Segmentation

Setting up the files

• All 16 slices of raw structural data should be in the segment directory: eg E2097S7I1-16.MR
• Run matlab 6 or later:
% matlab -nojvm
• Initialize and Crop the Volume in matlab

>> HC_InitVol
Enter the base name (i.e. E1000S3I): E2097S7I.MR
What is the image size: [512 512] (enter)
Do you want to crop the images [Y]: Y
Choose a crop rectangle by clicking at the top left and bottom right. Only crop once! (If you need to do it again, exit matlab and restart the procedure from the beginning.)
Enter the number of images: 16
Do you want to flip the images along the z-axis? [N] (hit enter for no ... unless they were acquired posterior = slice 1, anterior = slice 2. It is ALWAYS better to scan anterior as the 1st slice)
Enter output file name: [vAnatomy.dat]vAnatFSE512Crop.dat
 
• It just saved the file. Type exit to leave matlab.

• Make a backup copy of the crop information (in conversionDataKS.mat)
% cd crop
% cp conversionDataKS.mat conversionDataKS.FSE.mat
% cd ..
• Go back to matlab and create the interpolation files
% matlab
>> [v, vSize]=readVolume('vAnatFSE512Crop.dat');
>> [v2,vSize2]=HC_Interp(v,vSize,3,'linear');
>> writeVolume('vAnatFSE512Crop46.dat',v2,vSize2);
>> [v2,vSize2]=HC_Interp(v,vSize,7,'linear');
>> writeVolume('vAnatFSE512Crop106.dat',v2,vSize2);
>> exit

Doing the segmentation

The segmentation has two steps: the initial segmentation, and the interpolated segmentation. Interpolation is necessary to have near isotropic voxels for more accurate topology in the unfolding.

Segmentation can either be done only on a PC. Interpolating the segmentation requires shuffling files back to unix, but this definitely saves time in the long run.

1. The initial segmentation

• Export all of your vAnat.dat files to a PC (ftp using binary).
  
• Run the program mrGray
  
• Create a new project and name it after your subject.
• Select vAnatFSE512Crop.dat
  
• Figure out which slices to segment. If you scanned slices 3-13, then you will segment slices 2-12 (because mrGray starts its numbering at 0).
• Segment the white matter 1st (colored in beige), using a pixel size of 3 pixels minimum
• Select white matter with the checkbox at the right
• Another checkbox indicates the highlight size (use 3)
• The left mouse button highlights a pixel, and the right mouse button deselects pixels
• Then, segment the CSF (colored in yellow), imposing a thin boundary on the segmentation. (You want it thin because you will expand it after the interpolation step).
• Save the segmentation (File -> Classification -> Save as). Save early, save often. (I use filename class10 for the 1st version, then go to class11 so there is a backup).
• Here is an example of an initial segmentation:

• After you've finished and saved it, quit out of mrGray.

2. The interpolated segmentation

A) Choosing interpolation parameters

Your voxels will likely be anisotropic from the segmentation. For my scans, voxels were about .4 x .4 x 3.0 mm. The segmentation does not like this, so what I do is upsample along the z-axis. I used to upsample by a factor of  7, making the voxel size exactly .3902 x .3902 x .4285. Then I would segment this interpolated segmentation. This is a ton of work. A shortcut I developed is to interpolate the segmentation by a factor of 3, smoothen it out, and then do another round of interpolation by 7/3 to reach a final interpolation factor of 7. I found this last interpolation step rarely requires manual attention.

If you have the same imaging tools as I do, you can easily do the same procedure. If not, you will have to choose your own interpolation scheme, but will have to keep track of all the places where pixel size is specified. (1 - the unfolding step, 2- in HFM, enter the appropriate initial image sizes, 3 - when running mrAlign, enter the appropriate numbers).



B) Using matlab to do the interpolation:

If your filename is class11.dat, then ftp the file back to your UNIX computer, and do the following:
 
• Run matlab
% matlab -nojvm
• Execute the following commands to do nearest neighbor interpolation on the segmentation by a factor of 3
[v, vSize]=HC_readSeg('class11');
[v2,vSize2]=HC_Interp(v,vSize,3,'nearest');
HC_writeSeg('class20',v2,vSize2);
exit

C) Manually interpolate the segmentation with mrGray (This takes about 8 hours, and is best done in two 4-hour sessions).

