help with registration please

[William Simpson]

Thanks for offering to help!
Here are the two original pics plus an animated .gif that is poorly registered.

http://picpaste.com/AALtcc1-I1sU4DxK.png
http://picpaste.com/AALtcc2-eUb1IyDH.png
http://picpaste.com/AA_time_1_2_lt_cc_motion_bunwarp-Ij3FZv1v.gif

Thanks for any help!
Bill

[Stephan Saalfeld]

Hi Bill-Jodie,

to me it seems that you try to register two projections through a thick
3d volume in the projection plane.  That's not really possible without
further knowledge of the 3d-structure or otherwise separable components
of the signal.  If my impression is wrong, please correct me.

Best,
Stephan




On Tue, 2011-03-15 at 17:27 +0000, William Simpson wrote:
> Thanks for offering to help!
> Here are the two original pics plus an animated .gif that is poorly registered.
>
> http://picpaste.com/AALtcc1-I1sU4DxK.png
> http://picpaste.com/AALtcc2-eUb1IyDH.png
> http://picpaste.com/AA_time_1_2_lt_cc_motion_bunwarp-Ij3FZv1v.gif
>
> Thanks for any help!
> Bill

[William Simpson]

Thanks very much Stephan and Carnë for your help.

Yes, Jodie and I are both working on this.
Yes, the objects being imaged are 3D.

3D registration I can see is a different approach, where you try to
reconstruct a 3D model of the object being imaged, and bring the 3D
elements into registration in 3D space. But should it not be possible
to use transformations in the plane, especially if the depth of the
elements is small compared to the height and width of the object being
imaged?

In this case, I would say that the depth is 1/5 of the height and width.

I am thinking that a better example to post here would be:
- image1
- image2 = image1 after being translated, sheared, and rotated
- show you the resulting registered stack.

I have a feeling that the result will not be great, simply because we
do not know how to optimise the registration routines. I am sure that
these routines are excellent! We are ignorant, and that is why we are
asking for expert advice here.

Thanks again for any help!

Cheers
Bill
On Tue, Mar 15, 2011 at 9:40 PM, Stephan Saalfeld <[hidden email]> wrote:

> Hi Bill-Jodie,
>
> to me it seems that you try to register two projections through a thick
> 3d volume in the projection plane.  That's not really possible without
> further knowledge of the 3d-structure or otherwise separable components
> of the signal.  If my impression is wrong, please correct me.
>
> Best,
> Stephan
>
>
>
>
> On Tue, 2011-03-15 at 17:27 +0000, William Simpson wrote:
>> Thanks for offering to help!
>> Here are the two original pics plus an animated .gif that is poorly registered.
>>
>> http://picpaste.com/AALtcc1-I1sU4DxK.png
>> http://picpaste.com/AALtcc2-eUb1IyDH.png
>> http://picpaste.com/AA_time_1_2_lt_cc_motion_bunwarp-Ij3FZv1v.gif
>>
>> Thanks for any help!
>> Bill
>

[Stephan Saalfeld]

Bill,

> 3D registration I can see is a different approach, where you try to
> reconstruct a 3D model of the object being imaged, and bring the 3D
> elements into registration in 3D space. But should it not be possible
> to use transformations in the plane, especially if the depth of the
> elements is small compared to the height and width of the object being
> imaged?
>

You can do that but the structures that you see in one plane aren't
actually there.  That is, they do not have the expected topological and
geometric properties.  Simpler, you have parallaxes depending on
internal 3d-shears and tilts in the tissue, but all structures are
mapped into the same plane.  The consequence is that you see structure A
move left and structure B move right.  You would need to separate both
structures which is complicated from the 2d projection.  Conventional 2d
registration will randomly prefer either A or B (if at all), align it
and bring the other structure out of place by that.

Find what I could do with the new elastic stack align plugin that I have
announced yesterday:

http://picpaste.com/elastic-align-3LE3c0GU.gif

with the parameters:

http://picpaste.com/align-parameters-hm2JNSsM.png

It is clearly visible at various points that structures at various
depths move in different directions due to 3d deformation of the tissue.


> In this case, I would say that the depth is 1/5 of the height and width.

Which is a lot and leaves plenty of room for 3d distortion.

> I have a feeling that the result will not be great, simply because we
> do not know how to optimise the registration routines. I am sure that
> these routines are excellent! We are ignorant, and that is why we are
> asking for expert advice here.
>

It will not be great because 2d registration through warping a plane is
not the appropriate tool here.

Best,
Stephan

> Thanks again for any help!
>
> Cheers
> Bill
> On Tue, Mar 15, 2011 at 9:40 PM, Stephan Saalfeld <[hidden email]> wrote:
> > Hi Bill-Jodie,
> >
> > to me it seems that you try to register two projections through a thick
> > 3d volume in the projection plane.  That's not really possible without
> > further knowledge of the 3d-structure or otherwise separable components
> > of the signal.  If my impression is wrong, please correct me.
> >
> > Best,
> > Stephan
> >
> >
> >
> >
> > On Tue, 2011-03-15 at 17:27 +0000, William Simpson wrote:
> >> Thanks for offering to help!
> >> Here are the two original pics plus an animated .gif that is poorly registered.
> >>
> >> http://picpaste.com/AALtcc1-I1sU4DxK.png
> >> http://picpaste.com/AALtcc2-eUb1IyDH.png
> >> http://picpaste.com/AA_time_1_2_lt_cc_motion_bunwarp-Ij3FZv1v.gif
> >>
> >> Thanks for any help!
> >> Bill
> >

[William Simpson]

Hi Stephan,

> You can do that but the structures that you see in one plane aren't
> actually there.  That is, they do not have the expected topological and
> geometric properties.  Simpler, you have parallaxes depending on
> internal 3d-shears and tilts in the tissue, but all structures are
> mapped into the same plane.  The consequence is that you see structure A
> move left and structure B move right.  You would need to separate both
> structures which is complicated from the 2d projection.  Conventional 2d
> registration will randomly prefer either A or B (if at all), align it
> and bring the other structure out of place by that.
Points taken. Only problem: we do not have 3D information, just the
images taken at different times and from slightly different (unknown)
aspects.
:-(

> Find what I could do with the new elastic stack align plugin that I have
> announced yesterday:
>
> http://picpaste.com/elastic-align-3LE3c0GU.gif
>
> with the parameters:
>
> http://picpaste.com/align-parameters-hm2JNSsM.png
>
> It is clearly visible at various points that structures at various
> depths move in different directions due to 3d deformation of the tissue.

Thanks very much! Much better than our attempt.

Thanks very much Stephan for your help.

Bill