fidoptACOUSLIC-AI, the fetal abdomen on blind-sweep ultrasound
One frame of six blind ultrasound sweeps over a pregnant woman's abdomen, made by a novice with a low-cost probe, with the fetal abdomen drawn as an ellipse by a trained reader: one subject of ACOUSLIC-AI's training set (Sierra Leone).
Read before comparing (8)
- Frame. A case stacks six sweeps of 140 frames. The readers drew ellipses on the first and last frames of each run where the abdomen and its plane type were seen, and where it changed much; the frames between were interpolated. The case is the first frame of the stack drawn on an optimal plane, or on a suboptimal one when none is: the start of a run, drawn by hand. Of the 53 cases drawn, 35 are optimal planes and 18 suboptimal, 39 in the first sweep.
- Truth. A filled ellipse, not an outline: the clinical abdominal circumference measurement, set to zero outside the sector, so a frame near the sector's edge holds a cut ellipse. The page names no review of the training set's ellipses; the hidden test set's were reviewed by a sonographer of 37 years.
- Subjects. The circumference table names each case's subject: 300 cases of 281 subjects, 19 with two examinations, kept to one.
- Images. Novices with an hour of training sweeping blind by the obstetric sweep protocol, three transverse and three sagittal sweeps, with a Telemed MicrUs Pro-C60S on a smartphone, at three public health units in Sierra Leone: 744 x 562 pixels of 0.28 mm, grey.
- Tolerance. 2 mm, 7 pixels, as HC18.
- Start. On the 28 test frames Otsu's dark pixels score a Dice of 0.06 and a surface Dice of 0.05 at 2 mm; one constant mask, the development ellipses' pixel majority, scores 0.37 and 0.07.
- Outside numbers. The challenge ranks the abdominal circumference on a hidden test set from Tanzania and Radboudumc, not these frames; networks trained on the 300 public cases have seen the test cases.
- Overlap. No other dataset of the corpus holds fetal ultrasound; HC18's Radboud images (2014 to 2015, GE Voluson) are another acquisition, and SonoCorpus's MD5s match no ACOUSLIC stack 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
- Test cases
- 28 test subjects drawn with seed 2026 among the 281 subjects of the provider's training set, and 25 development subjects from the rest; the challenge's test set is hidden. A subject with two examinations gives the first by uuid (draw.held_out the 19 others). One frame a subject.
- Truth
- contours by experts the provider's mask on the chosen frame, its optimal (1) and suboptimal (2) planes merged: an ellipse drawn by hand around the fetal abdomen by one of two readers with about two years of blind-sweep experience, each reading half of the cases, on a frame they drew rather than interpolated (the challenge's dataset page)
- Source
- zenodo.org/records/12697994
doi.org/10.5281/zenodo.12697994
zenodo.org/records/11005384
acouslic-ai.grand-challenge.org/datasets - Licence
- CC BY-NC-SA 4.0 (the Zenodo deposit's licence)
25 cases for tuning, never tested
c0319132-d83e-4e3a-ad50-4b281f2dcebc, ae6eed2f-58ca-47c2-a25d-c0d6ed80e178, aa609674-742e-4bd3-9b07-fb501abeba3b, 97363a66-8aec-4b52-8797-514e8c98205b, de964751-ec9d-4c53-9da3-9d7f200987b1, 1dab0e37-e5e4-41d4-b132-32103b385c78, 833c09d1-7257-45fc-9d41-76060614481f, bd55bc61-5335-41af-920b-f66206c47502, 7d44ea4e-9079-4a4f-bd0e-2e27ff70e9ea, 7b06f498-6552-4d51-bf8a-11bb23b70a92, 7112c87c-fd2b-488e-bad7-f447f6b49caa, f8039e25-4652-440c-9476-b425f3fccb22, ae58606d-8adf-4552-b42e-c251f446485e, 8ba63007-395c-4f8b-b5a0-a5f17596588a, 1ff3221d-43d3-4065-9393-347f2d0edac0, 585178ca-483e-45fa-9c89-2dbd51ebb69e, ec65aa35-d54c-44da-80a4-883f631c345e, c9a235e5-f070-4b9f-afb9-149067423cd5, d234b2ec-e6e7-4336-93cd-288314639995, ba6acb2e-8a8f-4a21-a986-ee1c2a1d41fc, 21883c54-c2ac-4bc3-b557-aa6d41d51087, fe236600-a41c-4d26-b4ea-6d3ff109cf41, 3d902e8c-626f-46b6-8942-243b3270d929, af8fa0cf-5f02-47f5-b8dd-d215b049f264, 30d9ee24-7603-46c5-9d05-7d8a6d6a292e