fidoptMyoSegmenTUM thigh muscles
Water image of a chemical-shift-encoded MR of both thighs, the stack nearest the hip, of a healthy volunteer or a patient with a neuromuscular disease, with four muscle groups of each thigh drawn by hand and checked by a radiologist (MyoSegmenTUM_thigh).
10 test casesthighimage MRresearch only
Read before comparing (7)
- Cases. The water image of a six-echo chemical-shift-encoded MR of both thighs on a Philips Ingenia 3 T, the stack nearest the hip: 672 by 672 by 65 voxels of 0.78 by 0.78 by 4 mm for the volunteers, 512 by 512 by 30 of 0.82 by 0.82 by 4 mm for the patients (myotonic dystrophy type 2, limb-girdle dystrophy 2A, ALS). Values as delivered, no unit. The fat image and the fat fraction map, and the stacks nearer the knee, are not used. No mask.
- Header. The NIfTI files give 1 mm pixels, and on the patients 1 mm between slices; the stacks' own DICOM gives 0.78125 or 0.81875 mm and 4 mm, which prepare.py writes. The masks share the image's voxels; their headers carry the same wrong spacing and, on the patients, an origin 8 mm off.
- Labels. Quadriceps, sartorius, gracilis and hamstrings of each thigh, left being the patient's left. Where two masks claim a voxel, 18 voxels on HV001, the smaller muscle keeps it. On the volunteers the muscles are cut at the stack's lower edge, towards the knee, and end at their origins near the hip, above which the stack holds the unlabelled pelvis; on the patients' 30-slice stacks the quadriceps and sartorius run through all slices, the gracilis and hamstrings through all of them on P001 and stop 3 to 12 slices short of the top on P003 and P004.
- Truth. Drawn by hand slice by slice by one operator on the water and fat images and reviewed by a musculoskeletal radiologist (Schlaeger et al. 2018). HV010 and HV011 hold the same images with two different delineations, which agree at a Dice of 0.94 on the left quadriceps, 0.92 on the right hamstrings, 0.91 on the right gracilis and 0.88 on the left sartorius: the truth is good to that.
- Start. A mask of the whole image scores a Dice of 0.004 at the median over the test cases and muscles, an Otsu mask 0.018 (0.005 to 0.26, the quadriceps highest). TotalSegmentator 2.18.0's total_mr has no thigh muscle (its thigh_shoulder_muscles_mr task needs a licence) and writes only the femurs and gluteal muscles on HV001 and P001.
- Small set. 14 subjects: 10 test cases and 4 development cases, as for any set under 20 subjects.
- Overlap. No other dataset of the corpus holds MyoSegmenTUM images: SarcomaLimbMR's limb MR is another collection, and the screening's Huashan and Folkhälsan thigh set, which reuses MyoSegmenTUM, is not built.
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 subjects drawn with seed 2026, healthy volunteers and patients in proportion (7 and 3), and the 4 others for development, one session each. Of the 19 subjects five are held out: HV011, whose images are HV010's, and HV012 to HV015, whose water image is a DICOM without frame positions.
- Truth
- contours by experts the deposit's masks of the quadriceps, sartorius, gracilis and hamstrings of each thigh, drawn by hand slice by slice on the water and fat images by one operator and reviewed by a musculoskeletal radiologist (Schlaeger et al. 2018)
- Licence
- none stated: the OSF project carries no licence and is shared through a view-only link
4 cases for tuning, never tested
HV006, HV010, P002, HV002