computational methods in neuroimaging
DESCRIPTION
Computational methods in NeuroimagingTRANSCRIPT
Krishna Prasad MiyapuramCognitive Science & Computer Science
Indian Institute of Technology Gandhinagar
Computational Methods in Neuroimaging
19982000
20022004
20082011
2012
ElectronicsArtificial Intelligence
Cognitive Neuroscience
Neuroeconomics
Predictive coding
M.Tech.
M.Sc.
Journey of a thousand miles begins with a single step
Outline
• Imaging the Human Brain
• Physics of Functional MRI
• Classical analysis: Statistical Parametric Mapping
• Data Visualization
• Beyond Blobs: Functional Connectivity
• Machine Learning Methods
• Data Mining Techniques
The BIG Question
Cognitive Science
Psychology
Artificial Intelligence
Neuroscience
EducationLinguistics
Anthropology
Philosophy
What is the nature of human MIND?
The small Answer
Study the human BRAIN!
Neuroimaging Techniques
Frontal LobeParietal Lobe
Occipital Lobe
Temporal Lobe
Cerebellum
Basal ganglia
The Human Brain
Trees do not have an organ called “brain”.
AnteriorPosterior
Superior(Dorsal)
Inferior(Ventral)
Parts of the Brain
Temporal LobeOccipital Lobe
Frontal LobeParietal Lobe
Terminology – Planes and Sections
Coronal
Saggital
Axial
Axial / Horizontal Plane
Saggital Plane
Coronal Plane
3D imaging
The Nobel Prize in Physiology or Medicine for2003 jointly to
Paul C. Lauterbur Sir Peter Mansfield
"for their discoveries concerning magnetic resonance imaging“
http://www.nobel.se/medicine/laureates/2003/press.html
Sectional view of an MRI Scanner
Scanner room
Console room
Patient Table
RF (Head) coil
Gradient coil
Static magnetic field
Physics of MRI
Brain activity
Oxyhaemoglobin
Deoxyhaemoglobin
MRI signal intensity
Oxygen consumption
Cerebral blood flow
Rest (Normal blood flow)
Activation (High blood flow)
(A) (B)
(C)
Experimental Design
A B A B B A B A A B A
B BA A
BA
BA
B BA
BA
B
Input
Output
Input
Output
Process
BaselineTask condition
Data Analysis
fMRI time series
Statistical Parametric Map
Within-subject registrationslice-timing correction
RealignmentCoregistration
(structural to functional)
Between-subject registrationspatial normalization
Spatial smoothing
General Linear ModelDesign matrix
Parameter estimation
Statistical InferenceLinear Contrasts
Thresholding
Random Effects Analysis(Group analysis only)
Preprocessing
Statistical Analysis
Softwares for fMRI Analysis
Statistical Parametric Mapping
• SPM is a form of data reduction, condensing information (in a statistically meaningful way) from a number of individual scans into a single image volume that can be more easily viewed and interpreted.
SPM has an extensive web site at:http://www.fil.ion.ucl.ac.uk/spm
Image Processing
Within-subject registrationslice-timing correction
RealignmentCoregistration
(structural to functional)
Between-subject registrationspatial normalization
Spatial smoothing
Need for motion correction
• People move, even if they don’t realize!
(A) (B)
(D)(C)roll
pitch
yaw
x
y
z
Rigid body movement: 3 translation parameters
3 rotation parameters
Same location in the grid
Same location in the brain
Statistical Analysis
General Linear ModelDesign matrix
Parameter estimation
Statistical InferenceLinear Contrasts
Thresholding
Random Effects Analysis(Group analysis only)
Statistics: How?fMRI model setup
• A General Linear Model (GLM) is setup modellingthe control and test conditions as effects of interest.
y = Xb + e• The GLM is used to specify
the conditions in the form of a design matrix, which defines the experimental design and the nature of hypothesis testing to be implemented.
Specifying Contrasts– A contrast can be used to compare different
conditions in the study.
– The conditions that are of interest are given a positive value, such as 1, and conditions that are subtracted from the conditions of interest are given a negative value, such as -1.
Thresholding:During the assessment of
Results, height and extent thresholds are applied to determine significant activations.
Visualization
Glass Brain for Active-Rest Brain Slice picture for Active-Rest
Rendering onto Subject’s Anatomical Brain
• A High resolution anatomical image (dimensions: 128x128x160 , resolution 1.95 x 1.95 x 1 mm) is acquired.
