fidopt Interactive view
Spine, bones and joints, segmentation

PENGWIN 2026, femur

CT of one side of the hips and thighs of a patient with a femoral fracture before surgery, from seven Chinese hospitals, with the broken femur labelled (PENGWIN 2026 challenge, tasks 1 and 2, femur part).

16 test casespelvisimage CTresearch only
Look at the images
Read before comparing (7)
  • Pool. The 170 femur subjects (251 to 420) of PENGWIN 2026's 340 training subjects; their CTs show proximal femoral fractures, a few shaft fractures. The provider cut each CT to the injured side, 246 to 398 voxels wide, so one femur is in view, with the hip bone beside it; three test CTs (280, 281, 297) hold the whole thigh to the knee and one (409) the trunk.
  • Labels. One label, femurs: the femur in view, its 2 to 11 fragments merged (a median of 3 on the test CTs), as PENGWIN merges a bone's fragments. TotalSegmentator's femur_left and femur_right both count for it. The hip bone and other bones in view are not labelled; fragments under 500 mm³ are background. Its centre lies on the patient's right in 11 of the 16 test CTs.
  • Truth. The challenge's design document: an nnU-Net trained on about 1,000 mostly unbroken pelvic CTs drew the bones, three annotators with five years of experience or more refined them and cut the fragments in 3D Slicer, and two senior experts validated each case. Graded semi-automatic, as PENGWIN: the draft's prior may remain. No agreement between readers is published for 2026.
  • Images. The deposit's MetaImages unchanged, in Hounsfield units stored as integers or as integer-valued floats; most read -1023 at the lowest, 314 reads -2048 outside the field of view. Slices 0.8 or 1 mm apart, pixels 0.45 to 0.93 mm on the test CTs.
  • Overlap. The headers of all 340 training images and of PENGWIN 2024's 100 CTs were read: a pair counts as one scan when spacing and axes agree, the origins lie on one voxel lattice and the volumes overlap in the world. That check finds each PENGWIN 2024 CT as PENGWIN 2026 subject 001 to 100 (100 of 100; 2024 case 006 and 2026 subject 006 are equal voxel for voxel). Among the femur subjects it finds seven copies equal voxel for voxel (376, 377, 304, 382, 384, 385 and 313 repeat 252, 253, 257, 266, 270, 273 and 275), two scans cut into both sides (378 and 379 are the other sides of 255 and 256) and two pairs with different voxels on one lattice that may be one patient scanned twice (414 with 253, 321 with 307): the second of each is held out. It finds no femur CT on the lattice of a pelvic subject of either edition, nor of CTPelvic1K_Clinic's 20 built CTs.
  • Start. A mask of the whole image scores a Dice of 0.018 at the median of the test CTs (0.008 to 0.059), an Otsu mask 0.046. TotalSegmentator 2.18.0 (total) scores 0.911 on test case 263 (161 mL against the truth's 137 mL).
  • Outside numbers. None: the challenge ranks fragment instances on its hidden test set; the benchmark scores the merged femur, with the surface Dice at 1 mm.

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
16 test CTs drawn with seed 2026 among the 159 independent femur subjects of the public training set, and 14 development CTs. Eleven subjects are held out as another subject's copy, other side or likely second scan (draw.held_out). The challenge's 80 femur test CTs are not public; the pelvic subjects of the same release are not used.
Truth
contours by a program, verified by people the femur, the union of the deposit's fragments of it (values 151 to 200): a network's draft of the bones corrected by hand in 3D Slicer, the fragments cut by hand by one of three annotators, then validated by two senior experts; fragments under 500 mm³ are omitted
Licence
CC BY-NC-SA 4.0
14 cases for tuning, never tested

294, 419, 418, 332, 326, 268, 403, 325, 406, 312, 317, 303, 380, 292