fidoptMS lesions, Ljubljana
Brain MR of a multiple sclerosis patient, T1, contrast-enhanced T1, T2 and FLAIR co-registered as one four-channel image, the white matter lesions outlined by a consensus of three raters (Laboratory of Imaging Technologies, University of Ljubljana).
10 test casesbrainimage MRresearch only
Read before comparing (9)
- Cohort. 30 patients, 23 women and 7 men aged 25 to 64: 24 relapsing-remitting, 2 clinically isolated syndromes, 2 secondary and 1 primary progressive, one not stated (the archives' patient_info.csv). One examination a patient.
- Images. T1, contrast-enhanced T1, T2 and FLAIR, defaced, bias corrected and registered to the FLAIR by the provider, given as one four-channel image (t1, t1c, t2, flair) on the FLAIR's grid: 192 x 512 x 512 voxels of 0.8 x 0.47 x 0.47 mm, 16-bit values as 32-bit floats, without unit. The brain mask is the provider's and is given to methods. The raw sequences in their own space are not read.
- Truth. Three raters segmented each patient alone, then reached a consensus together; the paper ran the whole protocol twice, six months apart, and found the consensus far more repeatable than any single rater. Graded expert.
- Lesions. 0.33 to 31 mL on the 10 test cases (median 9.8 mL), 18 to 373 connected pieces; on FLAIR they read 1.25 to 1.6 times the median of the rest of the brain. All lie inside the brain mask but 7% of patient29's.
- Header. The consensus files' origin is half a voxel off the FLAIR's along each axis while their voxels are the FLAIR's (the lesions read brightest on the FLAIR unmoved): prepare.py writes them under the FLAIR's header.
- Room. The test masks moved one voxel keep a Dice of 0.59 to 0.91: many lesions are a few voxels across.
- Start. Labelling the whole brain mask scores a mean Dice of 0.017 on the test cases; Otsu's bright FLAIR voxels in the brain 0.019; the brightest 2% of the FLAIR in the brain 0.20 (surface Dice 0.22). TotalSegmentator has no lesion task.
- Overlap. MSLesionsMR (MSLesSeg, Catania) and MSLesionsBaghdad are other cohorts, no descriptor names this database, and no built image has the grid of any of the 20 cases (12,304 stores compared by size and spacing, a check that finds AbdomenDIRQA's case20 in MSD_Pancreas); HaNSeg comes from another Ljubljana institute and holds head-and-neck CT. The Shifts 2.0 challenge redistributes these patients (MSLUB), so a method trained on Shifts has seen them.
- Outside numbers. None: the paper reports rater variability, not method scores on these ten.
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
- 10 test patients drawn with seed 2026 among the 20 outside development, and 10 development patients, the last of the same seeded order: the database's 30 patients, one examination each. The provider publishes no split.
- Truth
- contours by experts the database's consensus of the white matter lesions: three expert raters each segmented the lesions with the laboratory's semi-automated contouring tools on the co-registered images, then revised them together in joint sessions until they agreed (Lesjak et al. 2018)
- Source
- lit.fe.uni-lj.si/en/research/resources/3D-MR-MS
lit.fe.uni-lj.si/data/research/resources/3D-MR-MS/3D-MR-MS.pdf
doi.org/10.1007/s12021-017-9348-7 - Licence
- CC BY-NC-ND (the deposit's page); CC BY (its LICENSE.txt and description PDF)
10 cases for tuning, never tested
patient12, patient07, patient11, patient05, patient25, patient04, patient26, patient02, patient09, patient28