|
| |
Glioma Volume Project - Generalized Procedures
Eigenimage Filter/Gram-Schmidt Orthogonalization
The information present in a sequence of images acquired
using different MRI protocols show both the spatial location of different
tissues and their contrast attributes determined by the protocols. The
Eigenimage filter is a linear filter that produces a single composite image,
termed an eigenimage, using the contrast differences from all images in the
sequence. In the eigenimage different tissues are segmented from each other
and the resulting pixel intensity for each tissue is weighted based on the
partial volume in each voxel. This results in enhanced visualization of the
tissue's extent and accurate/reproducible volume determination. Originally the
Eigenimage filter was derived using the solution to a generalized eigenvalue
problem. This derivation was simplified to show that the filter could be
determined using the Gram-Schmidt Orthogonalization procedure. By using the
Gram-Schmidt Orthogonalization an analytical solution is determined allowing
fast and accurate results.
In the application of the Eigenimage filter an operator must
define a representative location for each tissue to be segmented. The location
for each tissue is selected using a small sample region of interest (ROI)
within each tissue from the images in the sequence. The average gray level of
the ROI for each tissue is used to characterize the contrast between the
tissues from all MRI protocols used. These average gray levels define the
elements of a vector, termed a tissue signature vector. These signature
vectors are utilized in the Eigenimage filter to produce the weighting
coefficients used in the linear filter procedure. The steps used to
reproducible apply the Eigenimage filter to glioma volume determinations are
given below.
- Registration
Since the contrast differences from all MRI protocols
acquired are used to direct the segmentation using the Eigenimage filter the
images for all protocols must be registered prior to any processing. The
registration is only required if movement occurs between acquisition of each
sequence. The determination of the sequences to register is done by loading
one location for all sequences into one Browser in Eigentool. The image
sequences used at HFH for glioma studies include pre- and post-Gd T1 weighted
spin echo images, fluid attenuated inversion recovery (FLAIR), and short and
long echo-time (TE) T2 weighted images (see
Original Dataset). By
scrolling or looping through the sequences those that require registration can
be found. In addition, a ROI can be created that outlines known anatomy and
this outline can be displayed on all images to check alignment (see
Registration Misalignment). In most studies the post-Gd T1 weighted
image must be registered to the other sequences. In some studies,
approximately 20% at HFH, the FLAIR and T2 weighted images must also be
registered. The registration can be accomplished using several methods, e.g.
AIR and Head and Hat techniques. The procedures used for these two methods are
given below.
-
AIR Registration
Load all locations from one sequence to use as a basis set
into one Browser. Select a gray level range for the head that excludes the
background. Load the sequence to be registered into another Browser and
select a gray level range for the head from this sequence. Open the
Registration Dialog and select the AIR option. Input the lower value for the
gray level range in the options for t1 and t2, where t1 is the threshold
value for the basis set and t2 is the threshold value for the sequence to be
registered. Select the Match button and wait for the determination of the
translation/rotation process to be completed. When the translation/rotation
process is completed select the Reslice button. Verify the new image
sequence is registered. If it is not registered delete the new Browser and
select new gray level ranges for each image sequence and repeat the
registration process. Note if the registration is not acceptable it may be
necessary to scale the gray level of the tissues from the sequence to be
registered to the basis set. This can be accomplished by first creating an
ROI that encompasses the brain in the sequence to be registered. The tissues
under this ROI are then scaled to the same value as the tissue in the basis
set by using the Multiply by a Constant option in the Math dialog. Note, the
new sequence created by the scaling step may appear anomalous, but it is
only used to direct the determination of the translation/rotation matrix.
Select the Match button and use the scaled sequence as the registration set.
When the determination of the translation/rotation process is completed
select the Reslice button and apply this step to the original registration
sequence, not the scaled set. Verify the new image sequence is registered.
If the registration is acceptable repeat the procedure on any other
sequences that require registration. During the registration all locations
that contain the lesion should be noted so further processing is only done
on these locations to reduce the total processing time (see Step3: Lesion
Extent).
