building of statistical models

38
Guido Gerig UNC, September 2002 1 Building of statistical models Guido Gerig Guido Gerig Department of Computer Science, UNC, Department of Computer Science, UNC, Chapel Hill Chapel Hill

Upload: nissim-pugh

Post on 31-Dec-2015

33 views

Category:

Documents


0 download

DESCRIPTION

Building of statistical models. Guido Gerig Department of Computer Science, UNC, Chapel Hill. Statistical Shape Models. Drive deformable model segmentation statistical geometric model statistical image boundary model - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Building of  statistical models

Guido Gerig UNC, September 2002 1

Building of statistical models

Building of statistical models

Guido GerigGuido Gerig

Department of Computer Science, UNC, Department of Computer Science, UNC, Chapel HillChapel Hill

Page 2: Building of  statistical models

Guido Gerig UNC, September 2002 2

Statistical Shape ModelsStatistical Shape Models

• Drive deformable Drive deformable model segmentationmodel segmentation• statistical geometric model

• statistical image boundary model

• Analysis of shape Analysis of shape deformation deformation (evolution, (evolution, development, development, degeneration, disease)degeneration, disease)

Page 3: Building of  statistical models

Guido Gerig UNC, September 2002 3

Manual Image SegmentationManual Image Segmentation

IRIS segmentation tool: Segmentation of hippocampus/amygdala from 3D MRI data.

•Manual segmentation in all three orthogonal slice orientations.

• Instant 3D display of segmented structures.

•Cursor interaction between 2D/ 3D.

•Painting and cutting in 3D display.

•Open standard s (C++, openGL, Fltk, VTK).

Page 4: Building of  statistical models

Guido Gerig UNC, September 2002 4

SNAP: Segmentation by level set evolutionSNAP: Segmentation by level set evolutionSNAP (prototype): (prototype):

• 3D level-set evolution

• Preprocessing pipeline and manual editing

• Boundary-driven and region-competition snakes

Page 5: Building of  statistical models

Guido Gerig UNC, September 2002 5

Segmentation by level set evolution (midag.cs.unc.edu)

Segmentation by level set evolution (midag.cs.unc.edu)

Page 6: Building of  statistical models

Guido Gerig UNC, September 2002 6

Extraction of anatomical models: SNAP Tool: 3D Geodesic SnakeSegmentation by 3D level set evolution:• region-competition & boundary driven snake• manual interaction for initialization and postprocessing (IRIS)

free dowload: midag.cs.unc.edu

Page 7: Building of  statistical models

Guido Gerig UNC, September 2002 7

Modeling of Caudate ShapeModeling of Caudate Shape

M-rep

PDM

Surface Parametrization

Page 8: Building of  statistical models

Guido Gerig UNC, September 2002 8

Parametrized 3D surface modelsParametrized 3D surface models

Raw 3D voxel model

Ch. Brechbuehler, G. Gerig and O. Kuebler, Parametrization of closed surfaces for 3-D shape description, CVIU, Vol. 61, No. 2, pp. 154-170, March 1995

A. Kelemen, G. Székely, and G. Gerig, Three-dimensional Model-based Segmentation, IEEE TMI, 18(10):828-839, Oct. 1999

Smoothed object Parametrized surface

Page 9: Building of  statistical models

Guido Gerig UNC, September 2002 9

Surface ParametrizationSurface Parametrization

Mapping single faces to spherical quadrilaterals

Latitude and longitude from diffusion

Page 10: Building of  statistical models

Guido Gerig UNC, September 2002 10

Initial ParametrizationInitial Parametrization

a) Spherical parameter space with surface net, b) cylindrical projection, c) object with coordinate grid.

Problem: Distortion / Inhomogeneous distribution

Page 11: Building of  statistical models

Guido Gerig UNC, September 2002 11

Parametrization after OptimizationParametrization after Optimization

a) Spherical parameter space with surface net, b) cylindrical projection, c) object with coordinate grid.

After optimization: Equal parameter area of elementary surface facets, reduced distortion.

