fidoptISLES 2022, diffusion to FLAIR
The diffusion-weighted brain MR (b=1000) and ADC map of a patient with an acute or sub-acute ischaemic stroke, and the FLAIR of the same session, registered onto them by the corpus, as the target (ISLES 2022 training set, Munich and Bern).
Read before comparing (6)
- Pool. ISLES 2022'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.
- Source. The DWI at b=1000 and its ADC map as delivered, skull-stripped, as channels dwi and adc: Munich's resliced by the provider to 2 mm, Bern's as acquired (1.15 or 1.8 mm in plane, 4.8 mm slices).
- Target. The FLAIR of the same session, which the provider does not register: the corpus registers it rigidly onto the DWI (mutual information) and resamples it once onto the DWI's grid, so it is smoother than the FLAIR as acquired. The two skull-stripped brains overlap at a Dice of 0.996 at the median of the 38 cases (0.975 to 0.999), against 0.979 in the scanner frame; two development cases keep the scanner frame, where the registration did not improve that overlap: strokecase0207, whose registration drifted 13 mm, and strokecase0087, about 1 mm off. A re-registration of the built test cases finds at most 1.1 mm left (median 0.3). Graded automatic and ranked apart: what the registration leaves counts as error for every method.
- Intensities. No unit, and the FLAIR's scale changes with the scanner: its median in the mask runs from 149 to 1213 over the test cases. truth.dynamic_range is 1970, the median of the test targets' 99.9th percentiles inside the mask (271 to 2310), one range for the PSNR and SSIM, loose on the low-scale scanners.
- Baseline. Inside the mask of the test cases a constant at each target's median scores an MAE of 222 on average (31 to 365), a PSNR of 18.0 dB (12.2 to 33.1) and an SSIM of 0.401 (0.247 to 0.810).
- Overlap. ISLES2024's Munich patients include ISLES 2022 ones: each drawn Munich case's DWI was matched, after an affine registration, against the follow-up DWI of the 33 Munich patients of ISLES2024 and ISLES24_NCCTCTA_Syn. strokecase0076 is ISLES2024's stroke0036 (Pearson r 0.991) and is held out; the other 990 pairs reach 0.673 at most (the same check scores a moved and resampled ISLES 2022 DWI at 0.995). None of Bern's 52 ADC maps shares a grid with ISLESOutcomeMR's (ISLES 2017, Bern and Freiburg, 2005 to 2015), a check that finds ISLES 2016's 30 training maps in ISLES 2017 at r 1.000; a patient scanned in both periods cannot be ruled out without dates.
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
- The challenge's 150 test cases are hidden. 20 test cases drawn with seed 2026 among the 203 training cases left (16 Munich, 4 Bern, the centres in proportion) and 18 development cases; held out are ISLES2022's 46 test and development cases, whose DWI its methods are given, and strokecase0076, a patient of ISLES2024.
- Truth
- target image by a program the same session's FLAIR after the corpus's rigid registration and cubic B-spline resampling onto the DWI's grid, in the scanner's units
- Licence
- CC BY 4.0
18 cases for tuning, never tested
strokecase0203, strokecase0114, strokecase0207, strokecase0190, strokecase0221, strokecase0041, strokecase0087, strokecase0248, strokecase0103, strokecase0051, strokecase0171, strokecase0062, strokecase0195, strokecase0111, strokecase0012, strokecase0182, strokecase0074, strokecase0216