fidoptON-Harmony (OpenNeuro ds004712), Siemens T1w to Philips T1w
T1-weighted brain MR of one healthy adult on a Siemens Prisma in Oxford (source) and on a Philips Ingenia in Nottingham on another visit (target), the target rigidly registered onto the source by the corpus (ON-Harmony travelling heads).
10 test casesbrainsource Siemens T1w, target Philips T1w
Read before comparing (9)
- Subjects. 20 healthy adults, 13 men and 7 women aged 19 to 50, each scanned on six 3T scanners of three vendors at five UK sites (phase A 2018 to 2021, phase B 2023 to 2024). Two scanners all 20 visited are used: the Siemens Prisma with a 32-channel coil at FMRIB, Oxford, and the Philips Ingenia at the SPMIC, Nottingham.
- Source. The Siemens 3D MPRAGE (TR 2 s, TI 880 ms, flip 8 degrees), 1 mm, as delivered, defaced by the UK Biobank procedure (face set to 0); the scanner's values without unit.
- Target. The Philips 3D T1 TFE (TR 7 ms, flip 8 degrees), 1 mm, of another visit to another site, rigidly registered onto the source by the corpus (mutual information, transforms/) and resampled once by a cubic B-spline. Its values are the scanner's display values: dcm2niix wrote the phase A subjects in Philips' floating-point values, 40 to 130 times larger, and prepare.py scales them back with their sidecar's slopes.
- Registration. Graded automatic. SynthSeg 1.0 on the source and on the registered target agrees at a mean Dice of 0.93 (0.90 to 0.94 on the 18 cases) over its structures; a second Siemens visit of sub-13192 to the same scanner, registered the same way, agrees at 0.96. What is left combines the two scanners' contrast and geometric distortion with any registration error; it counts against every method alike.
- Mask. The head where both images hold data: the registered target above a twentieth of its 99th percentile, its largest part, holes filled slice by slice, where neither scan is 0; 2.5 to 3.7 litres, the neck included.
- Range. truth.dynamic_range is 2010, the median over the test cases of the target's 99.9th percentile inside the mask (1979 to 2110).
- Baseline. Inside the mask on the 10 test cases a constant at the target's median scores an MAE of 364 (324 to 383), a PSNR of 13.1 dB and an SSIM of 0.16; the source itself 262 (224 to 318), 14.7 dB and 0.71; the source mapped by the mean quantile curve of the other cases, learnt leave-one-out, 247 (200 to 322), 15.5 dB and 0.72. On the 8 development cases that mapping beats the constant on all 8 (MAE 243 against 339).
- Overlap. No descriptor names ON-Harmony. SpineGeneric's test pairs hold sub-oxfordFmrib04 (man, 42) and sub-oxfordFmrib09 (man, 32), scanned at FMRIB in 2019; no key links the two releases, and ON-Harmony's sub-13305 (man, 42) and sub-14229 (man, 35 at first scan, phase A) could be them, so both are held out.
- Public. The resource was released for harmonisation research: a model trained on ON-Harmony has seen these subjects.
No method has a result on every case of this dataset yet.
Case by case
Click a score to look at that result: the target image, and over it the method's synthetic image. Underlined scores have their image; for the others you see the case without the result.
About the data
- Test cases
- 10 test subjects drawn with seed 2026, then the other 8 as development, one visit pair each. The pool is 18 of the 20 participants: sub-13305 and sub-14229 are held out (draw.held_out) because they may be two of SpineGeneric's test subjects.
- Truth
- target image by a program the same subject's T1w on the Philips Ingenia, another day, after the corpus rigid registration and B-spline resampling onto the Siemens T1w, in the scanner's units
- Source
- openneuro.org/datasets/ds004712/versions/2.0.1
doi.org/10.18112/openneuro.ds004712.v2.0.1
doi.org/10.1162/imag_a_00042 - Licence
- CC0 1.0
8 cases for tuning, never tested
sub-10975, sub-03997, sub-14230, sub-14482, sub-15320, sub-03286, sub-16981, sub-16975