• This image is Segmented into Grey, White and CSF images.
• The subjects brain is extracted from the Grey and White matter images.
• The activations can now be rendered onto 3D anatomical image of the subject
Psychophysiological Interactions
• Slides from Roland Benoit, MfD 2007/8• Data from
– C. Buchel and K. Friston. Modulation of connectivity in visual pathways by attention: Cortical interactions evaluated with structural equation modelling and fMRI, Cerebral Cortex, 7: 768-778, 1997
• Figures from – K.J. Friston, C. Buchel, G.R. Fink, J. Morris, E. Rolls, and
R. Dolan. Psychophysiological and modulatory interactions in Neuroimaging. NeuroImage, 6:218-229, 1997
– Christian Ruff’s ppt “Experimental Design”
• Tutorial: http://www.fil.ion.ucl.ac.uk/spm/data/
Functional Connectivity
Functional IntegrationFunctional Segregation
Effective ConnectivityFunctional Connectivity
Attention
V1
V5
An Example
Set
source
target
stimuli
source
target
Two Interpretations
Context-sensitive connectivity Modulation of stimulus-specific responses
How it works: Interactions
V1 X Attention
How it is done: PPI & SPM5
• Estimate GLM
• Extract time series at Region of Interest
How it is done: PPI & SPM5
3. Deconvolve, Calculate Interaction, Reconvolve
How it is done: PPI & SPM5
3. Estimate new GLM
How it works: GLM
0 0 1
V1 Att V1XAtt
z = -9 mm
Multi Voxel Pattern Analysis
Problem Statement
• Over the past decade functional Magnetic Resonance Imaging (fMRI) has emerged as a powerful technique to locate activity of human brain while engaged in a particular task or cognitive state.
• We consider the inverse problem of detecting the cognitive state of a human subject based on the fMRI data.
• f : fMRI-sequence(t1,tN) CognitiveState
Dominant Cognitive StatefMRI SCAN 1 fMRI SCAN 2 fMRI SCAN N
fMRI time series (N >= 1)
f
Definition and Motivation
• What is Cognitive State?
• It is the state of a process/operation within the human brain that affects it’s mental contents.
• Motivation :
– Such functions could provide the basis for a new approach– to study human reasoning processes.
– Also deeper understanding of the functioning of human – brain could help us build more advanced AI systems.
Visuo Motor Sequence Learning
Visuo-Motor Mapping:• Association of various visual
instructions to appropriate actions.
• -Stopping at red traffic signal• -Driving slowly at speed• breaker
Sequence Learning:• Learning a task that requires
sequencing a number of actions to achieve a goal.
• -Driving a car• -Lacing a shoe
Visual
Instruction 1Motor Action 1
Visual
Instruction 2
Visual
Instruction 3
Visual
Instruction N
Motor Action 2
Motor Action 3
Motor Action N
Visuo-Motor Mappings
S
e
q
u
e
n
c
e
L
e
a
r
n
i
n
g
Position-to-Position Mapping
P
o
s
i
t
i
o
n
S
e
q
u
e
n
c
e
1
2
1
2
1
2
1 2
1
2
1
2
Visual Display Keypad Response
Position-to-Color Mapping
P
o
s
i
t
i
o
n
S
e
q
u
e
n
c
e
2
1
1
2
2
1
1
2
2 1
2
1
Visual Display Keypad Response
Visuo-Motor Tasks
VisualStimuli
ArbitraryMapping
(Position-to-Color)
Response
P2C
P2P
VisualStimuli
Trial & Error
Response
EarlyLearningLate
Learning
Classification Problem
P2P Vs P2C: Detect the following cognitive states
– “subject is paying attention only towards the position of the visual stimuli”
– “subject is paying attention towards the position and color of the visual stimuli”
Early Vs Late Learning: Detect– the following cognitive
states– “subject has learnt the V-M
sequence”– “subject is in the early
process of learning the V-M sequence”
Machine Learning Approach
• To estimate the function f: fMRI-sequence(t1, tN) -> CognitiveState we have explored the following machine learning techniques:
Gaussian Naïve Bayes (GNB) Classifier k-Nearest Neighbor (kNN) Algorithm Support Vector Machines (SVM)
• Single-Subject Classifier: Classifiers that are trained and tested with a single subject’s fMRI data.
