fidopt Interactive view
Brain MR, segmentation

Claustrum at 7 T

A brain-masked 7 T MP2RAGE image of a healthy adult volunteer at Maastricht University, 0.7 or 1.0 mm, with the claustrum on each side drawn by hand (Calarco et al. 2026).

10 test casesbrainimage MR
Look at the images
Read before comparing (7)
  • Pool. Two Maastricht 7 T sets of ten healthy adults, MP2RAGE UNI at 0.7 mm (Haast et al. 2024, 4 women, 28.6 years on average) and 1.0 mm (Kashyap et al. 2018, 5 women, 25.7). The paper's 0.5 mm set (MICA-PNI) is left out: its segmentations lie on an average of three runs that is not published.
  • Same volunteers. The papers give no ages or sexes a volunteer, so metadata cannot tell whether a volunteer is in both sets. Each 1.0 mm brain was registered rigidly to each 0.7 mm brain and correlated inside the brain: 0p7 sub-03 and 1p0 sub-08 (0.920), 0p7 sub-06 and 1p0 sub-09 (0.946), 0p7 sub-09 and 1p0 sub-06 (0.920) match, every other pair stays at or under 0.61 (median 0.53). Positive control: each 0.7 mm brain resampled to 1 mm, turned by 8 degrees and moved 6 mm found itself first, at 0.956 to 0.975. The three 1.0 mm images are held out; the 0.7 mm ones are kept.
  • Images. The repository's images byte for byte: background removed (MPRAGEised), bias-corrected, skull-stripped and defaced, values without unit, 0 or a few negative values outside the brain. No mask is given.
  • Truth. One rater drew each claustrum by hand in native space with a one-voxel brush, keeping partial-volume voxels judged mostly claustral, and a second rater checked and refined every segmentation. The second rater drew the left claustrum of one volunteer a set again: Dice 0.93 on average. Graded expert.
  • Labels. 1 the left claustrum, 2 the right, each on its side of the brain on every case. The test claustra hold 0.93 to 2.0 mL a side; a sheet one or two voxels thick, so one voxel of shift costs much of the Dice and the surface Dice at 1 mm says more.
  • Start. Over the 10 test cases a whole-image mask scores a mean Dice over the two labels of 0.0002 at the median, an Otsu mask 0.0025.
  • Overlap. No other dataset of the corpus names these sets, Maastricht or the claustrum; the corpus's other 7 T sets (HippoSubfields7T, MR3T7T_Syn) and UltraCortex (9.4 T) are other sites' volunteers.

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 volunteers drawn with seed 2026 among the 17, each set in proportion (6 at 0.7 mm, 4 at 1.0 mm), and the other 7 as development cases. One case a volunteer: three volunteers of the 1.0 mm set are held out as the same people as three of the 0.7 mm set.
Truth
contours by experts the left and right claustrum drawn by hand in each volunteer's native space in ITK-SNAP by one rater, with a one-voxel brush in the three planes, then checked and refined by a second rater (Calarco et al. 2026)
Licence
CC BY 4.0
7 cases for tuning, never tested

0p7_sub-09, 1p0_sub-03, 0p7_sub-01, 1p0_sub-05, 0p7_sub-03, 1p0_sub-04, 0p7_sub-04