GenSegDataset — A Unified 2D Medical Image Segmentation Benchmark

GenSegDataset is a standardized collection of 10 public 2D medical image segmentation datasets spanning 8 imaging modalities, re-packaged into a single consistent layout (identical directory structure, mask encoding, split files, and metadata) so that segmentation models — and generative mask-conditioned data augmentation methods — can be trained and compared across modalities with one data pipeline.

This repository redistributes standardized derivatives of existing public datasets. Each subset retains the license and citation requirements of its original source — see Licensing & Attribution and please cite the original works.

---

Overview

Subset	Modality	Anatomy / Target	Classes	Channels	Sample size	Protocol	Train / Val / Test
cvc_clinicdb	Colonoscopy	Polyp	2	RGB	384×288	official	490 / 61 / 61
kvasir_seg	GI endoscopy	Polyp	2	RGB	~622×529 (var)	official	800 / 100 / 100
fives	Retinal fundus	Vessel	2	RGB	2048×2048	official	480 / 120 / 200
busi	Breast ultrasound	Tumor	2	RGB	variable	5-fold (fold01–05)	545 / 78 / 157 (fold01)
refuge2	Retinal fundus	Optic disc & cup	3	RGB	~2124×2056	official	400 / 400 / 400
acdc_png	Cardiac MRI (2D slices)	RV / Myo / LV	4	grayscale	~240×256 (var)	official	136 / 210 / 380
idridd_segmentation	Retinal fundus	Diabetic-retinopathy lesions	6	RGB	4288×2848	5-fold (fold01–05)	43 / 11 / 27 (fold01)
pannuke_semantic	Histopathology (H&E)	Nuclei (5 types)	6	RGB	256×256	3-fold (fold01–03)	2722 / 2523 / 2656 (fold01)
medsegdb_isic2018	Dermoscopy	Skin lesion	2	RGB	256×256	holdout	2582 / 369 / 737
medsegdb_kits19	Kidney CT (2D slices)	Kidney region (binary)	2	grayscale¹	256×256	5-fold (fold01–05)	2832 / 479 / 705 (fold01)

¹ medsegdb_kits19 images are grayscale in content but stored as 3-channel PNG; read them as grayscale (IMREAD_GRAYSCALE) for true single-channel input.

For cross-validation subsets (busi, idridd_segmentation, medsegdb_kits19: 5 folds; pannuke_semantic: 3 folds) every fold reuses the same images under a different train/val/test partition. holdout and official provide a single fixed partition.

---

Directory layout

Each subset is shipped as a single archive <subset>.tar at the repo root; extracting it yields the structure below.

GenSegDataset/
  <subset>.tar               # download & extract -> <subset>/...
  <subset>/                  # (after extraction)
    metadata.json            # subset-level metadata
    manifest.jsonl           # one JSON line per image: relative image/mask paths
    <protocol>/              # e.g. official | fold01..fold05 | fold01..fold03 | holdout
      train/ val/ test/
        images/              # input images (.png)
        masks/               # segmentation masks (.png)
  README.md                  # this card

• Pairing: an image and its mask share the same file stem (images/<id>.png ↔ masks/<id>.png); manifest.jsonl also lists the pairing explicitly with paths relative to the subset root.
• Modality / channels: input images are RGB (3-channel) except acdc_png (true grayscale) and medsegdb_kits19 (grayscale content, 3-channel container).

Mask encoding

Masks are single-channel uint8 label maps with values 0 … C-1 (0 = background), not 0/255. Semantic meaning per index (confirmed against each subset's metadata.json and the standardization scripts):

Subset	Class indices
cvc_clinicdb, kvasir_seg	0 background · 1 polyp
fives	0 background · 1 vessel
busi	0 background · 1 tumor (multi-instance masks merged)
medsegdb_isic2018	0 background · 1 lesion
medsegdb_kits19	0 background · 1 foreground (kidney region, binary)
refuge2	0 background · 1 optic disc · 2 optic cup
acdc_png	0 background · 1 right ventricle · 2 myocardium · 3 left ventricle
idridd_segmentation	0 background · 1 microaneurysms · 2 haemorrhages · 3 hard exudates · 4 soft exudates · 5 optic disc
pannuke_semantic	0 background · 1 neoplastic · 2 inflammatory · 3 connective · 4 dead · 5 epithelial

---

Usage

Download & extract a subset

from huggingface_hub import hf_hub_download
import tarfile

p = hf_hub_download("GenSegDataset/GenSegDataset", "cvc_clinicdb.tar", repo_type="dataset")
tarfile.open(p).extractall("GenSegDataset")        # -> GenSegDataset/cvc_clinicdb/...

Direct file access (after extraction)

import cv2, glob, os

root = "GenSegDataset/cvc_clinicdb/official/train"
img = cv2.imread(f"{root}/images/0001.png")                       # RGB input
msk = cv2.imread(f"{root}/masks/0001.png", cv2.IMREAD_GRAYSCALE)  # label map 0..C-1

With datasets (once the Parquet/loader build is published)

from datasets import load_dataset
ds = load_dataset("GenSegDataset/GenSegDataset", "cvc_clinicdb")   # config = subset
sample = ds["train"][0]            # {"image": PIL.Image, "mask": PIL.Image, ...}

The datasets-loadable build (Parquet with embedded image/mask + a config per subset) is added on top of the raw file mirror; until then use direct file access.