-
Head and Hat
Registration
Load all locations from one sequence to use as a basis set
into one Browser. Open the ROI/VOI Dialog and select the Contour option under
the Draw menu. Select a gray level range for the image background and place
a seed point outside the head in the background. A multi-resolution
algorithm then determines the skin contour automatically from the seed
point. For some locations the seed point or gray level range may need to be
redefined to create the contour. The contours do not have to be exact and
minimal operator interaction is required to create acceptable contours. Once
the contour is created on the basis sequence the procedure is repeated on
the sequence to be registered, this sequence must be loaded into a new
Browser. Following the creation of contours on both image sequences, open
the Registration Dialog and select the Head and Hat option. Select the Match
button and wait for the determination of the translation/rotation process to
be completed. Note the residual value in the Information Window and select
the Match button again. Again note the residual value when the determination
of the translation/rotation process is completed. If the residual value has
increased from the previous value select the Match button again. Repeat the
Match procedure until the residual value decreases. When the residual value
decreases select the Back-up button and then select Reslice. Verify the new
image sequence is registered. If it is not registered delete the new Browser
and make new contours on both image sequences noting any major discrepancies
between the contours and the surface of the head. If the registration is
acceptable repeat the procedure on any other sequences that require
registration. During the registration all locations that contain the lesion
should be noted so further processing is only done on these locations to
reduce the total processing time (see Step 3:
Lesion Extent).
-
Noise Suppression
To enhance the performance of any segmentation routine
additive noise should be reduced. Eigentool has several smoothing filter
options that can be used for this purpose. For the glioma volume studies it
has been determined that an adaptive filter can be used to reduce noise while
maintaining partial volume information. The adaptive filter is performed using
the Average option in the Fit dialog. Note all sequences must be registered
prior to using the adaptive filter. The procedure to use the adaptive filter
requires all images for a single location to be loaded into one Browser. After
the images for one location are loaded the standard deviation from a ROI in
white matter must be determined. The ROI can be a simple box drawn on one of
the images. The standard deviation is used to specify the amount of noise
suppression and preservation of partial volume information that is done. For
the sequences acquired at HFH two times the average standard deviation from
all image sequences is used. This value is input into as sigma in the Fit
dialog. This value may need to be changed if other acquisition sequences or
variations in the number of image sequences changes. The options for the
adaptive filter are a 9x9x1 matrix size and the Average menu option from the
Fit dialog. Following the noise suppression procedure on the initial location
it can be automatically run without any operator interaction (i.e. in the
background of the computer processor) on all locations that contain the lesion
by selecting the Repeat function in Eigentool. Note the noise suppression only
needs to be done on the locations that contain the lesion (see Step3: Lesion
Extent).
-
Lesion
Extent
With all image processing the time required to obtain
results can be minimized if the processing steps are applied only to the
locations that contain useful information, i.e. contain the tissue of
interest. Therefore, determining the minimum number of locations to process
can significantly reduce the time required. In this project the lesion extent
can be determined by loading all locations from the FLAIR sequence into a
Browser and the locations that contain a FLAIR abnormality can then be noted.
It is recommended that at least one additional location extending past the
lesion in both directions should be included to ensure no lesion voxels are
excluded.
-
Segmentation
To segment the FLAIR and Gd components of the lesion one
location for all sequences following noise suppression are loaded into
Eigentool (see Original Dataset). The location selected should include both Gd enhancement and FLAIR
hyper-intensity visible in the images. The Segmentation dialog is then selected
and the operator must define all known tissues to be segmented. For the glioma
volume studies the tissues that must be defined are normal tissue (i.e. white
matter and gray matter), cerebrospinal fluid (CSF), Gd enhancement, and FLAIR
hyper-intensity. Note the number of tissues defined cannot exceed the number
of different MRI protocols acquired and used in the segmentation. A tissue is
defined by selecting a point within the images displayed using the Automatic
ROI option in the ROI/VOI dialog. The Automatic ROI option will grow a small
sample ROI based on the gray level under the point selected (see
Normal ROI,
CSF ROI,
Gd ROI and
FLAIR ROI). Note that these ROI do not have to connected or contiguous. The only limitations are that the
total ROI must only include the tissue type being defined. For example the
normal tissue ROI includes several areas in white matter and several areas in
gray matter but the ROI does not include any pixels from CSF or the lesion.
Following the creation of each sample ROI the tissue these ROI are defined as
a process in the Segmentation dialog. Once all ROI/processes are defined in the
segmentation dialog the Gram-Schmidt filter will create a separate eigenimage
for each tissue using the A-Ortho option and selecting the Filtr Img button
(see Segmented Images). If
the ROI selected for two or more tissues are too similar, i.e. have the same
contrast in all MRI protocols used, the segmented images may appear noisy.
