fidoptISLES 2016 and 2017, final infarct from acute MR
Acute diffusion and perfusion maps of a patient with an ischaemic stroke of the middle cerebral artery, with the final infarct a neuroradiologist drew on the 90-day follow-up (ISLES 2017 training set, Bern and Freiburg).
10 test casesbrainimage MR
Read before comparing (6)
- Pool. ISLES 2017's 43 training patients, which hold ISLES 2016's 30 (their ADC maps are identical): adults with an acute stroke of the middle cerebral artery (M1 or M2 occlusion), treated by thrombolysis or thrombectomy at Bern University Hospital or UMC Freiburg between 2005 and 2015, on a 1.5 T Avanto or 3 T Trio.
- Images. The provider's ADC and perfusion maps (Tmax, rCBF, rCBV, MTT, TTP, from Olea Sphere), the perfusion maps rigidly registered to the ADC, skull-stripped, 5 to 6.5 mm slices; the raw 4D perfusion series is left out. On the nine cases of 256 x 256 x 24 voxels the ADC reads about ten times lower than elsewhere (median 84 to 121 in the brain, against 808 to 1010): as delivered.
- Truth. Outcome, not what the images show: the final infarct drawn by one neuroradiologist on the 90-day T2 follow-up, which methods never see. On ISLES 2016's test cases two raters agree at a Dice of 0.58 (0.20 standard deviation); ISLES 2017's best entries reached about 0.3 on its test set. Ranked apart from lesion segmentation.
- Start. The whole brain as lesion scores a Dice of 0.067 at the median of the test cases, an Otsu mask of the ADC 0.039, ADC under 620 0.179, and Tmax over 6 s, the usual hypoperfusion rule, 0.242 (0.002 to 0.567). TotalSegmentator makes no stroke lesion.
- Channels. adc, tmax, rcbf, rcbv, mtt, ttp: a method that names adc alone is given the first channel; none of the corpus's methods reads the perfusion maps yet.
- Overlap. No other dataset of the corpus holds ISLES 2015 to 2018. Within the 43 no two ADC maps on one grid correlate above 0.44, and none shares a grid with ISLES 2022's 52 Bern ADC maps (ISLES2022, ISLES2022_Syn), a check that finds all 30 of ISLES 2016's training ADC maps in this set at r 1.000. A patient scanned again for ISLES 2022 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 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
- The record's ISLES 2017 test set has no public truth. 10 test cases drawn with seed 2026 among the 43 training patients, and 10 development cases, the last of the same seeded order; one case a patient.
- Truth
- contours by experts the final infarct, drawn on every axial slice of the 90-day follow-up T2 by a board-certified neuroradiologist in 3D Slicer (its largest extent, haemorrhagic transformation included, earlier infarcts excluded), on the acute ADC's grid
- Licence
- CC BY 4.0 (the Zenodo record); each file carries SICAS Medical Image Repository's notice: the Open Database License 1.0 for the database, the Database Contents License 1.0 for its contents
10 cases for tuning, never tested
training_28, training_15, training_39, training_47, training_38, training_37, training_16, training_42, training_44, training_1