MedSeg-7D: Seven Public Medical Segmentation Benchmarks (2D + 3D)

A curated bundle of seven public medical segmentation datasets, packaged with canonical leakage-free splits for the four datasets where one is needed (ACDC patient-level, BraTS2020 volume-level, CVC-ClinicDB video-level, plus seed-fixed image-level for the rest). All raw images and masks are retained at their original resolution; no resizing, no preprocessing baked in.

For the two volumetric MRI datasets (ACDC, BraTS), this release ships both 2D slice extracts and the original 3D NIfTI volumes, so users can choose 2D or 3D pipelines without re-downloading.

This is the dataset-only release that accompanied an evaluation-protocol audit of pixel- vs.\ latent-space diffusion augmentation for medical image segmentation. The bundle is reusable for any 2D medical-segmentation research, not just the original study.

Why this exists. Many existing medical-augmentation papers report non-comparable numbers because each uses a different (often undocumented) train/test split, and several datasets have hidden leakage if split at the image level (CVC same-video frames, ACDC same-patient slices, BraTS same-volume slices). This release fixes one canonical split per dataset so future work can be paired-comparable.

---

Contents

MedSeg-7D/
├── README.md
├── ACDC/                         (cardiac MRI, 100 patients)
│   ├── images/                   2D slices: patient<id>_frame<f>_slice_<s>.png
│   ├── masks/                    matching 2D-slice mask filenames (any-structure binary)
│   ├── 3D/                       ORIGINAL 3D NIfTI volumes (challenge layout)
│   │   ├── training/
│   │   │   ├── patient001/       Info.cfg + patient001_4d.nii.gz + frame01.nii.gz + frame01_gt.nii.gz + frame12.nii.gz + frame12_gt.nii.gz
│   │   │   └── ... (100 patients)
│   │   └── testing/              50 held-out patients (challenge test set)
│   └── split_info.json           CANONICAL patient-level split (seed=42, 80/20)
│   # 2D and 3D share the same 100-patient training cohort. The 3D side
│   # additionally ships the official 50 challenge test patients, which the
│   # 2D side does NOT include (we re-split the 100 train patients 80/20).
│
├── BraTS2020/                    (brain MRI FLAIR, 369 volumes → 22677 slices)
│   ├── images/
│   │   ├── volume_1/             volume_1_slice_<s>.png  (FLAIR channel, ~50-80 slices/vol)
│   │   ├── volume_2/
│   │   └── ... (369 vols)
│   ├── masks/
│   │   ├── volume_1/             matching whole-tumor binary mask
│   │   └── ... (369 vols)
│   └── split_info.json           CANONICAL volume-level split (seed=42, 295/74)
│   # NOTE: BraTS slices are nested into per-volume subdirectories because of
│   # HuggingFace's 10000 files-per-directory limit. Filenames preserve the
│   # original volume_X_slice_Y.png convention.
│
├── BraTS2021_3D/                 (brain MRI 3D NIfTI, 1251 patients — superset of BraTS2020)
│   ├── BraTS2021_00000/          5 NIfTI files: t1, t1ce, t2, flair, seg (4 modalities + GT)
│   ├── BraTS2021_00002/
│   └── ... (1251 patient dirs)
│   # IMPORTANT: This is BraTS *2021*, a SUPERSET of BraTS 2020. The 369
│   # volumes in our 2D `BraTS2020/` are a subset of the 1251 here. Patient
│   # IDs differ between the 2020 and 2021 releases, so split_info.json
│   # (volume-level for 2020) does NOT apply to BraTS2021_3D directly. Do
│   # not mix 2D and 3D Dice numbers.
│
├── BUSI/                         (breast ultrasound, 780 images)
│   ├── images/
│   └── masks/                    masks suffixed _mask.png
│
├── CVC-ClinicDB/                 (endoscopy polyp, 612 frames / 29 video sequences)
│   ├── PNG/
│   │   ├── Original/             RGB frames
│   │   └── Ground Truth/         binary masks
│   ├── TIF/                      original release format
│   ├── metadata.csv              per-frame metadata
│   ├── class_dict.csv
│   ├── pranet_split.json         PRIMARY: PraNet 550/62 image-level split (literature-standard)
│   └── video_split_seed42.json   ALTERNATIVE: leakage-free 23/6 video-level split (more rigorous)
│
├── Kvasir-SEG/                   (endoscopy polyp, 1000 images)
│   ├── images/                   RGB frames
│   ├── masks/                    binary masks
│   ├── bbox/                     bounding boxes (auxiliary)
│   ├── pranet_split.json         PRIMARY: PraNet 900/100 train/test split (literature-standard)
│   └── kavsir_seg_README.md      original release notes
│
├── REFUGE2/                      (fundus optic disc, 1200 images = 400 train + 400 val + 400 test)
│   ├── train/                    {images/, mask/}   400 images
│   ├── val/                      {images/, mask/}   400 images
│   └── test/                     {images/, mask/}   400 images
│
└── ISIC2018/                     (dermoscopy lesions, 2594 train + 100 val + 1000 test)
    ├── train/                    {images/, masks/}
    ├── validation/               {images/, masks/}
    └── test/                     {images/, masks/}

