fidoptIXI, TransMorph atlas-to-subject release
A skull-stripped, affinely aligned T1-weighted IXI brain (moving) and the shared CycleMorph atlas (fixed), from TransMorph's release, with automatic FreeSurfer aseg labels on both.
Read before comparing (7)
- Atlas. One atlas image and its label map are the fixed side of every case, public by design, as in TransMorph's own benchmark, which this dataset reproduces. A method may have been tuned on that atlas.
- Truth. Thirty FreeSurfer structures of the TransMorph comparison, automatic and not corrected by hand. Other label IDs stay in the files but are not scored.
- Geometry. The pickle release holds voxel arrays without scan affine, direction or voxel size. They keep the provider's order with identity geometry, one px a voxel: distances are in voxels, not millimetres. The intensities come normalised and the subjects affinely aligned; nothing more is applied.
- Baseline. With no registration the first ten test cases score a mean Dice of 0.34 over the thirty structures (0.28 to 0.40 a case).
- Identity. The release renames the subjects subject_N. Its demographic table, matched to IXI's own (IXI.xls) by sex, height, weight, dates of birth and scan and age, names 556 of its 562 rows with one IXI subject each.
- Overlap. By the demographic table the subjects are disjoint from IXI_T2PD's 60. Of the 58 release subjects left out as IXI_T2PD's, 19 were found by registering each IXI_T2PD T1 onto all 576 release images (0.88 to 0.97, against 0.82 for any other subject), 16 of them confirmed by the table and none contradicted; 39 more by the table alone.
- Exposure. These test subjects are a published benchmark, and many registration models may have trained on the provider's training split. Record a model's exposure when reading its results.
No method has a result on every pair of this dataset yet.
Pair by pair
Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.
About the data
- Test pairs
- The provider's test split taken whole, less the subjects this dataset cannot hold: 102 test subjects, one case a subject. 30 development subjects, in seeded order, from the 344 left in the training split. Left out of both pools: the 58 release subjects that are IXI_T2PD subjects, and the 21 whose IXI identity the demographic table does not give, which may include IXI_T2PD's five not found.
- Truth
- contours by a program Thirty anatomical structures used in the TransMorph comparison, retaining original FreeSurfer IDs. Other source IDs remain in the files but are not scored. The labels are automatic, not manually corrected.
- Source
- github.com/junyuchen245/TransMorph_Transformer_for_Medical_Image_Registration/blob/main/IXI/TransMorph_on_IXI.md
github.com/junyuchen245/Preprocessed_IXI_Dataset
brain-development.org/ixi-dataset - Licence
- CC BY-SA 3.0, explicitly included in the provider archive
30 pairs for tuning, never tested
subject_170, subject_450, subject_193, subject_494, subject_524, subject_52, subject_332, subject_242, subject_555, subject_122, subject_32, subject_178, subject_29, subject_382, subject_443, subject_463, subject_539, subject_149, subject_554, subject_438, subject_519, subject_534, subject_353, subject_361, subject_529, subject_249, subject_340, subject_278, subject_309, subject_118