itk_patch_segmentation

perform patch-based label fusion segmentation

itk_patch_segmentation <input> [output_labels] [options]

DESCRIPTION

itk_patch_segmentation performs patch-based regression segmentation using ITK-based image I/O. It supports multiple label fusion strategies including Support Vector Machine (SVM), Non-Negative Least Squares (NNLS), and exponential weighting for combining multiple atlas segmentations into a consensus segmentation of the input volume.

The tool compares local image patches between the input volume and a set of training volumes (atlases with known labels) to determine the most likely label for each voxel. The patch-based approach provides robust segmentation that is less sensitive to registration errors than voxel-wise label fusion.

This is the ITK-based version with support for multiple MINC and NIfTI formats. For the legacy MINC1-only implementation, see volume_patches.

OPTIONS

--train <file>

Specify a training library file listing atlas volumes and their labels.

--train2 <file>

Specify a second training library for two-level segmentation.

--nnls

Use Non-Negative Least Squares for label fusion weighting.

--svm

Use Support Vector Machine for label fusion.

--exp

Use exponential weighting for label fusion.

--verbose

Print detailed progress information.

--quiet

Suppress all output messages.

--clobber

Overwrite existing output files.

--mask <file>

Specify a mask to restrict segmentation to a region of interest.

--patch <n>

Set the patch radius in voxels (default: 1).

--search <n>

Set the search radius in voxels for finding matching patches (default: 1).

--cls <n>

Number of classes for segmentation.

--beta <value>

Regularization parameter controlling smoothness (default: 0.125).

--scaling

Enable intensity scaling between input and training volumes.

--threshold <value>

Set a threshold for patch similarity.

--discrete <n>

Number of discrete labels (default: 2).

--confidence

Output a confidence map alongside the segmentation.

--adist

Use absolute distance weighting.

--grading

Perform grading instead of segmentation.

--iter <n>

Maximum number of iterations (default: 50).

--extract

Extract labels from training library.

--top <n>

Use only the top N most similar training samples.

--float

Store output as 32-bit floating point.

--short

Store output as 16-bit signed integer.

--byte

Store output as 8-bit unsigned integer.

--box

Use box-shaped patches.

--ball

Use ball-shaped patches.

--prelabel <file>

Specify a pre-labeling volume for initialization.

--prob

Output probability maps for each label.

--groups <n>

Number of groups for cross-validation.

EXAMPLES

Segment using exponential weighting with a training library:

itk_patch_segmentation input.mnc output_labels.mnc --train library.txt --exp --mask brain_mask.mnc

Segment using SVM with larger patch and search radii:

itk_patch_segmentation input.mnc labels.mnc --train library.txt --svm --patch 2 --search 2

Segment with confidence output:

itk_patch_segmentation input.mnc labels.mnc --train library.txt --exp --confidence --beta 0.25

AUTHOR

Vladimir S. Fonov - McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University.

COPYRIGHTS

Copyright © 2012 by Vladimir S. Fonov

SEE ALSO

itk_patch_grading itk_patch_morphology