[MINC-users] N3

[elo@neurorx.com]

Hi Joel,

I'm not able to address all of your issues, but here are some aspects of
using nu_correct that you might find helpful:

A low number of iterations means that nu_correct is not applying a very
strong correction, and this may result in poor quality images. To rectify
this you might want to decrease the stopping condition for your scans.  I
find that -stop 2.e-5 works well, the default is 1.e-3.  When the stopping
condition is tightened however, the number of iterations may increase
greatly - so there is a tradeoff between correction quality and processing
time.

We in fact find that running nu_correct several times (with the same
parameters) yields the best results in terms of image uniformity.  This
may be because different gaussian modes are targeted and corrected.  For
our MRI images, applying repeated nu_corrects (about 6 times), produces
the best results.

Good luck,


Ernest Lo
NeuroRX Research
Montreal, Canada


> Date: Mon, 10 Oct 2005 17:16:00 -0500
> From: joel-bruss at uiowa.edu
> Subject: [MINC-users] N3
> To: minc-users at bic.mni.mcgill.ca
> Message-ID: <1128982560.434ae820b26f4 at webmail2.its.uiowa.edu>
> Content-Type: text/plain; charset=ISO-8859-1
> 
> Currently I have data from a GE 1.5T, and Siemens 1.5 and 3T (Avanto and Trio)
> scanners.  I am looking for a way to bias-correct these images so that they can
> be used with Gray/White segmentation and other volumetrics.  My data is in
> 16-bit Analyze format and I have a minc converter (ana2mnc).  When I run
> nu_correct, with default parameters (-V1.0) as well as specifying a mask of the
> brain image, I get very awful results.  In my 124-slice volume, there are
> regions where the image is totally washed out (5-10 slice runs) and others where
> the brain looks nice but the bias-correction happened outside of the skull. 
> When specifying for temp files to be saved as well as a tempdir, after
> converting the bias field, it looks like a perfectly uniform smoothing kernel
> (ovular/egg shape) which is definitely not what the bias appears as (more polar,
> specifically in the frontal lobe).  I've tried downloading the test data but am
> unsure of how to even work with it.  I've read Sled et al's IEEE paper but don't
> find it that useful in configuring nu_correct.  Also, I set the max iteration
> for 150 but it runs through 3 iterations (in about 2 minutes) and says that it
> is done.  Is this really possible?
> 
> Is there a help file, other than the README, or tutorial out there?  Any
> suggestions on how to configure nu_correct?  Should I use a different Minc
> converter?  Is it OK to use 16-bit images or do they need to be 8-bit?  Any help
> would be useful right now.
> 
> Thanks for your time
> 
> -Joel
> 
> 
> 
> 
> 
> ------------------------------
> 
> Message: 3
> Date: Tue, 11 Oct 2005 09:20:25 -0400
> From: Jason Lerch <jason at bic.mni.mcgill.ca>
> Subject: Re: [MINC-users] N3
> To: MINC users mailing list <minc-users at bic.mni.mcgill.ca>
> Message-ID: <434BBC19.90208 at bic.mni.mcgill.ca>
> Content-Type: text/plain; charset=ISO-8859-1; format=flowed
> 
> Make sure that the intensity histogram of your volumes is adequate for 
> N3 - which means that it should not lie between 0 and 1. If the actual 
> values don't matter to you, try multiplying the volume with a large 
> enough constant, then run N3.
> 
> Hope this helps,
> 
> Jason
> 
> joel-bruss at uiowa.edu wrote:
> 
> >Currently I have data from a GE 1.5T, and Siemens 1.5 and 3T (Avanto and Trio)
> >scanners.  I am looking for a way to bias-correct these images so that they can
> >be used with Gray/White segmentation and other volumetrics.  My data is in
> >16-bit Analyze format and I have a minc converter (ana2mnc).  When I run
> >nu_correct, with default parameters (-V1.0) as well as specifying a mask of the
> >brain image, I get very awful results.  In my 124-slice volume, there are
> >regions where the image is totally washed out (5-10 slice runs) and others where
> >the brain looks nice but the bias-correction happened outside of the skull. 
> >When specifying for temp files to be saved as well as a tempdir, after
> >converting the bias field, it looks like a perfectly uniform smoothing kernel
> >(ovular/egg shape) which is definitely not what the bias appears as (more polar,
> >specifically in the frontal lobe).  I've tried downloading the test data but am
> >unsure of how to even work with it.  I've read Sled et al's IEEE paper but don't
> >find it that useful in configuring nu_correct.  Also, I set the max iteration
> >for 150 but it runs through 3 iterations (in about 2 minutes) and says that it
> >is done.  Is this really possible?
> >
> >Is there a help file, other than the README, or tutorial out there?  Any
> >suggestions on how to configure nu_correct?  Should I use a different Minc
> >converter?  Is it OK to use 16-bit images or do they need to be 8-bit?  Any help
> >would be useful right now.
> >
> >Thanks for your time
> >
> >-Joel
> >
> >
> >
> >_______________________________________________
> >MINC-users at bic.mni.mcgill.ca
> >http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-users
> >  
> >
> 
> 
> 
> ------------------------------
> 
> _______________________________________________
> MINC-users mailing list
> MINC-users at bic.mni.mcgill.ca
> http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-users
> 
> 
> End of MINC-users Digest, Vol 3, Issue 8
> ****************************************
>