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

ISLES'24, non-contrast CT to CT angiography

Admission non-contrast head CT (source) and CT angiography (target) of a patient with an acute ischaemic stroke, the angiography registered rigidly onto the non-contrast CT and resampled onto its grid by the provider (ISLES'24 training set, Munich and Zurich).

12 test casesheadsource CT, target CTresearch only
Look at the images
Read before comparing (10)
  • Pool. The challenge's 149 public training patients, each treated by thrombectomy for an acute ischaemic stroke: 100 at the University Hospital of the Technical University of Munich, 49 at the University Hospital Zurich. The 96 test patients are hidden.
  • Images. The source is the admission non-contrast CT, defaced by the provider, in Hounsfield units: 0.32 to 0.49 mm in plane, 2 mm slices in Munich and 4 mm in Zurich, often tilted with the gantry; Zurich's pad outside their field of view with -8192. The target is the CT angiography of the same admission, cropped to the head by the provider, registered rigidly onto the NCCT (elastix or NiftyReg) and resampled onto its grid by linear interpolation. Both keep their values as delivered.
  • Alignment. Registered again rigidly on the skull (above 500 HU, correlation inside the mask, a check that brings back a known 1.5 degree and 2 mm motion to within 1.2 mm), the NCCT moves at most 0.15 to 1.5 mm over the head on the test cases (median 0.4 mm). Mutual information over the whole mask, where the contrast and the soft tissue weigh too, finds 0.6 to 3.9 mm (median 0.9 mm), always as a slight nod (0.1 to 1.4 degrees about the left-right axis) that the skull does not show, the largest on Zurich's 4 mm slices (stroke0142). Bone above 700 HU overlaps at a Dice of 0.68 to 0.94 (median 0.87), lowered also by the two scans' kernels. What is left of the motion counts as error for every method.
  • Mask. Where both scans hold the head (above -500 HU, opened by one in-plane voxel, the largest part, holes filled slice by slice), 2.2 to 3.1 litres on the test cases: it leaves out the air, the couch and what one scan's field of view or defacing removed and the other kept.
  • Baseline. Inside the mask on the 12 test cases, with KonfAI's criteria at CT's range of 4095, a constant at each CTA's median there scores an MAE of 208 HU (173 to 259), a PSNR of 19.5 dB (17.4 to 21.3) and an SSIM of 0.48 (0.41 to 0.52); the NCCT itself, the line to beat, 78 HU (59 to 95), 29.0 dB (26.9 to 31.5) and 0.87 (0.80 to 0.92).
  • What changes. Contrast in the arteries and veins, a slightly brighter brain (its median 8 to 17 HU above the NCCT's on three cases looked at), and another noise and kernel; the arteries' enhancement depends on each patient's timing and occlusion.
  • ISLES2022. Seven Munich patients' follow-up DWI is the scan of an ISLES2022 case (ISLES2024's notes): held out, with ISLES2024's own patients.
  • TopCoW and TopBrain. These come from Zurich's stroke centre in 2018 and 2019. No Zurich CT angiography here shares a pixel spacing with any of TopCoW's 125, and the eight that carry an acquisition time match none of TopCoW's times; the spacings may have been resampled, so the heads were compared too: each cut to its top 90 mm, registered rigidly (mutual information, up to eleven starts), then Pearson r of the values clipped to -100..1300 HU. Seven Zurich non-contrast CTs found their own angiography cut to a TopCoW-like braincase box at r 0.955 to 0.973 (the same angiography uncropped, 0.988 to 0.993). Against the corpus's 40 TopCoW and TopBrain angiographies, the 16 Zurich patients of this set and ISLES2024 reach at most 0.844 (all 640 pairs scored, stroke0159 against topcow_150 on the two angiographies, 0.63, where the check finds a patient's own angiography, cut and moved, at 0.97), the 49 at most 0.851. Most pairs of different heads never align (median r 0.05): the check rests on a same patient's pair aligning, as the seven controls did from the same starts.
  • Overlap. ISLES2024 holds 27 other patients of the same archive, as segmentation on the follow-up MR. VerSe's spine CTs also come from Munich: no image is shared.
  • Outside numbers. None: the challenge ranks infarct prediction, 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
12 test patients drawn with seed 2026 by centre, in proportion (8 Munich, 4 Zurich), among the 115 training patients outside draw.held_out, and 10 development patients, the last of the same seeded order. Held out: ISLES2024's 27 test and development patients, so that no patient is in both sets, and the seven Munich patients whose follow-up DWI is an ISLES2022 case.
Truth
target image as released the CT angiography of the same admission in Hounsfield units, registered rigidly onto the non-contrast CT and resampled onto its grid by linear interpolation by the provider (elastix or NiftyReg)
Licence
CC BY-NC-SA 4.0
10 cases for tuning, never tested

stroke0048, stroke0165, stroke0108, stroke0045, stroke0179, stroke0189, stroke0038, stroke0159, stroke0115, stroke0188