[ITK-users] Super-resolution resampling

[Gavin Baker]

Hello!

I have a time series of 3D data (relatively low resolution), captured in
sequence, with small positional changes (eg. translation). I would like
to perform a super-resolution resampling by first co-registering each
volumetric dataset (using rigid registration) in order to reduce noise
and improve detail.

Is there a registration process that is 1:N (fixed:moving)?

Or is the recommended method to pick a fixed image (ie. #0) and register
each 1..N individually to it?

Given a set of transforms that map each of the 1..N moving images back
to the fixed image for registration, is it possible to then resample the
volume at a higher spatial resolution, combining all image data? IOW
super-resolution resampling?

I tried searching for the above and didn't have much luck finding
relevant info.

Thanks -

:: Gavin
_____________________________________
Powered by www.kitware.com

Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html

Kitware offers ITK Training Courses, for more information visit:
http://www.kitware.com/products/protraining.php

Please keep messages on-topic and check the ITK FAQ at:
http://www.itk.org/Wiki/ITK_FAQ

Follow this link to subscribe/unsubscribe:
http://public.kitware.com/mailman/listinfo/insight-users

[Dženan Zukić]

Hi Gavin,

your plan sounds good! There is no 1:N registration, so you should proceed
with N 1:1 registrations. Pick one as a reference (#0 is good), register
all the other time points to it. You can initialize the k+1-st iteration by
the resulting transform of k-th registration to speed things up.

And yes, you can do super-resolution by resampling all these images onto a
higher resolution grid, e.g. same origin and direction, 2x higher size and
2x smaller spacing.

ITK has all the required classes for this process. Will you let us know how
satisfactory the result was? Ideally with some images :)

Regards,
DÅŸenan Zukić, PhD, Senior R&D Engineer, Kitware (Carrboro, N.C.)

Post by Gavin Baker
Hello!
I have a time series of 3D data (relatively low resolution), captured in
sequence, with small positional changes (eg. translation). I would like
to perform a super-resolution resampling by first co-registering each
volumetric dataset (using rigid registration) in order to reduce noise
and improve detail.
Is there a registration process that is 1:N (fixed:moving)?
Or is the recommended method to pick a fixed image (ie. #0) and register
each 1..N individually to it?
Given a set of transforms that map each of the 1..N moving images back
to the fixed image for registration, is it possible to then resample the
volume at a higher spatial resolution, combining all image data? IOW
super-resolution resampling?
I tried searching for the above and didn't have much luck finding
relevant info.
Thanks -
:: Gavin
_____________________________________
Powered by www.kitware.com
Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html
http://www.kitware.com/products/protraining.php
http://www.itk.org/Wiki/ITK_FAQ
http://public.kitware.com/mailman/listinfo/insight-users

[Gavin Baker]

Thanks, DÅŸenan -

I'll start with the N x 1:1 registration then.

I can see how to resample the moving image, with the transform applied,
as per the examples. However it is not clear how to _combine_ the N
images together for the super-resolution resampling. Or would it be a
two-step process, where each moving image is first resampled, and after
that they are averaged together?
Thanks -

:: Gavin

Post by Dženan Zukić
Hi Gavin,
your plan sounds good! There is no 1:N registration, so you should
proceed with N 1:1 registrations. Pick one as a reference (#0 is
good), register all the other time points to it. You can initialize
the k+1-st iteration by the resulting transform of k-th registration
to speed things up.>
And yes, you can do super-resolution by resampling all these images
onto a higher resolution grid, e.g. same origin and direction, 2x
higher size and 2x smaller spacing.>
ITK has all the required classes for this process. Will you let us
know how satisfactory the result was? Ideally with some images :)>
Regards,
DÅŸenan Zukić, PhD, Senior R&D Engineer, Kitware (Carrboro, N.C.)
On Tue, Jun 6, 2017 at 4:00 AM, Gavin Baker

Post by Gavin Baker
I have a time series of 3D data (relatively low resolution),
captured in>> sequence, with small positional changes (eg. translation). I
would like>> to perform a super-resolution resampling by first co-
registering each>> volumetric dataset (using rigid registration) in order to
reduce noise>> and improve detail.
Is there a registration process that is 1:N (fixed:moving)?
Or is the recommended method to pick a fixed image (ie. #0) and
register>> each 1..N individually to it?
Given a set of transforms that map each of the 1..N moving
images back>> to the fixed image for registration, is it possible to then
resample the>> volume at a higher spatial resolution, combining all image data? IOW>> super-resolution resampling?
I tried searching for the above and didn't have much luck finding
relevant info.
Thanks -
:: Gavin
_____________________________________
Powered by www.kitware.com
Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html
http://www.kitware.com/products/protraining.php
http://www.itk.org/Wiki/ITK_FAQ
http://public.kitware.com/mailman/listinfo/insight-users

[Dženan Zukić]

Hi Gavin,

if you want to avoid keeping N resampled images, you could have a sum of
resampled images which you divide by N at the end to get the average.

