fidopt Interactive view
CT to CT, another reconstruction, dose or contrast phase, synthesis

MCT-LTDiag, non-contrast to portal venous CT

The non-contrast CT (source) and the portal venous CT (target) of one four-phase liver examination of a patient with a liver tumour, the first registered onto the second by the provider (MCT-LTDiag, Peking Union Medical College Hospital).

30 test casesabdomensource CT, target CT
Look at the images
Read before comparing (6)
  • Alignment. The provider registered the non-contrast phase onto the portal venous one with ITKElastix's default B-spline map (deformable, mutual information, 20 mm grid, no rigid stage), checked by no one, and wrote it on that grid; nothing is registered or resampled here. Bone over 400 HU overlaps between the two at a median Dice of 0.87 on the test cases (0.63 to 0.93); on the three looked at, bones, kidneys and body outline fall together, bowel gas and the contrast in vessels and kidneys differ. Eight patients are held out: three whose non-contrast file lies on its own grid (230906d01, 240620a04, 240122d55), three whose non-contrast file is the raw scan in 16-bit integers, not registered by the provider (240112d26, 231206d15, 230906d08), one whose registered file holds no image over its top 30 slices (240112d44) and one with organs visibly displaced (230816b14).
  • Mask. The body on the source where both phases hold data: the registration writes exactly 0 outside the non-contrast scan's field, up to a fifth of some grids, and the target reads below -1024 outside its reconstructed circle; both are left out of the mask, and the images keep those values.
  • Images. The source is 32-bit floats from the provider's interpolation on every case, the target 16-bit Hounsfield units; slices 5 mm, pixels 0.56 to 0.98 mm. The target reads -1024, -2048 or -3024 outside its reconstructed circle.
  • Overlap. These 60 patients are other patients of the deposit than MCTLTDiag's, so no patient's portal venous CT is a segmentation input there and a target here.
  • Start. On the test cases a constant at the target's median in the body scores a median MAE of 153 HU (103 to 217), a PSNR of 23.9 dB and an SSIM of 0.26; the source as it is, 35.8 HU (25 to 65), 34.7 dB and 0.87. A method is to be compared with the source as it is, not with the constant: the room is from about 36 HU down to the 20 or so that registration noise leaves, plus the enhancement of vessels, kidneys, liver and tumours.
  • Outside numbers. None: the deposit has no synthesis benchmark.

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 tumour kind, then 30 development patients, the last of the same seeded order, the kinds in proportion; one examination a patient. The pool is the deposit's 517 patients less MCTLTDiag's 60 and eight whose non-contrast phase is not, or not usably, the provider's registration (draw.held_out). The deposit has no test set.
Truth
target image as released the portal venous phase as delivered (NIFTI/pvp.nii.gz), in Hounsfield units; the source is the non-contrast phase the provider registered onto its grid with ITKElastix (NIFTI/nc.nii.gz)
Licence
CC0 1.0
30 cases for tuning, never tested

230312a14, 240504e01, 240122d66, 231025c35, 240620a44, 231109b12, 240722e83, 240122d56, 230218a4, 231025c25, 230816b09, 240504e30, 231206d08, 240620a32, 240229c13, 240504b34, 240722e102, 230408a1, 240229c09, 240504b07, 240722e52, 230312a6, 231025c27, 230218b3, 240722e92, 230525d8, 240620a08, 240504b14, 240229c04, 240722e55