Automated BrainSuite Cortical Surface Extraction Pipeline

Overview - BrainSuite Cortical Surface Extraction

The workflow takes a skull stripped volume as input and runs a multi-stage process, extracting the cortical surface and splitting it into hemispheres.

Problem addressed by this workflow

Generating a shape model of the cortical surface from a 3D volume.

Detailed Workflow Usage & Specifications

Tissue Classification

The next step in the Brainsuite09 cortical surface identification sequence is tissue classification. This is performed by the Partial Volume Classifier (PVC). This process assigns an integer tissue label to each voxel in the image. These labels correspond to the type of tissue that is estimated to be in that voxel. PVC accounts for background voxels, cerebrospinal fluid, grey matter, and white matter. It also labels voxels that are composed of combinations of these voxels. The labels used are listed in Table PVC1. PVC can generate labelings with the six tissue classes or with a simpler 3 class tissue output.

The PVC algorithm uses the partial volume measurement model that was used for the bias field correction. However, it removes the spatially varying gain component since that is assumed to have been corrected by BFC. It does, however, include a spatial structural prior. This prior models the brain as a set of piecewise continuous regions of single tissue types, bounded by partial volume combinations. The prior encourages neighborhoods of voxels to be similar. The influence of this prior is controlled by the Spatial Prior.

Important: PVC should only be applied to skull-stripped brain images.

Note that the models used in PVC assume that only voxels containing CSF, GM, and WM remain in the image. Thus, you should always skull-strip your images before applying PVC.

Cerebrum Labeling

Once the brain volume has been extracted and classified, we need to separate it into cerebellum, cerebrum, and other structures so that we can identify the cerebral cortex. To achieve this, we apply a volumetric registration (AIR, Woods et al., 1998) to align a labeled atlas to the brain volume.

The Cerebrum labeler assumes that your brain volume is oriented into LPI (-x points left / -y points posterior/-z points inferior) coordinates (see front matter for a description of the coordinate system).

Cortex mask identification

Once the cerebrum, cerebellum, and brainstem have been labeled, we can extract the cerebral cortex using the cortical modeling tool. The default parameters, as shown, will select the white/grey boundary of the cerebrum. Additionally, you can control the tissue fraction used in the boundary decision. For example, using the 5% default, Brainsuite will select the boundary corresponding to at least 5% white matter in each voxel. This is combined with information from the atlas to produce a cerebral mask.

Mask Scrubbing Tool

The mask scrubbing tool removes small surface pits and bumps by examining voxel connectivity. With each iteration, it computes how many foreground voxels each foreground voxel has as a neighbor. If the threshold is not met, the foreground voxel becomes a background voxel. A similar process is applied to the background voxels.

Topology Correction

To be able to map the surface of your extracted cortical volume using 2D coordinates, the outer boundary must have spherical topology. The Topological Correction Algorithm (TCA) will make corrections to your volume to ensure this property. You can specify the minimum and maximum correction sizes to use. BrainSuite will begin by making its corrections using the minimum correction size as its upper limit, then iterate until it reaches the maximum correction speciied. This allows you to limit the size of individual change that has to be made to the volume to fix its topology.

Wisp Removal

The Wisp Removal tool removes thin, wispy structures that are likely to represent segmentation errors in the extraction process. It works by decomposing the binary mask into a graph, and separating weakly connected components. You can adjust the threshold used to break the wisps and the maximum number of iterations applied. This process should only take a few seconds.

Surface generation

The surface model generation consists of 3 steps.

Generation. This step produces an initial mesh surface from the cortex model produced in Stage 6. Split. This step will separate the initial cortical surface into left and right hemispheres, based on the labeling produced during the cerebrum labeling procedure. Parameterize. This step will induce a 2D coordinate system on each hemisphere by mapping them to the unit square.

URL: http://www.loni.ucla.edu/twiki/bin/view/CCB/PipelineWorkflows_BrainSuiteCorticalSurfaceExtraction