fidopt Interactive view
MR to CT, synthesis

DoseRAD2026, MR to CT, thorax and abdomen

0.35 T MR-Linac MR (bSSFP) of one patient and the CT to synthesise from it, the CT registered deformably to the MR by the provider (DoseRAD2026 training set, SynthRAD2025 centre B).

10 test casesthorax, abdomensource MR, target CTresearch only
Look at the images
Read before comparing (9)
  • Source. SynthRAD2025's centre B MR as delivered there: a ViewRay MRIdian 0.35 T MR-Linac, bSSFP (TrueFISP), 1 x 1 x 3 mm after SynthRAD2025's rigid registration, defacing, resampling and cropping. The proton subset's images are used, on that grid; the beam configurations and Monte Carlo doses are not.
  • Target. The planning CT registered deformably to the MR by the provider with ConvexAdam, every case looked at in three planes by two experts and obvious misregistrations left out. What that registration left misaligned counts as error for every method.
  • Air. In the abdomen the provider filled the CT's air cavities (below -200 HU) with the surrounding HU and wrote the MR's air cavities into it at -824 HU, blurred over one voxel. A method is scored against those cavities, not against the gas the CT saw.
  • Metal. The provider lists metal artefacts in 13 training patients; of those drawn, only development case 1ABB078.
  • Mask. Outside the intersection of the two body masks the provider set the CT to -1024 and the MR to 0. The mask is where either is above that value, holes filled; it differs from the provider's intersection by 0.2 % of its voxels or less on the cases looked at.
  • Checks. On 1ABB042, 1THB045 and 1THB058 the CT's bone lies on the MR's vertebrae, ribs and sternum, and the identity axes match the anatomy. The CT is in Hounsfield units, -1024 to about 3000.
  • Baseline. Inside the mask on the ten test cases, with KonfAI's criteria at CT's range of 4095, a constant at the target's median scores an MAE of 194 HU (108 to 321), a PSNR of 22.4 dB (18.5 to 26.3) and an SSIM of 0.30 (0.13 to 0.39); the thorax cases, mostly lung, are the hardest. The MR is not in Hounsfield units and gives no floor.
  • Exposure. SynthRAD2025's and DoseRAD2026's training cases are public with their CTs: a network trained on either may have seen these patients.
  • Overlap. The 75 ids were checked against every SynthRAD2025* descriptor. SynthRAD2025_MRCT holds 1THB002: on it the DoseRAD MR equals SynthRAD2025's voxel for voxel, only the origin moving, and the CT differs (deformed). SynthRAD2025_MRCT_Syn holds 11 other ids of the pool, none of the twenty used here, which its draw holds out. A later dataset built on SynthRAD2025 centre B keeps these twenty patients in their partition or leaves them out.

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
The provider's 40 test patients are withheld until March 2030, so 10 test patients are drawn with seed 2026 from the 75 of the training set, abdomen and thorax in proportion, then 10 development patients. Patient 1THB002 is left out: DoseRAD2026's images are SynthRAD2025 centre B's under the same patient ids, and SynthRAD2025_MRCT holds 1THB002 as a test case.
Truth
target image as released the patient's planning CT registered deformably to the MR by the provider (ConvexAdam, checked by hand in three planes, misregistered cases left out), air cavities of the abdomen corrected from the MR, in Hounsfield units
Licence
CC BY-NC 4.0
10 cases for tuning, never tested

1THB211, 1THB195, 1ABB155, 1THB048, 1ABB041, 1THB043, 1ABB036, 1THB027, 1ABB078, 1THB008