Approximate total size: ~18 GB.

---

Per-dataset details

1. ACDC — Cardiac cine-MRI

Modality	Cardiac cine-MRI (2D slices)
Original task	Multi-class cardiac structure segmentation
Patients / slices	100 / 1841
Mask classes	4 (background, RV, myocardium, LV) — preserved as in the original release
Split type (canonical)	Patient-level, 80 train / 20 test, seed=42
Split file	ACDC/split_info.json
Leakage risk	None at patient level. Slice-level random split would leak adjacent slices and inflate Dice ~5 points.
Source	ACDC Challenge (MICCAI 2017)
Reference	Bernard et al., IEEE TMI 2018
License	Original ACDC license; please refer to the original challenge website.

2. BraTS 2020 — Brain tumor MRI (FLAIR slices)

Modality	Brain MRI, FLAIR channel
Original task	Multi-class tumor segmentation
Volumes / slices	369 / 22677 (this release: FLAIR-only 2D slices)
Mask convention here	Whole-tumor binary (label 1+2+4 → 1)
Split type (canonical)	Volume-level, 295 train / 74 test, seed=42
Split file	BraTS2020/split_info.json
Leakage risk	None at volume level. Slice-level random would leak adjacent slices ~5 Dice points.
Note	Only the FLAIR modality is included. The original BraTS release has T1/T1ce/T2 in addition. If you need multi-modal data, fetch the original release.
Source	BraTS 2020 Challenge
Reference	Menze et al., IEEE TMI 2015; Bakas et al., 2017
License	Original BraTS license; please refer to the challenge website.

3. BUSI — Breast ultrasound

Modality	B-mode breast ultrasound
Original task	Lesion segmentation (benign / malignant / normal classes are also available)
Images	780
Mask convention	Binary foreground; mask filenames carry _mask.png suffix
Split type (canonical)	Image-level, 80/20, seed=42
Leakage risk	⚠️ The release does not publish patient IDs. Multiple images may come from the same patient. The image-level split is the community standard; "patient-level" cannot be verified from the release.
Source	BUSI Dataset (Cairo University)
Reference	Al-Dhabyani et al., Data in Brief 2020
License	CC-BY-4.0

