Top Banner
Wavelets and functional MRI Ed Bullmore Mathematics in Brain Imaging IPAM, UCLA July 21, 2004
45

Wavelets and functional MRI

Jan 04, 2016

Download

Documents

Wavelets and functional MRI. Ed Bullmore Mathematics in Brain Imaging IPAM, UCLA July 21, 2004. Overview. What’s a wavelet? Wavelet, fractal, brain Wavelets and functional MRI data analysis resampling or “wavestrapping” estimating noise and signal parameters hypothesis testing. - PowerPoint PPT Presentation
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Wavelets and functional MRI

Wavelets and functional MRI

Ed Bullmore

Mathematics in Brain Imaging

IPAM, UCLA

July 21, 2004

Page 2: Wavelets and functional MRI

• What’s a wavelet?

• Wavelet, fractal, brain

• Wavelets and functional MRI data analysis

- resampling or “wavestrapping”

- estimating noise and signal parameters

- hypothesis testing

Overview

(See Bullmore et al [2004] NeuroImage, IPAM issue, for more comprehensive literature review)

Page 3: Wavelets and functional MRI

Wavelets are “little waves”, defined by their scale j = 1,2,3…J and their

location k = 1,2,3…K = N/2j

…a family of orthonormal wavelets tesselates the time-frequency plane

Wavelets are little waves…

ψ φ

1)( dtt 1)( dtt

1)( dtt 0)( dtt

Page 4: Wavelets and functional MRI

Wavelets provide an orthonormal basis for multiresolution analysis

j

j

jkj

ktt

2

2

2

1)(,

k k

kjkjJj

kJkJ da ,,,, y

The scaling function multiplied by the approximation coefficients, plus the wavelets multiplied by the detail coefficients, recovers the signal losslessly

Page 5: Wavelets and functional MRI

Wavelets are adaptive to non-stationarity

“If we are interested in transient phenomena – a word pronounced at a particular time, an apple located in the left corner of an image – the

Fourier transform becomes a cumbersome tool.” Mallat (1998)

Page 6: Wavelets and functional MRI

Wavelets have decorrelating or whitening properties

)(2'

',', '22~,RH

jjkjkj kkdd

Correlation between any two wavelet coefficients will decay exponentially as a exponential function of distance between them depending on…

H = Hurst exponent of time series

R = regularity of wavelet (number of vanishing moments)

Page 7: Wavelets and functional MRI

Wavelets can be used to estimate fractal or spectral parameters

2

,,, )(1

1}Var{

kjjkj dd

Nd

The variance of wavelet coefficients within each scale of the

DWT is related to the scale j …

12, )2(Var

Hjjd .)12(}){Var(log ,2 jj jHcd or

H = Hurst exponent - simply related to the spectral exponent , e.g. for fractional Gaussian noise, = 2H-1…

…and the spectral exponent is simply related to the (Hausdorf) fractal dimension, D = T + (3-)/2

Page 8: Wavelets and functional MRI

Wavelet variance scaling in fMRI

Greater variance at coarser scales of the decomposition

Gradient of straight line fitted to log-linear plot = 2H+1

Page 9: Wavelets and functional MRI

The discrete wavelet transform is quick to compute

Mallat’s pyramid algorithm

– iterative high and low pass filtering of downsampled coefficients

– O(N) complexity, compared to O(N log(N)) for FFT

Page 10: Wavelets and functional MRI

• Fractal, scaling or scale-invariant, self-similar or self-affine structure is very common in nature, including biological systems

•A wavelet basis is fractal and so a natural choice of basis for analysis of fractal data

• Brain imaging data often have scaling properties in space and time

• Therefore, wavelets may be more than just another basis for analysis of fMRI data

Wavelet, fractal, brain

Continuous wavelet transform of fractal fMRI time series

Page 11: Wavelets and functional MRI

Biological fractals…• complex, patterned,

• statistically-self similar,

• scale invariant structures,

• with non-integer dimensions,

• generated by simple iterative rules,

• widespread in real and synthetic natural systems, including the brain.

