[Insight-users] Question on registration of cryogenic RGB images

Fucang Jia jiafucang@hotmail.com
Wed, 11 Dec 2002 22:35:16 +0800


Hi, Luis,

Thank you very much! Your reply let me understand the key of registration.


>From: Luis Ibanez <luis.ibanez@kitware.com>
>To: Fucang Jia <jiafucang@hotmail.com>
>CC: insight-users@public.kitware.com
>Subject: Re: [Insight-users] Question on registration of cryogenic RGB 
>images
>Date: Mon, 09 Dec 2002 08:27:15 -0500
>
>Hi Fucang,
>
>0) It looks like a single slice will have take
>    around 26Mb. (asuming 1 byte/ per RGB component)
>    Is this right ?

Yes, the images are 1byte per RGB component.

>1) It is not a problem to register images
>    with different resolutions. Just make sure
>    that the values of spacing are set correctly
>    (that is, the distance in millimeter between
>    pixel centers). Registration is performed in
>    physical space coordinates.

After looking up the maillists, I see (the translation in mm, not numbers of 
pixels).

>2) Don't worry about your old computer. It fits
>    the initial requirement set for the toolkit.
>    It may be slow, but it should be possible to
>    do it.

Yes, I make a MultiResMIRegistration on a Red component, the result is good 
enough. Though it needs about 900 Mb RAM, but it works, about 15 minutes, 
which is acceptable.

>3) For visuallization you may want to reduce the
>    resolution but probably some experimentation
>    will be needed here to find the right resolution
>    to visualize.
>    Note that you could use the multiresolution
>    framework

Yes

>4) About registering RGB color images. Bill Lorensen
>    made some validation studies and found that using
>    only the red channel is in general sufficient for
>    registering color images.
>
>    We haven't implemented RGB metrics yet. It is in
>    fact a matter of demand. If you find that a single
>    channel registration is not good enough, please
>    let us know so we can move the RGB metrics up in
>    the list of priorities.

Yes, after I look the red, green and blue component of the original image, I 
find them are similar, so red component is enough. The pixels' red , green 
and blue component has a high correlation, which make sure that one 
component is acceptable.

>5) By differences in intensities do you mean that
>    the brightness in the center one slice is different
>    from the brightness on the borders ?
>    or do you mean that the brightness of one slice
>    is different from the brightness of the next slice ?
>
>    If one image has a different brightness scale than
>    the others. The effect varies depending on the Metric.
>
>    a- MutualInformation is insensitive to brightness
>       changes.
>
>    b- MeanSquares will is sensitive to differences
>       in brightness between the two images being
>       registered
>
>    c- NormalizedCorrelation will compensate for
>       *multiplicative* changes in intensity but not
>       for additive changes in intensity
>
>    d- PatternIntensity will be sensitive to the
>       intensity changes. More sensitive to additive
>       changes than multiplicative changes.
>
>    BTW How big are the differences in brightness ?
>    a factor of 1.5 ? 1.2 ?

I have not test the difference in brightness, but I will compute the 
correlation of two images to make it clear. But here is a problem, when I 
use ImageRegistration to register two image, the registered image is all 
black, does ImageRegistration need two unsigned-short images?

>6) You don't need to segment the images before
>    registration.

Yes. I do.

>Please let us know if you have futher questions,
>
>
>Thanks
>
>
>    Luis
>
>
>===========================================
>
>Fucang Jia wrote:
>
>>Hi, everyone,
>>
>>At first I thank Luis's help.
>>
>>Recently I obtained 1399 cryogemic RGB color image, which were produced 
>>with a high-precision scanner.
>>These images have different resolution, which are as follows:
>>
>>1-66         2733 x 3179
>>67-89        2417 x 3010
>>90-544       2240 x 3010
>>545-644      2241 x 3011
>>645-673      2178 x 3011
>>674-682      2105 x 3041
>>683-702      2071 x 3041
>>703-989      2105 x 3041  ( same with 674-682th )
>>990-1034     2148 x 3086
>>1035-1286    2155 x 3110
>>1287         2163 x 3110
>>1288-1290    2294 x 3071
>>1291-1399    2155 x 3110
>>
>>The image difference is mainly translation and rotation. I want to align 
>>the second image to the first image, the third to the second, etc. So this 
>>is a 2D registration problem.
>>
>>The computer is somewhat old, PIII 500 with 640Mb Ram.
>>
>>Here is several problem I concern:
>>
>>1. The resolution is different, should I downsampled the image to the 
>>minimum 2071 x 3010? Or a smaller resolution
>>such as 500 x 750, because finally I want to visualize the three 
>>dimensional structure by Marching Cubes or Dividing Cubes in this 
>>computer.
>>
>>2. The image is RGB color, but most of registration algorithm is useful 
>>for gray scale image, the simplest method is convert RGB image to gray 
>>image, or only use one channel (red, green, blue), which method is more 
>>accurate? I noticed that Dr.Luis once said that he would like to implement 
>>a registration for color image, but I do not know whether this method was 
>>included in ITK already.
>>
>>3. The images were not homogenous in brightness, that is to say, the 
>>brightness of each image have more or least
>>difference, I think this will induce difficulty in registration. And the 
>>target area is very small contrast to the entire image. So should I first 
>>segment the image before registration?
>>
>>I am confused by these huge images, any advice would be very appreciated!
>>
>>Fucang
>>
>>_________________________________________________________________
>>Add photos to your e-mail with MSN 8. Get 2 months FREE*. 
>>http://join.msn.com/?page=features/featuredemail
>>
>>_______________________________________________
>>Insight-users mailing list
>>Insight-users@public.kitware.com
>>http://public.kitware.com/mailman/listinfo/insight-users
>>


_________________________________________________________________
MSN 8 with e-mail virus protection service: 2 months FREE* 
http://join.msn.com/?page=features/virus