Multiresolution analysis: theory and

applications
Analisi multirisoluzione: teoria e applicazioni
 Course overview

Course structure Contents

•  Review of Fourier theory
•  The course is about wavelets and •  Wavelets and multiresolution
multiresolution
•  Review of Information theoretic concepts
–  Theory: 4 hours per week
§  Tue. 8.30-10.30, room I •  Applications
§  Wed. 8.30-10.30, room I –  Image coding (JPEG2000)
–  Laboratory –  Feature extraction and signal/image analysis
§  Thu. 14.30-16.30 (Lab. Gamma) LM32
•  Wavelets and sparsity in neuroimaging
•  Exam
–  Theory: Oral (in general)
–  Lab: Evaluation of lab. sessions and questions
during the exam
–  Projects: only in case of diploma project
 Books
Stephane Mallat
(Ecole Polytechnique)
Martin Vetterli (EPFL)
 Telecommunications for Multimedia

Good news Bad news

•  It is fun! •  Some theoretical background is unavoidable
•  Get in touch with the state-of-the-art –  Mathematics
technology §  Fourier transform
§  Linear operators
•  Convince yourself that the time spent on §  Digital filters
maths&stats was not wasted §  Wavelet transform

•  Learn how to map theories into applications –  (some) Information theory

•  Acquiring the tools for doing good research!

“Scale”
“Scale”
“Scale”
 Brain tissue microstructure

●  The brain is principally

composed of a type of cells
called neurons.
●  A neuron is composed of a
cellular body called soma and a
tail called axon that is physical
link between the neurons.
●  The axons are usually group in
bundles called fibers.
●  In the brain the soma are
positioned in the cortex and are
generally called gray matter
(GM), while the fibers are
positioned in the central regions
and are called white matter
(WM).
Axons
bundles
 Magnetic Resonance Imaging

●  Standard MRI is the principal

non-invasive imaging technique
used for clinical purposes.

●  Using standard MRI techniques

is possible to distinguish
between GM, WM and CSF but
not the complex structure of
the White Matter fibers bundles.

●  To overcome this limitation,

using an additional pulse is
possible to obtain a different
type of images called Diffusion
Weighted MRI.
 Diffusion Weighted MRI

●  Diffusion MRI was born to observe the

diffusion of water molecules in soft
tissues.
●  The diffusion signal can be modelled using
some mathematical algorithms called
reconstruction techniques.
●  From the reconstructed signal is possible
to calculate numerous measures to
characterize the tissue and to calculate
the orientation of the fibers tract in the
voxel.
●  From the single voxel orientation profile is
possible to reconstruct the brain fibers
tracts topology, this operation is called
tractography.
 Objectives

●  Find the optimal reconstruction technique for Diffusion

MRI data
●  Definition of a standard criterion for validation
○  Synthetic data
●  Identification of new scalar indices as numerical
biomarkers of the structural properties of brain
tissues
○  Anatomically and biophysically plausible besides
being objectively measurable
○  Supporting and improving cortical connectivity
modeling
●  Uses of this indices features
○  Tissues characterization by pattern recognition
○  Patient vs Control classification
 Diffusion signal

●  Invented by Stejskal and Tanner (1965)

●  It exploits an additional sequence of pulses: Pulse

Gradient Spin Echo (PSGE) to measure the
attenuation of the signal due to the diffusion of
water in the soft tissues

●  Changing the gradient direction (u) and strength

(b-value) it is possible to obtain different volumes
called DWI, each one representing the attenuation
of the diffusion in the chosen direction

●  The b-value depends on the duration of the pulse 𝜏

and the pulse frequency q:
Sampling topologies
 From diffusion signal to water molecules pdf

●  The signal attenuation E(q) is related to

Ensemble Average Propagator (EAP) by a
Fourier relationship:

r: distance traveled by molecules in the unit

time
q: reciprocal vector

●  The EAP represents the probability of a net

displacement r in the unit time
 Continuous Analytical Basis for Diffusion Imaging

●  Continuous analytical basis besides SH have been proposed to find an accurate

mathematical description of the diffusion signal and its derivations

●  Analytical models aim at approximating the signal E(q) by a truncated linear combination
of basis functions Φ(q) up to the order N:

cj are the transform coefficients characterizing the signal. Usually these coefficients
are obtained by linear fitting, e.g. using regularized mean squares
 Continuous Analytical Basis for Diffusion Imaging

The principal advantages of Continuous Basis are:

●  Continuous analytical signal representation in q-space independently from the

acquisition sampling scheme

●  Possibility to calculate the EAP and the ODF analytically, obtaining an exact solution for
all the computations

Principal open issues:

●  Identification of the sampling topology
●  Identification of the optimal basis for signal approximation
 Simple Harmonic Oscillator based Reconstruction and Estimation

●  SHORE is a continuous analytical basis introduced by Ozarslan in 2009

●  The signal is approximated using a combination of orthonormal functions which are the
solutions of the 3D quantum mechanical harmonic oscillator
●  Separable solution (Merlet 2013): Laguerre Polynomials for the radial part and Spherical
Harmonics for the angular part
Wiring the brain
 Modeling and recognition of waveforms by
multiresolution methods with application to hdEEG

