patch_segmentation_pipeline.pl

RASCAL patch-based label fusion segmentation pipeline

patch_segmentation_pipeline.pl <input_t1.mnc> <output_prefix> --model-dir <dir> [options]

DESCRIPTION

patch_segmentation_pipeline.pl runs the RASCAL (Rapid Automatic Segmentation of the human Cerebral cortex using an Atlas Library) patch-based label fusion segmentation pipeline as described by Weier et al. (2014). The pipeline registers an input T1-weighted MRI volume to a set of pre-labelled atlas templates, then uses non-local patch-based label fusion to propagate anatomical labels from the atlas library to the subject.

The pipeline supports multiple registration backends including minctracc and ANTs. Preprocessing steps such as non-uniformity correction (N3/N4) and intensity normalization (NUYL) can be enabled. Graph-cut optimization (GCO) can be applied for spatial regularization of the segmentation. The model directory, which contains the atlas library, must be specified.

OPTIONS

--verbose
Print progress information during processing.
--qc
Generate quality control images.
--clobber
Overwrite existing output files.
--cleanup
Remove temporary files after processing.
--model-dir <dir>
Path to the model directory containing the atlas library. This option is required.
--subject <id>
Specify a subject identifier string for output naming.
--search <n>
Search radius in voxels for the patch-based label fusion. Default: 2.
--patch <n>
Patch radius in voxels for the patch-based label fusion. Default: 1.
--exclude <id>
Exclude a specific atlas subject from the library (e.g. for leave-one-out validation).
--classes <n>
Number of label classes in the atlas library. Default: 36.
--variant <name>
Name of the label variant to use from the model directory. Default: labels.
--threshold <f>
Confidence threshold for label assignment.
--compare <file>
Provide a manual segmentation for comparison and overlap statistics.
--short
Store output labels in short integer format.
--preprocess
Run preprocessing steps (intensity normalization, brain masking) before segmentation.
--minctracc
Use minctracc for nonlinear registration to the atlas templates.
--ants
Use ANTs for nonlinear registration to the atlas templates.
--build
Build the atlas library from the model directory.
--symmetric
Use symmetric registration.
--resample_labels_order <n>
Interpolation order for resampling label volumes. Default: 1.
--preselect
Preselect atlas subjects most similar to the input before label fusion.
--pairwise
Use pairwise label fusion strategy.
--refine
Refine the segmentation with additional iterations.
--qc_lut <file>
Specify a colour lookup table for quality control visualisation.
--nuc
Apply N3 non-uniformity correction during preprocessing.
--nuc3T
Apply non-uniformity correction optimised for 3T MRI scanners.
--no_nuyl
Disable NUYL intensity normalization.
--no_gco
Disable graph-cut optimization for label regularization.
--no_nl
Disable nonlinear registration.
--baa
Use BEaST-based brain mask for the analysis.
--list <file>
Specify the atlas sample list file within the model directory. Default: samples.lst.
--nomask
Do not apply a brain mask during processing.
--patch_normalize
Normalize patch intensities before comparison.

EXAMPLES

Run segmentation with the required model directory:

patch_segmentation_pipeline.pl subject_t1.mnc output_prefix --model-dir /path/to/atlas

Run with ANTs registration, QC output, and cleanup:

patch_segmentation_pipeline.pl subject_t1.mnc output_prefix \
    --model-dir /path/to/atlas --ants --qc --cleanup --verbose

Run with custom patch parameters and compare to manual labels:

patch_segmentation_pipeline.pl subject_t1.mnc output_prefix \
    --model-dir /path/to/atlas --search 3 --patch 2 \
    --compare manual_labels.mnc --clobber

AUTHOR

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

COPYRIGHTS

Copyright (C) Vladimir S. Fonov and McGill University. Licensed under the terms of the GNU General Public License version 3 or later.

SEE ALSO

itk_patch_morphology(1), hcag_segmentation_pipeline.pl(1)