whole-slide PANDA or AGGC22 endpoints.
A native whole-slide cohort of axillary lymph nodes with patient-linked slide labels for evaluating breast-cancer metastasis detection and weakly supervised learning.
A breast-cancer cohort of histology sections paired with spatial transcriptomics, created for patient-held-out prediction of local gene expression from tissue morphology.
A breast-cancer tumour-microenvironment benchmark combining WSI tissue regions, lymphocyte and plasma-cell annotations, tumour bulk, and slide-level TIL scores.
Cells, mitosis & TILsTissue segmentationbreast
195slidesWSIROIS subset
93slidesWSIBULK subset
82slidesWSITILS subset
Release
Public TIGER training release
Units
patients, slides, ROIs, cells
Modalities
whole-slide image, tissue annotation, cell annotation, outcome-linked TIL score
Formats
TIFF WSI, TIFF masks, XML annotations, CSV scores
Last verified
License / terms
CC BY-NC 4.0 for organizer data; TCGA-derived material retains TCGA terms
Use with care: Do not add the three subset counts and call the result an independent patient cohort: identities can overlap and the subsets carry different annotation products.
Additional cautions
WSIROIS includes 151 TCGA-BRCA slides and adapts BCSS and NuCLS annotations. It is not external validation for TCGA, BCSS, or NuCLS experiments.
All public data were standardized to approximately 0.5 µm/px; original TCGA annotations may not align directly with a differently scaled source slide.
Private leaderboard test sets are not part of the public download.
A gated uterine-pathology resource pairing native H&E whole-slide images with German diagnostic reports, English translations, slide-level descriptions, diagnostic categories, and case identifiers.
WSI-languageDiagnosis & gradinguterusfemale reproductive system
216casesclinical cases reported in arXiv v2
455slidesH&E WSIs and slide-level WSI–report pairs reported in arXiv v2
Release
Official Hugging Face release accompanying arXiv:2607.04020v2
Units
cases, slides, reports
Modalities
native H&E whole-slide image, German diagnostic pathology report, English report translation, slide-level textual description, slide-level diagnostic category
Formats
SVS whole-slide image, XLSX metadata
Last verified
License / terms
Data Use Agreement; users must sign in to Hugging Face and accept the gated dataset conditions
Provenance / overlap
Institute of Pathology, Technical University of Munich routine uterine-pathology cases
Use with care: The paper reports 216 cases and 455 slide-level WSI–report pairs; the live Hugging Face card separately displayed 674 GB, automatic gating, and a Data Use Agreement on 2026-08-20. Treat card size and access state as mutable and verify them again before use.
Additional cautions
Original German reports were translated to English with an offline Qwen-30B workflow and then manually quality-controlled. The English text is derived clinical language, not an independent report, and residual translation or alignment errors remain possible.
The slide-level paired release retains cases with at most three WSIs to reduce report-alignment ambiguity. Use the provided case identifiers for case-level splitting; never split slides from one case across train and evaluation.
The cohort comes from routine practice at one tertiary medical center and is dominated by normal or no-significant-abnormality pairs. It is a dataset description, not epidemiological prevalence or evidence of external generalization.
The paper's Methods text reports 0.25 µm/pixel, whereas Figure 2 displays 0.26 µm/pixel. Inspect slide metadata and pin the value actually used by the processing pipeline.
Only H&E SVS slides are reported; no IHC, special stains, or frozen sections are included. A public dataset page does not waive the gated DUA or permit redistribution.
A native H&E whole-slide subset linked by mapping files to a larger glioblastoma radiology, molecular, demographic, treatment, and outcomes collection.
CC BY 4.0 plus TCIA citation and data-usage requirements
Provenance / overlap
UPENN-GBM
Use with care: Only 34 subjects have pathology, despite the much larger radiology cohort. Never use the collection-header subject count as the WSI cohort size.
Additional cautions
Pathology and radiology identifiers are connected through explicit mapping files, and not every radiology subject has pathology or vice versa.
Multiple slides can belong to a pathology subject; preserve subject grouping.
Matched de novo and recurrent material may represent different disease time points. Define the temporal unit before multimodal or outcome analysis.
A lung-adenocarcinoma weakly supervised tissue-segmentation benchmark combining patch labels and dense validation/test pixels from institutional and TCGA slides.
Tissue segmentationDiagnosis & gradinglung
87slides67 institutional plus 20 TCGA source WSIs
10,091patchespatch-labelled training set
Release
WSSS4LUAD challenge release
Units
slides, patches, annotations
Modalities
patch image, patch label, dense tissue mask
Formats
PNG patches, pixel masks, patch-level labels
Last verified
License / terms
Challenge terms; no standalone dataset license stated on the reviewed official record
Provenance / overlap
Guangdong Provincial People's Hospital; TCGA-LUAD
Use with care: The released benchmark is patch and mask data derived from 87 WSIs, not a native-slide cohort.
