Image Analysis

Magnetic resonance imaging of the head provides high resolution images of soft tissue structures within the brain. These images provide a range of important information that can be used in functional imaging studies. We are particularly interesting in using anatomical MR information in reconstruction and interpretation of images from PET and EEG/MEG data. In our work on EEG/MEG we need to extract boundaries of the brain, skull and scalp for use in accurate modeling of the mapping between neural current sources and the measured MEG and EEG data. In developing anatomically constrained EEG/MEG inverse methods, we also need to extract a high resolution representation of the cortical surface to which sources are confined. Similarly in our PET work, we can use labeled tissue maps, generated from our MR data, to guide the formation of the PET images by incorporating these maps into statistical image models that are used in a Bayesian image reconstruction approach. To provide this information we are currently developing methods for the automated extraction of brain and other surfaces from MR data, parametric representations of the cortical surface, and automated 3D partial-volume tissue labeling of the extracted brain. Software for automated extraction of the brain volume from 3D T1-weigted MR scans is currently available through the software we developed, called BrainSuite.

Example of a brain volume automatically extracted from an MR image using our Brain Surface Extraction (BSE) software. This software extracts high and low resolution cortical surfaces from the data. The BSE software is available now. Click on the "BSE software" button for details. (Click image for larger 22k jpeg version of this image)

Skull / Scalp Extraction

We developed a new technique for segmentation of skull and scalp in human T1-weighted magnetic resonance (MR) images that generates realistic models of the head for EEG and MEG source modeling. Our method performs skull segmentation using a sequence of mathematical morphological operations. Prior to the segmentation of skull, we segment the scalp and the brain from the MR image. The scalp mask allows us to quickly exclude background voxels with intensities similar to those of the skull, while the brain mask obtained from our Brain Surface Extractor algorithm ensures that the brain does not intersect our skull segmentation. We find the inner and the outer skull boundaries using thresholding and morphological closing and opening operations. We then mask the results with the scalp and brain volumes to ensure closed and nonintersecting skull boundaries. We applied our scalp and skull segmentation algorithm to several MR images and validated our method using coregistered CT-MR image data sets. We observe that our method is capable of producing scalp and skull segmentations suitable for MEG and EEG source modeling in 3D T1-weighted human MR images. pdf

The Skull/Scalp Extraction is now a part of BrainSuite.

Figure: Segmentation of brain, skull and scalp from MRI and its corresponding CT on transaxial, sagittal and coronal slices respectively. First Row: Original MR Image. Second Row: Original CT data. Third Row: Segmentation of brain, skull and scalp from MRI. Fourth Row: Segmentation of brain, skull and scalp from CT.

Figure: Surface tesselations of the scalp, outer skull, inner skull, and brain from a T1 weighted MR Image. Click image to enlarge.

Cortical Surface Parameterization

Cortical surface parameterization has several applications in visualization and analysis of the brain surface. We developed a scheme for parameterizing the surface of the cerebral cortex. The parameterization is formulated as the minimization of an energy functional in the pth norm. The energy is called p-energy and the resulting minimizations are p-harmonic maps. Brain surfaces from multiple subjects are brought into common parameter space using the scheme. 3D spatial averages of the cortical surfaces are generated by using the correspondences induced by common parameter space. We observe that as the value of p increases, the resulting maps are more stable and that results in  better correspondence between brains so that the spatial average reveals a more realistic representation of the common cortical features. poster & pdf

Figure: Extracted cortical surfaces, their p-harmonic maps & their spatial averages. Click image to enlarge.

Figure: P Harmonic maps of cortical surfaces. They are obtained by minimizing the p-energy.

USC Biomedical Imaging Research Lab © 2004, designed by Dimitrios Pantazis