fidoptCHAOS MR T1/T2
T1-DUAL in-phase (fixed) and T2-SPIR (moving) MR of the upper abdomen of one patient, two series of one study, with the liver, kidneys and spleen drawn on each (CHAOS).
Read before comparing (8)
- Truth. Three radiology experts (10, 12 and 28 years of experience) drew every slice by hand, the majority vote being the label, once on the T1-DUAL and once on the T2-SPIR. These are two readings on two sequences: on the test pairs an organ is 0.90 to 1.73 times as large on the T2 as on the T1 (the kidneys a median 1.15).
- Slices. They lie 5.5 to 9 mm apart, so the labels are staircases along the body's axis. The two test pairs whose organs moved by less than 5 mm (case15, case21) overlap at a Dice of 0.90 and 0.87 with no registration: about the most a perfect field can reach here.
- Alignment. The two series share the study's frame and were taken 3.8 to 12.6 minutes apart on the test pairs, the T2-SPIR first; the provider says they are not registered. As delivered the organ centres lie 1.6 to 64 mm apart (median 19 mm), the T2's higher and further back, as when breathing out; the labels overlap at a Dice of 0.27 (case22) to 0.90.
- Start. Each image stays on its own grid, and the field is asked on the T1's grid. No rigid start and no known deformation are added: a fit on the organ centres would remove the breathing to be registered, and nothing outside the four organs is annotated to tell patient motion from organ motion. On case22, whose organs moved most, the spinal canals coincide.
- Field of view. It cuts an organ, the liver dome or a kidney's pole, on 6 of the 10 test pairs: no field brings back what one image does not hold.
- Contrast. The T2-SPIR suppresses fat, so fat bright on the T1 is dark on the T2 and the intensities do not map one to one. The T1-DUAL's opposed-phase echo is not used. 1.5 T Philips scanner, 1.36 to 2.03 mm in the slice.
- Overlap. Learn2Reg 2021's AbdomenMRCT used CHAOS MR only as unpaired auxiliary training scans, so multiGradICON, trained on that task, has seen these scans; its paired cases, which L2R_AbdomenMRCT holds, are eight TCGA patients of TCIA. No image is in both datasets.
- Cohort. The provider describes the organs as healthy, with no tumours, metastases or lesions at their borders. All patients come from Dokuz Eylul University Hospital, Izmir.
No method has a result on every pair of this dataset yet.
| Accuracy | Smoothness | Cost | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Change | HD95 mm | SDlogJ | Folds % | Time s | RAM GB | GPU GB | Images |
| No registration | 0.683 ±0.137 | 20.62 ±9.07 | 0.000 | 0.00 | ||||||
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
- 10 test pairs drawn with seed 2026 among the 10 labelled patients outside development, and 10 development pairs. Only 20 of the 40 MR patients have labels, the challenge's training set; the labels of the challenge's 20 test patients are withheld, and those patients are not used.
- Truth
- contours by experts the liver, right kidney, left kidney and spleen, drawn by hand on every slice of each sequence by three radiologists and merged by majority vote (the dataset's labels)
- Source
- zenodo.org/records/3431873
chaos.grand-challenge.org/Data
chaos.grand-challenge.org/Rules
doi.org/10.1016/j.media.2020.101950
learn2reg.grand-challenge.org/Learn2Reg2021 - Licence
- CC BY-NC-SA 4.0
10 pairs for tuning, never tested
case01, case02, case03, case10, case31, case20, case05, case19, case37, case32