Additional cautions
Twenty source WSIs are from TCGA-LUAD; audit overlap with TCGA pretraining or evaluation.
Follow the official split and distinguish weak patch labels from dense validation/test labels.
No standalone dataset license was found on the reviewed official challenge page.
Scanner-native or source-resolution slide files such as SVS, NDPI, or MRXS
Native-WSI cohort, subject to the record’s access and QC notes
Converted full-slide
A complete slide converted to another pyramidal or standardized format
Full-slide data; disclose conversion, resolution, and compression
Flat full-slide image
A downsampled or flattened JPEG/PNG representation of a complete slide
Full-slide representation, not scanner-native WSI
WSI linked separately
Metadata, labels, reports, or features with identifiers for slides hosted elsewhere
Linked resource; acquire and license the source WSIs separately
WSI-derived image / ROI
Patches, regions, fields of view, cells, masks, or section images extracted from source slides
WSI-derived benchmark, not an independent WSI cohort
This distinction is essential. For example, PatchCamelyon is derived from CAMELYON16,
NuCLS is nested within BCSS/TCGA material, and CoNIC is a processed child of Lizard.
Reporting them as independent external cohorts would create leakage.
The linked computational-pathology curriculum
explains how these tasks relate to WSI engineering, multiple-instance learning, foundation
encoders, cells and tissue, multimodal models, spatial methods, and clinical evaluation.
TUM-Uteria(2026, arXiv
preprint) describes 216 routine clinical cases and 455 H&E SVS–report pairs from one
tertiary center. Original German diagnostic reports were translated to English, aligned
to slides with LLM assistance plus manual and pathologist review, and released with case
identifiers for grouped splitting. The official Hugging Face dataset
card was automatically
gated by a Data Use Agreement and displayed 674 GB when checked on 20 August 2026.
Those are live-card observations, not immutable paper claims: re-check the manifest,
terms, size, and access state before use.
Provenance families you must not split across roles
Treat parent and derived resources as one overlap family until identifiers prove otherwise:
TCGA family: TCGA → TUPAC16, BCSS → NuCLS, plus portions or source material in
TIGER, OCELOT, MoNuSeg, MoNuSAC, PanNuke, Lizard, PathText, and other derived sets.
CAMELYON family: CAMELYON16 → PatchCamelyon; CAMELYON16 and CAMELYON17 also
have institutional and scanner structure that belongs in the split.
Lizard family: DigestPath, CRAG, GlaS, PanNuke, CoNSeP, and TCGA contribute to
Lizard; CoNIC is then sampled from Lizard.
Spatial family: ST-Net and HER2ST are upstream studies that may also be represented
through HEST. A standardized HEST sample is not a new independent patient.
Repeated scans: PAR and the multi-scanner AGGC22 subset contain multiple scans of the
same physical glass slide. MITOS-ATYPIA-14 depicts the same physical material on two
scanners. Split by patient/specimen/physical slide before scanner.
Challenge generations: REG2025 and REG2026 are separate rule sets, but later releases
may have institutional or material continuity. Do not assume independence without a
manifest audit.
source cohorts on each card is a warning edge, not a complete row-level deduplication
graph. When external validity matters, compare patient, case, slide, institution, and
source accession identifiers directly.
State the prediction target and unit: cell, patch, ROI, slide, specimen, case, or patient.
Pin the exact release, manifest revision, download date, official URL, and license/terms.
Build a hierarchy table linking patient → case → specimen → physical slide → scan →
ROI/patch/spot/report.
Split at the highest shared identity. Add site, scanner, or acquisition batch grouping
when the scientific claim requires it.
Resolve overlap with pretraining data and every source/derived cohort in evaluation.
Record stain, magnification or MPP, scanner, file conversion, tissue/QC filters, and
label provenance.
Keep official challenge splits and hidden-test rules intact; never tune on a public
leaderboard result and call it untouched test performance.
Report exclusions and missing modalities. “One cohort” rarely means every patient has
every slide, report, assay, and endpoint.
For spatial data, add registration accuracy and spot/section/patient nesting. For
WSI-language data, add text provenance, de-identification, synthetic-generation method,
slide-report alignment, and contamination checks. For survival or response, add index
date, censoring, treatment context, and endpoint definition.
The Atlas can include patches or fields of view only when their derivation from pathology
slides is documented and scientifically important. It does not classify resources such
as BreakHis, PathVQA, PathMMU, Quilt-1M, or OpenPath as WSI datasets: they are useful for
other pathology questions, but their images are microscopy crops, textbook figures, or
web/social-media material without a reliable slide and patient hierarchy.
“Comprehensive” here means broad, reviewed coverage with explicit boundaries—not a claim
that every historical pathology image collection is present. Dataset releases, licenses,
challenge rules, and access pages change. To propose a missing WSI family or a correction,
send the maintainer—or, if you have source-repository access, the dataset-suggestion
issue form—the official host, primary paper, release, hierarchy, and license evidence.
