Evaluating coregisteration of stacks

[Nas-5]

Hi all,

I have sets (different subjects) of MR stacks aquired
with 2 different sequences one is an EPI and the other
is an anatomical scan. For analysis I prefer to avoid
co-registration for most of the cases and do
registration only for those cases that would "really"
need to be co-registered. I would like to know how
should I decide if registration is necessary? Other
than just visually inspecting the stacks I would like
to know if there is any quantitative method for this
matter?

I would also like to ask another question: say you
have coregisterd two images how will you evaluate "the
goodness" of registration?

What tools ImageJ provid for this purpose?

Best,
Nas

__________________________________________________
Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around
http://mail.yahoo.com 

[Michael Doube-2]

Hi Nas,

You can test the goodness of registration by assessing the repeatability
of some outcome measure - say, the position and dimensions of the
transformed image, or whatever it is that you're interested in - some
result that you can only get in the registered space.  I found the sync
windows plugin useful for dealing with registered images.

HTH

Mike

Nas wrote:

> Hi all,
>
> I have sets (different subjects) of MR stacks aquired
> with 2 different sequences one is an EPI and the other
> is an anatomical scan. For analysis I prefer to avoid
> co-registration for most of the cases and do
> registration only for those cases that would "really"
> need to be co-registered. I would like to know how
> should I decide if registration is necessary? Other
> than just visually inspecting the stacks I would like
> to know if there is any quantitative method for this
> matter?
>
> I would also like to ask another question: say you
> have coregisterd two images how will you evaluate "the
> goodness" of registration?
>
> What tools ImageJ provid for this purpose?
>
> Best,
> Nas
>
> __________________________________________________
> Do You Yahoo!?
> Tired of spam?  Yahoo! Mail has the best spam protection around
> http://mail.yahoo.com 
>  

--
Michael Doube  BPhil BVSc MRCVS
PhD Student
Dental Institute
Queen Mary, University of London
New Rd
London  E1 1BB
United Kingdom

Phone [phone] ext 2681

[Michael Doube-2]

I just remembered the other thing: if you do repeat registrations and
save the results, then subtract one transformed image from the other
transform repeats, you will get black images if registration is
perfectly repeatable, or embossed-looking images if they're not. The
amount of 'embossing' is an expression of how repeatable your
registration is. You could do some maths on the embossed / error images
to quantify the amount of difference I guess.

Mike

Nas wrote:

> Hi all,
>
> I have sets (different subjects) of MR stacks aquired
> with 2 different sequences one is an EPI and the other
> is an anatomical scan. For analysis I prefer to avoid
> co-registration for most of the cases and do
> registration only for those cases that would "really"
> need to be co-registered. I would like to know how
> should I decide if registration is necessary? Other
> than just visually inspecting the stacks I would like
> to know if there is any quantitative method for this
> matter?
>
> I would also like to ask another question: say you
> have coregisterd two images how will you evaluate "the
> goodness" of registration?
>
> What tools ImageJ provid for this purpose?
>
> Best,
> Nas
>
> __________________________________________________
> Do You Yahoo!?
> Tired of spam?  Yahoo! Mail has the best spam protection around
> http://mail.yahoo.com 
>  

[Gabriel Landini]

On Thursday 03 August 2006 16:25, Michael Doube wrote:
> You can test the goodness of registration by assessing the repeatability
> of some outcome measure - say, the position and dimensions of the
> transformed image, or whatever it is that you're interested in - some
> result that you can only get in the registered space.

I am not sure that repeatability is directrelated to gly oodness of fit.
Suppose that the registration is consistently bad. By your argument the goodness of fit is good despite that the registration is bad.

I think that in the case of registration one has to come up with some measure of distance between the images being registered. Finding out how the registration algorithm works may give you a clue (since the algorithm has to search for an optimal solution).

Cheers,

Gabriel

[William O'Connell]

Nas,

Do the image stacks have images with the same dimensions? If so, you might try overlaying the two stacks as color images with the first stack in, say, Red, and the 2nd stack in, say, Blue. Mis-registrations will be obvious. I think you will need a 3rd stack of zero value images for the green channel. You can convert the 3 stacks to an RGB stack with "Image/Color/RGB Merge". Or you can use your original stacks with the ColorComparison plugin. Either way, the stacks must first be converted to 8 bit grayscale.

best, William O'Connell

 -------------- Original message ----------------------
From: Nas <[hidden email]>

