README.md

metadata

license: cc-by-4.0
task_categories:
  - image-segmentation
tags:
  - medical
  - histopathology
  - whole-slide-imaging
  - veterinary
  - canine
  - dermatology
  - oncology
size_categories:
  - n<1K

CATCH — CAnine CuTaneous Cancer Histology

350 whole-slide images of canine skin tumors (H&E), covering 7 tumor subtypes with dense multi-class region annotations by a veterinary pathologist. Mirrored for the MedOtter segmentation benchmark.

Wilm F., Fragoso M., Marzahl C., Qiu J., Puget C., Diehl L., Bertram C.A., Klopfleisch R., Maier A., Breininger K., Aubreville M. — Pan-tumor CAnine cuTaneous Cancer Histology (CATCH) dataset, Scientific Data 9, 588 (2022). doi:10.1038/s41597-022-01692-w

Data DOI: 10.7937/TCIA.2M93-FX66 · Source: TCIA CATCH collection

What this mirror contains — read before using

The originals are 350 Aperio .svs slides totalling 522 GB at 0.2533 µm/px, distributed by TCIA behind an Aspera plugin. This mirror stores each slide rendered at the 4 µm/px pyramid level — the exact resolution the CATCH paper's own segmentation baseline operates at (512×512 px ≙ 2048×2048 µm) — as a lossless PNG, paired with a 13-class indexed mask rasterized at the same level. It is a derived, downsampled representation, not the original WSIs.

For full-resolution work, use TCIA. The complete original polygon annotations are included here as CATCH.json so any other pyramid level can be re-derived.

Slides	350
Patients	282
Resolution	4.05 µm/px (pyramid level 2, ≈16× downsample, all 350 slides)
Image size	median 6025×4644, max 11812×6046, mean 29.7 Mpx
Polygons	12,424
Classes	13 (+ label 0 = unannotated)
Splits	train 245 / val 35 / test 70 (official, patient-level)

Splits

The official split from the authors' CanineCutaneousTumors repo, balanced at 35 / 5 / 10 slides per subtype. It is patient-level: no patient appears in two splits (verified across all 282 patients).

A slide's patient is the filename prefix <Subtype>_<NN> — 41 patients contribute 2 slides, 8 contribute 3, 2 contribute 4, and 1 contributes 6. Group on patient_id, not on slide.

Labels

mask is a single-channel uint8 PNG. Label 0 means unannotated, not background — CATCH has no background class by design, and the authors exclude unannotated tissue from training and evaluation. Treat 0 as don't-care, or synthesize a background class by Otsu-thresholding the white point per slide, which is what the paper's baseline does.

ID	Class	Group	Slides present
0	unannotated	—	350
1	Bone	Tissue	21
2	Cartilage	Tissue	4
3	Dermis	Tissue	322
4	Epidermis	Tissue	321
5	Subcutis	Tissue	246
6	Inflamm/Necrosis	Tissue	149
7	Melanoma	Tumor	50
8	Plasmacytoma	Tumor	50
9	Mast Cell Tumor	Tumor	50
10	PNST	Tumor	50
11	SCC	Tumor	50
12	Trichoblastoma	Tumor	50
13	Histiocytoma	Tumor	50

Bone (21 slides) and Cartilage (4 slides) are too rare for class-averaged metrics — the authors exclude both from their own baseline. Do the same.

Exactly one tumor class occurs per slide, and it always equals the slide's subtype (verified 350/350). tumor_class_id / tumor_class_name give it directly, so a binary tumor-vs-rest target needs no lookup.

Polygons are hierarchical — rasterize in file order

Annotations nest: the dermis encircles a tumor mass, and islands of normal dermis sit inside the tumor. Masks here are rasterized in COCO file order, which is the authors' documented sort — "polygons are sorted in increasing order of their hierarchy level, i.e. polygons enclosed by another will be read out after their enclosing polygon" — so a later fill correctly overwrites the region it sits within.

