NormalizeSkullStripped - MRC CBU Imaging Wiki
location: NormalizeSkullStripped

Normalizing with skull stripping

This is a short howto on normalizing brains using SPM and skull stripping. In our hands, this gives considerably better results than the standard method in SPM99 - which involve skull/scalp masking, but no skull stripping.

Setting up

You will need the binary for the skull stripping utility bet by Steve Smith; this is included for linux or Windows in the relevant MriCro distributions, or is part of the downloadable FMRIB software library. The "bet" or "bet.exe" binary file should go on your executable path. Next download the stripped_norm archive. In it you will find a matlab script "bet.m"; this needs to go into a directory on your matlab path. You will also find an analyze image "sbrain_avg152T1.hdr/img"; the .hdr and .img file should be copied into the SPM templates directory. For the record, I made this by: running BET on the ICBM 152 template image, 1mm resolution, with parameters "-f 0.7 -g -0.05", resampling the output to 2mm voxels, applying a mask hand-edited in MRIcro (included in the archive as brain_avg152T1.roi), and smoothing by 8mm.

Skull stripping

We start by skull stripping the structural image. To make this less error prone, and to reduce processing time, you may want to trim the structural image down to a bounding box which contains the brain and skull - for example down to the bottom of the brain stem. Our structural images by default contain a large amount of dead space in the Z direction - above the brain, and down into the neck. One way to trim the images is using the clipping functions in MriCro. If your image came with a .mat file (for example after conversion of a Bruker image using pvconv ), you may need to modify or discard the .mat file, as the image has now changed dimension, and the .mat file will no longer apply. Another option is to use my trim_img function to trim the images.Next, adjust the orientation of the image using the SPM display utility, so that it more or less matches the template, with the anterior commissure at 0,0,0. The structural should now have .mat file, if it did not before.

To skull strip, run the BET skull stripping binary using the SPM/matlab "bet.m" interface in this little archive. Run by typing "bet" at the matlab prompt. The advantage is that the SPM script deals with .mat files in a smooth way, saves the image with a predictable name in the same directory, and you have a GUI to select the images. For a first pass, accept the "fractional intensity threshold" default of 0.5 - see the BET page for details on this parameter.

You now have a skull stripped version of the structural, with an "ss_" appended to the file name, and a .mat file that is identical to the original image. Check the skull stripped image to make sure that stripping worked OK; if not, consider removing more of the space above the head and below the brain stem and repeating the procedure. You could also try changing the fractional intensity threshold - towards 1 if there is skull that has not been removed, towards 0 if too much is being stripped.

Coregistration to EPI

Next, coregister to the EPIs. Use the SPM Mutual information coregistration (SPM Defaults button - Coregistration - Use Mutual Information Registration). Coregister the skull stripped image to (for example) the mean EPI image, but without reslicing. Thus: Coregister - Coregister only, then choose the mean EPI as target, the skull stripped image as source image and the not-skull-stripped structural as Other Images. I suggest you don't reslice at this stage, because it will just create an ugly low resolution version of the structural which is of little use. It is usually easier to coregister the structural to the EPIs like this than the other way round (Structural as target), because, if the structural is the target, you will have to select all the EPIs as Other images, which is rather time consuming.


Using the SPM defaults turn OFF template brain masking (SPM Defaults button - Normalization - Defaults for parameter estimatation, then accept all defaults up to "Mask brain when Resistering" - set this to "No Brain Mask"). The image to determine the parameters is the stripped structural, the template image should be the sbrain_avg152T1.img included in this archive. Other images to reslice - for now, just the unstripped structural. Check that all worked well in the output windows. If it did, click on Normalize, select Write normalized, select the parameter set you just made ("ss_myimage_sn3d.mat" or similar), and select all the EPI images you want to reslice. You want to do the reslicing in two steps like this, because SPM masks out all the voxels in the normalized images, that were not available for any one of the images; if you reslice the structural at the same time as the smaller-field-of-view EPIs, you will get a structural with strange areas sliced out of it.Matthew Brett 17/10/02

CbuImaging: NormalizeSkullStripped (last edited 2013-03-07 21:24:03 by localhost)