This can occur when the ROI contain substantial amounts of partial volume
averaging of different tissues so the contrast between them is reduced. In
this case new ROI must be selected for these tissues. The eigenimages
themselves can be used to improve the sample ROI creation. Using the
eigenimages a gray level range for each tissue can be easily determined since
the Gram-Schmidt Orthogonalization process maps the gray level in the
eigenimages to a value of one (1.0) for the segmented tissue and all other
tissues defined are mapped to zero (0.0). The gray level range recommended to
use on the eigenimages is 0.8 to 1.2. These values can be increased or
decreased based on the ROI created and some editing of stray and artifact
pixels should be removed (see New
Normal ROI, New CSF
ROI, New Gd
ROI, and New FLAIR
ROI). As a
result of using the eigenimages to make new ROI it is possible to create ROI
that encompass more pixels from each tissue. By increasing the number of
pixels used for the ROI for each tissue reduces the statistical variation for
the average gray levels used for the signature vector determination. This in
turn results in an increased signal to noise ratio (SNR) for the resulting
eigenimages. The new ROI created are used by replacing the old processes in
the Segmentation dialog (see Clear Processes button) and selecting the Filtr
Img button again (see New Segmented Images). Note the new eigenimages may not look substantially
different from the original eigenimages, but the SNR should be improved and
therefore any quantitative analysis should be improved. Once the eigenimages
appear acceptable the segmentation can be run on all locations that contain
the lesion by selecting the Repeat function in Eigentool. The Repeat function
is run in the computer background so no operator interaction is required. In
addition to the segmentation of the tissues selected by the operator an
additional segmentation can be performed that will display all voxels that
contain tissues or artifacts that were not selected by the operator. This
image is termed an orthogonal image based on the mathematics used in its
creation. In the orthogonal image any tissue not selected by the operator,
e.g. fat, skin and other extra-cranial tissues, will be displayed. This may
include other intra-cranial features as well as possible lesion components.
The orthogonal image will only provide useful information if the total number
of tissues defined (i.e. processes defined in the Segmentation dialog) does not
exceed the number of MRI protocols used.
-
Volume Determination
For volume determination in the glioma studies the
eigenimages created will include some tissue/artifacts that are not the tissue
of interest (i.e. skin, fat etc. will be seen in the eigenimages). This is due
to the fact that not all tissues can be defined as processes to segment based
on the limitation of the number of MRI protocols acquired. As mentioned above
these pixels may be displayed in the orthogonal image if it is created, but
these pixels must not be included in the volume determination for the tissue
of interest. In order to limit the volume determination to the tissues of
interest (i.e. Gd enhancement and FLAIR hyper-intensity) a volume of interest
(VOI) is used to limit the calculation. The procedure to create the VOI
requires the eigenimages for all locations for the Gd-enhancement to be loaded
into one Browser and the eigenimages for the FLAIR hyper-intensity to be
loaded into a second Browser (see Substance Image
Set). The creation of the VOI is done using a simple
histogram analysis methodology based on the knowledge of the statistical
variation in the results from Gram-Schmidt Orthogonalization procedure. This
is based on the assumption that the noise distribution in MRI can be modeled
as zero-mean Gaussian noise (for signals >0), and since the Gram-Schmidt
Orthogonalization is a linear filter, the resulting noise in the eigenimages
will also be zero-mean Gaussian distributed. Based on these assumptions,
standard statistical methods can be used to determine an appropriate
confidence level for excluding the tissues that were removed in the
segmentation, e.g. excluding normal tissue, CSF and FLAIR hyper-intensity in
the Gd-enhancement eigenimage. The histogram analysis procedure requires a
small ROI to be drawn on one of the eigenimages in the tissue that was removed
(see Substance
Image Set with ROIs). From this ROI the standard deviation for the eigenimage pixel value
distribution around zero is determined using the Thr Peak option in the ROI/VOI
dialog under Image Processing (see Thr Peak in
the ROI/VOI dialog). For the glioma studies a threshold value of three
standard deviations above zero was chosen to provide a 99% confidence that all
tissues removed in the segmentation are not included in the VOI for the tissue
of interest (see Thr Peak Results). The initial VOI created by this processing step
will include tissues other than the Gd enhancement and FLAIR hyper-intensity
and so some editing will be required to remove these pixels from the VOI using
the Morphological and Erase Irregular options in the ROI/VOI dialog (see
Erase Irregular Results). This procedure must be performed on all locations for both the eigenimages segmenting the Gd enhancement and the FLAIR hyper-intensity. Once
a VOI for each feature on all locations is created the volume for each feature
and the overlapping pixels, i.e. partial volume pixels, between features is
determined automatically using the Substance dialog (see
Substance Results).
