SPM_Notes.txt

Notes on SPM 8
As directed by the wonderful Chris Magnano

Realign
-------
Estimate does all of the math but it doesn't do anything
Reslice takes the maths and then actually applies tit to the image
It's useful to have that file saved because then you can realign a lot of images
 -- say you have 50 images that you do the math on, you can reasonably assume
 -- that the 51 will not significantly affect the math, so you can merely apply
 -- the calculations that you have previously done
There are SPM tutorials on YouTube by Andrew Jahn, he also has a blog with posts
on SPM that we can look up and are a good intro (random grad student)
Select images from:
       -- Actual ADNI images can be found acmi/fmri/AD_T1,CN_T1,MCI_T1 which are
       -- the three catagories for the three different types of patients
       -- There are also nifti stripped images but there are poorly segmented
       -- images mixed in with the lot, so it might not be fully raliable to
       -- just parse straight through -- basically all the ones with the stupid
       -- prefixes are the ones that we want to avoid
       -- IBSR are the good images
Just doing check registration allows you to look at a bunch of images at once
sometimes the images in SPM are a little messed up since the metadata is a lit
tle bit wrong, but when you load the matrices into MATLAB they turn out just 
fine

IBSR - public database for manually segmented brain images for training set
     - if we loose them, we can easily just go back into google and kindly ask
       for them again

If you click on the top in the batch editor, it explains what is actually 
happening

The graphs that pop up are just telling you how well internally the things just
went

Results stored in the same directory you got the images in with your custom
prefixes

Keeps the first image constant and then changes the other images to match

Everything else in the spatial preprocessing works similar to realign

You can actually do everything at once by adding in different modules into the
batch editor, by clicking dependency in the  other modules, you can pipe the
results of one module into another

Realign just puts the iamges back into the space of the first image whereas
coregister intentionally forces all of the other images to be in the space and
orientation of the first image

Coregister
----------
You have a reference image that everything else is trying to fit into
The source image is what we fit to the reference (must be an MRI), but the other
images are ignored until they need to be giggled to fit the reference, and these
do not have to be MRI images (different algo)

Save the function which we can then edit the MATLAB script since the GUI doens't
really let us do exactly what we want, we then copied the functions of the job
file (automatically saved along with the script) and put that into a function
inside of the main file, and then you have to replace one of the inputs with the
parameter that you are passing into your new function

An actual example of how to do this, we should looke at
 -- chris13/scripts/batchattempt1
 - Before this, there was already a SPM segmentation of all the patient files
 - which saves nice mat files, including the segsn.mat and the inverted one
 - which does nice spatial normalization, and invert allows us to get the image
 - back into its native space
 - Those normalizations allow us to put the ATLAS file back into the MRI space
 - of the original image

Some ATLASes saved in the acmi/fmri/altAtlas folder, one for each patient
The ATLAS also grayscales the voxels, which are on a scale of 1-116 for each
anatomical region in the brain. This then allows people to identify brain
segments so that people can identify parts of the brain in the MRI space by
comparing it to the ATLAS space

What Chris did:
     Tissue segmentation with CRF then anatomical segmentation with SPM
     Looked at moving the ATLAS into the MRI space instead of moving the MRI
     into the ATLAS space

~600 ADNI images
18 IBSR images

When you do a SPM seg
 c1 white matter
 c2 gray matter
 c3 cerebral spinal fluid

When you do the SPM seg, it does all the realign and such included, but when
we're doing things manually, then we'll have to do the realign stuff (plus for
the IBSR files, which have much worse meta data, we have to do the realign
manually beforehand)

Display is better for one image, but it can only do one image

c1 iamges give you a probability map of where the white matter would be
 - If you want, you can manually go into MATLAB and view the three images
(c1,c2,c3) into the same images so you can see it all side by side

scripts/evalseg allows us to look at a segmented image and evaluate it compared
to a ground truth image

White mattle lesions in data/DHS/data in nifti images and white matter regions
 - these are flair images, but we can ask badya about getting white matter
 - lesion MRI scans to work with

NITRC.org is where we can get all of our brain scan files, and they might have
	  anatomical segmentation training files, but this was more basic than
	  the atlas

White matter lesion stuff is interesting (especially since the binary 
      segmentation leans toward ignoring all lesions)
So is using CRFs for doing anatomical segmentation






Code Notes

Pertinant files are /acmi/CRFGM_fastTune.m and /acmi/CRFGM_test.m and condor
scripts
Most other files are legacy

/scratch/cmagnan1 directory
  Consists of temp files created while running CRF script


Script Arguments

  CRFGM_fastTune:
    General method is leave-one-out cross-validation
    fold (int 1-5 for IBST): which of 5 folds for tuning to use
    minIter (int) and maxIter(int): number of iterations to run for tuning
    	    	      		    loop, which determines optimal number
				    of iterations for leave-out image via
				    5-fold

				    Typical values: minIter: 200
				    	    	    maxIter: 1000
    leaveOut(int): the image which will eventually be used for testing

     res(int): only use the res-th voxels (as in every other, every third)
     	       	    	    	   	  
					  Typical Value: 16 or 2 or 1




Comparing UGM and SPM8

	  look at eval_seg.m