dwidenoise exemplar database

[Lestropie]

Mentioned in #3023 but warrants its own Issue.

If we want to assess impacts of changes to denoising procedures, in terms of both performance and efficacy, we need a database of exemplar data against which to run different implementations / settings. Ideally if users of the software have data where they determined the current implementation to be suboptimal, they would contribute to such a database.

OSF provides 5GB for private, 50GB for public. Given we need multiple large 4D series, 5GB won't cut it. Requiring that all such data be public may however hinder contribution.

Main items that need to be included:

• A "conventional DTI" acquisition: ~ 35 directions, b=1000.
• A typical modern multi-shell DWI acquisition.
• A high-quality multi-shell DWI acquisition (eg. HCP-Lifespan CCP protocol).
• A DWI series with very few directions (eg. 12).
• A single-echo fMRI series.
• A multi-echo fMRI series.
• Data where k-space zero-filling was applied by the scanner.
• Data where gradient non-linearity distortion correction was applied by the scanner.
• Any other instances where existing denoising implementations have gone awry.

Where complex data are available, should additionally evaluate processing on magnitude data only.

[tsalo]

Here are a few datasets @mattcieslak and I have come across that might be useful:

1. Complex-valued DWI runs with different resolutions: https://openneuro.org/datasets/ds004666/versions/1.0.5
2. Complex-valued multi-echo fMRI with no-RF volumes: https://openneuro.org/datasets/ds005250/versions/1.1.4
   • 5 echoes, 2x2x2 mm voxels, 3T
3. Complex-valued single-echo fMRI: https://portal.conp.ca/dataset?id=projects/cneuromod
   • 3T
4. Complex-valued single-echo fMRI: https://openneuro.org/datasets/ds003427
   • Limited FoV (visual cortex), 7T

We also have some complex-valued DWI data basically using the HBCD protocol from a pilot that we might be able to share if that would be helpful.