fidoptM4Raw, one 0.3 T T2 repetition to the mean of the others
An axial T2-weighted brain MR of a healthy adult at 0.3 T: one acquisition as source and the mean of the same session's other repetitions of it as target (M4Raw).
Read before comparing (10)
- Task. Denoising: the target is the source's own acquisition repeated, averaged. The provider's demo averages all repetitions, the input among them; here the source is left out of the mean so that its noise is not in the target. The source is the repetition the demo takes as input, the n // 2-th of n.
- Target noise. The target keeps the noise of its mean: a fifth of a repetition's variance on a test subject, half on a development one, so a perfect denoiser does not score an MAE of 0, least of all on development.
- Alignment. The repetitions are one protocol run again in one session, nothing registered. Head motion between them is small on most subjects (the source's MAE 3.1 to 3.7 on 23 of 25), not on 2022101102, whose repetitions drift up to a pixel (0.94 mm) from the source by phase correlation (MAE 4.5), nor quite on 2022101302 (4.0). There it blurs the target and counts as error for every method, least for one that smooths.
- Images. 18 axial slices of 256 x 256 pixels, 0.9375 mm in plane, 5 mm thick and 6 mm apart, from a 0.3 T XinGaoYi scanner at Shenzhen Technology University (fast spin echo, TR 5500 ms, TE 128 ms). The values are the provider's root sum of squares in arbitrary units, as stored. The files give no orientation: the direction is the identity, and left and right cannot be told.
- Mask. The head on the target: above the triangle threshold, just over the noise floor, its largest connected part, closed by 3 pixels and filled slice by slice. It is made from the target and given to every method.
- Range. truth.dynamic_range is 91.3, the median of the 99.9th percentiles inside the mask of the 25 test targets (83.4 to 98.7): one range for the dataset's PSNR and SSIM.
- Baseline. Inside the mask on the 25 test cases a constant at the target's median scores an MAE of 10.1 (9.3 to 11.5), a PSNR of 16.5 dB and an SSIM of 0.19; the source as given 3.4 (3.1 to 4.5), 26.4 dB and 0.90, better than the constant on all 25: it is the line a denoiser must beat.
- Subjects. A session is a subject: the 208 sessions of the record (train, validation, test and motion zips) carry 208 different patient_id values, a field the same session's files share across zips (its GRE files).
- Other contrasts. The record also holds T1 and FLAIR repetitions of the same subjects; a subject sits in one dataset only, so they are not cases.
- Outside numbers. The record was made for denoising and reconstruction and publishes its test targets: a model trained on it may have seen these subjects. The provider asks that the test images not be used for tuning.
| Accuracy | Cost | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| # | Method | MAE | Change | PSNR dB | SSIM | Time s | RAM GB | GPU GB | Images |
| 1 | ANTs DenoiseImageCPUapp | 2.5 ±0.1 | −0.9 | 28.77 ±0.30 | 0.937 ±0.004 | 10 | 0.8 | ||
| 2 | DIPY non-local meansCPUapp | 2.8 ±0.1 | −0.7 | 27.80 ±0.29 | 0.931 ±0.004 | 8 | 1.0 | ||
| 3 | Gaussian blurCPUapp | 2.9 ±0.1 | −0.5 | 27.06 ±0.27 | 0.916 ±0.004 | 5 | 0.7 | ||
| Source as given | 3.4 ±0.1 | 26.41 ±0.35 | 0.899 ±0.008 | ||||||
The order averages each method's ranks on MAE, PSNR dB, SSIM, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 25 cases. Two neighbours are tied when a paired Wilcoxon test on the shared cases does not find the upper one better on MAE (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.
With the settings of the challenge
Run with the settings that the challenge's organisers or the method's authors chose for this task. Shown beside the table and not ranked with it: the results above use default settings.
| Accuracy | Cost | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| # | Method | MAE | Change | PSNR dB | SSIM | Time s | RAM GB | GPU GB | Images |
| Gaussian blur · sigma 0.55 mmCPUappchallenge settings | 2.5 ±0.1 | −0.9 | 28.99 ±0.35 | 0.940 ±0.005 | 5 | 0.7 | |||
MAE: mean absolute error of a synthetic image against its target, inside the body, in the image's units. PSNR: the same by signal-to-noise ratio, higher is better. SSIM: the same by structural similarity, 1 is perfect.
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
- The provider's 25 test subjects, taken whole, are the test cases, and 16 development subjects are drawn with seed 2026 among its 158 training and validation subjects. One case a subject; the 25 subjects the provider moved out for motion are left out.
- Truth
- target image as released the mean of the session's other T2 repetitions (five on a test subject, two on a development one), each the provider's root sum of squares of its four coils' images, on the source's grid by acquisition
- Source
- zenodo.org/records/8056074
doi.org/10.5281/zenodo.8056074
doi.org/10.1038/s41597-023-02181-4
github.com/mylyu/M4Raw - Licence
- CC BY 4.0 (the Zenodo record's licence field)
16 cases for tuning, never tested
2022091501, 2022062402, 2022070502, 2022061214, 2022090104, 2022070507, 2022090305, 2022090901, 2022061408, 2022092001, 2022091411, 2022061209, 2022092502, 2022091508, 2022070501, 2022061404