README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- image-segmentation
tags:
- medical
- fetoscopy
- endoscopy
- placenta
- vessel-segmentation
- fetal-surgery
- ttts
pretty_name: FetReg2021
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: train_unlabeled
    path: data/train_unlabeled-*
  - split: test_unlabeled
    path: data/test_unlabeled-*
dataset_info:
  features:
  - name: image
    dtype: image
  - name: mask
    dtype: image
  - name: image_id
    dtype: string
  - name: file_name
    dtype: string
  - name: video_id
    dtype: string
  - name: center_id
    dtype: string
  - name: center
    dtype: string
  - name: challenge_split
    dtype: string
  - name: task
    dtype: string
  - name: frame_index
    dtype: int32
  - name: clip_id
    dtype: string
  - name: clip_frame_index
    dtype: int32
  - name: sequence_index
    dtype: int32
  - name: num_frames_sequence
    dtype: int32
  - name: width
    dtype: int32
  - name: height
    dtype: int32
  - name: fetoplac_subject_id
    dtype: string
  - name: in_fetoplac_annotated
    dtype: bool
  splits:
  - name: train
    num_bytes: 654341601
    num_examples: 2060
  - name: test
    num_bytes: 242931356
    num_examples: 658
  - name: train_unlabeled
    num_bytes: 1959789850
    num_examples: 7411
  - name: test_unlabeled
    num_bytes: 588615799
    num_examples: 2225
  download_size: 3445042447
  dataset_size: 3445678606
---

# FetReg2021

Placental vessel segmentation in **in-vivo fetoscopy** — the MICCAI/EndoVis 2021
**FetReg** challenge dataset. Frames come from fetoscopic laser photocoagulation
for **Twin-to-Twin Transfusion Syndrome (TTTS)**, captured during 24 procedures
at two fetal-surgery centres, and are cropped square to the fetoscope field of
view.

There is only **one** FetReg edition (2021). "FetReg2022" is a citation-year
artifact: the challenge-findings paper appeared as a 2022 preprint about the
2021 challenge, and the data went fully public in June 2022.

## Contents

One row per frame. The complete official release is mirrored — both challenge
tasks, both splits, all 24 procedures, including the **real test ground truth**.

| split | task | videos | frames | masks |
|-------|------|--------|--------|-------|
| `train` | 1 — segmentation | 18 | 2,060 | ✅ |
| `test` | 1 — segmentation | 6 | 658 | ✅ |
| `train_unlabeled` | 2 — registration | 18 | 7,411 | ❌ |
| `test_unlabeled` | 2 — registration | 6 | 2,225 | ❌ |
| **total** | | **24** | **12,354** | 2,718 |

**Both labelled splits carry real ground truth**, so no train/test fallback is
needed. Task 2 ships **no ground truth of any kind** — no masks and no
homographies — by design; the challenge scored it with a proxy *N*-frame SSIM
metric. Those frames are kept here with `mask=None` so mosaicking and
semi-supervised work remain possible.

Video IDs run `Video001`–`Video025` with **`Video021` absent** (25 − 1 = 24).

## Classes

Mutually exclusive. Single-channel PNG, PIL mode `L`, **raw integer labels** —
no palette, no colour lookup. Verified exhaustively: all 2,718 masks contain
only `{0,1,2,3}`, with no anti-aliasing strays.

| id | class | pixel share | frames containing it |
|----|-------|-------------|----------------------|
| 0 | background | 88.18% | 2,718 / 2,718 |
| 1 | placental vessel | 9.01% | 2,691 / 2,718 |
| 2 | ablation tool | 1.35% | 901 / 2,718 |
| 3 | fetus | 1.46% | 376 / 2,718 |

The authors' own visualisation script (mirrored here as
`upstream_FetReg2021_segmentation_visualisation.py`) maps these to
black / red / blue / green.

