[Mristudio-users] transformations with AIR in Landmarker

[susumu]

Siewmin raised some important points and I hope other users could gain some
important information from this thread.

 

 I got the mean DWI and tensor file from the original 4D DWI. The tensor
file is as you've said, which contains  "Dxx, Dyy, Dzz, Dxy, Dxz Dyz" in one
file. In Landmarker, I loaded only the mean DWI of the subject and then the
tensor file with the "open .d" button. The 4D raw DWI is not used in
Landmarker or later. With the "linear" option in AIR, I could rotate the
tensor and the mean DWI to the template DWI without having to subsample the
input data first. The "nonlinear" option requires the subject data (mean DWI
and the tensor file) to be of the same dimension as the template (i.e 181,
217, 181 if ICBM81 DWI is used as template). This is noted in the pop up
message, and it would only continue once the input files are resampled to
the template. 

 

>It is always advised to do linear transformation first. To start
non-linear, you have to make sure that the brain locations and shapes are as
similar as possible. This is similar to non-linear fitting. You always have
to provide "initial" values close to the real solutions. Otherwise,
non-linear fitting could converge to a wrong solution (trapped by "local
minima").

 

Additionally, I tried affine in the linear registration option. The output
DWI of the registration is identical to the input of the DWIs if I used one
of my subject's DWI as the template. Using the 2nd order polynomial in
nonlinear option, an error message occurs "registration terminated due to a
calculation problem, please try a lower order model". It seems to only work
if the template is one of those provided in Landmarker.

 

> If you can send us two images you are having the problem, we can take a
look at why AIR doesn't work.

 

I then saved the rotated tensor file as one file in the .dat format and the
linearly registered 3D  meanDWI. In MRIview3D, I loaded this new 3D DWI and
open this new tensor.dat file with the open.dat button to re-calculates the
fa, colour map etc. I viewed the colour map output from this and it looks
pretty good, and as described previously. 

 

> This is the correct procedure.

 

I saved the fa and principal eigenvector file and used them to perform
fiber-tracking. The tracts that are manually extracted obtained looks
similiar to the fiber tracts from the original raw 4D DWI. 

 

> That is a good sign.

 

The FA values of these tracts are smaller uniformly in comparison, and the
assymetry preserved (i.e the left tract FA is smaller than the right tract
FA using the maps calculated from the original and rotated tensor). 

 

> Normalization requires pixel interpolation, which leads to decreased FA.
So you can not compare FA values with and without normalization.

 

If the non-linear option is used, is there a problem that the tensor is
subsampled first before it is re-oriented, as described above? 

 

> As I mentioned above, nonlinear should be AFTER linear transformation.
This means, the data are transformed multiple times and pixels are
interpolated multiple times. To avoid this, once you did linear (save
transformation matrix: T-linear) and nonlinear (save transformation matrix:
T-nonlinear), you can combine the T-linear x T-nonlinear using one of the
buttons in the LDDMM section. Once you combined all transformation matrices,
you can apply it only once to the original image in the native space.

 

> Another interesting thing is, there are two ways to transform images;
transform your image to an atlas or transform atlas (usually, binary
segmentation map) to your original data for automated segmentation. For the
latter approach, you don't have to transform (therefore interpolate) your
original data. When you transform the atlas (segmentation map), make sure to
use "nearest neighbor" (no interpolation), not the tri-linear interpolation.
The segmentation map consists of arbitrary numbers assigned to each segment
(e.g. the internal capsule is "1" and the corona radiata is "3"). If you
interpolate these two segments and obtain "2" in between these two segments
and if "2" means pons, you would get the pons in between the internal
capsule and the corona radiata. On the other hand, when you are transforming
your data to an atlas space, use the tri-linear interpolation. Otherwise,
you could get strange discontinuity artifacts.

 

Do you suggest using the affine or non-linear transformation of the DWIs and
the corresponding transformation of the tensor? Lastly, I'm not sure if the
colour map display is of an issue? It occurs both using the linear affine
option and the non-linear option.

 

> When everything is working right, the color map should look smooth and
clean. Maybe you are using "nearest neighbor" for the tensor transformation?

 

Thanks

 

Siewmin

 

 

 

 

On Sat, May 9, 2009 at 7:44 AM, susumu <susumu at mri.jhu.edu> wrote:

First, I tried rotating the tensor but had problems with the results. When I
look at the colour map from the resulting tensor, it looks ok generally, but
there are certain areas of blotchiness and some mixing of colours especially
in the sidelines of white matter regions (i.e there can be a striking red
line or a blotch of redness in the middle of a green regions within the
saggital stratum). These are the steps I took, and would like to check if it
is done correctly.

1. resample mean DWI of subject data (together with the 4D DTI data and the
tensor) to DWI of ICBM 81 template. Non-linear transform (2nd order
trilinear) the resampled meandwi of subject to dwi of template, and used
that to drive the transformation of the resampled 4D DTI and the resampled
tensor simultaneously. (The resampled tensor was loaded with the .dat
button.) I used MRIview 3D to load the normalised 4D DTI, followed by the
normalised tensor with the .dat button, and calculate the fa, colour etc
maps from the normalised tensor. I  tried various thresholding with AIR but
the colour maps problem remains.