> Hi all,
>
> I have sets (different subjects) of MR stacks aquired
> with 2 different sequences one is an EPI and the other
> is an anatomical scan. For analysis I prefer to avoid
> co-registration for most of the cases and do
> registration only for those cases that would "really"
> need to be co-registered. I would like to know how
> should I decide if registration is necessary? Other
> than just visually inspecting the stacks I would like
> to know if there is any quantitative method for this
> matter?
>
> I would also like to ask another question: say you
> have coregisterd two images how will you evaluate "the
> goodness" of registration?
>
> What tools ImageJ provid for this purpose?
>
> Best,
> Nas
>
> __________________________________________________
> Do You Yahoo!?
> Tired of spam?  Yahoo! Mail has the best spam protection around
> http://mail.yahoo.com 

[Nas-5]

Hi Gabriel,
   
  Thank you for your reply. Actually I'm trying to chose between few algorithms that I have based on their outcome. Therefore , as these algorithms either use different similarity functions (and as a result different optimization cost functions and criteria) or even if they chose the same similarity functions they vary in other modules in implementation. So I would need to compare them based on only the outcome and decide which algorithm and setting gives me the best results. As you have also mentioned I might need to come up with some measure of distance between the original image and the registered images from various algorithms... something like abolute difference maybe..
   
  Best,
  Nasim
  Gabriel Landini <[hidden email]> wrote:
  On Thursday 03 August 2006 16:25, Michael Doube wrote:
> You can test the goodness of registration by assessing the repeatability
> of some outcome measure - say, the position and dimensions of the
> transformed image, or whatever it is that you're interested in - some
> result that you can only get in the registered space.

I am not sure that repeatability is directrelated to gly oodness of fit.
Suppose that the registration is consistently bad. By your argument the goodness of fit is good despite that the registration is bad.

I think that in the case of registration one has to come up with some measure of distance between the images being registered. Finding out how the registration algorithm works may give you a clue (since the algorithm has to search for an optimal solution).

Cheers,

Gabriel


 
---------------------------------
How low will we go? Check out Yahoo! Messenger’s low  PC-to-Phone call rates.

[Nas-5]

Hi William,
   
  Thank you very much for your replies. Yes, images have same dimensions. I was not familiar with this feature of ImageJ. It really makes assessment of registrations very easy! It is very helpful.Thank you for this.
   
  Best,
  Nasim

William O'Connell <[hidden email]> wrote:
  Nas,

Do the image stacks have images with the same dimensions? If so, you might try overlaying the two stacks as color images with the first stack in, say, Red, and the 2nd stack in, say, Blue. Mis-registrations will be obvious. I think you will need a 3rd stack of zero value images for the green channel. You can convert the 3 stacks to an RGB stack with "Image/Color/RGB Merge". Or you can use your original stacks with the ColorComparison plugin. Either way, the stacks must first be converted to 8 bit grayscale.

best, William O'Connell

-------------- Original message ----------------------
From: Nas

> Hi all,
>
> I have sets (different subjects) of MR stacks aquired
> with 2 different sequences one is an EPI and the other
> is an anatomical scan. For analysis I prefer to avoid
> co-registration for most of the cases and do
> registration only for those cases that would "really"
> need to be co-registered. I would like to know how
> should I decide if registration is necessary? Other
> than just visually inspecting the stacks I would like
> to know if there is any quantitative method for this
> matter?
>
> I would also like to ask another question: say you
> have coregisterd two images how will you evaluate "the
> goodness" of registration?
>
> What tools ImageJ provid for this purpose?
>
> Best,
> Nas
>
> __________________________________________________
> Do You Yahoo!?
> Tired of spam? Yahoo! Mail has the best spam protection around
> http://mail.yahoo.com 

 
---------------------------------
Do you Yahoo!?
 Next-gen email? Have it all with the  all-new Yahoo! Mail Beta.

[Robert Dougherty]

Nas,

I just posted a new version of the plugin Convolve 3D on my web site.  It
computes correlations and, especially, correlation coefficients between
stacks.  It might do just what you need.  It only addresses translations,
though.  Rotations and scale changes are outside its scope.

Bob

Robert P. Dougherty, Ph.D.
President, OptiNav, Inc.
Phone [phone]
Fax [phone]
www.optinav.com

> -------------- Original message ----------------------
> From: Nas
> > Hi all,
> >
> > I have sets (different subjects) of MR stacks aquired
> > with 2 different sequences one is an EPI and the other
> > is an anatomical scan. For analysis I prefer to avoid
> > co-registration for most of the cases and do
> > registration only for those cases that would "really"
> > need to be co-registered. I would like to know how
> > should I decide if registration is necessary? Other
> > than just visually inspecting the stacks I would like
> > to know if there is any quantitative method for this
> > matter?
> >
> > I would also like to ask another question: say you
> > have coregisterd two images how will you evaluate "the
> > goodness" of registration?
> >
> > What tools ImageJ provid for this purpose?
> >
> > Best,
> > Nas