4. CVC-ClinicDB — Colonoscopy polyp

Modality	Colonoscopy (RGB endoscopy)
Original task	Polyp segmentation
Frames / video sequences	612 / 29
Mask convention	Binary polyp foreground
PRIMARY split (literature-standard)	PraNet's 550/62 image-level train/test, used by PraNet, Polyp-PVT, SANet, ESFPNet and most polyp papers
Primary split file	CVC-ClinicDB/pranet_split.json
ALTERNATIVE split (leakage-free)	Video-level, 23 train / 6 test sequences, seed=42 (489 frames train, 123 frames test)
Alternative split file	CVC-ClinicDB/video_split_seed42.json
Important note	The PraNet split is image-level and leaks same-video frames across train/test (CVC has 29 underlying sequences). Use it for direct comparison to literature; use video-level for honest leakage-free generalization numbers. The two are not directly cross-comparable in absolute Dice.
Source	CVC-ClinicDB
Reference	Bernal et al., Computerized Medical Imaging and Graphics 2015
License	Released for academic use; cite the original paper.

5. Kvasir-SEG — Colonoscopy polyp

Modality	Colonoscopy (RGB endoscopy)
Original task	Polyp segmentation
Images	1000
Mask convention	Binary polyp foreground
PRIMARY split (literature-standard)	PraNet's 900/100 train/test (specific file lists), used by PraNet, Polyp-PVT, SANet, ESFPNet and the entire polyp-segmentation literature
Primary split file	Kvasir-SEG/pranet_split.json
Leakage risk	The release does not publish per-procedure metadata. Image-level is the community standard; per-procedure leakage cannot be audited.
Note	Filenames in our release use .jpg (the original Kvasir-SEG extension); PraNet ships them as .png after conversion — basenames match exactly. Auxiliary bbox/ (bounding boxes) included from the original release.
Source	Kvasir-SEG
Reference	Jha et al., MMM 2020
License	CC-BY-4.0

6. REFUGE2 — Fundus optic disc

Modality	Fundus photography
Original task	Optic disc and cup segmentation
Images	1200 = 400 train + 400 validation + 400 test (full official challenge release)
Mask convention	Multi-class (BG / disc / cup) preserved; for binary disc segmentation, treat any non-background pixel as foreground
Split type	Pre-released 400/400/400 train/val/test split is preserved
Leakage risk	None — each image is from a different patient by protocol.
Caveat	Modern segmenters reach ≥99.9 Dice on optic-disc segmentation; this dataset is saturated for that task. Use only when you specifically need fundus / glaucoma data.
Source	REFUGE2 Challenge
Reference	Orlando et al., Medical Image Analysis 2020; Fang et al., Medical Image Analysis 2022
License	Original REFUGE2 license; please refer to the challenge website.

7. ISIC 2018 — Dermoscopy

Modality	Dermoscopy
Original task	Skin lesion segmentation (Task 1)
Images	2594 train + 100 val + 1000 test (this release: PNG-extracted from the original ISIC 2018 archive)
Mask convention	Binary lesion foreground (any-pixel > 0 → 1)
Split type	Pre-released train/validation/test split is preserved
Leakage risk	The release does not publish patient IDs. Multiple lesions per patient are possible but cross-lesion contamination is generally considered low risk.
Source	ISIC 2018 Challenge
Reference	Codella et al., 2019; Tschandl et al., Sci. Data 2018
License	CC-BY-NC-4.0 (HAM10000-derived images)

---

Comparison to literature and existing HuggingFace cards

We audited the most common split conventions in published segmentation papers (MICCAI / IEEE TMI / MIA / CVPR / NeurIPS) and the two existing HuggingFace community cards for the same datasets, then aligned our defaults where sensible. Summary:

