fidoptLower leg muscles, 3D ultrasound
Freehand 3D ultrasound of a healthy volunteer's calf, compounded from four to six tracked sweeps with the leg in a water bath, the soleus and both gastrocnemius heads labelled (Leg-3D-US, Nantes University and TU Munich).
10 test caseslimbsimage USresearch only
Read before comparing (9)
- Volunteers. 44 healthy volunteers aged 18 to 45, one calf each, 20 left legs and 24 right, which is which not given. No other dataset of the corpus holds these images.
- Images. 8-bit and isotropic, 0.260 to 0.275 mm on the test cases; the ultrasound fills 38 to 51% of the grid, the rest is 0. The sweeps' seams show as lines along the leg, and the deep soleus is dark on many volumes.
- Axes. The tracker's, not the body's: the leg's long axis is the third, and the lateral head lies on either side of the medial one (7 and 3 test cases, the two legs mirrored).
- Truth. An expert outlined the muscles on sparse 2D B-mode frames in Stradwin; the provider's ZOI method made them 3D, leaving about 10% of volume error at the borders, which an expert then corrected on the test and validation volunteers (under 4%). Graded expert; no second reading of the 3D labels is published.
- Development. val_01, val_02, val_03, val_04 and val_08 have corrected borders; train_17, train_18, train_22, train_30 and train_38 keep the uncorrected interpolation, so a score on those five says less.
- Stray values. 18 of the 44 maps also hold 250, 255 or both, thin pieces on the border of two muscles, as if two masks were added; test case test_12 has 6,103 voxels at 255. The corpus names only the three muscles, so those voxels count as no muscle.
- Checks. On the 20 cases every muscle is there, the labels lie in the ultrasound, and the soleus is deep and largest, the medial head larger than the lateral. TotalSegmentator has no ultrasound task, so no network check.
- Outside numbers. The paper's 10 test volunteers are these, but its numbers do not compare: it gives neither the HD95's unit, the tolerance nor the resolution. The test labels are public: a model trained on the whole archive may have seen them.
- Source errors. The page gives the paper's title with another paper's DOI (10.1371/journal.pone.0268550); the paper is 10.1007/s11548-024-03170-7. The page says a 40 mm probe at 2 to 10 MHz, the paper a 50 mm probe at 4 to 15 MHz.
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
- The provider's 10 test volunteers, all of them, are the test cases. The 10 development cases are its 5 validation volunteers and 5 of its 29 training volunteers. One volume a volunteer, so no volunteer is on both sides.
- Truth
- contours by experts the three muscles outlined by an expert on sparse 2D B-mode frames, made into 3D by a zero-order interpolation between the outlined frames and smoothed (the provider's ZOI method), the borders then corrected by an expert on the test and validation volunteers
- Source
- www.cs.cit.tum.de/camp/publications/leg-3d-us-dataset
github.com/Al3xand1a/segmentation-border-analysis
doi.org/10.1007/s11548-024-03170-7 - Licence
- GNU General Public License, no version given
10 cases for tuning, never tested
val_01, val_02, val_03, val_04, val_08, train_18, train_17, train_30, train_38, train_22