PathTrove data atlas

What public pathology data exists, and what it can really be used for

355

data resources

anatomical sites

13,490

report entries

16,936

source quotations

Scale and growth

83% of resources were released in 2020 or later

Resources per release year. 2026 only covers January to June.

2025 alone added 73 resources, more than everything released before 2019.

≤2017201820192020202120222023202420252026 H1

Most resources are datasets; one in four comes from a challenge or benchmark

Challenges and benchmarks often reuse images from existing datasets.

Resource type

2616727

DatasetChallengeBenchmark

Coverage

12 sites have 30 or more resources each; 7 sites have fewer than 10

A resource can be tagged with several sites.

Breast

102

Lung

Colorectum

Brain

Kidney

Prostate

Liver

Lymph node

Skin

Bladder

Stomach

Cervix

Bone

Esophagus

Blood

Ovary

Pancreas

Adrenal gland

Heart

Uterus

Head and neck

Soft tissue

Bile duct

Eye

Spleen

Pleura

Testis

Thyroid

Thymus

GI tract

≥ 30 resources10–29 resources< 10 resources

58 resources involve 23 rare diseases (Orphanet codes); glioblastoma has the most

The 12 rare diseases with the most resources; a resource can involve several.

Glioblastoma

Cholangiocarcinoma

Thymoma

Diffuse large B-cell lymphoma

Uveal melanoma

Adrenocortical carcinoma

Multiple myeloma

Cervical squamous cell carcinoma

Ependymoma

Clear cell ovarian carcinoma

Endometrioid ovarian carcinoma

Mucinous adenocarcinoma

Cases come mainly from Europe, East Asia and North America

A resource can include cases from several regions. 117 resources do not say whether they are single- or multi-center.

Europe

≈90

East Asia

North America

South Asia

Five more regions (Southeast Asia, Latin America, Middle East, Africa, Oceania) have fewer than 10 resources each.

Single or multiple centers

114124

Multi-centerSingle-center

Image objects and modalities

From 3D images to patches, counts differ by four orders of magnitude

Counts are image instances, not patients.

10K

100K

10M

100M

3D images

1,475

15 resources

Field-of-view images

13,235

3 resources

TMA cores

25,353

7 resources

Whole-slide images

343,865

136 resources

ROI images

11.14M

84 resources

Patches

28.71M

81 resources

HorizontalDot area

Beyond H&E, six other kinds of stain, imaging and molecular data

A resource can contain several modalities.

H&E

268

Special stains

IHC / mIHC

IF / mIF

Molecular data

Spatial transcriptomics

CODEX

Among resources with molecular data, SurGen has the most patients and is fully open; sequencing data mostly requires an application

Bars show patients per resource; tags show how open the original data is.

Example of controlled access: PTRC-HGSOC has 158 patients, with 120 WGS and 106 RNA-seq samples deposited in dbGaP.

SurGen (colorectal, MSI/KRAS status)

843Fully open

BOEHMK (ovary, DNA/mutations)

444Partially open

SegLungTCGA (lung, DNA/mutations)

411Fully open

MSKMINDProjectM (lung, DNA/mutations)

247Partially open

PD-L1 (lung, DNA/mutations)

217Partially open

TransNEO (breast, DNA and RNA)

168Partially open

PTRC-HGSOC (ovary, WGS and RNA-seq)

158Partially open

FUSCC bladder cancer (WES)

75Partially open

AURORA (metastatic breast cancer, DNA and RNA)

55Partially open

Two aggregated sets are far larger than the newly acquired spatial transcriptomics data: 6 new cohorts total 207 samples

Samples per resource (slides for STimage-1K4M). The aggregated sets collect published data and can overlap with the newly acquired resources, so the numbers cannot be added.

13spatial transcriptomics resources

207samples in the 6 newly acquired resources

HEST-1k (multi-organ)

1,276aggregated set · samples

STimage-1K4M (multi-organ)

1,149aggregated set · slides

ST-Net (breast)

68newly acquired

GSE210616 (breast)

43newly acquired

her2st (breast)

36newly acquired

STHELAR (multi-organ)

31partly reuses existing data

CellHIST-Bench (multi-organ)

25newly acquired

GSE111672 (pancreas)

23newly acquired

CAMEO-Lung

23processed from existing data

CAMEO-Thymus

19processed from existing data

Human DLPFC Visium (brain)

12newly acquired

CAMEO-Breast

7processed from existing data

iSCALE (stomach, brain)

5partly reuses existing data

Among fluorescence and CODEX multiplex imaging resources, two exceed 1,000 patients

Bars show patients per resource.