Regards,
DÅŸenan Zukić, PhD, Senior R&D Engineer, Kitware (Carrboro, N.C.)

Post by Gavin Baker
Thanks, DÅŸenan -
I'll start with the N x 1:1 registration then.
I can see how to resample the moving image, with the transform applied, as
per the examples. However it is not clear how to _combine_ the N images
together for the super-resolution resampling. Or would it be a two-step
process, where each moving image is first resampled, and after that they
are averaged together?
Thanks -
:: Gavin
Hi Gavin,
your plan sounds good! There is no 1:N registration, so you should proceed
with N 1:1 registrations. Pick one as a reference (#0 is good), register
all the other time points to it. You can initialize the k+1-st iteration by
the resulting transform of k-th registration to speed things up.
And yes, you can do super-resolution by resampling all these images onto a
higher resolution grid, e.g. same origin and direction, 2x higher size and
2x smaller spacing.
ITK has all the required classes for this process. Will you let us know
how satisfactory the result was? Ideally with some images :)
Regards,
DÅŸenan Zukić, PhD, Senior R&D Engineer, Kitware (Carrboro, N.C.)
Hello!
I have a time series of 3D data (relatively low resolution), captured in
sequence, with small positional changes (eg. translation). I would like
to perform a super-resolution resampling by first co-registering each
volumetric dataset (using rigid registration) in order to reduce noise
and improve detail.
Is there a registration process that is 1:N (fixed:moving)?
Or is the recommended method to pick a fixed image (ie. #0) and register
each 1..N individually to it?
Given a set of transforms that map each of the 1..N moving images back
to the fixed image for registration, is it possible to then resample the
volume at a higher spatial resolution, combining all image data? IOW
super-resolution resampling?
I tried searching for the above and didn't have much luck finding
relevant info.
Thanks -
:: Gavin
_____________________________________
Powered by www.kitware.com
Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html
http://www.kitware.com/products/protraining.php
http://www.itk.org/Wiki/ITK_FAQ
http://public.kitware.com/mailman/listinfo/insight-users

[Gavin Baker]

Thanks, Samuel,

Great point - using the N/2 sample makes a lot of sense. I'll start
with that example and see how I go. Any thoughts on my followup question
about the super-resolution resampling the N images together would be
most appreciated.
Regards -

:: Gavin

Post by Dženan Zukić
Hi Gavin,
One small addition, I would probably take the N/2 image to register
everybody else to, in order to minimize the maximal transformation
(might not matter in your case since it is only small transformations
but it could minimize errors due to resampling).>
https://itk.org/Wiki/ITK/Examples/Registration/ImageRegistrationMethod>
You will most likely want to use a different optimizer and you can see
in the code how to set the size etc of the output image in the
resampler.>
On Tue, Jun 6, 2017 at 9:16 AM, DÅŸenan Zukić

Post by Dženan Zukić
your plan sounds good! There is no 1:N registration, so you should
proceed with N 1:1 registrations. Pick one as a reference (#0 is
good), register all the other time points to it. You can initialize
the k+1-st iteration by the resulting transform of k-th registration
to speed things up.>>
And yes, you can do super-resolution by resampling all these images
onto a higher resolution grid, e.g. same origin and direction, 2x
higher size and 2x smaller spacing.>>
ITK has all the required classes for this process. Will you let us
know how satisfactory the result was? Ideally with some images :)>>
Regards,
DÅŸenan Zukić, PhD, Senior R&D Engineer, Kitware (Carrboro, N.C.)
On Tue, Jun 6, 2017 at 4:00 AM, Gavin Baker

Post by Gavin Baker
I have a time series of 3D data (relatively low resolution),
captured in>>> sequence, with small positional changes (eg. translation). I
would like>>> to perform a super-resolution resampling by first co-
registering each>>> volumetric dataset (using rigid registration) in order to reduce
noise>>> and improve detail.
Is there a registration process that is 1:N (fixed:moving)?
Or is the recommended method to pick a fixed image (ie. #0) and
register>>> each 1..N individually to it?
Given a set of transforms that map each of the 1..N moving
images back>>> to the fixed image for registration, is it possible to then
resample the>>> volume at a higher spatial resolution, combining all image
data? IOW>>> super-resolution resampling?
I tried searching for the above and didn't have much luck finding
relevant info.
Thanks -
:: Gavin
_____________________________________
Powered by www.kitware.com
Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html
http://www.kitware.com/products/protraining.php
http://www.itk.org/Wiki/ITK_FAQ
http://public.kitware.com/mailman/listinfo/insight-users

_____________________________________
Powered by www.kitware.com
Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html
http://www.kitware.com/products/protraining.php
http://www.itk.org/Wiki/ITK_FAQ
http://public.kitware.com/mailman/listinfo/insight-users
_______________________________________________
Community mailing list
http://public.kitware.com/mailman/listinfo/community

--
Samuel Gerber
R&D Engineer
Kitware, Inc.