structure
DESCRIPTION
Structure. Harvard. 1. Diffusion-based Registration. MIT. Utah. 2.Group Effect Maps. 3. Automatic Segmentation. 1. Shape and Atlas Based Segmentation. 1. DTI Processing. 2. Statistical Shape Analysis. 2. Surface Processing. 3, DTI Connectivity Analysis. 3. PDE Implementations. - PowerPoint PPT PresentationTRANSCRIPT
National Alliance for Medical Image Computing http://na-mic.org
Structure
National Alliance for Medical Image Computing http://na-mic.org
Core 1: OverviewHarvard
Georgia TechUNC
UtahMIT
Segmentation
Registration
Foundational Methods
Structural Features and Statistics
Connective Features and Statistics
1. Shape and Atlas Based Segmentation
2. Statistical Shape Analysis
3, DTI Connectivity Analysis
1. Diffusion-based Registration
2.Group Effect Maps
3. Automatic Segmentation1. DTI Processing
2. Surface Processing
3. PDE Implementations
1. Combined Statistical/PDE Methods1. Quantitative DTI Analysis
2. Cross-Sectional Shape Analysis2. Stochastic Flow Models
Figure 3: a) A rendering of a cortical surface, extracted from MRI, shows a degree of noise that significantlyaffects success ive processing. b) A feature-preserving, PDE-based filter smooths away small-scale noisewhile preserving sharp features such as the concave regions of the sulci.
(a) (b)
National Alliance for Medical Image Computing http://na-mic.org
Core 1: Overview
• Computational tools for image analysis– Extract anatomical structures at many
scales – Measure properties of extracted
structures– Determine connectivity between
extracted structures– Relate disease factors to
measurements
National Alliance for Medical Image Computing http://na-mic.org
Shape Analysis
• Developing pipeline protocols for population comparisons, jointly with UNC.
• Integrating discriminative analysis into the pipeline:– Shape-based classification
National Alliance for Medical Image Computing http://na-mic.org
EM Segmentation with Non-Stationary Tissue Priors
• Integrating into Slicer
M-Step
E-Step
Bias:Predict Error
Image
Correct Intensities
MF-Step:Regularize Weights
Estimate TissueProbability
Label Map
National Alliance for Medical Image Computing http://na-mic.org
Already in NAMIC Software
• Shape prior for segmentation– Leventon 2001– Added to ITK by others
• DTI visualization– O’Donnell (CSAIL), LMI
(BWH)– In VTK-based 3D Slicer
National Alliance for Medical Image Computing http://na-mic.org
Future work (6 months)
• Complete shape-based segmentation implementation– Insert into toolkit
• Shape based comparison and population analysis– Structural components– Tract components
National Alliance for Medical Image Computing http://na-mic.org
Q-Ball Imaging in Slicer
Estepar, Snyder, Kindlmann, Westin
National Alliance for Medical Image Computing http://na-mic.org
Automatic Thalamus Segmentation
LGNMGN
VLMD
VA
VLMD
VA
CM
Pu Pu
CMVL
VA
MD
Ziyan, Tuch
National Alliance for Medical Image Computing http://na-mic.org
1. Make QBALL availableCheck QBALL code into Slicer and VTK.
2. Does nonlinear registration boost stats?Measure power benefit of ITK nonlinear registration for FA group comparisons.
3. Are group comparisons based on the full tensor more sensitive?Implement and measure sensitivity of tensor-based group comparison method.
Future Work
National Alliance for Medical Image Computing http://na-mic.org
Utah Core 1 Activities
• Differential Geometry for DTI analysis
• Descriptive statistics of DTI• Hypothesis testing DTI• Interpolation and filtering of DTI
National Alliance for Medical Image Computing http://na-mic.org
Curved Tensor Geometry
• Natural geometry for tensor analysis
• Enforces positive eigenvalues
• Basis for statistics, interpolation, and processing
Space of 2x2 tensors: a bb c
National Alliance for Medical Image Computing http://na-mic.org
Descriptive Statistics
• Averages and Modes of Variation• Preserves natural properties
–Positive eigenvalues–Tensor Orientation–Tensor Size (determinant)
• Prototype implemented in ITK
National Alliance for Medical Image Computing http://na-mic.org
Hypothesis Testing
• Tests differences in diffusion tensors from two groups
• Uses full six-dimensional information from tensors
• Prototype implemented in ITK• Upcoming IPMI submission
National Alliance for Medical Image Computing http://na-mic.org
Interpolation and Filtering
• Interpolation of tensors–Based on weighted averages in
curved geometry• Filtering
–Anisotropic filtering based on curved geometry
• Implementation in progress
National Alliance for Medical Image Computing http://na-mic.org
Statistics Processing
Software• Different tensor geometries
can be defined• Each package can swap
in/out different geometries
TensorGeometry
LinearGeometry CurvedGeometry
Other?
