fidopt Interactive view
MR to CT, synthesis

TopCoW 2024, MR angiography to CT angiography

Time-of-flight MR angiography of the head of one stroke-centre patient (source) and the same patient's CT angiography (target), registered onto it by the corpus (TopCoW 2024 training set, University Hospital Zurich).

10 test casesbrainsource MR, target CTresearch only
Look at the images
Read before comparing (10)
  • Pool. TopCoW 2024's 125 labelled training patients, each with a time-of-flight MR angiography and a CT angiography, from the University Hospital Zurich's Stroke Center in 2018 and 2019. TopCoW and TopBrain, which register these two images, hold 40 of them: they are left out, so that no CT angiography a registration method is given there is a target here. 85 remain.
  • Images. The source is the MRA as delivered, 3D time of flight mostly at 3 T, about 0.3 mm in plane and 0.6 mm slices, a slab around the circle of Willis. The target is the CTA, in Hounsfield units, cut to the braincase by the provider (its front edge can cross the face), resampled linearly onto the MRA's grid. The two are separate examinations on separate scanners, how far apart the paper does not say.
  • Registration. The release registers nothing, their frames up to 1.5 m apart. The corpus registered the CTA rigidly onto the MRA as it computes an elastix-rigid start (mutual information inside each image's body, from a few starts, on images reduced to 512 voxels an axis), from the intensities alone; the CTA turns by 5 to 23 degrees on the test cases.
  • Check. The release's circle-of-Willis labels, drawn by people on both images, carried through the transform: over the classes both images hold, without the internal carotids, the mean Dice is 0.47 to 0.78 on the test cases (median 0.66) and 0.44 to 0.76 on the development cases, near TopCoW's own start (0.61 at the median). The arteries are 1 to 4 mm wide, so these Dice values leave about a millimetre of misalignment, and the skull and vessels fall on each other in the overlays of three test cases. What the registration leaves counts as error for every method: graded automatic.
  • Mask. The resampled CTA above -500 HU, its largest part, the holes of each axial slice filled, inside the CTA's own grid: the head within the MRA's slab, 1.8 to 2.6 litres on the test cases.
  • Contrast. The target is a contrast CT: its arteries are bright, as on the MRA, but its veins and sinuses are too, which time of flight does not show.
  • Baseline. Inside the mask on the 10 test cases, a constant at each CTA's median there (34 to 50 HU) scores an MAE of 176 (136 to 222) HU, a PSNR of 20.5 (18.8 to 22.3) dB and an SSIM of 0.60 (0.57 to 0.64), by the corpus's evaluation. On the 10 development cases, ImpactSynth_MR (TotalSynth's MR-to-CT model, a KonfAI app) scores an MAE of 97 to 126 HU (107 on average) where the constant scores 146 to 211 (176): a method faithful to the MRA beats the constant on every case.
  • Overlap. TopCoW and TopBrain hold the other 40 patients. ISLES24_NCCTCTA_Syn and ISLES2024 found no Zurich CT angiography of ISLES 2024 sharing a pixel spacing with any of TopCoW's 125. AneurysmTOF is another cohort.
  • Overlap, by image. Each of the 20 CTAs, cut to its top 90 mm, was registered on the skull to each of ISLES24_NCCTCTA_Syn's 7 Zurich CTAs and correlated: at most 0.82 (median 0.59), where each Zurich CTA cut to a TopCoW-sized box and moved by 20 degrees finds itself at 0.993 or more and its own non-contrast CT, another scan of the same head, at 0.968 to 0.986. ISLES2024's 9 other Zurich patients, whose CTAs the corpus does not build, are not checked by image.
  • Outside numbers. None: the challenge ranks vessel segmentation, not synthesis.

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
10 test patients drawn with seed 2026 among the 85 of the release's 125 labelled training patients that neither TopCoW nor TopBrain holds, then 10 development patients. The 5 validation pairs without labels are not used; the challenge's test patients are hidden.
Truth
target image by a program the CT angiography in Hounsfield units, registered rigidly onto the MR angiography by the corpus from the intensities and resampled once onto its grid, the transform kept in transforms/
Licence
opendata.swiss 'Open use. Must provide the source. Use for commercial purposes requires permission of the data owner.'
10 cases for tuning, never tested

topcow_058, topcow_067, topcow_084, topcow_159, topcow_041, topcow_068, topcow_136, topcow_056, topcow_160, topcow_076