Level
Sets for Image Segmentation
Level set
methods for image segmentation rely on an evolving closed surface defined
by a moving interface, the front, which expands from a point out into
the image, fitting itself to the region it is released within. The method
involves solving a set of Partial Differential Equations. We have published
in BIOSIGNALS 2008 a novel method for brain segmentation using
Level Sets. Comparison
of ground truth to the well-known FAST (developed by the team at the University
of Oxford) and our swarm + level set solution is shown below. Our
method has advantages over those based on image voxel classification.
The numerical nature of the level set method (level set equations are
solved using the finite element method) allows for a wide variety
of metrics of the surface to be calculated. The gradient across the interface
can be calculated easily, from which we can calculate curvature or other
geometric data if required.
3D Gabor
Wavelets for Image Registration
We have published
in Journal of Medical Image Analysis (Impact Factor 3.256) a novel
metod for evaluating image registration algorithms based on 3D Gabor wavelets.
A 3D Gabor wavelet is a sinusoidal wave modulated by a 3D Gaussian function
and can be defined as:
Diffusion
Tensor Imaging
Diffusion Tensor Imaging
(DTI) uses a set of DW images to reconstruct, on a voxel by voxel basis,
a tensor that characterizes the three-dimensional diffusion profile. The
eigenvalues of the diffusion tensor provide diffusivities along the orientations
defined by the respective eigenvectors. The direction of highest diffusivity
has been found to coincide with the principal fibre orientation in regions
with coherent fibres. Utilizing these orientations, reconstruction of
some white matter tracts is feasible using a process known as tractography
[4,5]. Despite its potential in clinical application, the single tensor
DTI model cannot describe diffusion profiles that arise from non-coherent
fibres coexisting in the same voxel.
We have published
in JMRI (impact factor 2.67) a two-tensor model to resolve fibre
crossings from conventional DTI datasets (32 diffusion sensitizing directions,
b=1000 s/mm2). This exploits the planar diffusion profile in regions with
fibre crossings [1] and applies a spatial regularisation scheme to reduce
noise artefacts. The regularisation scheme was based on relaxation labeling
and on a discrete set of basis directions. Orientational, anisotropy,
and diffusivity information could be resolved in regions of two fibre
crossings using full brain coverage scans acquired in less than 6 minutes.
We have published
in SPIE Medical Imaging 2008 a novel method for estimating parameters
of a multi-tensor model for DTI. The method uses Bayesian inference, see
image below.
We have submitted
for publication in ISMRM 2008 a paper describing a novel method
for brain fibre tracking. The method uses Bayesian inference and Markov
chain Monte Carlo (MCMC)
We have also
submitted for publication in ISMRM 2008 a paper describing a novel
method for brain fibre tracking.
Evolutionary
Computation for RF Pulse Design
Currently
images from a MR scanner needs to be 'enhanced' or 'corrected' before
they can be effectively used. It would make more sense to improve the
physical MR image reconstruction process in the first place. By incorporating
specific optimisation criteria to the pulse design process, we aim to
develop methods for rapid acquisition of high quality MR images.
RF pulses
are an essential part of MRI since they are the device by which nuclear
spins can be excited out of their lowest energy state, and be detected
to produce MR images. Pulses designed by the current methods suffer from
sensitivity to B1 inhomogeneity. Namely, the values of the magnetic fields
are not the same throughout the sample. The problems with pulse design
become much more significant in ultra-high fields, e.g., 7 T. At high-field,
SAR (specific absorption ration), a measure of the power absorbed in tissue
per unit mass, is limited and this creates some of the most difficult
challenges for imaging. Field inhomogeneities occur because of difference
in bulk magnetic susceptibility (BMS).
We are developing
new pulse design methods combining MR physics, mathematical functions,
and evolutionary computational algorithms. The methods will be tested
through computer simulation as well as on the Philips 7 T scanner housed
by the SPMMRC.
MRI
Classification
|