DescriptiveStatsTensorGeometry
HypothesisTestsTensorGeometry
InterpolationTensorGeometry
FilteringTensorGeometry
National Alliance for Medical Image Computing http://na-mic.org
Future Work (6 months)• Further develop tensor statistics—
make publicly available• Build prototypes of tensor filtering
and interpolation• Continue research into DTI
hypothesis testing–Methods–Exploratory Experiments
National Alliance for Medical Image Computing http://na-mic.org
UNC: Quantitative DTI Analysis
Guido Gerig, Isabelle CorougeStudents: Casey Goodlett and Clement Vachet
National Alliance for Medical Image Computing http://na-mic.org
Conventional Analysis: ROI or voxel-based group tests after alignment
Patient
Control
Quantitative DTI Analysis
UNC NA-MIC Approach:
• Quantitative Analysis of Fiber Tracts
• DTI Tensor Statistics across/along fiber bundles
• Statistics of tensors
Tracking/clustering
selection
FA FA along tract
National Alliance for Medical Image Computing http://na-mic.org
Example: Fiber-tract Measurements
Corouge, Isabelle, Gouttard, Sylvain and Gerig, Guido, "Towards a Shape Model of White Matter Fiber Bundles using Diffusion Tensor MRI" , Proc. IEEE Computer Society, Int. Symp. on Biomedical Imaging, to appear April 2004
Gerig, Guido, Gouttard, Sylvain and Corouge, Isabelle, "Analysis of Brain White Matter via Fiber Tract Modeling", full paper IEEE Engineering in Medicine and Biology Society EMBS 2004, Sept. 2004
uncinate fasciculus
uncinate fasciculus FA along uncinate
cingulum FA along cingulate
Major fiber tracts
National Alliance for Medical Image Computing http://na-mic.org
Processing Steps
• Tractography– Data structure for sets of
attributed streamlines• Clustering• Parameterization• Diffusion properties
across/along bundles• Graph/Text Output• Statistical Analysis
Slicer (?) ITK DTI Fiber Spatial
Object data structure (J. Jomier)
Normalized Cuts (ITK) B-splines (ITK) NEW: DTI stats in
nonlinear space (UTAH) Display/Files Biostatistics / ev. DTI
hypothesis testing (UTAH)
National Alliance for Medical Image Computing http://na-mic.org
ResultsFiberViewer Prototype
(ITK)• Clustering (various metrics)• Parametrization• FA/ADC/Eigen-value
Statistics• Uses SpatialObjects and
SpatialObject-Viewer• Used in two UNC clinical
studies (neonates, autism)• Validation: ISMRM’05
National Alliance for Medical Image Computing http://na-mic.org
Next 6 months• Methodology Development:
– DTI tensor statistics: close collab. with UTAH– Deliver ITK tools for clustering/parameterization to Core 2– Feasibility tests with tractography from Slicer– Deliver prototype platform to Core 2 to discuss integration
into Slicer
• Clinical Study: DTI data from Core 3– Check feasibility of tract-based analysis w.r.t. DTI resolution
(isotropic voxels(?)), SNR– Apply procedure to measure properties of:
• Cingulate (replicate ROI findings)• Uncinate fasciculus (replicate ROI findings)• Other tracts of interest
National Alliance for Medical Image Computing http://na-mic.org
UNC: Statistical Shape Analysis
Martin StynerStudents: Ipek Oguz and Christine Shun Xu
National Alliance for Medical Image Computing http://na-mic.org
Shape Analysis Pipeline• Clinical need: Localization of shape and volume
changes• 3D objects of spherical topology• Input: Segmentation from models or binary images• Modeling Steps:
– Individual surface models• Regularization• Correspondence
– Alignment via Procrustes & choice of scale– Skeletal description
• Structural subdivision• Statistical analysis of models
National Alliance for Medical Image Computing http://na-mic.org
Shape Analysis Pipeline• Thickness maps
– Distance to skeleton• Local shape analysis
– To template or template-free– Univariate Euclidean distance– Multivariate Hotelling T2 distance– Raw p values, t/T2-maps, effect-size– Conservative correction for Type II error
• MIT discriminative analysis complements our shape analysis well
• Visualizations of steps for QC
National Alliance for Medical Image Computing http://na-mic.org
Next 6 months• NAMIC toolkit development
– Standardization of IO & internal representation• With MIT & Georgia Tech
– Standardization of visualization tools– Automation of tools, transfer to standard
• Methodology development– Non-Euclidean shape metrics with permutation
tests– Probabilistic structural subdivision method– 3D visualization maps of statistical metrics
• Clinical: Shape analysis data from Core 3– Feasibility of shape analysis on data from Core 3– Caudate shape analysis on Brockton VA/Harvard
data
National Alliance for Medical Image Computing http://na-mic.org
Georgia Tech
Ramsey Al-Hakim Steven HakerDelphine NainEric PichonAllen Tannenbaum
National Alliance for Medical Image Computing http://na-mic.org
Anisotropic active contours
• Add directionality
National Alliance for Medical Image Computing http://na-mic.org
Curve minimization
• Calculus of variations– Start with initial curve– Deform to minimize energy– Steady state is locally optimum
• Dynamic programming– Choose seed point s– For any point t, determine globally
optimal curve t s
Registration,Atlas-basedsegmentation
Segmentation
National Alliance for Medical Image Computing http://na-mic.org
Synthetic example (3D)
National Alliance for Medical Image Computing http://na-mic.org
L2 Bases Functions Local Shape Analysis
Our goal is to build more localized shape priors that can handle surfaces with high frequencies (high curvatures) and learn the local variations from the training set.
We propose to compare different L2 bases. In particular, we would like to investigate the use of multiscale shape analysis and learn localized shape statistics from the data using bayesian statistics.
The applications are shape prior for segmentation, registration and classification.
National Alliance for Medical Image Computing http://na-mic.org
Example: Some Local Variations
Finding local variations in Prostate data at different frequency levels and spatial locations
Low Frequency
High Frequency
National Alliance for Medical Image Computing http://na-mic.org
Segmentation of Area 46 Using Fallon’s Rules-I
Ramsey Al-HakimBME Undergraduate
National Alliance for Medical Image Computing http://na-mic.org
Segmentation of Area 46 Using Fallon’s Rules-II
Ramsey Al-HakimBME Undergraduate
National Alliance for Medical Image Computing http://na-mic.org
Work in Next Six Months
• Choice of anisotropic conformal factors for DTI-tractography.
• Comparison of L2 bases for shape analysis (application to caudate).
• Making Fallon’s rules more automatic for segmentation.
National Alliance for Medical Image Computing http://na-mic.org
Structure