mtnormalise: Stabilise field outside brain

[Lestropie]

Have seen reasonably high-quality data that following mtnormalise have replicated the bright-FoV-corner artifact prevalent with the original mtbin. Currently the field could be made quite extreme to appropriately correct intensities near the edge of the brain, that then become problematic once the field is extrapolated to the rest of the image.

It should be possible to influence the field estimation in such a way to encourage or enforce it to be 0.0 in the logarithmic domain beyond the brain. I can foresee this being possible in three different ways:

1. Regularisation toward 0.0 in all voxels.
   Pessimistic about the efficacy of this in isolation; making it strong enough to mitigate such artifacts may already result in over-regularisation.
2. More tailored: For each voxel in the FoV not in the user-specified mask, compute the geodesic distance to the mask. Then apply a regularisation whose strength scales with that distance. This should help prevent regularisation from under-estimating field fluctuations near the edge of the brain, since voxels both within and near the mask won't be regularised, but mitigate bright FoV corners.
   Slightly unusual in that the behaviour of the optimisation might depend on the extent of the image FoV outside the
   brain.
   It's possible that with such regularisation a greater polynomial order may be required.
3. Use a basis other than polynomial that results in a boundary value problem constrained to 0.0 on the surface of some surrounding cuboid.
   Eg. DCT?