FetoPlac — Fetoscopic Placental Vessel Segmentation (Bano et al., 2020)

Re-hosted mirror of the Fetoscopy Placenta Dataset released alongside "Deep Placental Vessel Segmentation for Fetoscopic Mosaicking" (Bano et al., MICCAI 2020). Intended for use with the EasyMedSeg benchmark.

Why this mirror exists

The canonical UCL WEISS host (weiss-develop.cs.ucl.ac.uk/fetoscopy-data/...) went offline when the WEISS lab merged into the Hawkes Institute and the open-data subtree was retired. The full archive is not available on Zenodo, OSF, or Kaggle, and the broader FetReg2021 Synapse release is a different (gated, multi-class) dataset. This HF mirror is built from the byte-identical Internet Archive Wayback snapshot (web.archive.org/web/20220412011703id_/..., ZIP magic + size verified against the original Apache Content-Length header).

Composition

Two splits with unified schema:

Split	Frames	Source	Mask type
test	483	6 surgeries	manual binary GT
unannotated	950	6 continuous clips	U-Net pseudo-labels
All	1433

Per-video frame counts:

Video	test (GT)	unannotated	Native resolution
video01	121	400	470 x 470 (test); 448 x 448 (ua)
video02	101	200	540 x 540 / 448
video03	39	50	550 x 550 / 448
video04	88	100	640 x 640 / 448
video05	37	100	640 x 640 / 448
video06	97	100	720 x 720 / 448

The paper protocol is 6-fold leave-one-video-out cross-validation on the 483 GT frames — video_id is exposed in every row to make this reproducible. There is no fixed train/val/test partition in the original release; we expose all 483 GT frames under the test split because EasyMedSeg evaluation reads from test by default.

Schema (both splits)

Column	Type	Description
image_id	string	Frame stem (e.g. anon001_02785)
image	Image	Source RGB frame (PNG bytes, native size)
mask	Image	Vessel mask (PNG bytes). GT for test, pseudo for unannotated
fov_mask	Image	Per-video circular field-of-view ROI mask (1-bit PNG)
video_id	string	video01 .. video06
mask_type	string	"gt" for test, "predicted" for unannotated

Mask decoding

The upstream PNG masks use non-canonical pixel encodings: per-video the dominant non-background value is 25 or 33 (an artifact of the PixelAnnotationTool palette index), plus minor anti-aliasing values (1, 2, 3, 15) at <0.5% of pixels. The canonical decoding rule is:

binary_vessel = (np.array(mask)[..., 0] > 0).astype(np.uint8)

Pseudo-labels in the unannotated split are continuous probability maps with ~256 unique grayscale values; threshold at the value required by downstream usage.

FoV mask

The 6 per-video FoV masks are circular endoscope field-of-view stencils (boolean), embedded per-row for convenience. Apply via:

fov = np.array(row["fov_mask"]).astype(bool)
masked_image = np.array(row["image"]) * fov[..., None]

License

CC BY-NC-SA 4.0 (Attribution-NonCommercial-ShareAlike 4.0 International), as stated in the original Read me.txt shipped with the zip. Research / non-commercial use only.

Citation

@inproceedings{bano2020deep,
  title     = {Deep Placental Vessel Segmentation for Fetoscopic Mosaicking},
  author    = {Bano, Sophia and Vasconcelos, Francisco and Shepherd, Luke M. and
               Vander Poorten, Emmanuel and Vercauteren, Tom and Ourselin, Sebastien and
               David, Anna L. and Deprest, Jan and Stoyanov, Danail},
  booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2020},
  series    = {Lecture Notes in Computer Science},
  volume    = {12263},
  pages     = {763--773},
  publisher = {Springer},
  year      = {2020},
  doi       = {10.1007/978-3-030-59716-0_73}
}