Page 12: Wavelets and functional MRI

Fractal cardiologyApparently complex, but simply generated, dendritic anatomy

Statistically self-similar (self-affine) behavior in time

1/f-like spectral properties

Page 13: Wavelets and functional MRI

Brain images often have fractal properties in space (and time)

Biological

Synthetic

Page 14: Wavelets and functional MRI

Allometric scaling and self-similar scaling in brain structure

Zhang & Sejnowski (2000) PNAS 97, 5621-5626

Kiselev et al (2004) NeuroImage 20, 1765-1774

Page 15: Wavelets and functional MRI

Power law scaling of local field

potentials

Leopold et al (2003) Cerebral Cortex

Page 16: Wavelets and functional MRI

Power law scaling of fMRI time seriesRaw fMRI

Images, Time

MR

sig

na

l

0 20 40 60 80 100 120

10

40

10

60

10

80

11

00

Movement-corrected fMRI

Images, Time

Co

rre

cte

d M

R s

ign

al

0 20 40 60 80 100 120

-20

-10

01

02

0

Brownian head drift

Images,Time

He

ad

mo

vem

en

t, m

m

0 20 40 60 80 100 120

0.0

0.2

0.4

0.6

0.8

Raw fMRI spectrum

Frequency

Sp

ect

ral D

en

sity

10 20 30 40 50 60

05

01

00

15

02

00

25

03

00

Movement-corrected fMRI spectrum

Frequency

Sp

ect

ral D

en

sity

10 20 30 40 50 60

05

01

00

15

02

00

25

03

00

1/f spectral scaling

Log Frequency

Lo

g S

pe

ctra

l De

nsi

ty

1.0 1.5 2.0 2.5 3.0 3.5 4.0

-20

24

6Head movement is often a slow “drift”, which tends to exaggerate low frequencies in fMRI spectra; but, even after movement correction, fMRI time series often demonstrate power law scaling of spectral power -

S{f} ~ f or log(S{f}) ~ log(f)

Page 17: Wavelets and functional MRI

• Resampling or “wavestrapping”

* Michael Breakspear

• Time series modeling and estimation

* Voichita Maxim

• Hypothesis testing

* Jalal Fadili and Levent Sendur

Wavelets and fMRI data analysis

Page 18: Wavelets and functional MRI

Time series is autocorrelated or colored

Its wavelet coefficients are decorrelated or white … exchangeable under the null hypothesis

The resampled series is colored

Bullmore et al (2001) Human Brain Mapping 12, 61-78

Resampling or “wavestrapping”

Page 19: Wavelets and functional MRI

-25

-20

-15

-10

-5

0

5

10

15

20

-25

-20

-15

-10

-5

0

5

10

15

20

25

-15

-10

-5

0

5

10

15

20

-20

-15

-10

-5

0

5

10

15

-10

-5

0

5

10

15

20

-20

-15

-10

-5

0

5

10

15

20

-20

-15

-10

-5

0

5

10

15

20

25

30

35

-1

0

1

2

3

4

5

6

-5

-4

-3

-2

-1

0

1

2

3

d1

d2

d3

d4

d5

d6

s6

DWT

iDWT

Time series resampling in the wavelet domain or 1D “wavestrapping”

Time Wavelets

Observed

Resampled

NB: Boundary correction issues inform choice of Daubechies wavelets – which have most compact support for any number of vanishing moments

Page 20: Wavelets and functional MRI

1D wavestrapping preserves fMRI autocorrelation function

If the exchangeability assumption was not satisfied, then the wavestrapped interval for the ACF would be unlikely to include the observed ACF…

…the wavestrapped data would have a whiter spectrum and a flatter ACF than the observed data.

