How to sort images?
wlad.og...@freenet.de
Hello group,
I am trying to convert a set of MR images obtained from some GE scanner
and saved in DICOM format into a 3D volume. The test image set has 15
images and all of them have the same Study Instance UID (0020,000D),
Series Instance UID (0020,000E), Study ID (0020,0010), Series Number
(0020,0011), Acquisition Number (0020,0012) and Frame of Reference UID
(0020,0052) tags. Based on this information I have grouped them into the
same volume, but this decision was false as I have found further that
they have different Image Orientation tags (0020,0037).
So, my questions are:
1. Is this DICOM conformant when images with the same Frame of
Reference UID have different Image Orientations?
2. If yes, which information should one use for correct sorting of
images into spatially consistent volumes (3D matrices of pixels)?
Thanks in advance
Regards
Wlad
----- Posted via NewsOne.Net: Free (anonymous) Usenet News via the Web -----
http://newsone.net/ -- Free reading and anonymous posting to 60,000+ groups
NewsOne.Net prohibits users from posting spam. If this or other posts
made through NewsOne.Net violate posting guidelines, email ab...@newsone.net
David Clunie
> So, my questions are:
> 1. Is this DICOM conformant when images with the same Frame of
> Reference UID have different Image Orientations?
Yes ... this just means that the slices are not parallel to each
other, for example one might have coronal images in one series and
axial images in another series but the same FoR UID, or in the same
series one might have multiple different angled slices through the
lumbar spine, aligned along the axis of each inter-vertebral disk
level, or one might have a series of rotating para-coronal to para-
saggital slices through the knee.
The same FoR UID indicates in all these cases that the values for
position and orientation are comparable (i.e. are specified in the
same co-ordinate space with the same origin and axes, etc.)
> 2. If yes, which information should one use for correct sorting of
> images into spatially consistent volumes (3D matrices of pixels)?
One can only easily construct a 3D matrix if the slices are parallel
and contiguous, obviously. Normally, one would acquire such slices
in this manner if 3D reconstruction is envisaged.
Otherwise one has to create a 3D matrix by sampling the non-parallel
slices into the matrix you want ... this is not a trivial problem
and there may be holes that don't interpolate well depending on how
well the non-parallel slices cover the space of interest.
If one just wants to sort slices in order (e.g. to present them as
individual 2D images) then one can still use the position (top-left-
hand-corner) if the slices don't intersect, or any other appropriate
reference point in the images. For some acquisitions it is better to
sort by the angle (e.g. the rotating knee acquisition I described)
derived from the orientation direction cosines.
david
--
David A. Clunie mailto:dcl...@comview.com
Development Director, Medical Imaging Products http://www.comview.com/
ComView Corporation Work 914-332-4800 Fax 208-445-5867
220 White Plains Road, 5th Floor Home 570-897-7123 Fax 570-897-5117
Tarrytown NY 10591 http://www.dclunie.com/
wlad.og...@freenet.de
questions. I'm searching for a method which I can use to
find subset of images that was made during one scan (acquisition?)
and can be used for an easy 3D reconstruction (without re-sampling
and interpolation). But from your reply I see that such an easy
reconstruction is not always possible. Is it true?
Suppose I will use re-sampling and interpolation for 3D reconstruction. How
can I select all those images that I can use for such a reconstruction? Which
tags should I check for this purpose?
David Clunie
There is no simple answer to this ... it depends entirely
on the acquisition.
Frankly, I would simply insist that for 3D reconstruction,
the slices in the series to be reconstructed all be acquired
parallel, contiguous and with the same thickness and spacing
(in CT one might want to relax the same thickness requirement
but rarely in MR).
I don't see much practical advantage in allowing otherwise,
since most applications where non-parallel slices are obtained
during the same acquisition sample the space very sparsely and
would not result in a very meaningful reconstruction.