The purpose of this work was to focus on a particular

pathology, namely temporal lobe epilepsy, in order to detect,
analyze and model the so-called interictal spikes.
JPEG2000
Mathematical tools
 Introduction

•  Sparse representations: few coefficients reveal the information we are looking for.
–  Such representations can be constructed by decomposing signals over elementary waveforms
chosen in a family called a dictionary.
–  An orthogonal basis is a dictionary of minimum size that can yield a sparse representation if
designed to concentrate the signal energy over a set of few vectors. This set gives a geometric
signal description.
§  Signal compression and noise reduction
–  Dictionaries of vectors that are larger than bases are needed to build sparse representations of
complex signals. But choosing is difficult and requires more complex algorithms.
§  Sparse representations in redundant dictionaries can improve pattern recognition, compression and noise
reduction

•  Basic ingredients: Fourier and Wavelet transforms

–  They decompose signals over oscillatory waveforms that reveal many signal properties and
provide a path to sparse representations
 Signals as functions

•  CT analogue signals (real valued functions of continuous independent variables)

–  1D: f=f(t)
–  2D: f=f(x,y) x,y
–  Real world signals (audio, ECG, pictures taken with an analog camera)
•  DT analogue signals (real valued functions of discrete variables)
–  1D: f=f[k]
–  2D: f=f[i,j]
–  Sampled signals
•  Digital signals (discrete valued functions of DT variables)
–  1D: y=y[k]
–  2D: y=y[i,j]
–  Sampled and discretized signals
 Images as functions

•  Gray scale images: 2D functions

–  Domain of the functions: set of (x,y) values for which f(x,y) is defined : 2D lattice [i,j] defining
the pixel locations
–  Set of values taken by the function : gray levels

•  Digital images can be seen as functions defined over a discrete domain {i,j: 0<i<I, 0<j<J}
–  I,J: number of rows (columns) of the matrix corresponding to the image
–  f=f[i,j]: gray level in position [i,j]
 Example 1: δ function

⎧1 i= j=0
δ [i, j ] = ⎨
⎩0 i, j ≠ 0; i ≠ j

⎧1 i = 0; j = J
δ [i, j − J ] = ⎨
⎩0 otherwise
 Example 2: Gaussian

Continuous function
x2 + y 2
1 2σ 2
f ( x, y ) = e
σ 2π

Discrete version

i2 + j2
1 2σ 2
f [i, j ] = e
σ 2π
Example 3: Natural image
Example 3: Natural image
 The Fourier kingdom
+∞

•  Frequency domain characterization of signals F (ω ) = ∫ f (t )e− jωt dt

−∞
+∞
jωt
f (t ) = ∫ F (ω )e dt
−∞

Signal domain

Frequency domain
The Fourier kingdom
Gaussian function
The Fourier kingdom
rect function

sinc function
 2D Fourier transform

+∞
− j (ω x x + ω y y )
fˆ (ω x , ω y ) = ∫ f ( x, y ) e dxdy
−∞
+∞
1 j (ω x + ω y )
f ( x, y ) = ∫ fˆ (ω x , ω y ) e x y d ω x d ω y
4π 2 −∞

1
∫∫ f ( x, y ) g * ( x, y ) dxdy = fˆ (ωx , ω y )gˆ * (ωx , ω y ) d ω x d ω y
4π 2 ∫∫
Parseval formula

2 1 ˆ
2
f = g → ∫∫ f ( x, y ) dxdy = 2 ∫∫ f (ωx , ω y ) d ωx d ω y Plancherel equality
4π
The Fourier kingdom
Wavelets

Wavalet in signal (time or space) domain

Wavalet in frequency (Fourier) domain

Wavelet representation
Wavelet representation
Wavelets are good for transients

scalogram
 Wavelets&Pyramids
N
N

N
Wavelets&Pyramids
Wavelets&Pyramids
Wavelets&Pyramids
Wavelets&Filterbanks

H ↓2

H ↓2

G ↓2

H ↓2

G ↓2
G ↓2
 Wavelets&Filterbanks

↑2 H

↑2 H
↑2 G

↑2 H

↑2 G
↑2 G

Very efficient implementation by recursive filtering

 Fourier versus Wavelets

Fourier Wavelets
–  Basis functions are sinusoids –  Different families of basis functions are
§  More in general, complex exponentials possible
–  Switching from signal domain t to §  Haar, Daubechies’, biorthogonal
frequency domain f –  Switching from the signal domain to a
§  Either spatial or temporal multiresolution representation
–  Good localization either in time or in –  Good localization in time and frequency
frequency §  Information on both the sharpness of the
§  Transformed domain: Information on the transient and the point where it happens
sharpness of the transient but not on its –  Good for any type of signal
position
–  Good for stationary signals but unsuitable
for transient phenomena
 Applications

•  Compression and coding

–  Critically sampled representations (discrete wavelet transforms, DWT)
•  Feature extraction for signal analysis
–  Overcomplete bases (continuous wavelet transform, wavelet frames)
•  Image modeling
–  Modeling the human visual system: Objective metrics for visual quality assessment
–  Texture synthesis
•  Image enhancement
–  Denoising by wavelet thresholding, deblurring, hole filling
•  Image processing on manyfolds
–  Wavelet transform on the sphere: applications in diffusion MRI