Pablo-Subcort-Structure
From wiki
Jump to navigationJump to search
All source code:
Media:segmentation_src.zip contains the matlab source code and the shell scripts Code required to do shape and image statistics training
svn+ssh://your-cs-login@cvs.cs.unc.edu/cvs/proj/midag/svn/matlab/commonQuantileCode svn+ssh://your-cs-login@cvs.cs.unc.edu/cvs/proj/midag/svn/matlab/quantileStats svn+ssh://your-cs-login@cvs.cs.unc.edu/cvs/proj/midag/svn/matlab/shapeStat
e.g. on linux: svn co svn+ssh://styner@cvs.cs.unc.edu/cvs/proj/midag/svn/matlab/commonQuantileCode n
Installed under /tools/Pablo/matlab: add this folder to matlab : Set Path... -> add with Subfolders
Empty starting directory with config files:
Media:newSeg.zip contains the config files to start the segmentation
The Pablo installation is located in /tools/Pablo/linux64
copy "newSeg" folder Media:newSeg.zip This is a folder which contains all the files required to start the segmentation process. It has the configuration files for Pablo in the 'config' folder, a 'param.txt' file which serves as the configuration for the shape statistics computed in Matlab, and a blank 'binary_raw3' folder for holding the original binary segmentations that m-reps must be fit to. In general the param.txt file should not need to be edited.
put binary segmentations in newSeg\binary_raw3
- compress_raw3.script can be used to take raw3.gz files, gunzip them, and
recompress them using Pablo
cd "newSeg"
from matlab call:
binary_fits('binary_raw3', 'template_model.m3d')
arguments:
1) path to directory with binary segmentations
2) initial m3d model
This file does two stages of m-rep fitting. In the first stage, the template
model specified as the second argument to the script is used as the initial
model. The initial model is then deformed without any use of PGA shape
statistics to fit each of the binary segmentations. The results of this
stage are output to 'stage_1\output_m3d'.
After the first stage is complete, all of the files in 'stage_1\output_m3d' are
used to create a mean m-rep with PGA statistics. This is done in Matlab
by calling the 'calculatePGA' m-file. The results are stored in 'stage_1\pga'.
The file containing the mean m-rep and PGA modes is called SymMean.m3d.
For the second stage, the SymMean.m3d file from the shape statistics is used
as the initial model to the binary fitting. This stage uses the statistics
which generally produces a better fit than the first stage. The results of
the second stage are stored in 'stage_2\output_m3d'. These represent the
final m-reps fit to the binary images.
The numbers, such as the surface distance match, output by Pablo for each
binary fit are stored in log files, either 'stage_1/log' or 'stage_2/log'.
In case this is used for simple shape analysis of the m-rep properties of already existing
segmentations, it is advised to run also both stages. It is expected that the second stage
regularizes the atom correspondence.
training_X_regions('..\gray_raw3',50,'trained_local_50');
arguments:
1) path to directory with grayscale images
2) number of training samples to use (chooses randomly)
3) path to output directory (script will create it)
Can be either training_global_regions or training_local_regions for image statistics
This file computes both image and shape statistics for a training set which
will be used as input to a segmentation. It first creates a directory called
'training' inside the directory given in the third agrument, and randomly selects
a given number (second argument) of m3d files from
'binary_fitting\stage_2\output_m3d' and copies them to 'training\m3d'. With
these, it computes a mean with PGA modes and stores the result in 'training\pga'.
It also computes image statistics based on the randomly-selected training set and
stores them in 'training/hist_stats'. The image statistics are first computed
for individual files using Pablo, then the Matlab command 'doSPETraining'
aggregates them and stores the main resulting file as
'hist_stats\trained\trainedSPE-from-XX.spe', where XX is the number of training
samples.
segment('train_local_50', '..\gray_raw3\2368-004-01_10_T1_crop.raw3','2368-004-01');
arguments
1) path to training directory
2) path to target grayscale image
3) path of output (script will append '.m3d')
This file runs a segmentation. It looks in the given training directory for
the mean/PGA m3d file in 'training\pga\SymMean.m3d' and uses it as the input
model to the segementation. It also looks for the training image statistics
in '\training\hist_stats\trained\trainedSPE-from-XX.spe' . Finally, it uses the
grayscale image given as the second argument as input for Pablo to run the
segmentation.
For comparing output segmentations to original segmentations:
export_segs_to_byu('gray_raw3','hippoL/seg_train_50');
Arguments: 1) directory with grayscale images
2) directory with m-rep segmentations (.m3d) to convert
This script takes all the .m3d files in the second argument and converts them to .byu.
It needs an image file to export them in the correct world coordinates, so the first
argument points to a directory containing grayscale images. It assumes all the
grayscale images are of the same dimension and voxel size, so it always uses the
first file in that directory to get the world coordinates. It uses the executable
m3dToByu.exe (Windows only), which uses Pablo code to export an m3d to byu. This part
can easily be done manually, but a scriptable version is in development.
byu2gipl.script This file converts all the .byu files in the current working directory to both .iv and .gipl. It uses MeshConvert to go from .byu to .iv. It then uses surf2vol to go from .iv to .gipl. The call to surf2vol uses hardcoded values of 162x162x188 for the dimensions and 0.75x0.75x0.75 for voxel size.
volume_overlap('/Autism/autism/longitudinal/all_structures/mrep/raw3','*caudL_pp_crop.gipl.gz',
'seg_local_50','dice_local_50.txt');
Arguments: 1) directory containing original .gipl segmentations
2) string to pick which original segmentations to use
3) directory containing new .iv segemntations
4) output file name
This function outputs the volume overlap from the original segmentation surfaces
to the new segmentations. It assumes both the original segmentations and new
segmentations are in .gipl format (the script byu2gipl.script will take .byu files
of m-rep segmentations and convert them to .gipl). It then calls ImageStat and
concatenates the result of "grep DiceCoef" on ImageStat's output to the user-specified
output file.
surface_dist('/Autism/autism/longitudinal/all_structures/mrep/raw3','*caudL_pp_crop.gipl.gz',
'seg_local_50','sd_local_50.txt');
Arguments: 1) directory containing original .gipl segmentations
2) string to pick which original segmentations to use
3) directory containing new .iv segemntations
4) output file name
This function outputs the surface distances from the original segmentation surfaces
to the new segmentations. It assumes the original segmentations are in .gipl format and
converts them to .iv. It assumes the new segmentations are already in .iv (the script
byu2gipl.script will leave both .iv and .gipl versions of the new segmentations). It
then calls MeshValmet and concatenates the tail of MeshValmet's output file to the
user-specified output file.
To make plots of the output, use plot_surface_dist.
plot_volume_overlap('dice_local_50.txt');
Arguments: 1) input file (should be a file output by volume_overlap.m)
This function just plots all the volume overlap coefficients computed by volume_overlap.m
in ascending order. Use the Matlab command 'hold all' between calls of plot_volume_overlap
to plot multiple lines on the same graph.
plot_surface_dist('sd_local_50.txt');
Arguments: 1) input file (should be a file output by surface_dist.m)
This function plots all the surface distance numbers computed by surface_dist.m
in descending order. It takes the 9th column from the output of surface_dist.m, which
is the mean absolute distance. It also multiplies the numbers by 0.75 to convert
from voxels to millimeters since each voxel is 0.75 mm. Use the Matlab command 'hold all'
between calls of plot_surface_dist to plot multiple lines on the same graph.