⚠️ Do not sort by the area field instead. area is the shoelace area, so a polygon drawn as a ring around a tumor reports a smaller area than the blob it encloses, while a standard polygon fill (cv2.fillPoly, PIL ImageDraw.polygon) fills its outer boundary solid. Sorting area-descending therefore paints the ring last and buries the tumor completely. Measured on this data, it destroys the entire tumor annotation on Plasmacytoma_08_1, Trichoblastoma_31_2 and Trichoblastoma_34_1.

File order reproduces the per-class slide presence of the source polygons exactly for all 13 classes across all 350 slides; area-descending does not.

Columns

image · mask · slide · stem · subtype · patient_id · split · scanner · tumor_class_id · tumor_class_name · width · height · mpp · downsample · level0_width · level0_height · annotated_frac · classes_present

annotated_frac is the fraction of canvas carrying a label (median ≈ 0.50; much of the remainder is glass, not untraced tissue).

Overlap with other datasets — leakage warnings

• Multi-Scanner Canine Cutaneous SCC (Zenodo 7418555) re-scans 44 of CATCH's 50 SCC slides on 4 additional scanners (220 images, also distributed at 4 µm/px). It is not joinable by name — its files are renumbered scc_01…scc_44 and the shipped COCO/SQLite carry no CATCH provenance — but it is joinable by 4 µm/px image dimensions: each of its 44 Aperio-CS2 images matches exactly one CATCH SCC slide, 44/44 with zero ambiguity. The six SCC slides not re-scanned are SCC_08_1, SCC_11_1, SCC_12_3, SCC_16_1, SCC_27_1, SCC_28_1; exposure by split is train 30/35, val 5/5, test 9/10, over 29 of the 33 SCC patients.
• MIDOG++ / MIDOG 2022 Domain 4 is 50 canine cutaneous mast cell tumor cases from the same archive, scanner and resolution as CATCH's 50 MCT slides. The reuse is undocumented and no cross-reference ID exists — the naming schemes are not joinable. Treat the MCT subset as potentially contaminated if you also use MIDOG.
• CCMCT / MITOS_WSI_CCMCT (32 canine cutaneous MCT WSIs) may likewise overlap the MCT subset. Also unjoinable.
• No overlap with human histopathology sets (TCGA-derived, PanNuke, MoNuSeg, MoNuSAC, NuCLS, CoNIC, CAMELYON, …) — different species.

⚠️ Web summaries claiming "CATCH is on Zenodo as 4 µm/px TIFFs" are wrong; that record is the 44-slide Multi-Scanner SCC derivative, not CATCH.

Annotation provenance

One annotation tier is released. Pathologist M. Fragoso drew ~82% of the annotations; the remainder was drawn by three medical students and then reviewed for correctness and completeness by M. Fragoso. The distributed SQLite has exactly one entry in its Persons table — a single merged layer, with no algorithmic pre-annotation. Two further veterinary pathologists annotated one ROI on each of the 70 test slides for an inter-rater study; that data was never published and is not part of this dataset.

Reported reliability (paper Table 3, generalized conformity index): tumor 0.8514, epidermis 0.7512. Dermis/subcutis are the weakest pair, and inflammation/necrosis vs tumor is the other main confusion axis.

License

CC BY 4.0, as stated for all three data rows on the TCIA collection page, the authors' designated distribution channel.

⚠️ Discrepancy, disclosed for transparency: the licenses block inside the official CATCH.json declares Attribution-NonCommercial-NoDerivs 2.0. This appears to be a COCO-export template default rather than a deliberate choice, and it is contradicted by TCIA's own Data Access table and by the CC BY 4.0 paper. We treat the TCIA statement as controlling. If your use is commercial or derivative-heavy, verify with the authors first.

Users must also abide by the TCIA Data Usage Policy. Please cite the paper and the data DOI above.

Reproducing this mirror

Level 2 of each remote .svs is read via HTTP byte-range requests — a .svs is a pyramidal TIFF, so pulling only that level costs 0.79% of each file (4.1 GB total instead of 522 GB) and needs no Aspera client and no login. CATCH.json is then rasterized in file order at the same level.