fidoptISLES'24 final infarct, DWI and ADC
Follow-up brain diffusion MR of a patient 2 to 9 days after thrombectomy for an acute ischaemic stroke: the DWI trace image and the ADC map, registered by the provider onto the admission CT, as one two-channel image, with the final infarct masked (ISLES'24 training set, Munich and Zurich).
14 test casesbrainimage MRresearch only
Read before comparing (10)
- Pool. The challenge's 149 public training patients, each treated by thrombectomy for an acute ischaemic stroke with near-complete or complete recanalisation (mTICI 2c or 3): 100 at the University Hospital of the Technical University of Munich, 49 at the University Hospital Zurich.
- Images. The follow-up MR, 2 to 9 days after the admission CT: the DWI trace image and the ADC map, which the provider registered onto the admission NCCT (affine, elastix or NiftyReg), resampled onto its grid and skull-stripped (HD-BET). Grids are the NCCT's: 0.33 to 0.54 mm in plane, 2 mm slices in Munich (0.8 mm on stroke0040), 4 mm in Zurich (2 mm on stroke0137), often tilted with the gantry. The native MR is not released; neither the admission CT nor the clinical tables are given, nor a brain mask.
- Values. As delivered, as 32-bit floats. The ADC's scale differs between scanners: its median in the brain runs from 0.0004 to 103,711 on the 27 cases built. On 8 of them (stroke0040, 0081, 0082, 0085, 0091, 0106, 0109, 0187, all Munich) the region outside the MR's field of view holds non-zero values, not 0.
- Truth. The ISLES'22 ensemble (DeepISLES) segmented the final infarct on the follow-up DWI; an experienced neuroradiologist of Munich or Zurich inspected each mask, and the poor ones were revised by a trained medical student and a neuroradiology resident; how many is not said. On the seven same-scan pairs with ISLES2022 (held out), these masks agree with ISLES'22's expert masks at a Dice of 0.84 to 0.96 (median 0.92), an optimistic figure since DeepISLES learned from ISLES'22.
- Lesions. 1.3 to 151 ml on the 14 test cases (median 12.9 ml), up to 29 separate parts; on stroke0171, not drawn, the mask is one voxel. Every mask lies inside the brain; on the 27 cases built the masked tissue is brighter on the DWI than a shell six voxels thick around it (1.2 to 2.2 times its median), and darker on the ADC on 26 (not stroke0002, 184 ml). The first ten test masks moved by one voxel in plane keep a Dice of 0.91 to 0.97 (mean 0.94).
- Start. The whole brain given as the lesion scores a Dice of 0.001 to 0.19 on the test cases (median 0.014). TotalSegmentator makes no stroke lesion.
- ISLES2022. ISLES'24's Munich patients include ISLES'22 patients. Each of the 100 was compared with ISLES2022's 36 Munich cases (affine registration of the DWIs, Pearson r inside the brains): the follow-up DWI of stroke0008, 0011, 0022, 0028, 0052, 0062 and 0075 is the DWI of ISLES2022's strokecase0050, 0237, 0065, 0122, 0003, 0040 and 0225 (r 0.990 to 0.997), held out here and in ISLES24_NCCTCTA_Syn. The 33 Munich patients of the two sets, checked again with the padding outside the brain removed and more starts, reach at most r 0.668 with the other cases, every pair's brains aligned or told apart by sex and age; a known pair, an ISLES2022 DWI moved and resampled onto a tilted CT grid, scores 0.994. ISLES2022_Syn's 31 Munich cases were compared with these 33 patients in the same way: only its strokecase0076 matches, stroke0036 here (r 0.991, the next 0.673), and ISLES2022_Syn holds it out.
- TopCoW and TopBrain. Zurich's stroke centre also gave TopCoW and TopBrain their patients. No Zurich CT angiography of ISLES'24 shares a pixel spacing or an acquisition time with any of TopCoW's, and the heads of the 16 Zurich patients of this set and ISLES24_NCCTCTA_Syn match none of the corpus's 40 TopCoW and TopBrain angiographies (r at most 0.844, against 0.955 to 0.973 for a patient's own scans; the check is in ISLES24_NCCTCTA_Syn's notes).
- Overlap. ISLES24_NCCTCTA_Syn takes other patients of the same archive, none of these 27. VerSe's spine CTs also come from Munich: no image is shared, and whether a person is in both cannot be told.
- Outside numbers. None: the challenge ranks infarct prediction from the admission CT, not segmentation on the follow-up MR.
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
- 14 test patients drawn with seed 2026 by centre, in proportion (9 Munich, 5 Zurich), among the 140 training patients outside draw.held_out, and 13 development patients, the last of the same seeded order. Held out: stroke0141 and stroke0162, whose masks are empty, and seven Munich patients whose follow-up DWI is the scan of an ISLES2022 case (notes). The challenge's 96 test patients are hidden.
- Truth
- contours by a program, verified by people the dataset's mask of the final infarct on the follow-up DWI, made by the ISLES'22 ensemble (DeepISLES), inspected by an experienced neuroradiologist of Munich or Zurich, the masks of poor quality revised by a trained medical student and a neuroradiology resident, then brought with the DWI onto the admission CT's grid by the provider
- Source
- zenodo.org/records/17652035
arxiv.org/abs/2408.11142
arxiv.org/abs/2408.10966
github.com/ezequieldlrosa/isles24
isles-24.grand-challenge.org
arxiv.org/abs/2403.19425 - Licence
- CC BY-NC-SA 4.0
13 cases for tuning, never tested
stroke0066, stroke0147, stroke0055, stroke0187, stroke0156, stroke0085, stroke0081, stroke0151, stroke0073, stroke0017, stroke0161, stroke0002, stroke0106