Title: Biomedical Imaging
1Biomedical Imaging
- 100 million Americans undergo diagnostic imaging
procedures every year. Lately functional imaging
is added as a new routine clinical imaging
procedure. - There will be a bigger expansion in biomedical
imaging technology in the 21st century. - There is a growing demand in developing new
imaging instruments, novel imaging procedures,
and data processing. - The Decade of the Brain signed by President
Bush, 1990-1999. - The National Institute for Biomedical Imaging
and Bioengineering (NIBIB) signed by President
Clinton, 2000.
2Medical Imaging
- Anatomical Imaging
- Vs.
- Functional Imaging
3Brain Imaging Techniques
- Michasel I. Posner and Marcus E. Raichle, Images
of Mind, New York, Scientific American Library,
1997
4Functional Brain Imaging
- Anatomical Imaging Ever since imaging
- Functional Imaging last 10 years
- 3 Requirements for Functional Imaging
- Scanner (Camera) Scanner, Coils etc
- Data Analysis Signal Image Processing
- Experimental Protocol (Neuroscience)
Primary Motor Area
First-order motion
V5
MEG
fMRI
5How the Brain Works
- Functional Segregation
-
- Functional Integration
6Functional Brain Imaging
7PET SPECT
- Positron Emission Tomography (PET) and SPECT
- Regional Cerebral Blood Flow
- Regional Cerebral Glucose Metabolism
PET Language Functions
8Functional MRI
- Principles
- Processing Methods
- Problems and Solutions
- Applications
9fMRI Experimental Setup
10Spatiotemporal Resolution
- Spatial Resolution
- Speed of image acquisition
- Image signal/noise (SNR)
- Magnetic field gradient strength
- Static magnetic field
- Effective spatial resolution of the BOLD effect
3mm - Temporal Resolution
- Number of slices acquired
- Slow blood flow changes
- Effective temporal resolution a few seconds
11Echo Planar Imaging (EPI)
- High-speed imaging
- First proposed by Mansfield.
- After one or multiple rf excitation (single-shot
EPI or multi-shot EPI), a series of gradient
pulse is followed, thus MR signals are created by
a series of gradient reversals or oscillations. - Allows imaging of multiple slices in reasonable
short time. - fMR images are acquired with gradient echo EPI
sequences.
12Effective Transverse Relaxation Time, T2
- T2 spin-spin relaxation time
- The time to reduce the transverse magnetization
- Pure T2 is a molecular effect
- Effective decay of transverse magnetization
- Molecular interactions (pure T2)
- Magnetic inhomogeneity (T2,inhomogeneity)
- 1/T21/T21/T2,inhomogeneity
- 1/T21/T2??H/2
13T2 Difference in BOLD fMRI
Signal intensity
MR Signal Intensity ? Tesla
activated
resting
13 at 1.5T
Time
14Neuro-Physiology
EEG
fMRI
15fMRI BOLD
- Sensory, motor, or cognitive task ? Localized
increase in neural activity ? Increased metabolic
rate ? Local vasodilation ? Increased blood
volume ltlt Increased blood flow (50) gtgt metabolic
rate increase (5) ? Decreased ratio
deoxy-Hb/oxy-Hb ? Less spin dephasing from
magnetic inhomogeneities caused by deoxy-Hb ?
Increase in T2 ? Increased MR signal. - Deoxy-Hb paramagnetic
- Oxy-Hb diamagnetic
16Anatomical MRI vs. Functional MRI
Spin-echo Gradient-echo
Spin-echo Gradient-echo
17fMR Data Analysis
- Movement Correction
- Activaty Detection
- Model-driven Processing
- Data-driven Processing
18Time-series Images
19Motion Correction
- Subjects always move in the scanner.
- If correlated with stimulation, artifacts show up
as activated regions. - Registration
- Software AIR by Wood et al.
- Motion models
- translation-rotation, polynomials, etc
- Coregistration costs
- Transformation
20fMRI Data Processing
- Head/Brain Motion Correction
- Image Registration Software
- N-th Order Polynomials
- Voxel Similarity Measures
- Mean Squared Difference (MSD)
- Pearson product-moment cross correlation (NCC)
- Mutual Information (MI)
- Normalized Mutual Information (NMI)
- Entropy of Difference Image (EDI)
- Modified Pattern Intensity (MPP)
- Ratio Image Uniformity (RIU)
- Modified Ratio Image Uniformity (MRIU)
- Brain Activity Detection
- Model-Driven Processing
- Data-Driven Processing
t1 sec
t2 sec
T.-S. Kim et al., IEEE TNS Nucl. Med. Img.
Sci. (2000) Mag. Res. Med (2002)
21fMRI Model-driven Processing
LCC
SPM
(cross correlation)
(general linear model)
22Cross-Correlation Thresholding
- Linear Correlation Coefficient (LCC)
Activation Map
23Statistical Parametric Mapping
- SPM
- A voxel by voxel hypothesis testing approach
- Developed at The Wellcome Department of Cognitive
Neurology Institute of Neurology, University
College London - http//www.fil.ion.ucl.ac.uk/spm/
24The Basics of SPM
25Data-Driven Processing
- No Reference Function ? No assumption of brain
activity - Why
- May reveal brain activity that cannot be detected
with model-driven methods - Clustering
- Principle Component Analysis (PCA)
- Independent Component Analysis (ICA)
- Blind Source Separation Problem
- BS Infomax (Bell Sejnowski, 1995)
- FastICA (Hyvärinen, 1999)
- Mixture Density ICA (J. Jeong T.-S. Kim, 2002
Xu. L, 1997)
26Typical Experimental Designs
- Block (or Steady-State) fMRI
- Measure evoked hemodynamic responses due to
multiple stimuli or events.
