PipelineOverview.txt

Research Pipeline Overview
Summer 2014
Teo Gelles & Andrew Gilchrist-Scott
Prof. Ameet Soni

Last Updated: 08/01/2014

Directory: /sonigroup/summer2014/tgelles1/brainseg2014/files/



Our project pipeline works as follows:

1)  We start with a dataset consisting of MRI images in the .nii format,
    located in /sonigroup/fmri/AD_T1, /sonigroup/fmri/CN_T1, and
    /sonigroup/fmri/MCI_T1, consisting of 92, 102, and 203 images
    respectively.  These are all images from the ADNI database.  We
    also have another dataset from the IBSR database located in
    /sonigroup/fmri/IBSR_nifti_stripped/

2)  We use a CRF PGM, largely copies from Chris Magnano's work last summer,
    to learn and create tissue segmentations for the images.  The CRF learns
    on the IBSR dataset, and places the ADNI tissue segmentations in
    /sonigroup/summer2014/ADNI_tissues/

3)  We skullstrip the images, using our script ./files/scripts/skullstrip.py,
    which is a wrapper around the freesurfer software package installed on
    our computers, available at
    https://surfer.nmr.mgh.harvard.edu/fswiki/DownloadAndInstall
    The skullstripped images are placed in
    /scratch/tgelles1/summer2014/ADNI_stripped/

4)  The images come with a substantial black margin around the brains,
    which is detrimental to the rest of our pipeline.  So, we use the
    cropBlack.m script in ./files/SLIC/cropBlack.m to remove said margin
    and place the cropped files in
    /scratch/tgelles1/summer2014/ADNI_cropped/

5)  We perform supervoxelation on the cropped images, using the SLIC algorithm
    which we implemented for the 3D .nii file format.  Classically, the SLIC
    algorithm does not use the exact number of supervoxels specified as input,
    instead it chooses a close value that better fits the size of the relevant
    brain.  However, this is problematic for most statistical analysis, so
    we slightly modified it to use precise numbers of supervoxels for all brains.
    Ultimately, we decided to initially distribute the supervoxels in a uniform
    cube throughout the images.  To modularize the acts of generating the supervoxels
    and getting statistical features from the supervoxels, we had to save multiples
    data types for each brain, namely the supervoxel labelling for each image
    ("labels" or "slic"), a more visual representation of the labelling ("border"),
    information about the supervoxel centers ("centerinfo"), the amount of image
    cropped in step 4 ("cropoffset"), and the original image at the resolution
    chose (almost always full resolution) ("x"). We have many different output
    directories for our supervoxelations depending on the exact calls made, as follows:

    Classical SLIC with varying numbers of inputs and algorithm-chosen number
    of supervoxels:
    /scratch/tgelles1/summer2014/slic/
    Naming Convention: dataType-brainType-numSV-shapeParam-resolution-numIterations.nii
					 
    SLIC with supervoxels arranged in a 5x5x5 grid regardless of image size:
    /scratch/tgelles1/summer2014/slicExact125
    Naming Convention: dataType-brainType-numSV-shapeParam--numIterations.nii
    	   	       (resolution is always 1)

    Same as above, with supervoxels in 4x5x6 grid:
    /scratch/tgelles1/summer2014/slicExact120

    Same as above, with supervoxels in 5x6x7 grid:
    /scratch/tgelles1/summer2014/slicExact210

    Smae as above, with supervoxels in a 6x7x8 grid:
    /scratch/tgelles1/summer2014/slicExact504

6)  Integrated with the supervoxelation process is the generating of statistical
    features, using ./files/SLIC/getSLICFeatures.m.  These features are
    saved within the features/ subdirectory of the supervoxelated images themselves,
    for examples /scratch/tgelles1/summer2014/slicExact125/features/.  They
    are in the csv format.  Features in a format to send to Ameet Soni's
    colleague Sriraam Natarajan are saved as .txt files in a custom format.

7)  The final step in the pipeline is to determine whether the generated
    features correlate with Alzheimer's prediction.  This step remains ongoing,
    using multiple different analysis techniques.  The relevant code is contained
    in either the ./files directory if we developed it ourselves or in
    ./other_software if we used a ready-made package.