• Ftp the above classification back to the PC
  
• Open mrGray
• Create a new project again
• This time, select vAnatFSE512Crop46.dat as the volume anatomy file (46 slices).
• Hit the menu button File -> Classification -> Load, and select class20.
• Hit Overlays for gray, white, and CSF.
• Define the VOI properly (the VOI determines where gray matter will be grown)
• Select VOI crossbars
• The crossbars can be moved with the right mouse button, and report coordinates (R, G, B) in the upper right hand corner of the window.
• Then use the right mouse button, and the axial view, to move the VOI and make sure all of the proper slices are selected. If you scanned slices 3-13, and before segmented slices 2-12, now you will want to include only slices 6-36 (subtract 1 and multiply by 3). Make sure the VOI is set to this.
• Interpolate the segmentation so that it is smooth. Be sure to use a pixel size of 3 (not smaller, or you're just wasting time attending to variance)
• Do the white matter first:
• Leave slices 6, 9, 12, 15, 18, 21, 24, 27, 30, 33, and 36 alone because you defined these already. Make the transition between the interevening slices smoother.
• You will have to conceptualize how the white matter is distributed in 3D. Don't trust the interpolated volume, linear interpolation cannot substitute human conceptualization. That is, visualize how the structure looks in 3D. Sometimes white matter will be discontinuous inplane but continuous throughplane.
• If a 3D volume is available, use it to guide you as to how the white matter is arranged.
• THE ABSOLUTE MOST IMPORTANT THING IS TO MAKE SURE THE WHITE MATTER IS DEFINED TOPOLOGICALLY CORRECTLY! What does this mean? That there are no handles, i.e. direct connections between things that should be continuous only along the cortical manifold (eg opposite sides of gyri.)
• Save the classification (again, early and often, eg class21).
• For the CSF:
• Make it smoother from slice to slice.
• Extend the CSF to form a complete boundary to the segmentation. Make it VERY thick on the outside so when gray matter is grown, it does not grow outside the segmentation through slices.
Here is an example of interpolating between 3 slices (6, 7, and 8), with the adjusted CSF:
18 (6)  
19      
20      
21 (7)
22      
23      
24 (8)
• Check and make sure there are no handles by checking the topology Classification -> Topology -> Analysis. If there are, try to correct them.
• Grow the gray matter. Select Classification -> Gray Layers
• Set to 18 if your pixel size is 0.4 mm (set to less if your pixel size is greater, such that # of layers * pixel size <= 7.2)
  
• Also set gray contention settings on (this will make there be no holes in the gray matter).
• Finally, hit Classification -> Grow Gray Matter
  
• Check through EACH slice; make sure that gray matter has not leaked out past the boundaries or is not connected. Check this by making only the gray matter visible. It should all be green and continuous. If it is not, then select Classification -> Delete -> All Gray, and add some CSF or make adjustments where needed.
• After you've fixed all errors, save the classification. Don't bother saving the gray matter, there is one more computational step to do.

D) Final interpolation

The interpolation actually has to be done one more time, but manual labor is rarely needed as long as you did a careful and topologically correct job in step b above.
• Ftp that latest classification back to the UNIX computer. Run matlab to do the interpolation
% matlab
[v, vSize]=HC_readSeg('class21');
[v2,vSize2]=HC_Interp(v,vSize,7/3,'nearest');
HC_writeSeg('class30',v2,vSize2);
exit
• Ftp class30 back to the PC so we can extract and save the gray matter w/ mrGray (note now we have 106 slices)
• Again create a new project, this time with vAnatFSE512Crop106.dat as the anatomy volume. Load the classification class30.
  
• Adjust the VOI (subtract 1 and then multiply by 7). If you acquired slices 3-13, segmented slices 2-12, did the interpolated segmentation on slices 6-36, you would now want slices 14-84.
• Turn the gray matter contention settings OFF and set the gray matter layers as above.
• Grow the gray matter (Classification -> Grow Gray Matter)
• Select the gray matter on the left side by ctrl-left clicking (ctrl-right click to de-select)
• Scroll through to make sure it looks good (ie nothing is missing). It is very easy for extraneous gray matter to exist, so be sure to turn off the display for white matter and CSF, and just look at the gray.
• Save it (File -> Save Selected Gray, [filename: grayleft])
• IMPORTANT: Find the coordinates of a gray matter point. Using the crossbars, find a point volumetrically near the middle of the segmentation (ie slice 49), and near the middle of the manifold (ie near the subiculum-parahippocampal border), and adjust the crossbars until it is on a pixel of gray matter. WRITE DOWN THESE COORDINATES, the unfolding algorithm needs 'em.
• Do the exact same procedure for the right side.
• NOTE: If your unfolded maps when all is said and done do not look right on, it is possible you may need to manually edit this high-resolution segmentation to make it smooth. The reason for this is the following: the segmentation itself gets interpolated by the unfolding method, and this interpolation can occasionally miss subtle details of the segmentation. The workaround is to make the segmentation foolproof enough that interpolation will not miss anything: that is, be conservative. Have the white matter be thick so there are no handles (spots where one side of a sulcus is continuous through plane with the other side of the same sulcus), and make sure the transitions from slice to slice are gradual, not bigger than the thickness of the white matter itself. Also have the CSF be thick for the same reason.