Research good subsampling scheme

[halirutan]

We need an excellent subsampling scheme that we can use to validate the Shepp-Logan phantom data.

• IDEA: Is it possible to find the most appropriate subsampling scheme by using a fully sampled dataset and iterating over different subsampling using Loraks for the reconstruction? Then, we compare reconstruction results and validate different subsampling by quality.

[halirutan]

We had this idea that it could use a fully-sampled dataset, send it through the SVD and calculate the low-rank loss. After the backwards pass, we will have gradients at the input tensor. Now we think this might be somehow an indicator about what voxels can be left out. This could be a data-driven method to find good sub-sampling masks.

Since this will likely depend on the particular dataset, it might be a good idea to run this for several datasets and see if a general pattern emerges. Additionally, the question will be if such a subsampling scheme can even be implemented on the scanner and would save time.