Dataset	Mainstream literature default	HuggingFace community card	Our default	Verdict
CVC-ClinicDB	PraNet's 550/62 image-level files (de facto standard since 2020)	Angelou0516/CVC-ClinicDB: 80/10/10 image-level, ESFPNet split	PraNet 550/62 (pranet_split.json) as primary; video-level 23/6 (video_split_seed42.json) as leakage-free alternative	✅ Matches PraNet exactly + adds a leakage-audit option that nobody else ships
Kvasir-SEG	PraNet's 900/100 file list (de facto standard)	kowndinya23/Kvasir-SEG: 880/120 (no test)	PraNet 900/100 (pranet_split.json)	✅ Matches PraNet exactly
BUSI	Image-level random; growing minority does 5-fold + de-duplication (BUS-Set, Med Phys 2023, documents duplicate leakage)	n/a	Image-level 80/20 seed=42	Matches majority; flag: BUSI release has documented duplicates, and patient IDs are not public, so true patient-level splits are not possible
ISIC 2018	Official 2594/100/1000 OR pooled 80/20	varies	Official 2594/100/1000 preserved	Matches official challenge split
REFUGE2	Official 400/400/400 (train/val/test domain-shift design)	varies	Official train/val/test preserved	Matches official
ACDC	Patient-level; TransUNet 70/10/20 of the 100 train OR nnU-Net 5-fold patient CV	rarely correct on HF	Patient-level 80/20 seed=42 (in split_info.json); 3D side also ships official 100-train + 50-test challenge layout	Stricter than the careless cards; consistent with TransUNet/nnU-Net practice
BraTS 2020	Volume-level; nnU-Net 5-fold patient CV is the most-cited recipe	rarely correct on HF	Volume-level 80/20 seed=42 (295/74)	Matches the careful camp; nnU-Net's 5-fold is a reasonable alternative on the same volumes

Mainstream papers we cross-checked: PraNet (Fan et al., MICCAI 2020), Polyp-PVT (Dong et al., 2021), ESFPNet (Chang et al., 2024), BUS-Set (Thomas et al., Med Phys 2023), TransUNet (Chen et al., 2021), SwinUNet (Cao et al., 2022), nnU-Net (Isensee et al., Nat. Methods 2021).

Notable disagreements with HuggingFace community cards

• kowndinya23/Kvasir-SEG (880/120) merges the test fold into validation, making it non-comparable to PraNet's 900/100. Ours preserves test/val separation.
• Angelou0516/CVC-ClinicDB does image-level 80/10/10 without flagging the same-video frame leakage that affects all 3 splits. We add an explicit video-level split for leakage-free evaluation.
• Neither HuggingFace card we found exposes patient-level splits for ACDC or BraTS — we provide them via split_info.json.

When to not use our defaults

• If you must directly compare to PraNet/Polyp-PVT numbers, use their released 1450/test files (not in this bundle, but reproducible from the raw images here).
• If you need nnU-Net 5-fold CV on ACDC or BraTS, regenerate folds with the standard nnU-Net recipe — our 80/20 split is a single-fold approximation.
• If you need BraTS 2021 (1251 volumes) instead of 2020 (369), the 3D version is shipped under BraTS2021_3D/ (subset of 2020 patients is included; new 2021-specific patients are added).

---

Recommended use

For paired-comparison evaluation across methods, lock to the canonical splits in this release:

import json, os
from huggingface_hub import snapshot_download

ROOT = snapshot_download("MaybeRichard/MedSeg-7D", repo_type="dataset")

# ACDC (patient-level)
info = json.load(open(os.path.join(ROOT, "ACDC", "split_info.json")))
train_patients = set(info["train_patients"])
# enumerate slices, check patient ID in filename to assign train/test

# BraTS (volume-level) — slices are nested under per-volume subdirs
info = json.load(open(os.path.join(ROOT, "BraTS2020", "split_info.json")))
train_volumes = set(info["train_patients"])  # key name retained from original
# To enumerate all training slices:
#   for vol in train_volumes:
#       for img_path in glob.glob(f"{ROOT}/BraTS2020/images/{vol}/*.png"):
#           ...

