Pediatric T1 MRI with automatic regional labels

Pediatric T1 MRI with and without previous ADHD diagnosis

T1 MR of the brains of two different children from OpenNeuro ds002424, with FastSurfer anatomical labels on both; the test children have a parent-reported previous ADHD diagnosis.

10 test pairsbraininter-patientfixed MR, moving MRaligned once by the benchmark

Look at the images

Read before comparing (7)

Diagnosis. ADHD_diagnosis is the participant table's parent- or guardian-reported previous diagnosis at ses-T1, not a clinical adjudication. Controls, with no previous diagnosis reported, appear only in development: a population difference to weigh when tuning.
Images. Only the defaced ses-T1 T1-weighted scan is used; the study has no longitudinal image pairs. ds003500 is not mixed in, since its subjects' independence from these has not been established.
Truth. FastSurfer v2.5.4's automatic subcortical and DKT labels, resampled nearest-neighbour onto each native scan grid. They are not manual pediatric annotations: the dataset is kept apart and says nothing of clinical segmentation accuracy.
Headers. 14 source files (sub-01, 02, 04, 05, 06, 08, 09, 10, 11, 12, 14, 25, 30, 32) carry an sform that differs from their qform by up to 35 mm at the corners. FastSurfer reads the sform, ITK the qform, so labels.py gives FastSurfer a copy with the sform equal to the qform, and the labels lie on the image as ITK reads it.
Alignment. Each pair starts from an intensity-based elastix rigid transform, no label used, written into the moving image's and its labels' headers. Both images keep their native voxels and grids.
Baseline. From that start the ten test pairs have a mean Dice of 0.4117 (0.3137 to 0.4693) over the 95 labels and an HD95 of 5.8295 mm (4.3369 to 8.3677). ConvexAdam (MIND) results measured on an earlier label-centroid affine build do not describe these cases.
Source. The control draw reads the provider's participants.tsv (SHA-256 0ebb4dff7b0bfc926eab5e6c83a6d3e1509efa756a14829c7be53c3e6c404bc8); the T1 files are pinned in source_checksums.json.

		Accuracy			Smoothness		Cost
#	Method	Dice	Change	HD95 mm	SDlogJ	Folds %	Time s	RAM GB	GPU GB	Images
1	ConvexAdam (MIND)GPUapp	0.609 ±0.008	+0.197	3.69 ±0.18	0.060	0.04	12	2.6	2.9
2	anatomix + ConvexAdamGPUapp	0.593 ±0.012	+0.182	3.93 ±0.37	0.055	0.02	29	3.0	12.9
33–4	DemonsCPUapp	0.485 ±0.048	+0.073	5.30 ±0.86	0.054	0.00	9	1.5
43–4	multiGradICONGPUapp	0.588 ±0.125	+0.176	8.54 ±11.37	0.228	9.95	97	3.2	12.5
	No registration	0.412 ±0.035		5.83 ±0.83	0.000	0.00

The order averages each method's ranks on Dice, HD95 mm, SDlogJ, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 10 pairs. Two neighbours are tied when a paired Wilcoxon test on the shared pairs does not find the upper one better on Dice (5%): the places a method could hold are then shown under its own. The best value of each column is in colour. Click a row to see its images, a column title to sort by it. Time and memory measured on Laptop.

Dice: how much the organ contours overlap, 1 is perfect. HD95: distance between the contours' surfaces. SDlogJ: how irregular the deformation is, 0 is perfectly smooth. Folds: share of the image where the deformation folds over itself.

Pair by pair

Measure

Dice	1	2	3	4	5	6	7	8	9	10
ConvexAdam (MIND)	0.612	0.606	0.616	0.613	0.614	0.621	0.603	0.602	0.615	0.582
anatomix + ConvexAdam	0.584	0.595	0.603	0.609	0.610	0.609	0.566	0.575	0.611	0.574
Demons	0.431	0.403	0.432	0.562	0.576	0.470	0.403	0.528	0.490	0.554
multiGradICON	0.092	0.642	0.646	0.656	0.648	0.641	0.647	0.632	0.655	0.619
No registration	0.314	0.463	0.380	0.404	0.469	0.460	0.363	0.416	0.435	0.412

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

The first test case of Pediatric T1 MRI with and without previous ADHD diagnosis

The first test caseClick to look at every case.

Test pairs: The 35 ses-T1 subjects with ADHD_diagnosis == 1 were ordered with seed 2026:ADHDChildrenMR:subjects into 17 disjoint pairs, one subject unused; 10 pairs were drawn for test with seed 2026:ADHDChildrenMR, the other 7 kept for development. Three more development pairs take the first six of the 44 controls (ADHD_diagnosis == 0) in SHA-256 order with salt 2026:ADHDChildrenMR:controls. The 40 subjects are distinct.
Truth: contours by a program FastSurfer v2.5.4 FastSurferCNN aparc.DKTatlas+aseg.deep, generated locally from each native defaced T1w image; subcortical structures and DKT cortical parcellation. Automatic labels are resampled to the input grid by nearest neighbor. Pediatric domain validity has not been established by manual annotation.
Source: openneuro.org/datasets/ds002424/versions/1.2.0
doi.org/10.18112/openneuro.ds002424.v1.2.0
doi.org/10.1016/j.dib.2019.105801
Licence: CC0 1.0, as declared by the source dataset_description.json

10 pairs for tuning, never tested

sub-79__sub-12, sub-73__sub-66, sub-25__sub-04, sub-13__sub-29, sub-33__sub-74, sub-71__sub-67, sub-57__sub-06, sub-27__sub-41, sub-65__sub-52, sub-68__sub-23