## Columns

| column | notes |
|--------|-------|
| `image` | RGB PNG, square, 271–720 px (see caveats) |
| `mask` | mode-`L` PNG, values 0–3. **`None` in the two `*_unlabeled` splits** |
| `image_id` / `file_name` | upstream stem / filename, unchanged |
| `video_id` | `Video001`–`Video025`. **Group on this** — one procedure, one patient |
| `center_id` / `center` | `I`/`UCLH` or `II`/`IGG` — 12 procedures each |
| `challenge_split` | `train` or `test`, meaningful for the unlabeled splits too |
| `task` | `segmentation` (Task 1) or `registration` (Task 2) |
| `frame_index` | original frame number in the source video. **Task 1 only**; `None` for clips, whose numbering was reset upstream |
| `clip_id` / `clip_frame_index` | e.g. `CLIP01` and the index within it. **Task 2 only** |
| `sequence_index` | 0-based position within this `(video_id, task)` sequence, ordered by the parsed integer |
| `num_frames_sequence` | length of that sequence |
| `width` / `height` | this frame's own size — it is *not* constant per video |
| `fetoplac_subject_id` | `anonNNN` when this procedure also appears in `MedOtter/FetoPlac`, else `None` |
| `in_fetoplac_annotated` | `True` if this exact frame is in FetoPlac's 483-frame GT set. `None` for clips, where the original frame number is unrecoverable |

## ⚠️ Overlap with MedOtter/FetoPlac

`MedOtter/FetoPlac` (Bano et al., MICCAI 2020) is a **UCLH-only subset of these
same procedures**, not an independent dataset. Joining on the original video
frame number embedded in both naming schemes — FetoPlac `anon{NNN}_{FFFFF}.png`
↔ FetReg `Video{NNN}_frame{FFFFF}.png` — gives **five of FetoPlac's six subjects
contained at 100%, 445 of its 482 unique GT frames (92%), every one of them
inside FetReg's TRAIN split**:

| FetoPlac subject | FetReg video | containment |
|---|---|---|
| `anon001` | `Video001` | 120 / 120 |
| `anon002` | `Video002` | 101 / 101 |
| `anon003` | `Video003` | 39 / 39 |
| `anon005` | `Video007` | 88 / 88 |
| `anon012` | `Video019` | 97 / 97 |
| `anon010` | — | not in FetReg |

Use `fetoplac_subject_id` to exclude at the procedure level (the safe
granularity) and `in_fetoplac_annotated` for frame-level precision. **Do not
evaluate FetoPlac against a model trained on FetReg train, and never split these
procedures across train and test.** FetoPlac's binary vessel masks are also a
*different annotation* of the same pixels, so agreement between the two is not
independent evidence.

No overlap with the other EndoVis-family datasets — Endovis2017/2018 are
porcine robotic surgery, CholecSeg8k / m2caiSeg / Endoscapes2023 are
laparoscopic cholecystectomy. The only shared lineage is the EndoVis umbrella.

## Corrections to the upstream documentation

Every count here was measured from the archive's bytes. Five upstream numbers do
not survive that check; the values in this mirror are the measured ones.

1. **`Video016` train clip has 593 frames**, not the README's 493. 593 is what
   makes the README's own 7,411 train-clip total add up.
2. **`Video025` test Task 1 has 110 labelled frames**, not the README's 100.
   110 is what makes the README's own 658 test total add up.
3. **`Video025` test clip has 292 frames**, not the 272 in the README and paper
   Table 2 — so the test-clip total is **2,225** (not 2,205) and the grand total
   **9,636** (not 9,616).
4. **Paper Table 2's `Center` column swaps `Video018` and `Video019`.**
   Figures 4 and 5 both give `Video018 = II`, `Video019 = I`, and the FetoPlac
   overlap proves `Video019` is UCLH independently. This mirror uses the figures.
5. **Paper Table 2's per-class `Occurrence(frame)` column is row-shifted from
   `Video020` downward** — its `Video025` entry (648/320/83) is in fact the
   test-set column totals. The measured per-video occurrence ships in
   `class_map.json`.

Also: the README labels both *Test* subsections `Train_FetReg2021_Task*`
(copy-paste); the real directories are `Test_...`. And the 2021 descriptor's
claim of **three** centres including University Hospital Leuven is stale —
the final paper and the released archive both have two.