- Event fMRI
- Measure an evoked hemodynamic response due to a
single stimulus or event.
27Block fMRI
- Voluntary self-paced right index finger flex
- Linear cross correlation (cc) analysis
- Threshold ccgt0.33
28Event fMRI
29ApplicationSpatiotemporal Localization of Alpha
Activity Source in fMRI and EEG Using Mixture
Density ICA
30Independent Component Analysis (ICA)
- The goal of blind source separation in signal
processing is to recover independent source
signals (e.g., different people speaking, music
etc.) after they are linearly mixed by an unknown
medium, and recorded at N sensors (e.g.,
microphones). - The concept of independent component analysis
(ICA) as maximizing the degree of statistical
independence among outputs using contrast
functions. In contrast with decorrelation
techniques such as Principal Component Analysis
(PCA), which ensures that output pairs are
uncorrelated. ICA imposes the much stronger
criterion that the multivariate probability
density function (p.d.f.) of output variables - Finding such a factorization requires that the
mutual information between all variable pairs go
to zero. Mutual information depends on all
higher-order statistics of the output variables
while decorrelation only takes account of
2nd-order statistics.
31Signal Mixing Unmixing Example
Mixed Sources
?
Unmixed Sources
Sources
32Basics of ICA
- Blind Source Separation Problem
- Data Model
- Xobserved data, Amixing matrix,
Ssources - Assumptions
- Source density is NOT Gaussian
- Linear Mixing
- Find an unmixing matrix W by making components of
S sparse, nongaussian and independent, such that
-
AW-1 - Study Sources, S and mixing matrix, A to find
hidden components in measurements
33ICA Estimation Principles
- Two ICA Estimation Principles
- Nonlinear Decorrelation Find W such that
components of S and their non-linear transformed
components are uncorrelated. - Maximum Nongaussianity Find W such that
components are the maximally nongaussian. - ICA Limitations
- Logistic BS Infomax handles only super-gaussian
- Extended Infomax handles both sub- and
super-gaussian - FastICA handles both sub- and super-gaussian
- Mixture Density handles any different types of
density with parameterized nonlinearity
functions.
34Independent Component Analysis
x As
- Original ICA(Bell,1996) preselected gi( ) for
supergaussian - Extended ICA(Lee,1998) preselected gi( ) for
super or subgaussian - ICA with mixture density model Adjustable gi( )
for any density
35What is Alpha Activity?
- Rhythmic (8-10 Hz in EEG)
- Produced during awake and relaxed state
- Intermittent burst
- Unknown hemodynamics response in fMRI
- Sources ?
36Experimental Protocol
- Conditions
- - With closed eyes
- - Relaxation awake and relaxed, thus
producing Alpha - - Mathematical Tasks Perform pre-assigned
math operation, thus suppressing Alpha - Protocol for fMRI EEG
37Results of ICA in EEG
Measured EEG Signals
Temporal ICs
x1(t)
x19(t)
EEG measurements, xi(n)
Alpha power
index of electrode
38Mixture Density ICA for fMRI
- Assumption
- fMR Images as weighted sum of unknown spatial
maps (ICs)
- ICA decompose fMRI data into spatially
independent components maps
39Localization in fMRI and EEG
- Spatial Scalp Maps of Temporally Independent
Sources in EEG
- 2 Spatially Independent Maps of
fMRI
ICA only
Model- and data- driven methods
J. Jeong T.-S. Kim, submitted.
40Mixture Density ICA for EEG
- Assumption
- EEG channel data as weighted sum of temporal ICs
- ICA decompose EEG data into temporally
independent components maps
41Selection of alpha components in EEG
Ratio of alpha power to background power
SNR of components
Selected alpha dominant components
Spatial distribution of alpha comps
42Results of ICA in fMRI
cc 0.48
cc 0.44
Spatially independent maps
Associated Time courses
43Results of ICA in fMRI
72
cc .57
6
cc .46
7
cc .42
44Result of ME Localization
w/o ICA (left) and w/ ICA (right)
45fMRI Noise and Artifacts
- Problem negative and positive false activation
in functional images - Periodic Sources (physiological fluctuations)
- Cardiac and respiratory pulsation
- CSF pulsation
- Nonperiodic Sources
- MR system instability
- Magnetic susceptibility change
- Gross head motion
- Signal drifts and shifts
46Summary
- Image local changes in blood oxygen concentration
- Blood Oxygenation Level Dependant (BOLD) contrast
- Superior spatial resolution 2 mm
- Low temporal resolution 12 sec due to slow
blood hemodynamics - Provides both anatomical and functional
information - No radioactive material injected
- Artifacts due to physiological noise components
- Movement-related artifacts
- Signals originates from tissue and draining veins