Hippocampal subfields on 7T T2w MR (Beijing)
7 T T2-weighted MR around both hippocampi of a healthy young adult, with seven subfields drawn by hand on every coronal slice of each side (Chu et al., Scientific Data 2025, figshare+ 26075713).
10 test casesbrainimage MR
Read before comparing (10)
- Image. The provider's processed 7T T2w the subfields were drawn on, not the raw scan: a 3D SPACE at 0.4 x 0.4 x 1.0 mm, smoothed, N4-corrected and upsampled to 0.7 mm slices, in 64-bit floats without unit. The signal fades towards the bottom of the temporal lobes on some subjects.
- Crop. Each case is cut out of the whole head, voxels and scanner frame unchanged: the box of both hippocampi's labels widened by 16 mm on every side. The box comes from the truth: a method is handed the hippocampi's neighbourhood, never told where in it they are. The labels fill 1.3 to 1.9% of a case; no mask is given.
- Truth. The protocol of Wisse et al. for 7T MRI, drawn voxel by voxel in ITK-SNAP on coronal slices; the last slices, where the subfields can no longer be told apart, are one label, the tail. Two experts made the final labels of the 20 subjects together, by consensus.
- Ceiling. Each expert segmented ten subjects twice, a month apart: an expert agrees with herself at a Dice of 0.81 for the entorhinal cortex, 0.80 for CA2 and 0.875 for CA1 and the subiculum, and slightly less with the other expert. A method cannot be shown to be right beyond that.
- Labels. prepare.py merges the provider's two maps, 1 to 7 left and 8 to 14 right. On the 20 cases they never overlap, each holds its seven labels, and each left map lies on the patient's left by the header, whose own left and right nothing in the record checks.
- Size. Test medians run from 1.4 ml for CA1 to 0.06 ml for CA2, about 540 voxels. On CA2 and CA3 the Dice turns on a few voxels, and the surface Dice at the default 1 mm, 2.5 voxels in plane, forgives much of a structure 1 to 2 mm thick.
- Stray marks. 16 of the 40 provider maps hold marks away from the hippocampus. Those outside the box go with the rest of the head; those inside, 1 to 9 voxels each on 12 cases, stay and can lengthen a case's HD95 for their label.
- Names. Neither structures.json nor TotalSegmentator's open tasks hold these structures, so the names are the dataset's own and no network check was run. The entorhinal cortex covers this protocol's slices only and the tail the last few slices, not L2R_HippocampusMR's hippocampus_tail: the names are kept apart on purpose.
- Outside numbers. None: the paper reports no segmentation score on these labels. The cases are public with their labels, so a network trained on this record has seen the test cases.
- Overlap. The record also holds each subject's 3T and 7T T1w, T2w and resting-state fMRI, listed in the screening for registration and synthesis: a dataset built from them must keep this partition (test: sub-01, 02, 03, 05, 07, 10, 12, 17, 18 and 20). No other dataset of the corpus holds these subjects; MR3T7T_Syn is other people.
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 caseClick to look at every case.
- Test cases
- 10 test subjects drawn with seed 2026 among the record's 20, healthy right-handed volunteers of 18 to 25 years; the other 10 are the development cases, so every subject is used and none is left.
- Truth
- contours by experts the dataset's manual subfields of the hippocampal formation, seven on each side (subiculum, CA1, CA2, CA3, CA4 with the dentate gyrus, entorhinal cortex, hippocampal tail), traced voxel by voxel in ITK-SNAP on every coronal slice of the 7T T2w after the protocol of Wisse et al. (2012), by two experts together, who settled their differences by discussion
- Source
- doi.org/10.25452/figshare.plus.26075713.v1
api.figshare.com/v2/articles/26075713
ndownloader.figshare.com/files/50079657
ndownloader.figshare.com/files/50052801
doi.org/10.1038/s41597-025-04586-9 - Licence
- CC BY 4.0
10 cases for tuning, never tested
sub-16, sub-08, sub-14, sub-15, sub-19, sub-13, sub-11, sub-04, sub-09, sub-06