To see the effect of NRR above in a bit more detail, the next experiment was performed. It would also be interesting to see the pace of improvement as a function of scale (resolution). So, a multi-scale approach was again used: 3 levels, 10 iterations in level 2, then 2 iterations in level 1, and 1 iteration at the finest level. There would be 13 frames of NRR to look at and the experiment would be relatively long.
It is only the last (NRR) stages that are of interest here: many beneficial
changes are made in the first 5 iterations of NRR, the next 5 iterations
(still at coarse resolution) make small changes to only one small
region of the brain. At a lower resolution (level 1), there is change
yet it is not very visible to the eye if it is a good change. The
same goes for the last stage at level 0. It seems to almost make things
As above, but 6 iterations at level 1, and 6 at level 0. This experiment was run to reassure that conclusion were not made at haste. At level 1 there are little 'vibrations' that don't affect the non-reference image much. At level 0 there is a little more change, but again there is only a mere movement, up to a pixel or two in extent.
The conclusion is that going to finer resolutions will consume a great deal of time and produce little gain. There is significant change, but not necessarily good change.