fidoptThyroid lobes, 3D ultrasound
A 3D ultrasound volume of one thyroid lobe of a healthy volunteer, compounded from a tracked freehand sweep, with the lobe outlined by a medical expert with livewire (OpenCAS thyroid dataset, Wunderling et al. 2017, Magdeburg).
10 test caseshead and neckimage USresearch only
Read before comparing (6)
- Volunteers. Eight healthy volunteers, both lobes each, recorded with a GE Logiq E9 on two days of 2016 (six on 8 February, two on 7 July). The page speaks of eight DICOM files, one lobe each; the archive holds 16 volumes, two a volunteer, paired by PatientID.
- Images. 8-bit B-mode volumes as MeVisLab compounded them from the tracked sweeps, isotropic, 0.163 to 0.238 mm, 378 to 567 slices of about 250 x 350 pixels; the ultrasound fills 37 to 67% of the grid, the rest is 0. Each file also holds a second map on the same grid (anti-correlated with the image, a fifth of it at 255, no truth on it), left out. The axes are the array's: the carotid lies on either side of the lobe, and which lobe is left or right is not given.
- Truth. One expert's livewire outline of the lobe, no second reading. The paper scores three semi-automatic methods against it.
- Tolerance and start. The default 1 mm, five voxels: the truth grown or thinned by one voxel keeps a Dice of 0.94 to 0.98 and a surface Dice of 1.0 at 1 mm (0.999 at one voxel). On the 10 test lobes (4.6 to 12.9 ml) the whole grid scores a Dice of 0.05, the ultrasound field 0.10 and its bright half (above Otsu's threshold) 0.24. A lobe can reach the edge of the field (D13).
- Seen. TG3K (Gong et al. 2023) is 3,585 2D frames cut from these 16 volumes, those where the gland fills more than 6% of the frame, split 9:1 by frame: a method trained on TG3K (TRFE+, SAMUS, SonoBase) has seen every volunteer.
- Overlap. No other dataset of the corpus holds thyroid ultrasound or these volunteers; TN3K (a hospital in China) and DDTI (IDIME, Bogotá) are patients of other centres.
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
- Both lobes of 5 of the 8 volunteers are the 10 test cases, both lobes of the other 3 the 6 development cases. The volunteers (DICOM PatientIDs 11 to 16, 11112222 and 1111133333) are put in the order of sha256("2026:ThyroidLobeUS:<PatientID>"), fidopt's shuffled() with seed 2026, and the first 5 are test. Fewer than 20 subjects, so 10 test cases; fidopt draw cannot keep a volunteer's two lobes on one side, so the lists are fixed here.
- Truth
- contours by experts the lobe segmented with livewire on the compounded volume by the medical expert who also made the sweeps (OpenCAS's dataset page)
- Source
- archive.iar.kit.edu/opencas/index.html%3Fq=node%252F29.html
archive.iar.kit.edu/opencas/index.html
archive.iar.kit.edu/opencas/data/thyroid.zip
doi.org/10.1117/12.2254234
doi.org/10.1016/j.compbiomed.2022.106389 - Licence
- CC BY-NC-SA 3.0 (OpenCAS, for all its datasets); Wunderling et al. 2017 must be cited
6 cases for tuning, never tested
D05, D06, D09, D10, D15, D16