Algorithm:GATech
Contents
Ongoing Projects
White Matter Tractography
Introduction
We want to extract the white matter tracts from Diffusion Weighted MR data. The idea is to use directional information in a new anisotropic energy functional based on Finsler geometry.
Use Case
I'd like to segment neural fibers.
Link to Project Page
Researchers
- Georgia Tech: Eric Pichon, John Melonakos, Xavier Le Faucheur, Allen Tannenbaum
- Harvard: C-F Westin
Conformal Flattening
Introduction
We want to develop new flattening methods for better visualizing neural activity from fMRI brain imagery. Our technique is based on conformal mappings which map the cortical surface onto a sphere in an angle preserving manner.
Use Case
I'd like to flatten a structure, such as the brain, for visualization.
Link to Project Page
Researchers
- Georgia Tech: Shawn Lankton, Allen Tannenbaum
- Harvard: Steven Haker, Ron Kikinis
ITK Bayesian Classifier Image Filter
Introduction
This ITK filter is a segmentation algorithm that utilizes Bayes's Rule along with an affine-invariant anisotropic smoothing filter.
Use Case
I'd like to segment a volume or sub-volume into 'N' classes in a very general manner. I will provide the data and the number of classes that I expect and the algorithm will output a labelmap with 'N' classes.
Link to Project Page
Programming Week 1: Bayesian Classifier Image Filter
Researchers
John Melonakos
Luis Ibanez, Karthik Krishnan (core 2 collaborators - Kitware)
Stochastic Methods for Segmentation
Introduction
To develop new stochastic methods for implementing curvature driven flows for various medical tasks such as segmentation. This will be used as an alternative to level set methods and has certain advantages including the ability to explicitly take into account noise models.
Use Case
General image segmentation.
Link to Project Page
Researchers
Delphine Nain, Samuel Dambreville, Tony Yezzi, Gozde Unal, Allen Tannenbaum
Rule Based Segmentation Slicer Modules
Introduction
This Slicer module implements our semi-automatic segmenation algorithms for various brain structures. These algorithms are based on expert neuroanatomist (core 3) rules. Our programs drastically reduce the time it takes to segment various brain structures. We are currently working on the following brain areas: DLPFC, DPFC, Nucleus accumbens, Putamen.
Use Case
I'd like to rapidly segment and visualize a brain area by clicking on a convenient gui.
Links to Project Page
Programming Week 1: DLPFC Slicer Module
Programming Week 2: Rule Based Slicer Module
Researchers
Delphine Nain - Slicer Leader
John Melonakos - Bayesian Classification, Sulci Extraction
Ramsey Al-Hakim - DLPFC, Striatum
Tauseef Ur Rehman - DPFC
Shawn Lankton - Putamen
Jim Fallon (core 3 collaborator - UCI)
Martha Shenton (core 3 collaborator - Harvard)
Multiscale Shape Analysis
Project Description
We are investigating the use of spherical wavelet basis functions to represent shapes and learn a multiscale shape probability prior from a population of shapes. The goal is to build more localized shape priors that can handle surfaces with high frequencies (high curvatures), such as the caudate nucleus. The applications are shape prior for segmentation, registration and classification.
Links to Project Page
3D Shape Analysis Using Spherical Wavelets
Researchers
- Delphine Nain (GT, Core 1)
- Allen Tannenbaum (GT, Core 1)
- Steven Haker (BWH)
- Aaron Bobick (GT)
Relationship to other NA-MIC partners
shape analysis pipeline (Martin Styner, UNC, Polina Golland, MIT).
Brain Surface Registration
Project Description
Registering brain models to one another is currently done in a variety of ways. We are researching a method to perform this operation automatically and elegantly by solving PDE’s which produce one to one maps from one surface to another. Deep sulci of the brain will be used as landmarks during this registration operation.
Completed Projects
The Fast Marching algorithm has been integrated into the Slicer.
As described in:
A statistically based flow for image segmentation Eric Pichon, Allen Tannenbaum, and Ron Kikinis. Medical Image Analysis, 8(3):267-274, September 2004. [1]
the algorithm is versatile, fast, relatively simple to implement, and semi-automatic. It is based on minimizing a global energy defined from a learned non-parametric estimation of the statistics of the region to be segmented. Implementation details are discussed in the aformentioned publication. Also A new unified set of validation metrics is proposed that is used to validate the algorithm both on artificial and real MRI images. The algorithm performs well on large brain structures both in terms of accuracy and robustness to noise.
A user-oriented tutorial for the Fast Marching algorithm is available at: [2]
--User:Eric@ece.gatech.edu 15:33, 6 Dec 2004 (EST)
ImageSmooth Module
ImageSmooth module performs 2D and 3D smoothing of images. It works on the principle of κ(1 / 3),κ(1 / 4) smoothing of the level lines of an image. κ(1 / 3) performs smoothing for each of the slices in the 2D plane while κ(1 / 4) performs volumetric smoothing.
AffineSegment -- 3D segmentation using Affine Invariant Surface Flow
3D segmentation using Affine Invariant Surface Flow.
To segment a volume :
- Define a label for the segmented data : by clicking on the 'Label' button.
- Define some seed points : by creating some fiducials inside (not on the border of) the region of interest. Fiducials can be created by moving the pointer to the desired region and pressing the 'p' key. See the Fiducial module documentation for more on using fiducials.
- Choose the value of Inflationary term. If you dont know what to choose, just leave the default value
- Choose the initial Size of the starting sphere. You might want to start with a reasonable size of the sphere so that you are not outside the surface to start with nor is the starting sphere very very small (this will lead to making a lot of iterations to expand to reach the boundary) If you are not satisfied with the region covered by the initial sphere, press 'Reset' and you can start all over again
- Start expansion of the surface : by clicking on the 'Expand' button. The volume of the surface will be expanded by the value right of the expand button. Increase this value to segment a bigger object. (Typically, 100 iterations is good number to start with, if the target region is not very big. If the expansion did not go far enough, press 'Expand' again. Continue untill you have all of the region covered. Dont bother about leaks. Once you have finished with expansion, now press 'AffineContract'. This will smooth out the surface and will contract where required.
- Typically, 5-10 iterations are enough for this part. --User:Yogesh.rathi@bme.gatech.edu 16:34, 9 Dec 2004 (EST)