Procedures specific to this project:
-
Load a single, central location
displaying good gadolinium and flair enhancement for all restored sequences of
the patient study.
-
Select
Analyze->Segmentation from the
Main Window.
-
Select
Automatic in the Region of Interest/Volume of
Interest Dialog.
-
Select an image
that best displays the white and gray matter.
-
Press the left button over the white matter area contralateral to the abnormal anatomy.
Press the left button over a gray matter contralateral to the abnormal anatomy.
Ensure that the approximate areas of gray and white matter regions of interest
are the same. Additional areas far from the lesion, but not contralateral
may be used, if necessary.
-
Enter normal
for the process name in the Filter Dialog, where r1 is the default name.
-
Press
Select Process in the Filter Dialog.
Note
Select the image that best displays the cerebrospinal fluid and
repeat Steps 3b, 4, and 5 for cerebrospinal fluid. Use csf for the process name.
-
Select the
post-gadolinium image.
-
Press the left button over the brightest
gadolinium-enhanced area.
-
Enter gd for the
process name in the Filter Dialog.
-
Press
Select Process in the Filter Dialog.
Note
Repeat Steps 7, 8, and 9 for the brightest abnormal
area, in the flair image, that is not gadolinium-enhanced. Use flair
for the process name.
-
Press
Fltr Img in the Filter Dialog.
-
Press
Clear Processes in the
Filter Dialog.
-
Use the
left button in the Browser Window to display an eigenimage.
-
Select
Slice in the Region of Interest/Volume of Interest
Dialog.
-
Enter 0.8 for
Lo and enter 1.2 for Hi.
-
Press
Clear to remove any region of interest.
-
Press the left button anywhere on the image.
If the region does not cover the anatomy that
corresponds to the eigenimage, repeat steps b and c but use a
larger value for Hi. You can use up to 2.0 for the value of Hi.
-
Clean up the region using
Erase Irregular,
Ero4 , and/or
Era4. The region should not have stray points and should contain only
the anatomy that corresponds to the eigenimage, e.g. white matter,
gray matter, etc.
-
Select the
process name, in the Filter Dialog, that
corresponds to this eigenimage, if not already displayed.
-
Press
Select Process in the Filter Dialog.
Note
Repeat Steps 12-16 for each eigenimage.
-
Press
Fltr Img in the Filter Dialog.
-
Press
Repeat in the Filter Dialog. Select the range of
locations determined above and Press OK in the Repeat Dialog.
-
Load the eigenimages for the
gadolinium-enhanced process,
gd, into Browser 1, load the eigenimages for the flair process,
flair, into Browser 2. Load the orthogonal images, ortho,
into Browser 3, if orthogonal images were processed, otherwise ignore steps
involving Browser 3.
-
Link Browsers 1, 2, and 3.
-
Use the
left button in the Browser Window to display the central location.
-
Select
Irregular in the Region of Interest/Volume of
Interest Dialog.
-
Use the left button draw a large region on the flair image
which includes only normal brain anatomy. Do not include any tumor volume.
-
Press
Fili in the Region of Interest/Volume of Interest Dialog.
-
Select
Roi/Voi processing.
-
Select
All Images (2D).
-
Select
Largest Peak below
Thr Peak.
Note
See example Roi/Voi dialog to
verify all settings for Thr Peak.
-
Press
Thr Peak.
-
Link only Browsers 1 and 2.
-
If there is an
orthogonal images set, select
Edit->Clear->Volume on the
display containing
the orthogonal images, which should be Browser 3.
-
Select
Morphological processing.
-
Enter 15 for
Size and press
Era4.
-
Select
Erase Irregular in the Region of Interest/Volume of
Interest Dialog.
-
Use the left button on the display image to trace around any unwanted
regions, then press the right button in the region you desire to remove
regions of interest from. Ensure the drawn irregular region is closed upon
itself, otherwise the entire region of interest will be erased.
-
Any small regions
that remain may be edited using
Era 4,
Erase Irregular, or any other
region editing operation.
-
Repeat Step 14
for all locations by using the
left button in the Browser Window
to select each location, starting from the central location and proceeding
outward.
Note
In locations where there is no gadolinium enhancement, the entire
region of interest must be removed.
Delete the browser
containing the orthogonal images before continuing with Substance.
-
Select
Analyze->Post-Processing->Substance->Browsers
from the Main Window.
-
Press
Substance to create the combination images.
|