Page 12: Building of  statistical models

Guido Gerig UNC, September 2002 12

Optimization: Nonlinear / ConstraintsOptimization: Nonlinear / Constraints

Page 13: Building of  statistical models

Guido Gerig UNC, September 2002 14

Shape Representation by Spherical Harmonics (SPHARM)

Shape Representation by Spherical Harmonics (SPHARM)

),(

),(

),(

),(

z

y

x

r

),(),(0

K

k

k

km

mk

mk Ycr

mzk

myk

mxk

mk

c

c

c

c

Page 14: Building of  statistical models

Guido Gerig UNC, September 2002 16

Reconstruction from coefficientsReconstruction from coefficients

Global shape description by expansion into spherical harmonics: Reconstruction of the partial spherical harmonic series, using coefficients up to degree 1 (a), to degree 3 (b) and 7 (c).

Page 15: Building of  statistical models

Guido Gerig UNC, September 2002 17

Importance of uniform parametrizationImportance of uniform parametrization

Page 16: Building of  statistical models

Guido Gerig UNC, September 2002 19

Parametrization with spherical harmonicsParametrization with spherical harmonics

1

3

7

12

Page 17: Building of  statistical models

Guido Gerig UNC, September 2002 20

Correspondence through Normalization Correspondence through Normalization

Normalization using first order ellipsoid:

• Spatial alignment to major axes

• Rotation of parameter space.

Page 18: Building of  statistical models

Guido Gerig UNC, September 2002 22

3D Natural Shape Variability: Left Hippocampus of 90 Subjects

3D Natural Shape Variability: Left Hippocampus of 90 Subjects

Page 19: Building of  statistical models

Guido Gerig UNC, September 2002 23

Computing the statistical model: PCAComputing the statistical model: PCA

Page 20: Building of  statistical models

Guido Gerig UNC, September 2002 26

Major Eigenmodes of Deformation by PCAMajor Eigenmodes of Deformation by PCA

PCA of parametric shapes PCA of parametric shapes Average Shape, Major Average Shape, Major EigenmodesEigenmodes

Major Eigenmodes of Major Eigenmodes of Deformation define shape Deformation define shape space space expected variability. expected variability.

Page 21: Building of  statistical models

Guido Gerig UNC, September 2002 27

3D Eigenmodes of Deformation3D Eigenmodes of Deformation

Page 22: Building of  statistical models

Guido Gerig UNC, September 2002 28

Set of Statistical Anatomical ModelsSet of Statistical Anatomical Models

Page 23: Building of  statistical models

Guido Gerig UNC, September 2002 32

Correspondence through parameter space rotation

Normalization using first order ellipsoid:

•Rotation of parameter space to align major axis

•Spatial alignment to major axes

Parameters rotated to first order ellipsoids

Page 24: Building of  statistical models

Guido Gerig UNC, September 2002 33

Correspondence ctd.Correspondence ctd.

Rhodri Davies and Chris Rhodri Davies and Chris TaylorTaylor

• MDL criterion applied to shape population

• Refinement of correspondence to yield minimal description

• 83 left and right hippocampal surfaces

• Initial correspondence via SPHARM normalization

• IEEE TMI August 2002

Page 25: Building of  statistical models

Guido Gerig UNC, September 2002 34

Correspondence ctd.Correspondence ctd.

Homologous points before (blue) and after MDL refinement (red).

MSE of reconstructed vs. original shapes using n Eigenmodes (leave one out). SPHARM vs. MDL correspondence.

Page 26: Building of  statistical models

Guido Gerig UNC, September 2002 35

Model BuildingModel Building

Medial Medial representation representation for shape for shape populationpopulation

Styner, Gerig et al. , MMBIA’00 / IPMI 2001 / MICCAI 2001 / CVPR 2001/ MEDIA 2002 / IJCV 2003 /

VSkelTool

Page 27: Building of  statistical models

Guido Gerig UNC, September 2002 36

VSkelToolPhD Martin StynerVSkelToolPhD Martin Styner

Surface

PDM

Voronoi

Voronoi+M-rep

M-rep

M-rep

M-rep+Radii

Implied Bdr

Caudate

Population models:

•PDM

•M-rep

Page 28: Building of  statistical models

Guido Gerig UNC, September 2002 37

II: Medial Models for Shape AnalysisII: Medial Models for Shape Analysis

Medial Medial representation representation for shape for shape populationpopulation

Styner and Gerig, MMBIA’00 / IPMI 2001 / MICCAI 2001 / CVPR 2001/ ICPR 2002

Page 29: Building of  statistical models

Guido Gerig UNC, September 2002 38

Common model generationCommon model generation

Training populationTraining population

CommonCommon modelmodel

Study population

...

