fidoptOAIZIB-CM
Sagittal 3D DESS MR of the right knee of two adults of the Osteoarthritis Initiative at baseline, paired by us, with the femur, the tibia and their cartilage labelled (OAIZIB-CM).
Read before comparing (10)
- Knees. All 507 are right knees, so no image is flipped, of adults aged 45 to 79 scanned at baseline on a Siemens 3T Trio with the OAI's DESS. They span Kellgren-Lawrence grades 0 to 4, so a pair may join a knee without radiographic osteoarthritis to a severe one (046_348, 349_043, 367_080); one test subject has no grade.
- Truth. Experienced users at the Zuse Institute Berlin revised a statistical shape model's segmentations of the femur, the tibia and their cartilage. OAIZIB-CM splits the tibial cartilage into medial and lateral by a script, splitTC.m, which also drops isolated clusters. No agreement between raters is given.
- Field of view. 140 mm square: the femur and tibia labels stop where it cuts the shafts, at heights that depend on how the knee sat in the coil, and a piece of shaft one image holds and the other does not overlaps nothing whatever the field. The patella, fibula, menisci and patellar cartilage are not labelled.
- Cartilage. The cartilage labels are thin, 9 to 26 cm³ for the femoral and 0.7 to 3.5 cm³ for each tibial cartilage on the first 40 subjects, so a one-voxel shift changes their Dice substantially.
- Alignment. Each knee keeps its native grid. A starting translation, found from MR intensities and body masks only, is written in the moving image and label headers without resampling. Pose, knee bend and anatomy are left to the methods.
- Baseline. With no registration after this start, the 25 test pairs score a median Dice of 0.4158 (0.1877 to 0.6013) and a median HD95 of 9.43 mm.
- Labels. OAI-ZIB gave its masks without a pose; OAIZIB-CM gives each mask its image's header, and prepare.py checks that each label map lies on its image's grid where a right knee puts it.
- Images. 16-bit DESS intensities with no common scale. No mask is given.
- Overlap. KneeMR_Seg's cases are these pairs' fixed images, its test and development subjects KneeMR's; no other dataset of the corpus holds these knees.
- Outside numbers. None: the pairing is ours.
| Accuracy | Smoothness | Cost | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Change | HD95 mm | SDlogJ | Folds % | Time s | RAM GB | GPU GB | Images |
| 1 | ConvexAdam (MIND)GPUapp | 0.667 ±0.077 | +0.262 | 6.81 ±1.91 | 0.103 | 0.40 | 21 | 3.3 | 6.2 | |
| 2 | DemonsCPUapp | 0.511 ±0.063 | +0.106 | 9.08 ±1.70 | 0.090 | 0.00 | 19 | 3.2 | ||
| No registration | 0.405 ±0.045 | 10.52 ±1.65 | 0.000 | 0.00 | ||||||
The order averages each method's ranks on Dice, HD95 mm, SDlogJ, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 25 pairs. Two neighbours are tied when a paired Wilcoxon test on the shared pairs does not find the upper one better on Dice (5%): the places a method could hold are then shown under its own. The best value of each column is in colour. Click a row to see its images, a column title to sort by it. Time and memory measured on Laptop.
Dice: how much the organ contours overlap, 1 is perfect. HD95: distance between the contours' surfaces. SDlogJ: how irregular the deformation is, 0 is perfectly smooth. Folds: share of the image where the deformation folds over itself.
Pair by pair
Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.
About the data
- Test pairs
- 25 test pairs drawn with seed 2026 among the 253 pairs, then 22 development pairs from the 228 left; each person is in one pair only. The 507 subjects are paired two by two in sha256("2026:KneeMR:<subject>") order, and one subject is left out.
- Truth
- contours by experts the dataset's labels: femur, femoral cartilage, tibia, medial and lateral tibial cartilage (OAI-ZIB's labels, the tibial cartilage split in two by a script)
- Source
- huggingface.co/datasets/YongchengYAO/OAIZIB-CM
zenodo.org/records/14934086
doi.org/10.1016/j.media.2018.11.009
doi.org/10.12752/4.ATEZ.1.0
doi.org/10.1016/j.media.2023.103035
github.com/YongchengYAO/CartiMorph
nda.nih.gov/oai - Licence
- CC BY-NC 4.0, over OAI images (the OAI's data use agreement) and ZIB segmentations for scientific use only
22 pairs for tuning, never tested
241_261, 474_167, 266_313, 156_357, 274_050, 082_084, 119_228, 132_321, 113_419, 222_092, 311_272, 037_008, 103_362, 415_464, 329_169, 288_421, 359_055, 198_192, 136_067, 314_206, 041_005, 501_086