• identical resampling at all voxels preserves the spatial correlations in each slice

• permuting blocks of coefficients, or cyclically rotating coefficients within scale, are alternative resampling strategies

• experimental power at frequencies corresponding to a scale of the DWT may be preserved under wavestrapping

Page 21: Wavelets and functional MRI

Fourier- compared to wavelet-based fMRI time series resampling

Laird et al (2004) MRM

Wavelet resampling is a superior method to Fourier-based alternatives for nonparametric statistical testing of fMRI data

Page 22: Wavelets and functional MRI

2 or 3D discrete wavelet transform• multiresolution spatial filtering of time series statistic maps

• spatially extended signals are losslessly described by wavelet coefficients at mutually orthogonal scales and orientations

yx

x

xy

y

increased scale j, fewer wavelet coefficients {wj}

2D (i)DWT

spatial map of GLM coefficients {}

Page 23: Wavelets and functional MRI

2-dimensional wavestrappingBreakspear et al (2004) Human Brain Mapping, in press

2D wavelet coefficients can be reshuffled randomly, in blocks, or rotated cyclically within each scale (left)

Wavestrapping can be restricted to a subset of the image (right)

Page 24: Wavelets and functional MRI

2D wavestrapping preserves spatial spectrum (autocorrelation structure)

Different versions of 2D wavestrapping algorithm preserve observed spatial spectrum (top left) more or less exactly.

Random shuffling of 2D coefficients (top right), block permutation (bottom left) and cyclic rotation of coefficients (bottom right) show increasingly close convergence with observed spatial spectrum

• relative lack of underpinning theory on decorrelating properties of >1D processes

Page 25: Wavelets and functional MRI

3 or 4D spatiotemporal “wavestrapping” Breakspear et al (2004) Human Brain Mapping, in press

The only nonparametric method for inference on task-specific functional connectivity statistics?

Page 26: Wavelets and functional MRI

GLM with white errors; OLS is BLU estimator

y = X +

{} iid ~ N(0,= I2)

Estimating noise and signal parameters

But fMRI errors are generally colored or endogenously autocorrelated

• autoregressive pre-whitening

estimate unsmoothing kernel

• pre-coloring

apply smoothing kernel

tKK

KKXKy

NXKyK

),0(~,11

Autoregressive pre-whitening is more efficient but may be inadequate in the case of long-memory errors…

Page 27: Wavelets and functional MRI

Autoregressive pre-whitening may not always work well

• FMRI data may frequently have significant residual autocorrelation after attempted pre-whitening by AR(1) and AR(3) models

• A serious side-effect of inadequate pre-whitening is loss of nominal type 1 error control in tests of /SE()

Page 28: Wavelets and functional MRI

Fractional Gaussian noise as a model for functional MRI errors

For fMRI, we are interested in the GLM with fractional Gaussian noise errors

y = X +

{} iid ~ N(0,)

• Gaussian noise is the increment process of Brownian motion…

…fractional Gaussian noise is the increment process of fractional Brownian motion

• fGn is a zero mean stationary process characterised by two parameters, H and

Page 29: Wavelets and functional MRI

Some fractional Gaussian noises

H < 0.5, anti-persistent

H = 0.5, white

H > 0.5, persistent or 1/f

Page 30: Wavelets and functional MRI

Estimators of fractional Gaussian noise parameters in time and wavelet domains

Hurst exponent Variance

A maximum likelihood estimator in the wavelet domain (wavelet-ML) gives the best overall performance in terms of bias and efficiency of estimation of H and 2: Fadili & Bullmore (2001); Maxim et al (2004)

Page 31: Wavelets and functional MRI

Maximum likelihood estimation of fGn parameters in wavelet domain

Likelihood function of G, a fractional Gaussian noise, includes the inverse of its covariance matrix

S is assumed to be diagonalised in the wavelet domain, considerably simplifying its inversion

Page 32: Wavelets and functional MRI

Wavelet ML estimation of signal and fGn parameters

Page 33: Wavelets and functional MRI

Resting fMRI data looks like fGn (after head movement correction)

Freq

uenc

y

fGn

fBm

fGnfBm

+16 mm = 0

0

0.12

-1 1-2 0

Fre

quen

cy

Spectral Exponent,

“Raw” fMRI

Movement corrected fMRI

Page 34: Wavelets and functional MRI

FMRI noise as specimens of fGn

• High variance, antipersistent noise in lateral ventricles

• anti-persistence often located in CSF spaces

• High variance, persistent noise in medial posterior cortex

• persistence often located symmetrically in cortex

Page 35: Wavelets and functional MRI

Alzheimer’s disease associated with abnormal cortical fGn parameters

• fGn parameter maps can be co-registered in standard space and compared between groups using a cluster-level permutation test