---

Standardization methodology

All subsets were converted to the unified layout above with a shared pipeline:

1. Format unification — images/masks re-encoded to .png; masks remapped to a contiguous 0 … C-1 label space.
2. Fixed, reproducible splits — official splits used where they exist (cvc_clinicdb, kvasir_seg, fives, refuge2, acdc_png); otherwise fixed k-fold (busi, idridd_segmentation, medsegdb_kits19: 5-fold; pannuke_semantic: official 3-fold) or a fixed holdout (medsegdb_isic2018). Splits are frozen in manifest.jsonl so results are reproducible.
3. Metadata — each subset carries metadata.json and a per-image manifest.jsonl.

The collection was assembled to benchmark (a) 2D segmentation backbones and (b) generative, mask-conditioned data-augmentation methods under one consistent interface.

---

Licensing & Attribution

This repository contains standardized derivatives of the datasets below. Each subset is governed by its original license; users must comply with the source terms and cite the original publications. Source links:

Subset	Source
cvc_clinicdb	CVC-ClinicDB (Bernal et al., 2015)
kvasir_seg	Kvasir-SEG (Jha et al., 2020)
fives	FIVES (Jin et al., 2022)
busi	BUSI (Al-Dhabyani et al., 2020)
refuge2	REFUGE / REFUGE2 (Orlando et al., 2020; Fang et al., 2022)
acdc_png	ACDC (Bernard et al., 2018)
idridd_segmentation	IDRiD (Porwal et al., 2018, 2020)
pannuke_semantic	PanNuke (Gamper et al., 2019, 2020)
medsegdb_isic2018	ISIC 2018 / HAM10000 (Codella et al., 2019; Tschandl et al., 2018)
medsegdb_kits19	KiTS19 (Heller et al., 2019, 2021)

---

Citation

If you use GenSegDataset, please cite this collection and the original source dataset(s) you use.

@misc{gensegdataset2026,
  title        = {GenSegDataset: A Unified 2D Medical Image Segmentation Benchmark},
  author       = {<authors>},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  note         = {Standardized collection of 10 public 2D medical segmentation datasets}
}

Original-source BibTeX (please verify before camera-ready)

@article{bernal2015cvcclinicdb,
  title={WM-DOVA maps for accurate polyp highlighting in colonoscopy},
  author={Bernal, Jorge and S{\'a}nchez, F Javier and Fern{\'a}ndez-Esparrach, Gloria and Gil, Debora and Rodr{\'i}guez, Cristina and Vilari{\~n}o, Fernando},
  journal={Computerized Medical Imaging and Graphics}, volume={43}, pages={99--111}, year={2015}}

@inproceedings{jha2020kvasirseg,
  title={Kvasir-SEG: A segmented polyp dataset},
  author={Jha, Debesh and Smedsrud, Pia H and Riegler, Michael A and Halvorsen, P{\aa}l and de Lange, Thomas and Johansen, Dag and Johansen, H{\aa}vard D},
  booktitle={MultiMedia Modeling (MMM)}, year={2020}}

@article{jin2022fives,
  title={FIVES: A fundus image dataset for artificial intelligence based vessel segmentation},
  author={Jin, Kai and Huang, Xingru and Zhou, Jingxing and others},
  journal={Scientific Data}, volume={9}, year={2022}}

@article{aldhabyani2020busi,
  title={Dataset of breast ultrasound images},
  author={Al-Dhabyani, Walid and Gomaa, Mohammed and Khaled, Hussien and Fahmy, Aly},
  journal={Data in Brief}, volume={28}, year={2020}}

@article{orlando2020refuge,
  title={REFUGE Challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs},
  author={Orlando, Jos{\'e} Ignacio and Fu, Huazhu and others},
  journal={Medical Image Analysis}, volume={59}, year={2020}}

@article{bernard2018acdc,
  title={Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: Is the problem solved?},
  author={Bernard, Olivier and Lalande, Alain and others},
  journal={IEEE Transactions on Medical Imaging}, volume={37}, number={11}, year={2018}}

@article{porwal2020idrid,
  title={IDRiD: Diabetic retinopathy -- segmentation and grading challenge},
  author={Porwal, Prasanna and Pachade, Samiksha and others},
  journal={Medical Image Analysis}, volume={59}, year={2020}}

@article{gamper2020pannuke,
  title={PanNuke dataset extension, insights and baselines},
  author={Gamper, Jevgenij and Koohbanani, Navid Alemi and others},
  journal={arXiv:2003.10778}, year={2020}}

@article{codella2019isic2018,
  title={Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the ISIC},
  author={Codella, Noel and Rotemberg, Veronica and others},
  journal={arXiv:1902.03368}, year={2019}}

@article{tschandl2018ham10000,
  title={The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions},
  author={Tschandl, Philipp and Rosendahl, Cliff and Kittler, Harald},
  journal={Scientific Data}, volume={5}, year={2018}}

@article{heller2021kits19,
  title={The state of the art in kidney and kidney tumor segmentation in contrast-enhanced CT imaging: Results of the KiTS19 challenge},
  author={Heller, Nicholas and Isensee, Fabian and others},
  journal={Medical Image Analysis}, volume={67}, year={2021}}

---

Maintenance notes

• Class indices/names above were confirmed from each subset's metadata.json, the actual mask label values, and the standardization scripts (tools/process_downloaded_segmentation_datasets.py). Most subsets' metadata.json omits an explicit num_classes/modality field (the two medsegdb_* subsets include them).
• Read acdc_png and medsegdb_kits19 as grayscale even though some files are stored as 3-channel containers.