fidopt Interactive view
CBCT to CT, synthesis

COBRA2026, CBCT to CT, pelvis

Cone-beam CT of one radiotherapy patient's pelvis, reconstructed by the provider from the measured projections, and the planning CT to synthesise from it, registered deformably onto the CBCT by the provider (COBRA2026 training set).

30 test casespelvissource CBCT, target CTresearch only
Look at the images
Read before comparing (8)
  • Pool. COBRA2026 holds the CBCT projections, geometry, planning CT and reconstructions of 867 radiotherapy patients treated for pelvic cancers at six European centres, A to D on Elekta linacs and F and G on Varian ones. Only the 692 training cases are public; the provider withholds its 52 validation and 123 test cases until March 2032.
  • Source. cbct_rtk.mha as delivered: the CBCT the provider reconstructed from the measured projections with RTK's FDK on the clinical reconstruction's grid, its attenuation turned into CT numbers by one fixed formula, without scatter correction for the Elekta centres. Its values are not Hounsfield units, least at F and G: inside the mask on the test cases it misses the target by 149 to 211 HU at A, 82 to 151 at B to D, 238 to 269 at F and 174 to 375 at G.
  • Target. ct_def_masked.mha as delivered: the planning CT registered to that CBCT by elastix, rigid then a B-spline on the IMPACT similarity, -1024 outside the CBCT's field of view. Where the deformation missed the gas pockets, the provider filled them with soft tissue or took them from a synthetic CT a network made from the CBCT: there the target is partly a network's output. Centre B's planning CTs had contrast, which the provider suppressed inside the bowel.
  • Registration. The provider looked at every case's checkerboard overlays and kept the cases it found aligned; its training cases passed a looser check than its validation and test cases. What the registration left misaligned counts as error for every method. On A172, D030, G025 and F020 bones and body outline fall on the CBCT's.
  • Mask. The body inside the field of view, from the source alone: the CBCT above -500 inside the provider's field-of-view mask without the couch, its largest part, its holes filled slice by slice; 10% to 68% of the grid. The couch and the parts of the CT outside the CBCT's field are left out.
  • Baseline. Inside the mask on the 30 test cases, at CT's range of 4095, a constant at the target's median scores an MAE of 94 HU (73 to 111), a PSNR of 27.3 dB (22.0 to 29.3) and an SSIM of 0.41 (0.33 to 0.54). The CBCT itself scores 172 HU (82 to 375), 26.3 dB and 0.19, better than the constant on 5 of the 30: the body is mostly soft tissue near the constant, and the CBCT's offsets cost more.
  • SynthRAD. The SynthRAD2023 pelvis sets come from three Dutch centres, UMC Utrecht among them, and Utrecht is a COBRA centre (likely A, whose CTs are described in Dutch). No COBRA training CT has the acquisition of a SynthRAD2023 pelvis CT or of a SynthRAD2025 Task 2 CT: 25 share tube current, exposure, voltage and pixel size with one, none also the exposure time and the slice count. The same comparison pairs SynthRAD2023's 2PA004 with 1PA157, which correlate 0.91 inside the bodies after a rigid fit where two patients read 0.48: one planning CT in both of its tasks, so SynthRAD_MRCT_Syn must keep 1PA157 out on any redraw, 2PA004 being a SynthRAD_CBCTCT test case.
  • Outside numbers. None: the COBRA2026 challenge ranks CBCT reconstruction from the projections, which the corpus does not use.

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
30 test patients drawn with seed 2026 by centre in proportion among the 692 of the training set, the provider's validation and test sets being withheld until March 2032: 5 from each of A, B, C and D, 4 from F and 6 from G. Then 30 development patients, the last of the same seeded order, the centres in the same proportion.
Truth
target image as released the patient's planning CT registered deformably onto the CBCT by the provider (elastix, rigid then a B-spline on the IMPACT similarity, its gas pockets matched to the CBCT's), masked to the CBCT's field of view, in Hounsfield units
Licence
CC BY-NC 4.0
30 cases for tuning, never tested

G061, A162, C213, D040, B110, F130, G163, A066, C009, D357, B066, F022, G158, A243, C182, D531, B050, G164, F020, A069, C178, D334, B111, G167, A156, C036, D054, F059, G168, B080