Do you have a specific application in mind or are you just
trying to solve a general problem ?
david
--
wlad.og...@freenet.de
Our company develops computer aided surgery system. We are actually
putting such restrictions as your have described on input data for our
system,
but for some reasons we would like to provide our customers with a
possibility
to use acquisitions which do not conform to our scan protocols. So I'm
searching
for a method which I can use for the most complete 3D reconstruction and which
is
safe in sense that it do not introduce errors by grouping incomparable slices
together.
For example we have some acquisitions each containing 10 axial, 10 coronal and
10 sagittal slices
and I see that teoretically we can use this acquisitions for 3D
reconstruction, but I do not know
how to test whether it is safe.
Regards
Wlad
>Hi Wlad
>
>There is no simple answer to this ... it depends entirely
>on the acquisition.
>
>Frankly, I would simply insist that for 3D reconstruction,
>the slices in the series to be reconstructed all be acquired
>parallel, contiguous and with the same thickness and spacing
>(in CT one might want to relax the same thickness requirement
>but rarely in MR).
>
>I don't see much practical advantage in allowing otherwise,
>since most applications where non-parallel slices are obtained
>during the same acquisition sample the space very sparsely and
>would not result in a very meaningful reconstruction.
>
>Do you have a specific application in mind or are you just
>trying to solve a general problem ?
>
>david
Stuart Swerdloff
One theoretical basis for utilizing the information you have available is to
convert the
information back in to the (Spatial) Frequency domain, populate the Frequency
Domain
with your now "irregularly" spaced samples, pick a convolution kernel you can live
with (the
broader the kernel, the noisier the edges of the images will become), resample the
Frequency
Domain so that it has become regularly spaced, apply the Inverse Fourier Transform,
and then
divide on a point by point basis the resulting volume with the inverse transform of
your convolution kernel.
Not particularly fast or simple, however, it does provide a theoretical basis for
"being safe", so long
as the FoR is truly consistent.
You get to pick up the additional Frequency Domain information from the orthogonal
acquisitions,
you have a clear definition of how you are giving up Signal/Noise (the
interpolation in the Frequency Domain). It also makes it a little clearer why
doing three samplings (acquisitions) that are poorly
spaced on a particular axis doesn't give anywhere near the benefit of having
sampled consistently
well across the entire space.
A straight forward interpolation (volume of overlap or convolution based)
in the coordinate space domain will give you similar results,
but it gets a little more awkward to define exactly where you win, where you lose,
and why.
Stuart
wlad.og...@freenet.de wrote:
...
David Clunie
One thing to beware of when using slices from different
acquisitions (as I presume these orthogonal sets are), is
to be sure that the acquisition parameters are identical and
the system is tuned the same, otherwise the range of signals
and the contrast weighting in the slices from different
acquisitions will not be comparable. Since most systems run
some sort of automatic tuning to adjust transmit and receive
gain and center frequency between acquisitions, even if the
pulse sequence is the same and the TE and TR are the same,
the images may be quite different. Same goes for saturation
bands, flow and motion artifact, etc.
Secondly, only 10 slices in each plane seems inadequate ...
normally for a 3D reconstruction of a cube one would use at
least 64 or 128 slices in order to try to get as close to
isotropic voxels as possible.
If this is for a planning system, unless you are able to tolerate
errors of a magnitude measured in centimeters rather than
millimeters, I can't see how this can possibly work (safely).
david
wlad.og...@freenet.de wrote:
>
> Hi David
>
> Our company develops computer aided surgery system. We are actually
> putting such restrictions as your have described on input data for our
> system,
> but for some reasons we would like to provide our customers with a
> possibility
> to use acquisitions which do not conform to our scan protocols. So I'm
> searching
> for a method which I can use for the most complete 3D reconstruction and which
> is
> safe in sense that it do not introduce errors by grouping incomparable slices
> together.
>
> For example we have some acquisitions each containing 10 axial, 10 coronal and
> 10 sagittal slices
> and I see that teoretically we can use this acquisitions for 3D
> reconstruction, but I do not know
> how to test whether it is safe.
>
> Regards
>
> Wlad