• Preliminary evidence for significantly increased persistence of fMRI noise in lateral and medial temporal cortex, precentral and premotor cortex, dorsal coingulate and insula in patients (N=10) with early AD, compared to healthy elderly volunteers (N=10)

Maxim et al (2004)

Page 36: Wavelets and functional MRI

Hypothesis testing

V ~ 20,000 voxels

iV = E(FP), so if

i = 0.05, E(FP) = 1,000

• Bonferroni (i) = 0.05/V

• False discovery rate (FDR)

(i) >= 0.05/V

• GRF (i) < 0.05/V

for small d.f.

Page 37: Wavelets and functional MRI

Why wavelets for multiple hypothesis testing for spatial statistic maps?

• Compaction (few large, many small coefficients)

•Multiresolution properties obviate difficulties concerning optimal choice of monoresolution Gaussian smoothing kernel (MFT)

• Various strategies, including wavelet shrinkage, available to reduce the search volume prior to hypothesis testing (enhanced FDR)

• Known dependencies of neighboring coefficients under the alternative hypothesis (Bayesian bivariate shrinkage)

• Non-orthogonal wavelets exist and may be advantageous for representing edges and lines (dual tree complex wavelet transform)

Page 38: Wavelets and functional MRI

Wavelet shrinkage algorithms to control false discovery rate

• Take 2D or 3D wavelet transform of spatial {} maps

• Calculate P-value for each detail coefficient

• Sort the P-values in ascending order and find the order statistic i such that

• Calculate the critical threshold

• Apply hard or soft thresholding to all coefficients and take the inverse wavelet transform to recover the thresholded map in space

Page 39: Wavelets and functional MRI

Wavelet FDR control is multiresolution• Wavelet FDR = 0.01 retains activation in bilateral motor cortex and ipsilateral cerebellum

• Spatial FDR = 0.01, after monoresolution smoothing with Gaussian kernel, retains activation in contralateral motor cortex with FWHM = 6 mm but loses cerebellar activation with FWHM = 18 mm

Page 40: Wavelets and functional MRI

Enhanced wavelet FDR control is more powerful than “vanilla” wavelet FDR

Enhanced FDR = means FDR (and FWER) are also <= (Shen et al, 2002)

Page 41: Wavelets and functional MRI

Enhanced FDR control in wavelet domain

Shen et al (2002); Sendur et al (2004)

Generalised degrees of freedom is estimated by Monte Carlo integration to reduce the search volume for hypothesis testing

Page 42: Wavelets and functional MRI

Bayesian bivariate shrinkage and enhanced FDR control

222

111

nwy

nwy

yyy

yy

w

n

122

21

222

21

^

1

3

Say w1 is the child of w2…

MAP estimator of w1 is…

Test only coefficients which survive shrinkage by this rule

Page 43: Wavelets and functional MRI

Complex (dual tree) wavelet transform may be

advantageous for mapping edges

Complex wavelet coefficients are estimated by dual tree algorithm – their magnitude is shift invariant

Page 44: Wavelets and functional MRI

Conclusions

• Wavelets are versatile tools, particularly applicable to analysis of scaling, scale-invariant or fractal processes in time (or space, or both)

• Brain images often have scaling or 1/f properties and wavelets can provide estimators of their fractal parameters, e.g. Hurst exponent

• Specific wavelet applications to fMRI data analysis include:

• Resampling (up to 4D)

• Estimation (robust and informative noise models)

• Hypothesis testing (multiresolution and enhanced FDR control)

Page 45: Wavelets and functional MRI

Thanks

John Suckling

Mick Brammer

Jalal Fadili

Michael Breakspear

http://www-bmu.psychiatry.cam.ac.uk

Levent Sendur

Raymond Salvador

Voichita Maxim

Brandon Whitcher

[email protected]

Human Brain Project, National Institute of Mental Health and National Institute of Biomedical Imaging & Bioengineering

Wellcome Trust

GlaxoSmithKline plc