fidopt Interactive view
MR to CT, synthesis

Learn2Reg 2021 AbdomenMRCT, MR to CT

Abdominal MR (source) and CT (target) of one patient, the challenge's paired images on its 2 mm grid, the MR registered onto the CT by the corpus (Learn2Reg 2021, AbdomenMRCT).

3 test casesabdomensource MR, target CTresearch only
Look at the images
Read before comparing (8)
  • A case is one of the eight patients of the task with both an MR and a CT (TCGA patients of TCIA, listed on the challenge's datasets page): the challenge's T1 MR and CT, both resampled by the organisers onto one grid of 192 x 160 x 192 voxels of 2 mm and roughly aligned, not voxel for voxel. They are read from L2R_AbdomenMRCT as the corpus built it: the challenge's archive is behind a password on the organisers' cloud, so prepare.py takes the built pairs (FIDOPT_DATA, FIDOPT_DEV), checked by fidopt verify. The MR is 0 past its own field of view, which covers 117 to 189 of the grid's 192 axial slices.
  • Overlap. The images are L2R_AbdomenMRCT's (and those of its two derived segmentation sets, L2R_AbdomenMRCT_SegCT and _SegMR): the 3 test CTs, the targets here, and the 5 development CTs are the moving images every method receives in L2R_AbdomenMRCT and the inputs of L2R_AbdomenMRCT_SegCT, kept on purpose with the same split so that a patient is test or development everywhere; only 3 test cases follow from this. This dataset cannot be derived from it with "from" (a synthesis case taken from a pair needs the pair aligned voxel for voxel, and the challenge's are not): it is built from the same eight patients. As L2R_AbdomenMRCT says, TotalSegMR's training split holds other MR series of three of these patients.
  • The CT is the reference and is never resampled: the corpus registered the MR onto it once and keeps the transform. Registration recipe, as in Valentin Boussot's published MR to CT registrations of SynthRAD 2023 and 2025 (huggingface.co/datasets/VBoussot/synthrad2023-impact-registration): IMPACT-Reg 1.8.7's preset Elastix_IMPACT_Static (presets VBoussot/ImpactReg at 09d2a191), elastix-IMPACT 1.1.0 (cu130): a rigid stage on mutual information, then a cubic B-spline over four resolutions driven by IMPACT on TotalSegmentator M730 (its last decoder layer at the two coarse levels, its segmentation head at the two fine ones) and MIND, Static mode, with mutual information and a bending-energy penalty. One option differs from the preset's defaults: final_grid_spacing=10 (mm; the preset's default is 20). The start is the organisers' alignment (the identity). The registration reads the CT with its voxels outside the body set to -1024 (the couch out of the networks' view), the MR as delivered; its fixed mask is the MR's field of view 10 mm inside its edge as the start carries it, less 15 mm at each end of the CT's grid; its moving mask the MR's field of view 10 mm inside its edge.
  • transforms/<case>.npz holds each case's box and its transform as the registration composed it (the start, the B-spline, the registration's rigid stage; the B-spline's coefficients in hundredths of a mm, as 16-bit integers), 2.4 MB for the eight cases, and prepare.py applies it: building again gives the same files without registering.
  • Each registration was checked on the challenge's own labels of the liver, the spleen and the kidneys, drawn on both images (L2R_AbdomenMRCT's truth), carried onto the CT's grid inside the MR's field of view. The abdomen moves between the two scans, taken apart in time with other breathing and other filling of the bowel, and the organisers' alignment leaves it far off: the mean Dice of the four organs goes from 0.25, 0.20 and 0.49 at the start to 0.89, 0.83 and 0.90 after the registration on the three test cases, and from 0.11 to 0.59 at the start to 0.81 to 0.89 after it on the development cases (three organs on case_0001 and case_0002, whose CT labels miss the left kidney). Organ by organ on the test cases: liver 0.40, 0.39, 0.74 to 0.95, 0.90, 0.91; spleen 0.11, 0.22, 0.49 to 0.87, 0.87, 0.90; right kidney 0.25, 0.07, 0.37 to 0.91, 0.77, 0.90; left kidney 0.25, 0.11, 0.35 to 0.83, 0.78, 0.88. The registration's rigid stage alone reaches 0.49, 0.13 and 0.66: the B-spline does most of the work. No case was dropped. What it leaves (the kidneys of 0007 at 0.77 and 0.78; the bowel, whose gas and contents differ between the two days) counts as error for every method alike.
  • The B-spline folds: 0.3, 2.8 and 10.4% of the mask has a Jacobian determinant of 0 or less on the three test cases (0008, 0006, 0007), 1.8 to 13.2% on the development cases, in the lower abdomen among the bowel loops. On 0007 it reaches 31% of the mask in the second quarter of its slices from the feet (38 slices more than 20% folded), where the registered MR is smeared and does not show the tissue the CT holds; the mask does not leave the folded voxels out. Found by the review of 2026-10-15.
  • The mask is the CT's body (above -500 HU, opened by 2 mm in the slice, its largest component, each slice's holes filled) where the MR's field of view reaches both through the whole transform and through its rigid part alone: past the MR's own slices the B-spline is extrapolated, not registered, and its stretched MR is left out. 19 to 22 litres on the test cases. The MR's slices are thick and resampled by the organisers to 2 mm: it is smooth along the body's axis.
  • Inside the mask on the three test cases, a constant at the CT's median there (-89, -44 and -62 HU; the image fidopt check compares a method with) scores an MAE of 107, 102 and 101 HU (103 on average), a PSNR of 26.7 to 27.8 dB (27.2) and an SSIM of 0.35 to 0.43 (0.40), by the corpus's evaluation. The source is an MR in arbitrary units: no MAE of it as it is.

No method has a result on every case of this dataset yet.

Case by case

Click a score to look at that result: the target image, and over it the method's synthetic image. Underlined scores have their image; for the others you see the case without the result.

About the data

Test cases
L2R_AbdomenMRCT's split kept: its three test pairs (the challenge's validation pairs 0006 to 0008) are the test cases and its five development pairs (the training pairs 0001 to 0005) the development cases. One case a patient: the eight paired patients of the task.
Truth
target image by a program the patient's CT as the challenge delivers it on its 2 mm grid, cut to the box the MR reaches, in Hounsfield units; the MR registered onto it by the corpus with IMPACT-Reg 1.8.7 (Elastix_IMPACT_Static, final B-spline grid 10 mm), the transform kept in transforms/
Licence
CC BY 3.0 (TCIA TCGA CT) and CC BY-NC-SA 4.0 (CHAOS MR)
5 cases for tuning, never tested

case_0001, case_0002, case_0003, case_0004, case_0005