 

> First of all, I think you don't have to resample DTI images before AIR.
AIR can take data with different matrix and pixel sizes and the output
images have the same matrix/pixel dimensions as the template (in your case,
181x217x181 of ICBM81).

> By saying "4D DTI", do you mean raw 4D DWI before tensor calculation? if
so, I don't think you want transform raw DWIs. If you want to do anything to
raw DTI, you also want to transform the corresponding gradient table, which
could be messy. Our tools do not provide methods to transform raw DWIs.

> When you say, "tensor" data, I assume you are talking about 4D file that
contains "Dxx, Dyy, Dzz, Dxy, Dxz, Dyz" images in one file. If so, please
make sure that you load this 4D tensor file by the "open .d" button in
Landmarker. Once read in Landmarker, the 4D tensor file is disassembled to 6
3D data for Dxx, Dyy, Dzz, Dxy, Dxz, Dyz.

> Once the AIR is done using mean DWI (your data) - mean DWI (template), the
6 3D tensor element data should be stored all together into one file as a
"transformed 4D tensor file".

> This transformed 4D tensor file needs to be read into DtiStudio to
re-calculate new transformed FA, color, vec, etc.

> The blotch-looking colormap could happen by; 1) transform the color map
directly (you should recreate color maps using the recalcuated FA and vector
files from the transformed tensor file or 2) you transformed a vector file
(transformed vector files should be obtained from the transformed tensor
file. Vector can not be directly transformed).

 

Second, I tried without rotating the tensor. I calculate the scalar and
colour maps from the original 4D DTI data, aligned the maps to AC-PC line,
draw rois on them and invert transform the rois back to native dti space to
extract the relevant white matter tracts. I had two problems here. 1)  with
orienting the subjects scalar DWI to  another particular subject's DWI with
rigid transformation in AIR and 2) Invert transform ROIs drawn on the
registered colour map back to the original DTI space. These are the steps I
took:

 

> Maybe you are transforming raw DWI data? Again, you should not transform
raw DWI and recalcuate tensor. You have to first calculate tensor (throw
away the raw DWIs) and transform the tensor.

 

1. A subject's data which is acquired // to AC-PC is chosen and other
subjects not acquired //to AC are registered to this. Again the meanDWI is
used for source and target images, this time using rigid transformation in
AIR in Landmarker. However, there is no difference detected after the
transformation has taken place. However, if I change the target image from a
subject's DWI to the mean DWI of the ICBM 81 template, it works but the
source data would be subsampled to that of the ICBM template. I'm not sure
why it didn't work in the 1st case? 

 

> As mentioned above, AIR not only adjust the image orientation and
translation, but also the matrix / pixel size..

> Rigid transformation of AIR is sometimes very cranky. We are working on
alternative methods now. please try different threshold or affine.

 

Additionally, using either another subject or the template DWI as target
image, would yield the an output file (AlignlinearOutput.air) when the
transformation has been performed, which I saved the
invert_AlignlinearOutput.air. However, when I clicked on these .air files in
the folder it is saved as an installer package, I get the following error
message " An error has occurred.The application could not be installed
because the AIR file is damaged. Try obtaining a new AIR file from the
application author." I have installed the air application and placed it in
the same folder, together with the alignlinear.exe file. Is it possible to
view the transformation matrix files saved or did I not installed the air
application appropriately?

 

> Maybe Xin can answer this question better, but I'm not sure why you want
to open *.air??

> Once you save the *.air, you can load it back to Landmarker and apply to
images by using the "open M" button in the LDDMM section.

 

2. ROIs are drawn on the transformed AC-PC aligned colour maps with
ROI-editor. I would like to check if it's OK to draw rois on the transformed
colour map? These ROIs would later be used as the start ROI and stop ROI to
extract the corresponding white matter tract with DTI studio) 

 

> Drawing ROIs on color maps is fine. ROI is simply 1/0 information about
which pixels you chose and the color maps are simply what is displayed to
aid your ROI drawing. 

> If you draw ROI in the atlas space and have a plan to back-transform your
ROI files to the native space, there are several things you have to know;

1) Atlas space is 1x1x1 and your native space could be 2-3 mm. The ROI
defined in the thin 1 mm slice could be lost after transformation to the
thick 2-3mm slice. Therefore, you have to create 3D ROIs in the atlas space
with at least 3mm thick

2) Use nearest neighbor interpolation when you back-transfer. You don't want
to get a pixel with 0.3 or 0.4 after transformation. You want 1/0

 

3. I  tried to invert-transform these ROIs back to the original DTI space
with the previously saved invert_matrix.air file. How is that done in
Landmarker and if the previous error message needs to be resolved before I
can used the matrix files? 

 

> Xin, how to do the invert-transformation of *.air?

 

Susumu

 

 

 

 


_______________________________________________
Mristudio-users mailing list
Mristudio-users at mristudio.org
http://lists.mristudio.org/mailman/listinfo/mristudio-users

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.mristudio.org/pipermail/mristudio-users/attachments/20090510/2b26ed7d/attachment-0001.html