## Caveats

- **Resolution varies per video *and* within a video.** Task 1 sizes span
  320–720 px. `Video010` is the one sequence that changes mid-video: 17 frames
  at 622×622 and 83 at 638×638. Any code assuming one size per `video_id`
  will break.
- **A video's Task 2 clip is not the same geometry as its Task 1 frames** —
  `Video023` is 320 px in Task 1 but 271 in its clip; `Video022` 400 vs 673;
  `Video012` 320 vs 277. Do not reuse a Task 1 size for a clip.
- **Only `Video010`'s Task 1 filenames carry a doubled prefix**
  (`Video010_frame0Video010_00000.png`); the other five test videos use the
  clean `Video{NNN}_frame{NNNNN}.png` form. One regex does not cover both. Task 2
  clip indices and zero-padding are likewise inconsistent (`CLIP00`/`01`/`04`/
  `09`; 4-digit in some train videos, 5-digit in others; train clips start at 1,
  test clips at 0), so this mirror orders on the parsed integer via
  `sequence_index`.
- **Severe class imbalance.** Tool and fetus are ~1.4% of pixels each and absent
  from most frames; `Video012` contains no fetus at all. Per-class scores are
  unstable, and a metric that rewards a correctly-empty class will inflate them.
- **Every image was annotated once**, so no inter-rater agreement is computable.
  The pipeline was tiered — 4 researchers annotated 7 videos, a commercial team
  with clinical background annotated 17, then 2 researchers verified and 2 fetal
  medicine specialists signed off — but it converged to this single mask set,
  which is the gold standard.
- **Known annotation-completeness caveat.** The authors of **TTTSNet**
  (Płotka et al., *Med. Image Anal.* 2025) re-annotated FetReg's 18 training
  procedures, stating that these masks "omit small placental vessel segmentation
  and include incomplete labels for larger vessels". Their release is
  vessel-only, by different authors, under CC BY 4.0 — it is *not* FetReg ground
  truth, but it is a real caveat for vessel-recall comparisons.
- Frames are pre-cropped square to the fetoscope field of view (an upstream
  authorial choice). No field-of-view mask ships with FetReg.

## Fidelity

Image and mask bytes are copied **verbatim** from the UCL deposit — no
re-encode, no resize, no relabelling. Only the container changed (per-video
directories → parquet) and metadata columns were added.

## License

**CC BY-NC-SA 4.0**, inherited from the source deposit. ShareAlike applies: this
reformatted derivative carries the same license. Non-commercial use only.

## Source

- UCL Research Data Repository: https://rdr.ucl.ac.uk/articles/dataset/_b_FetReg_Largescale_Multi-centre_Fetoscopy_Placenta_Dataset_b_/30417166
- DOI: `10.5522/04/30417166.v1`
- Challenge: https://www.synapse.org/Synapse:syn25313156 (EndoVis 2021 sub-challenge)

Note: the URLs cited in the papers and on Synapse
(`weiss-develop.cs.ucl.ac.uk`, the UCL WEISS open-data page,
`fetreg2021.grand-challenge.org`) are all dead or redirected — WEISS was folded
into the UCL Hawkes Institute and the data moved to the RDR deposit above.

## Citation

Both are requested by the upstream README.

```bibtex
@article{bano2024fetreg,
  title   = {Placental vessel segmentation and registration in fetoscopy:
             Literature review and MICCAI FetReg2021 challenge findings},
  author  = {Bano, Sophia and Casella, Alessandro and Vasconcelos, Francisco and
             Qayyum, Abdul and Benzinou, Abdesslam and Mazher, Moona and
             Meriaudeau, Fabrice and others and Moccia, Sara and Stoyanov, Danail},
  journal = {Medical Image Analysis},
  volume  = {92},
  pages   = {103066},
  year    = {2024},
  doi     = {10.1016/j.media.2023.103066}
}

@article{bano2021fetreg,
  title   = {FetReg: Placental Vessel Segmentation and Registration in
             Fetoscopy Challenge Dataset},
  author  = {Bano, Sophia and Casella, Alessandro and Vasconcelos, Francisco and
             Moccia, Sara and Attilakos, George and Wimalasundera, Ruwan and
             David, Anna L and Paladini, Dario and Deprest, Jan and
             De Momi, Elena and Mattos, Leonardo S and Stoyanov, Danail},
  journal = {arXiv preprint arXiv:2106.05923},
  year    = {2021}
}
```