Hi Oori,

To be more specific, I expect methods for smoothing to go through
some changes as data is acquired at higher and higher resolution.
Your question about the noise is a good one to which I expect a
fairly affermative answer, but I do not know enough to say for
certain.

This subject seems to have gotten much more complicated at the
higher resolutions and with multi-channel coils. Multi-channel
coils can have large shading artifacts, would confuse the issue
on when to blur.

So I should no longer be making strong statements about when to
do it. As usual, the answer should start with "it depends"...

---

But one reason that blurring before scaling has been important
is because the gray matter tends to be close to the edge of the
brain. If a brain mask is not applied, then signals outside the
brain get blurred in. This is not a big deal if blurring happens
before scaling, but it is if scaling happens first.

Outside the brain (and even at edge voxels) there can be large
signal changes in voxels with lower means. Lower means implies
that the scaling operation amplifies these signals (relative to
voxels with higher means). If blurring happens after scaling,
then these amplified noisy signals will get undue weight, and
since they tend to have a large % change, the noisy signal will
tend to dominate.

---

However, with high resolution and multi-channel coil artifacts,
it seems that methods of blurring will need to go through some
corresponding changes. For instance, maybe blurring time series
data is not a good idea to begin with (assuming high-resolution
data).

- rick