Hi,
Thanks for your help - it's clarified a few things. However, I'm still
having a bit of trouble. Maybe you can tell me where I'm going wrong.
I've done what you suggested, trying to create a weight in the input
space.
I drew a mask around there areas of signal loss (orbital frontal and
inferior temporal areas) and saved this out as a binary image file (in
analyze format), with the mask valued at 0 and the rest (including
non-brain) as 1. I then included this image as the input weight using
the flirt GUI when registering the epi to the subjects T1, but it
appears to have made little difference to the actual registration. 
I then tried again, this time drawing a mask around the ventricles and
weighting them as 2, the signal dropout as 0, and the rest as 1, but
again, there appears to be little difference in the registration
results.
Am I missing something?
Thanks again,
Alex

-----Original Message-----
From: Mark Jenkinson [mailto:[log in to unmask]] 
Sent: Friday, October 01, 2004 10:27 PM
To: [log in to unmask]
Subject: Re: [FSL] Registration woes

Hi Alex,

I agree with Tim.  Especially in that it is very important to
distinguish
masking from weighting - they are quite different operations.

I would say that there isn't much advantage in smoothing the cost
weights
though, so you could easily ignore that step.

Also, I never recommend drawing weights based on a non-weighted
registration
(using mricro's linked view or anything else) as it is only likely to be
much good
if the registration is good, which is not going to be true when you
really need
the weights the most.  Either draw it directly on the functions where
you can
see the signal loss (and use the -inweight) or draw a general one in the
appropriate
areas on a structural or standard-space image (and use the -refweight
option).
The latter saves a lot of time if you are doing this many times on
different
scans.

It is also worth noting that the larger the ignored area in the weights
(voxels
with zero or near-zero values) then the less robust the initial search
phase of
the registration becomes.  So for weights with relative large areas
being set
near zero it is best to use "no search" either via the GUI pull-down
options
or by using -nosearch in the command line.  This will then require that
the
initial positions of the images are no more than say 20-30 degrees out
of
alignment, but as long as the slicing (sagittal, axial, oblique, etc) is
the same
or can be made the same (or nearly the same) using avwswapdim then
there should be no problems.

Hope this is also useful.

All the best,
    Mark



Tim Behrens wrote:

>Hi Alex,
>
>I'll have a go at answering these - Mark can correct me when I get it
>wrong..
>
>
>
>>1 - Would I have to mask just the initial highres, or both the initial
>>highres and task epi? If the latter, would I just mask the middle
volume,
>>as this is what is used in the registration? If so, my 4D file
contains
>>160 volumes. Would the middle correspond to the 79th, or the 80th?
>>
>>
>
>the cost function mask is a separate 3D image in reference (-refweight)
or
>input (-inweight) space ( you can actually apply both weightings if you
>like, but they have to be in the right spaces). Note that Cost function
>weighting is different from masking the original images, as masking
>introduces artificial edges which will drive the registration - usually
>not a good idea.
>
>
>
>
>>2 - The CBU page recommends co-registering the epi to the structural
and
>>using the "yoke" function in MRIcro to identify points where there is
>>signal dropout in the epi relative to the structural. These are the
>>portions that should be masked out. If I register the unmasked epi to
the
>>structural, wouldn't this yield a (relatively) poor correspondance
between
>>the images, so that using this as the basis for identifying signal
dropout
>>would not be the best way to go about it? Are there any alternative
ways
>>for idetifying which areas to mask out?
>>
>>
>>
>
>If you use -inweight, your cost function weight mask will be directly
in
>input space so you can just downweight the areas with susceptibility
>artefact, and upweight the boundaries you trust ( lateral walls of the
>ventricles etc.. ). If you wanted to you could apply a _generic_ MNI
space
>mask to the reference space (-refweight) which would do the same thing
-
>downweight occipital pole etc and upweight the ventricle walls. I guess
>this would be slightly less accurate than doing it individually for
each
>input volume, but I would think you could downweight pretty big areas
>without really detracting from the registration, so I would think this
>should be fine.
>
>
>
>
>>3 - Following from 2, would there be any problems in doing a straight
>>image subtraction to identify the areas to be masked? It seems to me
that
>>this would be plagued by the same problems as in 2, but that it would
be a
>>quicker way to go about it?
>>
>>
>
>I think this is answered by the previous point (??)
>
>
>
>>4 - The CBU page recommends smoothing the mask image (8mm in
accordance
>>with the SPM default). Would I also need to do this with FEAT? Would I
use
>>5mm, as this is the FEAT default (or alternatively, whatever level I
set
>>it to)?
>>
>>
>>
>
>I would think smoothing the mask image would be a good idea. However,
the
>smoothness here is different from the smoothness in Feat - here we are
>smoothing because of uncertainty in the location of the structures of
>interest, not due to GRF theory etc. In this case, I would thing 8mm
>smoothing of the mask image would be fine.
>
>
>
>>5 - As an aside, can someone tell me why the average epi template in
SPM
>>would be a poor choice as an intial highres? Does the fact that it's
not
>>the same brain as your subject introduce more problems than the
advantage
>>of having the added contrast?
>>
>>
>>
>
>Usually going to an initial highres, you would want to be able to
choose a
>low number of degrees of freedom to get a really robust registration.
If
>you use a reference brain which is a different global shape to the
input
>brain, you will need a high number of degrees of freedom even to get
close
>to a good registration, so I would think it would not be particularly
>advantageous to choose an averaged brain as an initial highres.
>
>Hope this is useful (and accurate!!)
>
>
>T
>
>
>
>
>
>
>
>>Sorry if some of these questions seem a bit basic - still trying to
get my
>>head around all of this!
>>Thanks heaps, I really appreciate your ongoing help!
>>Alex
>>
>>
>>