fidopt Interactive view
PET to CT, synthesis

Multimodal-HC, NAC PET to CT

FDG PET without attenuation correction (source) and the low-dose CT reconstructed for its attenuation correction (target), from the vertex to mid-femur, of one healthy adult's PET/CT session on a total-body scanner (Multimodal-HC, Rigshospitalet).

10 test caseswhole bodysource PET, target CT
Look at the images
Read before comparing (10)
  • Cohort. 100 healthy adults from the Danish Blood Donor Study, scanned at Rigshospitalet, Copenhagen; the open release holds 80, ten in each sex and age group (18-34, 35-49, 50-69 and 70-99 years).
  • Alignment. CT and PET share the table and the scanner's frame, on the same 531 slices of 2 mm; the static PET is the last 20 minutes of a 70-minute acquisition starting about 2 minutes after the CT. prepare.py brings the PET onto the CT's grid by linear interpolation in that frame, 0 past its grid; nothing is registered.
  • Source. The static PET without attenuation correction, on voxels of 1.65 x 1.65 x 2 mm, in the scanner's PROPCPS (proportional to counts per second), not Bq/mL. Uncorrected, the skin reads higher than the tissue beneath it, and a halo of counts surrounds the body.
  • Target. The CT reconstructed for the attenuation correction: 120 kVp, kernel Bf37f, no contrast agent. Read through each file's header scaling, as float32, its values are the scanner's whole Hounsfield units quantized by that scaling. Both images are cut to the mask's bounding box plus 10 voxels.
  • Mask. The provider's TotalSegmentator body (version 2.5.0, task body) on this CT, where the PET holds data: 40.9 to 95.0 litres a test case, without the table or the faces blanked before release. The PET's zeros take at most 25 mL out of a case's body.
  • Check. The PET's body matches the mask at a Dice of 0.906 to 0.932 on the ten test cases, its skin spreading a few millimetres outward. Its brain edge lies within 6 mm of the CT's inner skull at the vertex and 1.5 mm at the sides: no head moved. Overlays put the trunk and the head on their CT on all 20 cases.
  • Limbs. A forearm or a hand often lies elsewhere in the PET than in the CT: clearly on sub-006, slightly on five other test cases and five development cases. No method can place it from the PET, so it counts against every method. No case was dropped.
  • Breathing. The CT is a snapshot under free breathing; the static PET averages 20 minutes, about 52 to 72 minutes later. The providers leave out the slices within 4 cm of the top of the liver; the corpus scores the whole mask, so a diaphragm that moved counts against every method.
  • Baseline. Inside the mask on the ten test cases, a constant at the CT's median there (-89 to 38 HU) scores an MAE of 131 (93 to 153) HU, a PSNR of 24.3 (22.6 to 26.8) dB and an SSIM of 0.40 (0.31 to 0.58). The source, in PET units, has no MAE as it is.
  • Outside numbers. None scored as here: the providers score the attenuation map and the PET corrected with it, and the article's U-Net reports an MAE of 84.8 HU in a region it does not state. Every test case is released with its CT, so a network trained on this release has seen it.

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

The first test case of Multimodal-HC, NAC PET to CT
The first test caseClick to look at every case.
Test cases
10 test subjects drawn with seed 2026 by sex among the 80 of the open release (5 women, 5 men), then 10 development subjects (5 and 5). The provider holds the other 20 back as a test set, to be released in 2027. A later dataset of the same subjects, such as their MR and CT for registration, keeps each in its partition or leaves it out.
Truth
target image as released the subject's low-dose CT reconstructed for the PET's attenuation correction, in Hounsfield units as delivered, on its own grid cut to the body; the PET of the same session is brought onto it through the scanner's frame, which both images share
Licence
CC BY 4.0
10 cases for tuning, never tested

sub-013, sub-001, sub-095, sub-010, sub-062, sub-020, sub-024, sub-009, sub-067, sub-008