# Kvasir-SEG (PraNet 900/100, literature standard)
info = json.load(open(os.path.join(ROOT, "Kvasir-SEG", "pranet_split.json")))
train_files = set(info["train_files"])  # 900 file basenames (.jpg)
test_files  = set(info["test_files"])   # 100 file basenames (.jpg)

# CVC-ClinicDB (PraNet 550/62, literature standard — has same-video leakage!)
info = json.load(open(os.path.join(ROOT, "CVC-ClinicDB", "pranet_split.json")))
train_files = set(info["train_files"])  # 550 frames as <n>.png
test_files  = set(info["test_files"])   # 62 frames as <n>.png

# CVC-ClinicDB (video-level 23/6, leakage-free alternative)
info = json.load(open(os.path.join(ROOT, "CVC-ClinicDB", "video_split_seed42.json")))
train_seqs = set(info["train_sequences"])

For the 3D NIfTI versions:

import nibabel as nib
import os

# ACDC 3D (original challenge layout, 100 train + 50 test patients)
patient_dir = os.path.join(ROOT, "ACDC", "3D", "training", "patient001")
img = nib.load(os.path.join(patient_dir, "patient001_frame01.nii.gz")).get_fdata()
gt  = nib.load(os.path.join(patient_dir, "patient001_frame01_gt.nii.gz")).get_fdata()
# img shape: (H, W, num_short_axis_slices); gt has 4 classes (0=BG, 1=RV, 2=Myo, 3=LV)

# BraTS 2021 3D (1251 patients, 4 modalities + GT each)
pat = os.path.join(ROOT, "BraTS2021_3D", "BraTS2021_00000")
flair = nib.load(os.path.join(pat, "BraTS2021_00000_flair.nii.gz")).get_fdata()
seg   = nib.load(os.path.join(pat, "BraTS2021_00000_seg.nii.gz")).get_fdata()
# seg has 4 classes (0=BG, 1=necrotic, 2=edema, 4=enhancing); whole-tumor = (seg > 0)

For BUSI, the only dataset without a packaged split file, use a seed-fixed image-level 80/20 split:

import numpy as np
def get_image_level_split(n_images, seed=42, train_ratio=0.8):
    perm = np.random.RandomState(seed).permutation(n_images)
    n_train = int(n_images * train_ratio)
    return perm[:n_train], perm[n_train:]

(BUSI's release does not include patient IDs, so a true patient-level split is not possible. See per-dataset notes for caveats.)

---

Known caveats and good practices

1. Never use slice-level random split for ACDC or BraTS. Same-patient adjacent slices end up on both sides and inflate Dice ~5 points. Always read split_info.json.

2. CVC image-level split is leaky. Same-video frames cross train/test. Use the video-level split (video_split_seed42.json) for clean evaluation. Use image-level only for direct comparison to legacy literature, and label such results as "leakage-audited / auxiliary".

3. BUSI / Kvasir / ISIC do not provide patient IDs. Image-level random is the de-facto community standard; do not claim "patient-level independent" — there is no metadata to verify it.

4. REFUGE2 saturates at ~99.9 Dice. Don't use it as a downstream evaluator for augmentation studies; use it only when you need a fundus / optic-disc task specifically.

5. Mask conventions vary across datasets. Some are multi-class (ACDC: 4 classes; BraTS original: 4 classes; REFUGE2: 3 classes). For binary segmentation, use mask > 0. The released masks here keep the original multi-class labels where applicable (no information lost), so users can choose to binarize as needed.

6. All images and masks are at original resolution. No pre-processing baked in; you can resize per your protocol.

---

Citation

If this release is useful, please cite both the original dataset papers (see per-dataset references above) and the evaluation-protocol audit that produced these canonical splits:

@inproceedings{medseg7d2026,
  title  = {An Evaluation-Protocol Audit of Pixel- vs.\ Latent-Space Diffusion
            Augmentation for Medical Image Segmentation},
  author = {Anonymous},
  booktitle = {NeurIPS 2026 (E\&D Track)},
  year   = {2026}
}

License

This release does not redistribute datasets that are not already publicly available. Each dataset retains its original license; consult each per-dataset section above. The split metadata files (split_info.json, video_split_seed42.json) are released under MIT.