fidoptFUGC, the cervical lips on transvaginal ultrasound
A 2D transvaginal ultrasound image of the uterine cervix with its anterior and posterior lips outlined, drafted by a fine-tuned SAM and refined by an obstetric sonographer: one examination of the FUGC challenge (ISBI 2025, Jinan University), its first image.
9 test casespelvisimage US2D, distances in px
Read before comparing (9)
- Values. 1 is the anterior lip and 2 the posterior lip: the paper's Figure 1 names the region between the probe and the canal the anterior lip, and 1 lies above 2, nearer the probe, in all 440 labelled images. Neither the deposit nor the paper writes the values down.
- Truth. One sonographer refined every mask, with no second reading and no agreement figure. The anterior lip is the easier: the challenge's best Dice was 0.946 on it and 0.865 on the posterior lip.
- Subjects. The paper gives neither the number of women nor how images were split, and the files carry no patient id; consecutive numbers are frames of one examination, which can correlate as low as 0.76 (test_0083 to test_0098), far below the 0.882 a near copy keeps (control: an image cropped by 2% a side and noised finds its original first in 298 of 300, at 0.882 to 0.973). On grey thumbnails of 128 x 80 pixels, consecutive images at 0.75 or more, or any two at 0.882 or more, make 33 runs in the test split; each boundary between runs was then looked at, and runs showing the same cervix, myometrium and bladder were joined, leaving 9 examinations of 9 to 73 images (test_0001 to 0016, 0017 to 0025, 0026 to 0098, 0099 to 0117, 0118 to 0135, 0136 to 0154, 0155 to 0171, 0172 to 0240, 0241 to 0300), which correlate with one another at 0.783 at most. The 140 labelled training and validation images make 7 examinations the same way; three that correlate with a test examination above 0.783 are left out: train_0028 to 0049 (0.859 with test_0172 to 0240), val_0001 to 0006 (0.811) and train_0050 (0.787); the four kept reach 0.781 at most and were looked at beside their closest test examinations. No two of the 890 files are the same byte for byte.
- Size. 9 test and 4 development examinations: the 440 labelled images are frames of few women, so the dataset is small whatever its image count.
- Images. GE Voluson E10 and E8, after bladder emptying, cropped to the cervix and stored at 544 x 336 pixels, a tinted grey in colour. No pixel size is given, so distances are in pixels.
- Tolerance. 2 pixels, about 0.4% of the width, as KvasirSEG.
- Start. On the 9 test images Otsu's dark pixels score a mean Dice of 0.07 over the two lips and a surface Dice of 0.08 at 2 pixels; one constant label map, the development cases' pixel majority, scores 0.60 and 0.10, since the cervix always lies under the probe at the top of the image.
- Outside numbers. The challenge ranked methods on all 300 test images, these 9 among them; their labels are public with the images since August 2025.
- Overlap. No other dataset of the corpus holds cervical ultrasound. JNU-IFM, also from Jinan University, holds intrapartum transperineal videos from a Youkey D8, not these images, and SonoCorpus's MD5s match no FUGC image in another of its datasets.
No method has a result on every case of this dataset yet.
Case by case
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
The first test caseClick to look at every case.
- Test cases
- 9 test images, one an examination: the provider's 300 test images fall into 9 examinations (runs of consecutive numbers whose thumbnails correlate at 0.75 or more, joined by eye where the mother's anatomy is the same; notes), and the lowest-numbered image of each is kept, the other 291 held out. 4 development images, the first of each examination of the 140 labelled training and validation images that resembles no test examination; the other 136 are held out.
- Truth
- contours by a program, verified by people the provider's masks of the anterior (1) and posterior (2) cervical lips: drafted by an ultrasound-adapted SAM, then refined in the Pair software by one obstetric sonographer with over ten years of experience (Bai et al. 2026)
- Source
- zenodo.org/records/16893174
doi.org/10.5281/zenodo.16893174
zenodo.org/records/14305302
arxiv.org/abs/2601.15572
doi.org/10.1109/TMI.2026.3666364 - Licence
- CC BY 4.0 (the Zenodo deposit's licence)
4 cases for tuning, never tested
train_0001, val_0019, val_0013, val_0007