[Insight-users] Re: Using RequestedRegion

[Luis Ibanez]

Hi Rakesh,

This sounds strange.

However, notice that you must differentiate between the time
that it takes to compute *one* value of the image metric and
the time that it takes to complete the *full* registration.

Setting the FixedImageRegion should *reduce* the time that
it takes to compute *one* iteration of the optimizer, but
depending on your image, it may *increase* the total number
of optimizer iterations needed to achieve convergence.


Please post to the list the number of iterations that it took
you to complete the registration in both cases. I would guess
that it took you a lot more iterations in the case where the
FixedImageRegion was set.


If that is the case, what we could conclude is that the region
that you selected as FixedImageRegion does not contain enough
shape specific information as to drive the optimizer, and that
you probably should include other nearby regions in order to
give better definition to the cost function.


For example, if you where registering the lungs, and you select
a FixedImageRegion that is fully inside the parenchyma, you
are probably in need of including a small section of the ribs,
which will provide a clear basin in the topography of the cost
function.


Please take a look again to the number of iterations used in both
cases and let us know what you find. It will also help us to know
about the Metric that you are using, the Optimizer, and the actual
size of the FixedImageRegion and the BufferedRegion.



    Thanks



       Luis



---------------------------------------
Mullick, Rakesh (GE, Research) wrote:
> Hi Luis,
> 
>  
> 
> I have been experimenting with the use of SetFixedImageRegion for 
> Registration and observing that when I set a region the registration 
> takes longer than when I use the BufferedRegion. This is in Release 
> mode. Will the overhead be so much as to slow down the registration. I 
> am using about 1/10 of the overall volume to drive registration and 
> observing a 50% increase in clock time to register the same data.
> 
>  
> 
> Any thoughts or past threads are welcome.
> 
> Many thanks.
> 
>  
> 
> regards,
> rakesh
> ______________________________________________________________________
> *Rakesh Mullick*, Ph.D., Senior Scientist    *EMail*: 
> Rakesh.Mullick at ge.com <mailto:Rakesh.Mullick at ge.com>
> 
>  
>