fidoptKvasir-SEG, colonoscopy polyps outlined by hand
A white-light colonoscopy photograph, 2D and in colour, with its polyps outlined: pre-segmented by a junior doctor and a PhD student, then corrected by a gastroenterologist (Kvasir-SEG, Bærum Hospital, Norway).
Read before comparing (9)
- Images. White-light JPEGs of Bærum Hospital's routine colonoscopies of 2008 to 2016, read from HyperKvasir's OSF deposit, in colour, one vector of red, green and blue a pixel, mostly about 620 x 530 pixels. No pixel size is given, so distances are in pixels. The text the endoscope burnt into the frame is part of the image; a mask may hold several polyps.
- Pool. OSF holds all 1000 images but only 348 masks, and Simula's own server, which holds them all, is refused by Python's requests for an incomplete certificate chain: the pool is the 348 with a mask on OSF.
- Twins. The files carry no patient, examination or video identifier, and Kvasir-SEG holds several frames of one polyp. Each test image was compared with FCBFormer's 900 training and validation images and with the corpus: the eight closest thumbnails looked at, and the date and clock burnt into the frame, read by OCR then by eye, matched against every image of the same month.
- Left out. Four of the 19 show a polyp of FCBFormer's training images; the others share a day with a training image of the same recorder, or have a half-hidden day or clock that could be a training examination's. prepare.py's EXCLUDED gives each reason. Where no date can be read only the thumbnails speak, so frames of one examination can still hide.
- Within the split. No development image carries a test image's date or polyp. Test cases cju2trbpkv0c00988hxla5dzz and cju2zwg05a0oy0801yr73ig7g share a month and possibly a day, two hours apart, and are kept as two examinations.
- Truth. Test case cju3ykamdj9u208503pygyuc8 is a pale frame whose mask is two small red patches, 2% of the image, beside a fold, not the stalked shape below them.
- Tolerance. 2 pixels, about 0.3% of the width, as ISIC2018's 10 on its 3072-pixel photographs. On ten test images of an earlier draw, the truth grown by two pixels keeps a Dice of 0.98 but scores a surface Dice of 0 at one pixel and 1 at two; smoothed by about 6 pixels it scores 0.96 and 0.99; Otsu's dark pixels a Dice of 0.02.
- Outside numbers. Kvasir-SEG has no test split of its own and its images are public with their masks. FCBFormer's published checkpoint and anything trained on its split have seen none of the test images; networks trained on PraNet's split (900 of the 1000) or on HyperKvasir may have.
- Overlap. No other dataset of the corpus holds Kvasir or HyperKvasir images.
No method has a result on every case of this dataset yet.
Under a licence that is not open
Methods whose code or weights are released for non-commercial use, or under share-alike or custom terms: shown for reference, ranked nowhere.
Dice: how much the organ contours overlap, 1 is perfect. Surface Dice: the share of the contours' surfaces within a tolerance of each other, 1 is perfect. HD95: distance between the contours' surfaces.
Case by case
| Dice | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 |
|---|
Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.
About the data
- Test cases
- 18 test images: the 37 of FCBFormer's 100 test images (Sanderson and Matuszewski 2022) with a mask on OSF, less 19 whose polyp or examination FCBFormer's training or validation images may hold. 30 development images drawn with seed 2026 among the 308 other images with a mask, once three possible repeats of another image's patient are left out. One image a case.
- Truth
- contours by experts the provider's polyp masks, drawn as outlines on Labelbox: pre-segmented by a junior doctor and a PhD student, then verified and corrected image by image by a gastroenterologist (HyperKvasir's data paper, Borgli et al. 2020; Kvasir-SEG's paper says 'manually annotated by a medical doctor and then verified by an experienced gastroenterologist')
- Source
- osf.io/mh9sj
doi.org/10.17605/OSF.IO/MH9SJ
doi.org/10.1038/s41597-020-00622-y
datasets.simula.no/kvasir-seg
doi.org/10.1007/978-3-030-37734-2_37
arxiv.org/abs/1911.07069
github.com/ESandML/FCBFormer/tree/c3531990ddfac6b1ae6153dbfa0ad921ff63dedd - Licence
- CC BY 4.0 (HyperKvasir's OSF deposit, its license.txt and licence field); Simula's Kvasir-SEG page restricts the same files to research and education
30 cases for tuning, never tested
cju2zp89k9q1g0855k1x0f1xa, cju2y5zas8m7f0801d34g5owq, cju2x7vw87mu30878hye2ca0m, cju2hlm19vjjf0801o69qnber, cju2i8br1vqtd08784u6vmcjk, cju1hhj6mxfp90835n3wofrap, cju3tsh4lfsok0987w6x3a0v1, cju3ul8dogf1z09872y2ecowp, cju5cky5xb0ay0801oxet697t, cju5ew4h9cqaf0818rrczkmqh, cju1gkndf6yi10801o1qnje19, cju2u73dj53oz0878486k8k4b, cju2u2b9o4zvp08788qb9nqxj, cju2z6ez69g4u0801qwt088lw, cju2igw4gvxds0878808qj398, cju1b3zgj3d8e0801kpolea6c, cju2yw4s7z7p20988lmf2gdgd, cju3x7xsaijq80818f0psavav, cju1b75x63ddl0799sdp0i2j3, cju2zo0fwzv580988qlijd2xa, cju2zr3c3vwb00993jn06bbaz, cju1cvkfwqrec0993wbp1jlzm, cju5f26ebcuai0818xlwh6116, cju5eftctcdbj08712gdp989f, cju1euant5l960878iqj5vvto, cju0vtox5ain6099360pu62rp, cju2t2ivz43i10878zeg8r1br, cju3y9difj6th0801kd1rqm3w, cju2p91qir00k08350ddfif0w, cju2s2527pfyr0993l3h1149a