+ All Categories
Home > Documents > Multivariate Textures/Information Theory in Brain...

Multivariate Textures/Information Theory in Brain...

Date post: 22-Mar-2018
Category:
Upload: tranthien
View: 217 times
Download: 1 times
Share this document with a friend
26
1 Multivariate Textures/Information Theory in Brain Imaging Karl Young University of California, San Francisco Department of Radiology and Biomedical Imaging ADVANCED STATISTICAL CONCEPTS FOR MULTIMODAL MRI: THEORY AND APPLICATIONS Center for Imaging of Neurodegenerative Diseases June 19, 2010
Transcript

1

Multivariate Textures/Information Theory

in Brain Imaging

Karl Young

University of California, San Francisco

Department of Radiology and Biomedical Imaging

ADVANCED STATISTICAL CONCEPTS FOR MULTIMODAL MRI: THEORY AND APPLICATIONS

Center for Imaging of Neurodegenerative Diseases

June 19, 2010

2

The Challenge

Simultaneous availability of large data sets from a large number of medical imaging modalities (i.e. high dimensional feature space)– structural images (structural MRI, DTI,...)

– functional images (PET, fMRI, pMRI,…)

– metabolite images (MRSI)

Combined analysis is complex and hard to interpret

3

Proposed Approach Use texture/complexity analysis methods derived from

computer vision, machine learning, information theory, and nonlinear time series analysis.

Advantages:

– Automated (“objective” in distinction to region based methods)

– Global (in distinction to VBM, TBM)

– Multimodal (multivariate – simultaneous use of co-registered images is straightforward)

– Multiresolution (Generalizes texture measures -Haralick measures,… - and measures from nonlinear dynamics - fractal dimension, multifractal specrum, dynamical entropies, statistical complexity,… - )

4

Proposed Approach Use texture/complexity analysis methods derived from

computer vision, machine learning, information theory, and nonlinear time series analysis.

Disadvantages: – Texture/complexity measures are abstract and though

statistically efficacious can be difficult to causally associate with interpretable clinical outcomes.

– Sensitivity of texture/complexity measures can vary based on quality of preprocessing steps such as registration, interpolation, and segmentation of images

5

Background

Texture/complexity analysis in images is based on analysis of “higher order statistics” (i.e. relies on multiple image values rather than single image values as in VBM):

– Autocorrelation function: is 2nd order statistic

– Third order moment: is 3rd order statistic

6

Background

Related methods from nonlinear time series analysis (and generalized to image analysis) are based on analysis of correlation structure in time series

In both standard texture analysis and nonlinear time series, use of higher order statistics and correlation structure are attempts to account for nonlinear relationships

7

Background – Texture Analysis

Early work (1970’s) in texture analysis via computer vision research was done by Robert Haralick et al, e.g.:

R. M. Haralick, K. Shanmugam, and I. Dinstein, "Textural features for image classification,” IEEE Trans. Syst., Man, Cybern, vol. SMC-3, pp. 610-621, Nov. 1973.

Haralick defined a heuristically based set of texture measures that was first used to provide automatic region classification in remotely sensed (satellite) images

Long delay between original development of methods and application to medical imaging (maybe due to the fact that some of the early research was classified)

A number of recent applications in areas such as tumor detection from images

8

Background – Nonlinear Time Series Analysis

Early work (1970’s) demonstrated utility of multiresolution analysis (e.g. measuring fractal dimension and multifractal spectrum of chaotic systems)

Attempts (1990’s) to reconstruct nonlinear models from chaotic time series for optimal prediction led to rigorous classification of typical nonlinear dynamical systems in terms of information and computational complexity theories

Generalization of the application of these methods from time series to the analysis of higher dimensional data (e.g. images) proved difficult

9

Background – Combined Approach

Though not possible to directly apply rigorous time series methods to images, application of Haralick style heuristic methods led to hybrid approach for sensitive, interpretable diagnostic classification via image complexity measures:

– K. Young, Y. Chen, J. Kornak, G. B. Matson, N. Schuff. Summarizing Complexity in High Dimensions. Physical Review Letters 94:098701:1-4 (2005).