Two Shape Analyses - New insights, findings

Model building

Boundary: SPHARMMedial: m-rep

Page 30: Building of  statistical models

Guido Gerig UNC, September 2002 40

1. Shape space from training population1. Shape space from training population

• Variability from training populationVariability from training population• Major PCA deformations define Major PCA deformations define

shape space covering 95%shape space covering 95%• Variability is smoothed Variability is smoothed • Sample objects from shape spaceSample objects from shape space

1.1.

2.2.

3.3.

Page 31: Building of  statistical models

Guido Gerig UNC, September 2002 41

2. Common medial branching topology2. Common medial branching topology

a. Compute a. Compute individual medial individual medial branching branching topologies in topologies in shape spaceshape space

b. Combine medial b. Combine medial branching branching topologies into topologies into one common one common branching branching topologytopology

Page 32: Building of  statistical models

Guido Gerig UNC, September 2002 42

2a. Single branching topology2a. Single branching topology

Fine sampling of Fine sampling of boundaryboundary

Compute inner Compute inner Voronoi diagramVoronoi diagram

Group vertices into Group vertices into medial sheets medial sheets (Naef)(Naef)

Remove Remove unimportant unimportant medial sheets medial sheets (Pruning)(Pruning)

98% vol. overlap98% vol. overlap

Page 33: Building of  statistical models

Guido Gerig UNC, September 2002 44

2b. Common branching topology 2b. Common branching topology

Define common frame for spatial Define common frame for spatial comparisoncomparison

TPS-warp objects into common TPS-warp objects into common frame using boundary frame using boundary correspondencecorrespondence

Spatial match of sheets, paired Spatial match of sheets, paired Mahalanobis distanceMahalanobis distance

No structural (graph) topology No structural (graph) topology matchmatch

Warp topology Warp topology using SPHARM using SPHARM correspondence correspondence

on boundaryon boundary

MatchMatch MatchMatch

Match Match whole whole shape shape spacespace

Initial topology Initial topology (average case)(average case)

For all For all objects in objects in shape shape spacespace

Final topologyFinal topology

Page 34: Building of  statistical models

Guido Gerig UNC, September 2002 45

3. Optimal grid sampling of medial sheets3. Optimal grid sampling of medial sheets

Appropriate sampling Appropriate sampling for modelfor model

How to sample a How to sample a sheet ?sheet ?

Compute minimal grid Compute minimal grid parameters for parameters for sampling given sampling given predefined predefined approximation error approximation error in shape spacein shape space

Page 35: Building of  statistical models

Guido Gerig UNC, September 2002 46

3a. Sampling of medial sheet3a. Sampling of medial sheet

Smoothing of Smoothing of sheet edgesheet edge

Determine medial Determine medial axis of sheetaxis of sheet

Sample axisSample axisFind grid edgeFind grid edgeInterpolate restInterpolate rest m-rep fit to object m-rep fit to object

(Joshi)(Joshi)

1

2

Page 36: Building of  statistical models

Guido Gerig UNC, September 2002 47

3b. Minimal sampling of medial sheet3b. Minimal sampling of medial sheet

2x62x6 3x63x6 3x73x7 3x123x12 4x124x12

• Find minimal sampling given a predefined approximation errorFind minimal sampling given a predefined approximation error

norm. MAD error vs sampling

0.14

0.08

0.0530.048

0.075

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

2x6 3x6 3x7 3x12 4x12

MA

D /

AV

G(r

ad

ius)

Page 37: Building of  statistical models

Guido Gerig UNC, September 2002 48

Medial models of subcortical structuresMedial models of subcortical structures

Shapes with common m-rep model and implied boundaries Shapes with common m-rep model and implied boundaries of putamen, hippocampus, and lateral ventricles. of putamen, hippocampus, and lateral ventricles.

Each structure has a single-sheet branching topology. Each structure has a single-sheet branching topology.

Medial representations calculated automatically.Medial representations calculated automatically.

Page 38: Building of  statistical models

Guido Gerig UNC, September 2002 56

Medial models of subcortical structuresMedial models of subcortical structures

Shapes with common topology: M-rep and implied boundaries of putamen, hippocampus, and lateral ventricles.

Medial representations calculated automatically (goodness of fit criterion).