SegPath (multi-organ)

1,583patients · 158,687 patches

cell-niches-data (lung)

1,168patients · 4,429 TMA cores

CRC FFPE CODEX (colorectal)

35patients · 251 ROIs

CODEX-HCC (liver)

15patients · 37 markers · 646 images

Pairing and tasks

138 resources link images across stains or modalities

Resources by type of correspondence.

Pixel-aligned images support registration and virtual staining; different images of the same case link morphology to omics or other images.

Pixel-aligned

Different images of the same case

Partly aligned regions

Algorithm-generated pairs

Several markers on one section

Classification and segmentation dominate; clinical and molecular prediction tasks are rare

Resources per officially declared task; a resource can declare several. Pink: clinical and molecular prediction tasks.

Classification

197

Segmentation

120

Detection

Survival

Regression

Generation

VQA

Retrieval

Registration

Counting

Captioning

Staining

Reasoning

Molecular prediction

Treatment response

Clinical and acquisition information

Along the patient timeline, data exists from diagnosis to follow-up outcomes

Each node is a separate resource count.

Diagnosispathology

Biomarkers / receptors

Stage

variable

Grade

variable

Treatmentbaseline, perioperative, therapy

Pre-treatment / baseline

Intra- / perioperative

Treatment

Follow-upafter treatment

Post-treatment / follow-up

Overall survival

outcome

Recurrence / relapse

outcome

187 resources have no structured clinical variables; 168 do, with 694 fields in total

Horizontal: clinical fields per resource; bar height: number of resources.

Each field's meaning and values are written in the report, so the table can be searched by variable.

187

012–34–67–1011+

Structured clinical fields per resource

19DLBCL-Morph (lymphoma): IHC, FISH, IPI, survival

12HMU-CRC-Hist550K (colorectal cancer): TNM, MSI, OS

239 resources record the scanner maker; 23 contain slides from several scanners

Resource counts; a resource can record several scanner makers.

Multi-scanner designs range from controlled rescans (Multi-Scanner SCC, skin squamous cell carcinoma: 44 samples × 5 scanners) to routine archives.

Leica / Aperio

3DHISTECH

Hamamatsu

Olympus

Zeiss

Nikon

Philips

Others: Akoya 4 · Huron 3 · KFBio 3 · Ventana 2

Specimen sampling and processing

FFPE

100

Surgical resection

Biopsy

Frozen section

Quality, origin and access

Quality control is mostly manual review; 92 resources do not describe any

78 resources name what they checked, most often annotation quality, then staining, focus, registration and label consistency.

Reported quality control

101766992

ManualManual + automatedOnly some steps documentedAutomated only 9Not described

Most labels are new, but 138 resources reuse existing labels

Image origin and label origin are recorded separately: new labels on reused images are common.

Where the labels come from

20710132

NewPartly reuse existing labelsProcessed from existing labelsReorganized existing labels 5

Where the images come from

205506228

Newly acquiredPartly reuse existing imagesProcessed from existing imagesReorganized existing images

140 resources reuse existing images and may share cases with the datasets they came from

185 fully open, 149 partially open, 14 not open; most downloads still need extra steps

How open the original data is

53.2%42.8%4.0%

Fully openPartially openNot open 4.0%

按本图已分类记录计算占比。

Download test from mainland China (September 2026)

44.6%45.1%8.2%2.2%

Direct download worksNeeds a proxyCannot downloadOther 2.2%

Of the 30 that cannot be downloaded, 19 first require registration or approval

按本图已分类记录计算占比。

Data is hosted on many platforms, not just a few big ones

Counted by each guide's primary hosting platform. "Other" covers Dataverse, Grand Challenge, NCI GDC, NCBI GEO, Mendeley Data, EBI BioStudies and institutional servers.

Zenodo

Hugging Face

Figshare

TCIA

GitHub

Google Drive

AWS S3

Kaggle

Other

Try data search →