• Multiple-Subject Classifier: Classifiers that are trained with fMRI data of multiple subjects and tested with data of a new subject,
1. Very high dimensional data (184707 voxels/features).
2. Variation in shapes and sizes of brain across human subjects.
3. Variation in the level of fMRI activity across subjects.
Major Challenges
Feature Selection
• Select the n most discriminating voxels (Discrim) : Voxels are selected based on their ability to distinguish one target class (Cognitive States) from the other.
• – Select the n most active voxels (Active) : Voxels are selected based on their ability to distinguish either target class (Cognitive States) from the baseline condition.
• – Select the n feature pairs whose correlation discriminates the target classes (CorrPair) : Voxel pairs are selected based on the ability of their correlation to discriminate the target classes.
• We observed Poor performance of Discrim and Active features and relatively better performance of CorrPair features for multiple-subject classifiers.
20
40
60
80
t=12 t=36 t=72 t=12 t=36 t=72 t=12 t=36 t=72 t=12 t=36 t=72
GNB KNN (k=5) KNN (k=9) SVM
200
474
88
92
96
100
50 100 200 50 100 200 50 100 200 50 100 200
GNB KNN (k=5) KNN (k=9) SVM
Discrim
Active
Sin
gle
Su
bje
ctM
ult
iple
Su
bje
ct
No. of features
Classification Accuracy (%)
No. of features
Time Interval
CorrPair Feature Selection
Feature Selection
P2P Vs P2C Classification Study
97
100
40
55
70
85
50 100 200 50 100 200 50 100 200 50 100 200
GNB KNN (k=5) KNN (k=9) SVM
Discrim
Active
20
40
60
80
t=12 t=36 t=72 t=12 t=36 t=72 t=12 t=36 t=72 t=12 t=36 t=72
GNB KNN (k=5) KNN (k=9) SVM
200
445
Sin
gle
Su
bje
ctM
ult
iple
Su
bje
ct
No. of features
Classification Accuracy (%)
No. of features
Time Interval
CorrPair Feature Selection
Feature Selection
Early Vs Late Learning Study
Interim Conclusions
• The problem of detection of cognitive states in such a high dimensional feature space is feasible when right choice of features is made along with suitable methods for representation of data.
• Overall much better performance of single-subject classifiers over the multiple-subject classifiers.
• We were unsuccessful in learning a classifier function for “Four-Way Classification Study", the question that we can detect all the cognitive states is yet to be answered.
P2P
P2C
Early Late
P2P Early P2P Late
P2C Early P2C Late
Imagery Conditioning
• Imagery: Mental States like those that arise during perception but occur in the absence of immediate sensory input.
• What occurs in your mind when you see the following word
Imagery Conditioning
• Were your mental contents like this
– Has shape (round)
– Has colour (brown)
– Is type (cake)
•
Neither of the above …
Umm! It’s yummy!
Pylyshyn
Kosslyn
OR
OR
•
Experimental Paradigm
Perception Imagination
Reward
No Reward
Poisson ITI (mean 4 sec)
++
2 sec1 sec
3 sec
NothingScrambled
PictureMoney
BillNothing
Scrambled Picture
Money Bill
What did you See?
Reward Predicting Responses
Activation in Midbrain is greater for stimuli predicting reward than control stimuli irrespective of the reward being Perceived or Imagined
Predicting the imagined contents
• Support Vector Machine classifier trained on midbrain activation from visually presented trials successfully predicts whether the participant is imagining a reward or control picture
Participants
Mid
bra
in a
ctiv
atio
n (
clu
ster
ave
rage
)
Visual presentation CS+ CS-Imagination CS+ CS-
Data mining: Meta Analysis
Summary
• Computational Tools are indispensible for neuroimaging
• Classical Analysis uses a standard framework for functional localization
• We can ask questions about functional Integration (a.k.a. effective connectivity)
• Machine Learning Methods have made the reverse inference of cognitive states possible
• Further advances in Computational data mining techniques are to bring in a revolution in Neuroinformatics
Thank you
http://cogs.iitgn.ac.in
The Nobel Prize in Physiology or Medicine for2003 jointly to
Paul C. Lauterbur Sir Peter Mansfield
"for their discoveries concerning magnetic resonance imaging“
http://www.nobel.se/medicine/laureates/2003/press.html
The small Answer
Study the human BRAIN!