[Insight-users] Intensity-based registration with low contrast images.

[Rupert Brooks]

Hi Vikren,

I see Luis just beat me to it!   I just ran an ITK based registration
on these images, using the NormalizedCorrelation Metric, and obtained
a good registration.  As Luis did, I also had to crop off the black
border.

I got these parameters to map using a translation transform from the
fixed image to the moving image. 12.3474 -4.05349
This appears reasonable to me, but i didnt spend too long looking at
it.  I was also able to get a similar result with
MattesMutualInformation, using 50 bins.  (i didnt try other numbers of
bins)

Because the images are low contrast i would suggest that you be
careful not to lose data along the way.  So, avoid compressed formats
like JPG, and avoid converting to 8-bit if the original images are in
16bit. (Probably you just converted to 8-bit compressed for posting
anyway.)

Because mutual information uses an evenly spaced set of histogram
bins, these images might be more difficult to register with MI.  It
worked when i tried it, but i am almost certain its just registering
the fairly bright head, to the dark background, and not actually using
any of the detail inside.  In any case, if these images are
representative of your problem, i think MI is unnecessary.

Cheers,
Rupert B.

On Wed, Jul 23, 2008 at 11:08 AM, Luis Ibanez <luis.ibanez at kitware.com> wrote:
>
> Hi Vikren,
>
> Thanks for the clear description of your problem and for posting the
> images in an accessible place.
>
> The problem in your case, is not the low contrast of the images
> (although that certainly doesn't help...). The main problem is
> related to how these images have been composed.
>
> --
>
> I just ran on them the example ImageRegistration9
> (without modifications), an obtained a reasonable registration.
>
> Looking closer at your images, the problem is that you seem to have
> copy/pasted an image on top of a larger canvas.
>
> That is, your images are of size 512x512, with a background of value 0.
> Insight this region, there is a rectangle from (25,23) to (488,488) at
> a intensity level close to 20 at the top, and with a gradient that
> grows upto 45 at the bottom of the image.
>
> Inside that region you have the projection of the head.
>
>
> The rectangular region on the background is preventing the registration
> from being driven by the data from pixels inside the head.
>
>
> You can get rid of this problem in several ways.
>
>
>  a) you can crop the images in order to contain only the
>     region (24,24) to (488,488) before you feed the images
>     into the registration.
>
> or
>
>
>  b) inside the registration you can set the FixedImageRegion
>     to the region (24,24) to (488,488).
>
>
>
>  Regards,
>
>
>      Luis
>
> ----------------------
> Vikren Sarkar wrote:
>>
>> Greetings,
>>  I am trying to see if it is possible to perform intensity-based
>> registration on some very low contrast images that I have. I tried posting
>> this question a few days ago but realize there were some problems with the
>> website I was using. Hopefully photobucket will be better. The following two
>> images were acquired using an electronic portal imaging device. The only
>> difference is that one image is translated compared to the other.
>>  File 1 : http://i336.photobucket.com/albums/n333/poohvee/fixed.jpg
>>  File 2 : http://i336.photobucket.com/albums/n333/poohvee/moving.jpg
>>  Being new to ITK, I first started with several of the example
>> registration codes but none seem to work with the low contrast level. Luis
>> also suggested I try to run the images through edge detection. I tried that
>> approach as well as thresholding before registering but could not get any
>> results with any of the example registration codes. Finally, I did try the
>> example code using mutual information (example 2 in image registration) and
>> it did not work either.
>>  Does anyone have any other suggestion I can use to artificially increase
>> the contrast before registering or can someone suggest a metric that would
>> be better suited for the task? The final aim would be to try to use this
>> with 3D datasets but I would like to start with 2D. Thank you.
>>
>> Vik
>>
>>
>> ------------------------------------------------------------------------
>>
>> _______________________________________________
>> Insight-users mailing list
>> Insight-users at itk.org
>> http://www.itk.org/mailman/listinfo/insight-users
>
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>



-- 
--------------------------------------------------------------
Rupert Brooks
McGill Centre for Intelligent Machines (www.cim.mcgill.ca)
Ph.D Student, Electrical and Computer Engineering
http://www.cyberus.ca/~rbrooks