fidopt Interactive view
Lesions and tumours, segmentation

ISLES 2022

Skull-stripped diffusion-weighted brain MR (b=1000) of a patient with an acute or sub-acute ischaemic stroke, the lesions traced and approved by neuroradiologists (ISLES 2022 training set, Munich and Bern).

24 test casesbrainimage MRresearch only
Look at the images
Read before comparing (9)
  • Pool. The challenge's 250 training cases, one MR session of one patient each, from clinical routine: 198 at the University Hospital of the Technical University of Munich, 52 at the University Hospital of Bern. The 150 test cases, which add Hamburg-Eppendorf, are hidden. For its test set the paper calls Munich's and Hamburg's scans acute to early sub-acute, after revascularisation therapy, and Bern's hyper-acute to acute, before it.
  • Images. A case is the DWI at b=1000, on whose grid the masks are drawn. The ADC map and FLAIR, which the annotators read and the challenge's entries take, are not given: a score here comes from the DWI alone. Every image is skull-stripped, zero outside the brain; no mask is given. Values are as delivered, without a unit.
  • Grids. The provider resliced Munich's DWI to 2 x 2 x 2 mm on 193 of its 198 cases; test case strokecase0137 is 1.35 x 1.35 x 4.4 mm. Bern's comes as acquired, 1.15 or 1.80 mm in plane with 4.8 mm slices.
  • Truth. A 3D U-Net trained on earlier Munich tracings pre-segmented the scans; trained medical students corrected it or traced from scratch, a neuroradiology resident revised every mask, and one of three neuroradiologists with more than ten years in stroke imaging approved it, a doubtful small lesion going to a majority of three readers. The reviewers did not know which masks started from the network.
  • Protocol. Every lesion of at least two 2-mm voxels with a definite low ADC signal is traced on the DWI, artefacts told apart on the FLAIR; a very early lesion bright on the DWI with nothing on the FLAIR is traced too.
  • Agreement. Ten cases traced again by two other neuroradiologists agree with the dataset's masks at a mean Dice of 0.90 and 0.86, and with each other at 0.83 (0.57 to 0.98). The paper adds that tracings often drew an infarct's fading border too tight, and that showers of small embolic lesions, frequent here, are the hardest to trace completely.
  • Lesions. 0.04 to 482 ml over the 247 training cases with a lesion (median 7.2 ml), up to 126 lesions; test case strokecase0117 holds 0.05 ml. A case's Dice is that of its largest lesions, so missing small ones costs little; the challenge's lesion-wise scores are not computed. The first ten test masks moved one voxel along one axis keep a Dice of 0.54 to 0.89 (mean 0.73).
  • Checks. Every mask built holds only 0 and 1 on its DWI's grid. On the first 20 built every lesion voxel lies inside the brain, on DWI signal 1.3 to 2.5 times the median of the rest of the brain.
  • Outside numbers. None: the challenge ranked its entries on the hidden test set, from the three images.

No method has a result on every case of this dataset yet.

Case by case

Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.

About the data

Test cases
24 test cases drawn with seed 2026 among the 225 outside development of the 247 with a lesion, the two centres in proportion: 19 from Munich, 5 from Bern. 22 development cases, the last of the same seeded order. Three training cases with an empty mask (strokecase0150, strokecase0151, strokecase0170) stay out: a Dice needs a lesion.
Truth
contours by experts the dataset's mask of the stroke lesions on the DWI, pre-segmented by a network, corrected by trained medical students and a neuroradiology resident, then reviewed and approved by one of three neuroradiologists with more than ten years in stroke imaging
Licence
CC BY 4.0
22 cases for tuning, never tested

strokecase0091, strokecase0003, strokecase0166, strokecase0031, strokecase0247, strokecase0237, strokecase0217, strokecase0176, strokecase0049, strokecase0235, strokecase0112, strokecase0164, strokecase0205, strokecase0122, strokecase0071, strokecase0089, strokecase0219, strokecase0036, strokecase0050, strokecase0130, strokecase0075, strokecase0222