> I was wondering if you could give me some ideas on how to accomplish the
> task that I have here.
> What we need is a ratio GM:WM:CSF from a specified voxel within a brain
> from a high resolution( about 1x1x1 mm) structural scan of a single
> subject. Do I have to normalize the brain prior to segmentation?
You can segment the images in native space. The prior probability images will
still be overlayed, but only using an affine transform, rather than any
nonlinear warping. An affine transform is only a parameterisation of
rotations, translations, zooms and shears, and does not include the
distortions (warps, deformations) that would be estimated by nonlinear
registration.
> I
> understand that I need to use prior images for segmentation, but, correct
> me if I am wrong, those prior images are just averages of lots of segmented
> brains?
They are averages of lots of segmented images that have been masked and
smoothed by 8mm.
% _______________________________________________________________________
% /APRIORI/ DIRECTORY
% Images in this directory represent the a priori probabilities of the
% voxels in a spatially normalized (9-parameter affine) brain image
% belonging to a particular tissue type.
% _______________________________________________________________________
%
% apriori/gray.mnc, apriori/white.mnc & apriori/csf.mnc
% ---------------- ----------------- ---------------
% Images supplied by Alan Evans, MNI, Canada (ICBM, NIH P-20
% project, Principal Investigator John Mazziotta). Original 1mm
% resolution images were icbm_avg_151_gm.mnc, icbm_avg_151_wm.mnc
% and icbm_avg_151_csf.mnc. Images were masked using
% average_305_mask_1mm.mnc, reduced to 2mm resolution and smoothed
% using an 8mm FWHM Gaussian. These images represent the
% probabilities of finding gray matter, white matter or cerebro-
% spinal fluid at any point. These volumes are used largely for
% image segmentation, although they can be used as templates for
% spatial normalization. 151 subjects were used to create each
% volume.
%
% apriori/brainmask
% -----------------
% Image derived from average_305_mask_1mm.mnc, which was originally
% supplied by Alan Evans, MNI, Canada (ICBM, NIH P-20 project,
% Principal Investigator John Mazziotta). The original image
% contained ones and zeros, where ones represented voxels that were
% part of the brain. It was subsequently smoothed using an 8mm
% FWHM Gaussian. This volume can be used to weight the spatial
% normalization so that the final solution is not influenced by
% voxels outside the brain.
%
> And without prior images it will just take longer to segment the brain?
The probability of a voxel being grey matter is determined from the
combination of a load of probabilities, similar to:
p(GM | position, intensity) = p(intensity | GM) * P(GM | position) /
((p(intensity | GM) * P(GM | position)
+ (p(intensity | WM) * P(WM | position)
+ (p(intensity | CSF) * P(CSF | position)
+ (p(intensity | other) * P(other | position) )
This roughly translates into the probability of a voxel being grey matter
given its intensity and position within the brain, is the probability of
observing that intensity from grey matter (according to the estimated
intensity distribution of GM) times the probability of getting grey matter at
that position (according to the prior probability maps). This is then
normalised so that (for that voxel) the sum of probabilities over all tissue
types is one.
> If I do need to normalize the brain before segmentation, where can I get
> the normalization matrix so that I could adjust the voxel position, size
> and shape also? Is it on the normalization output page under linear
> (affine) component?
> What about values contained in _sn.mat file?
> Thank you beforehand for your input!
The _sn.mat contains a bunch of parameters that tell SPM how to warp a brain
into a standard space. You can use the normalise button, selecting "write
normalised only". This will ask you for an _sn.mat file for each subject,
followed by the images to write normalised.
The contents of the sn.mat file are an affine transform matrix, and some
parameters that describe a nonlinear warp. The affine transform is
represented on the normalization output page, but the nonlinear deformations
are usually represented by too many parameters (about 1000) to be shown.
Best regards,
-John
|