fidopt Interactive view
Colon wall on transabdominal ultrasound, 2D segmentation

C-TRUS, colon wall in transabdominal ultrasound

A transabdominal B-mode image of the colon of a patient with ulcerative colitis, 2D, with the colon wall outlined by a gastroenterologist (C-TRUS, University of Münster).

10 test casescolonimage US2D, distances in pxresearch only
Look at the images
Read before comparing (8)
  • Images. 827 grey JPEGs of 580 x 360 pixels from a Toshiba Xario, frames of the examination sequences of 13 patients; a depth scale is burnt in but no pixel size is given, so distances are in pixels. The provider rates each image's quality high, medium or low; the 13 case images are 3 high, 8 medium and 2 low.
  • Masks. 319 images, from the phase of finding the colon, have an empty mask and are never cases. Masks are JPEGs, nearly black and white: a pixel above 127 is wall.
  • Twins. 39 pairs of images are byte for byte the same frame of one patient: 31 with a mask on both copies, 14 of them with two different masks (one pair's agree at a Dice of 0.79), 5 with a mask on one copy only and 3 with none. Both copies are kept out of the cases.
  • Truth. The paper thanks six physicians for annotations and the group's DGVS page speaks of four experts; who drew which mask is not given. On a thin wall, the experts' moderate agreement caps what a score can show. Test case patient_03's mask is a short piece of the wall in view, 714 pixels.
  • Not prostate. C-TRUS stands for colon wall segmentation in transabdominal ultrasound; SonoCorpus's metadata labels it prostate by mistake. No other dataset of the corpus holds colon ultrasound, and SonoCorpus's MD5 manifest finds none of these images in its other 52 datasets.
  • Outside numbers. Case images are taken from SonoCorpus's test list, as KvasirSEG takes FCBFormer's test split, so SonoBase has trained on none of them; but its training split (333 images), like the provider's five folds, holds other images of all 13 patients, and every image is public with its mask.
  • Start. On the 10 test images, the whole image scores a Dice of 0.065, Otsu's bright and dark pixels 0.077 and 0.056; the wall covers 3.4% of an image on average. TotalSegmentator has no ultrasound task.
  • Tolerance. 2 pixels: the truth grown or thinned by one pixel keeps a Dice of 0.93 and 0.92 and a surface Dice of 0.97 and 1.0 at 2 pixels.

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
10 test patients drawn with seed 2026 out of the 13, the other 3 for development. One case a patient: one of the patient's images in SonoCorpus's test list, with a mask and no byte-identical twin, the first in a seeded order (prepare.py).
Truth
contours by experts the provider's masks (labels/<image>.jpg), the colon wall outlined by gastroenterologists experienced in intestinal ultrasound; the paper measures an average consensus Dice of 0.61 between experts (Leenings et al., ASMUS 2024)
Licence
none stated (the repository has no licence; issue #2 asking for one is unanswered)
3 cases for tuning, never tested

patient_07, patient_09, patient_10