fidoptTopBrain 2025/2026
Time-of-flight MR angiography (fixed) and CT angiography (moving) of the head of one stroke-centre patient, the arteries of the whole brain labelled on both (TopBrain training set, its 2025 labels).
Read before comparing (12)
- Alignment. The two scans are separate examinations whose scanner frames differ by 185 to 1488 mm, with rotation. The moving CTA's and labels' headers carry one rigid start estimated from intensities and body masks only, turning by 4.54 to 32.54 degrees on the 20 patients; both images keep their native grids.
- Baseline. With no registration after that start the ten test pairs score a median Dice of 0.442 (0.367 to 0.515) and an HD95 of 5.80 mm (3.49 to 7.62).
- Truth. One doctoral researcher, trained by a senior neurosurgeon, labelled both images in virtual reality with TopCoW's workflow; later images were pre-labelled by a network retrained as labels accumulated, then corrected voxel by voxel, the hard cases checked by the neurosurgeon. No agreement between annotators is given.
- Extents. A branch is labelled as far as each image shows it: the CTA's M3 holds 1.3 to 1.5 times the MRA's volume, its P3-P4 1.2 to 3.2 times, so a perfect field leaves part unmatched. Arteries are 1.2 to 4 mm wide, and a 1 mm shift halves a small one's overlap. Only classes both maps hold are scored, 21 to 26 a test pair.
- Partial. The internal carotid and ophthalmic arteries are labelled over different lengths on the two images, the MRA's carotids holding 8 to 9 times the CTA's volume. They form a partial set, shown, never ranked: median Dice 0.115 and HD95 37.14 mm on the test pairs after the start. The other 30 classes are ranked.
- Removed. The CTA's veins and the MRA's external carotid branches share the values 35 to 42 on the two maps: prepare.py sets them to 0, so veins are not measured.
- Images. The MRA is about 0.3 mm in plane with 0.6 mm slices, the CTA about 0.40 to 0.55 mm in plane and 0.3 to 0.75 mm through it; both are defaced, the CTA in Hounsfield units, the MRA without unit. Where the MRA slab extends below the CTA, part of a vertebral artery cannot be matched.
- Cohort. Patients of the University Hospital Zurich's Stroke Center admitted for a suspected stroke, with no large aneurysm in the circle of Willis and many large vessel occlusions and severe stenoses. Neither paper says how far apart the two examinations were.
- Overlap. The images are TopCoW 2024's, byte for byte. Five of the 25 patients are in datasets/TopCoW (001, 002, 003 and 014 among its development cases, 008 among its test cases) and stay there, so no image is in two datasets.
- Labels version. The August 2026 release adds 36-class labels whose making is not documented; the corpus uses the 2025 labels the preprint describes.
- Outside numbers. None: the challenge ranks segmentation, not registration.
- Licence. Commercial use needs the prior permission of the University Hospital of Zurich's Department of Neurology.
| Accuracy | Smoothness | Cost | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Change | HD95 mm | SDlogJ | Folds % | Time s | RAM GB | GPU GB | Images |
| 11–2 | ConvexAdam (MIND)GPUapp | 0.632 ±0.026 | +0.184 | 5.34 ±1.10 | 0.131 | 2.08 | 29 | 5.0 | 13.5 | |
| 21–2 | NiftyRegCPUapp | 0.604 ±0.030 | +0.156 | 5.92 ±1.06 | 0.594 | 3.34 | 228 | 4.5 | ||
| No registration | 0.448 ±0.037 | 5.67 ±1.11 | 0.000 | 0.00 | ||||||
The order averages each method's ranks on Dice, HD95 mm, SDlogJ, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 10 pairs. Two neighbours are tied when a paired Wilcoxon test on the shared pairs does not find the upper one better on Dice (5%): the places a method could hold are then shown under its own. The best value of each column is in colour. Click a row to see its images, a column title to sort by it. Time and memory measured on Laptop.
Dice: how much the organ contours overlap, 1 is perfect. HD95: distance between the contours' surfaces. SDlogJ: how irregular the deformation is, 0 is perfectly smooth. Folds: share of the image where the deformation folds over itself.
Pair by pair
Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.
About the data
- Test pairs
- 10 test pairs drawn with seed 2026 among the 20 training patients that TopCoW does not hold, the 10 others for development. The other 5 training patients stay in TopCoW; TopBrain's 20 test patients are not public.
- Truth
- contours by experts partial by experts the dataset's 2025 voxel labels of the brain's arteries on each image, without the internal carotid and ophthalmic arteries: basilar and vertebral arteries, superior, anterior inferior and posterior inferior cerebellar arteries, posterior (P1-P2, P3-P4), middle (M1, M2, M3) and anterior (A1-A2, A3) cerebral arteries, posterior and anterior communicating arteries, anterior choroidal arteries, and a third A2 and A3 where there is one (30 classes)
- Source
- zenodo.org/records/21972006
topbrain2025.grand-challenge.org/data
topbrain2026.grand-challenge.org
doi.org/10.64898/2026.05.28.26354312
zenodo.org/records/15692630
arxiv.org/abs/2312.17670
opendata.swiss/en/terms-of-use - Licence
- opendata.swiss 'Open use. Must provide the source. Use for commercial purposes requires permission of the data owner.'
10 pairs for tuning, never tested
topcow_006, topcow_026, topcow_021, topcow_024, topcow_004, topcow_011, topcow_012, topcow_016, topcow_007, topcow_020