– K. Young, N. Schuff. Measuring Structural Complexity in Brain Images. Neuroimage. 39(4):1721-1730 (2008).

– K. Young, A. Du, J. Kramer, H. Rosen, B. Miller, M. Weiner, N. Schuff. Patterns of Structural Complexity in Alzheimer’s Disease and Frontotemporal Dementia. Human Brain Mapping. 30(5):1667-77 (2009).

10

Fundamental Construct Fundamental to the generation of texture and complexity

measures is the generation of a co-occurrence matrix from an image or set of co-registered images

Co-occurrence matrices can be thought of as joint histograms

As joint histograms the co-occurrence matrices are density estimates of joint probability density functions associated with the correlation structure of the multivariate, co-registered images

Given the density estimates the texture/complexity measures can be generated using information and computational complexity theories

11

“Future” Conditioned on “Past” - P(q2|q1)

“Past” Conditioned on “future” - P(q1|q2)

Co-occurrence matrix <->Joint Histogram P(q2,q1)

q1q2

12

Use of Co-occurrence Matrices Image co-registration - joint entropy or mutual

information of joint intensity histogram is used as optimization measure for co-registration algorithm

Texture analysis – generated for base and offset pixels over image or region and used to estimate texture measures

Nonlinear time series analysis – estimated joint conditional distribution over past and future sequences

Hybrid method - generated for two regions (arbitrary but with fixed size and offset) over image or region of segmented multivariate image and used to estimate texture/complexity measures

13

For image registration:

14

15

16

For standard texture analysis:

d

17

0

0 1

1

0 1

2525 75 75

q1

q2q1

q2

counts

For time series:…000110101…

18

For Generation of Multivariate Texture/Complexity Measures

q1

q2

Question: How does P(q2|q1) vary as function of q1,q2, and shape ?

19

Multivariate Texture/Complexity Analysis Proceeds in 4 Stages

I. Choice of appropriate Feature Space (e.g. structural MRI, DTI, MRSI, or combinations)

II. Segmentation (Clustering) of Feature Space

III. Generation of Texture/Complexity Estimates

IV. Classification Based on Texture/Complexity Estimates (e.g. supervised or unsupervised)

20

Stage I – Choice of Appropriate Feature Space

Co-registered, warped, interpolated/smoothed

images

Structural

DTI

MRSI

Voxel Grid

Structural

MRSI

DTI

Feature Space

21

Features from all voxels

Stage II - Segment Feature Space (I.e. Find Clusters) and Map Cluster

Values Back to Voxel Grid

MRSI

Structural

DTI

Map clustered features

Back to voxel grid

22

Stage III - Generation of Complexity Estimates

Generate co-occurrence matrix/joint distribution by parsing labeled image (just as in standard texture analysis)

Calculate texture/complexity measures from co-occurrence matrix/joint distribution

23

Stage IV – Classification Based on Texture/Complexity Estimates

Classification using texture/complexity estimates in a supervised or unsuprevised learning algorithm (pick your favorite -LDA, SVM, RVM, Bayes nets, Random Forest,…)

24

3 Class LDA Results For AD, FTD, CN Using Entropy, Statistical Complexity,

Excess Entropy on structural MRI

25

3 Class LDA Results For AD, FTD, CN Using Entropy, Statistical Complexity,

Excess Entropy on structural MRI

26

Conclusion

Texture/Complexity Measures Provide Sensitive/Specific Statistics For Studying Subtle Global, Nonlinear Effects From Images

Collaborators – Yue Chen, John Kornak, Gerald Matson, Michael Weiner, Wang Zhan, Grant Gauger, and Norbert Schuff


Recommended