[Insight-users] Number of spacial samples Mattes MI metric

[Luis Ibanez]

Hi Jef,

The percent of samples is relative to the number of
pixels inside the "fixed image region" that you provide
to the registration as the section of the fixed image
that will taken into account for computing the metric.

It is not unusual to end up using a large number of samples,
certanly 100K to 1M samples can be common. Note that the
more samples you take, the more computing time you will need,
there is therefore a trade-off between Metric smoothness,
convergence, and computation time.

In your parameter fine-tunning process, you may want to start
with a large percent, until you get your registration to work,....

then reduce the number of samples... until it stops working  :-)

That will tell you what's the critical threshold for number
of samples in your particular image registration problem.

You will finally choose a number of samples at a safe value
above that critical threshold.

You may not want to play with the number of bins at the same
time that you play with the number of samples. If you add more
bins, then you will need more samples in order to get a decent
estimation of the Joint Entropy. In general you can get away
with relatively small histograms (30 to 50 bins per image).


Keep going... registration is a lot of fun !!


   Regards,


      Luis



-----------------
jef vdmb wrote:

> Hi,
> 
> I am working on a 2D- 3D registration problem using the Mattes 
> implementation of Mutual Information. One of the parameters to be set is 
> the number of spacial samples. The Softwareguide says:
> 
> The number of spatial samples to be used depends on the content of the 
> image. If the images are smooth and do not contain much detail, then 
> using approximately 1 percent of the pixels will do. On the other hand, 
> if the images are detailed, it may be necessary to use a much higher 
> proportion, such as 20 percent.
> 
> Does this mean 1 or 20 percent of the total amount of pixels in BOTH 
> images (fixed as well as moving)? This would mean a very high number of 
> samples (~ 30000 to ~800000).
> 
> Do I have to choose a number of histogram bins that is equally augmented?
> 
> Thanks you for any advice you can give me,
> 
> Jef
> 
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
> 
>