fidoptTopCoW 2024
Time-of-flight MR angiography (fixed) and CT angiography (moving) of the head of one patient of a stroke centre, with the arteries of the circle of Willis labelled on both (TopCoW 2024 training set).
Read before comparing (10)
- Truth. Two doctoral researchers trained by a senior neurosurgeon labelled the vessels in virtual reality, each on half of the patients, viewing both images, uncertain cases going to the neurosurgeon or a panel; after the first 260 images a network pre-labelled and people corrected. On five patients the annotators agree at a Dice of about 0.90 a class, 0.76 to 0.89 for the communicating arteries and the third A2.
- Label ends. Arteries leaving the circle are labelled up to their first major bifurcation, so where a label ends is a judgement on each image.
- Carotids. The internal carotids are labelled along their whole course on the MRA but only from C7 on the CTA, the MRA's holding 3 to 4 times the volume on the first 13 cases built. No field can make those extents equal, so they form a separate label set (dice_ica, hd95_ica), shown, never ranked; the board ranks the other eleven classes.
- Small vessels. The arteries are 1 to 4 mm wide, so a one-millimetre shift halves a small artery's overlap. The communicating arteries and the third A2 hold a few tens of mm³ and are often on one image only: only the classes both maps hold are scored, 7 to 10 a test pair.
- Alignment. The scans come from separate scanners, frame origins up to about 1.5 m apart, both cropped to the braincase. The moving header carries one rigid start from the intensities and body masks, no vessel label used, turning the CTA by 9.58 to 38.67 degrees; a translation alone left gross rotation (Dice 0.014, HD95 8.21 mm).
- Baseline. After the start the ten test pairs score a median Dice of 0.614 (0.498 to 0.745) and an HD95 of 1.61 mm (0.97 to 3.28); the carotid set 0.348 and 10.14 mm.
- Images. Each keeps its native grid: the MRA, 3D time of flight mostly at 3 T, about 0.3 mm in plane and 0.6 mm slices; the CTA, in Hounsfield units, 0.37 to 0.59 mm and 0.625 to 0.75 mm. The two are separate examinations, how far apart the paper does not say.
- Cohort. Patients of the University Hospital Zurich's Stroke Center in 2018 and 2019 with a suspected stroke, median age 74, without a large aneurysm in the circle or a rare variant the 13 classes cannot describe.
- Overlap. TopBrain labels more vessels on 25 of these patients: the five here (008 in test; 001, 002, 003 and 014 in development) stay, and TopBrain holds the 20 others, which a larger draw here must avoid.
- Outside numbers. None: the challenge ranks segmentation, not registration.
| Accuracy | Smoothness | Cost | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Change | HD95 mm | SDlogJ | Folds % | Time s | RAM GB | GPU GB | Images |
| 11–2 | ConvexAdam (MIND)GPUapp | 0.730 ±0.028 | +0.118 | 1.74 ±0.49 | 0.131 | 2.04 | 38 | 5.1 | 13.1 | |
| 21–2 | NiftyRegCPUapp | 0.746 ±0.028 | +0.135 | 2.06 ±0.84 | 0.426 | 3.24 | 203 | 4.4 | ||
| No registration | 0.612 ±0.050 | 1.73 ±0.46 | 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 patients drawn with seed 2026 among the 105 eligible, then 10 different development patients. The 20 patients TopBrain holds are left out of the pool.
- Truth
- contours by experts ica by experts the dataset's voxel labels of the circle of Willis on each image, without the internal carotid arteries: basilar artery, posterior, middle and anterior cerebral arteries, posterior and anterior communicating arteries, and a third A2 where there is one (11 classes)
- Source
- zenodo.org/records/15692630
topcow24.grand-challenge.org
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_001, topcow_002, topcow_003, topcow_139, topcow_051, topcow_009, topcow_014, topcow_163, topcow_031, topcow_140