fidopt Interactive view
Stroke lesion, DWI and ADC, segmentation

SOOP acute stroke lesions

Brain diffusion MR of a patient admitted with an acute ischaemic stroke: the b=1000 trace image and the ADC map as one two-channel image, the acute lesion traced by hand (SOOP, OpenNeuro ds004889).

30 test casesbrainimage MR
Look at the images
Read before comparing (9)
  • Cohort. Acute stroke encounters of 2019 and 2020 at Prisma Health-Upstate, Greenville, South Carolina. Version 1.1.2 holds 1,715 people, 1,451 with a mask of the acute lesion, imaged within 30 days of admission, most within 48 hours.
  • Images. The trace image and the ADC map share one acquisition and grid, so the mask fits both; one NIfTI vector image holds them (channels dwi, adc). Clinical axial slices 6 mm apart (5 mm on three patients), 0.49 to 1.2 mm in plane, on seven scanner models, defaced with the skull kept. The T1 and FLAIR, on other grids, are not given, nor is a mask.
  • Values. prepare.py applies the ADC's header scale, so both channels are 32-bit floats. The ADC's unit differs between scanners (its median in the brain 772 to 13,480 on the 20 cases built first): no common unit can be assumed. On 21 pool patients the file named ADC is not marked as an ADC map by its scanner; none is among the 20 built first.
  • Truth. One tracing a patient, by one of three trained raters, slice by slice in MRIcroGL on the ADC map. No second tracing, radiologist review or agreement is published, so the grade 'expert' means one trained rater. Masks stored as 2 or 3 are written as 1; chronic masks are not built.
  • Lesions. 0.25 to 147 ml on the 30 test cases (median 4.3 ml), ten under 1.5 ml. On 6 mm slices a small lesion is a few voxels: the first ten test masks moved by one voxel in plane keep a Dice of 0.69 to 0.92, moved by one slice 0.0 to 0.64. The surface Dice at 1 mm is strict here.
  • Checks. On 19 of the 20 cases built first the lesion is bright on the trace image and dark on the ADC map, as an acute infarct is. On development case sub-423 the mask lies on tissue with an old lesion's signal, kept as the dataset gives it. Left and right rest on the headers, not checked against anatomy.
  • Outside numbers. None: no segmentation score is published on these masks, and TotalSegmentator makes no stroke lesion.
  • Overlap. ISLES2022, the other acute stroke set, comes from Munich and Bern. The Aphasia Recovery Cohort (ChronicStrokeT2) is the same group's chronic cohort in Columbia: no shared identifier, other scanners, so no shared image, but a person could be in both with other scans.
  • ARC check. The 20 patients built first were compared with ARC's 228 subjects on dates, demographics, lesion side and volume: of 21 possible pairs with ChronicStrokeT2's subjects, 18 lesions lie apart and the other three are different heads. Of the 40 later candidates, checked against ChronicStrokeT2's 42 subjects, twelve stayed possible and are left out of the pool (MAYBE_IN_ARC). Whether other patients are in ARC cannot be told from the public files.

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

The first test case of SOOP acute stroke lesions
The first test caseClick to look at every case.
Test cases
30 test patients drawn with seed 2026 by scanner model, in proportion, among the 1,407 outside development, and 30 development patients, the last of the same seeded order. The pool, 1,437 patients with an acute lesion mask and an ADC map on the trace image's grid, leaves out sub-235 and sub-1233 (ADC map on another grid), the five with a chronic mask only and twelve who may be in ChronicStrokeT2.
Truth
contours by experts the dataset's mask of the acute lesion (desc-lesionAcute), traced by hand, slice by slice, in MRIcroGL on the ADC map by one of three trained raters of the authors' laboratory and stored on the trace image's grid; one tracing a patient, no review or agreement published
Licence
CC0
30 cases for tuning, never tested

sub-1537, sub-1024, sub-499, sub-1206, sub-829, sub-1417, sub-1655, sub-1204, sub-423, sub-1156, sub-1392, sub-963, sub-726, sub-1336, sub-799, sub-267, sub-925, sub-1223, sub-1498, sub-171, sub-1039, sub-508, sub-1025, sub-244, sub-181, sub-1391, sub-840, sub-1487, sub-50, sub-1351