Analysis of Functional MRI Timeseries Data Using Signal Processing Techniques

Similar documents
INDEPENDENT COMPONENT ANALYSIS APPLIED TO fmri DATA: A GENERATIVE MODEL FOR VALIDATING RESULTS

New Approaches for EEG Source Localization and Dipole Moment Estimation. Shun Chi Wu, Yuchen Yao, A. Lee Swindlehurst University of California Irvine

Journal of Articles in Support of The Null Hypothesis

SPM8 for Basic and Clinical Investigators. Preprocessing. fmri Preprocessing

Role of Parallel Imaging in High Field Functional MRI

FMRI Pre-Processing and Model- Based Statistics

Functional MRI in Clinical Research and Practice Preprocessing

Basic fmri Design and Analysis. Preprocessing

CS 229 Final Project Report Learning to Decode Cognitive States of Rat using Functional Magnetic Resonance Imaging Time Series

HST.583 Functional Magnetic Resonance Imaging: Data Acquisition and Analysis Fall 2006

EPI Data Are Acquired Serially. EPI Data Are Acquired Serially 10/23/2011. Functional Connectivity Preprocessing. fmri Preprocessing

1 Introduction Motivation and Aims Functional Imaging Computational Neuroanatomy... 12

The organization of the human cerebral cortex estimated by intrinsic functional connectivity

Clustered Components Analysis for Functional MRI

FMA901F: Machine Learning Lecture 3: Linear Models for Regression. Cristian Sminchisescu

Unsupervised Learning

SPM8 for Basic and Clinical Investigators. Preprocessing

Norbert Schuff VA Medical Center and UCSF

FMRI data: Independent Component Analysis (GIFT) & Connectivity Analysis (FNC)

Cognitive States Detection in fmri Data Analysis using incremental PCA

A Model-Independent, Multi-Image Approach to MR Inhomogeneity Correction

Spectral Classification

A NEURAL NETWORK BASED IMAGING SYSTEM FOR fmri ANALYSIS IMPLEMENTING WAVELET METHOD

CS/NEUR125 Brains, Minds, and Machines. Due: Wednesday, April 5

Independent Component Analysis of fmri Data

Effect of age and dementia on topology of brain functional networks. Paul McCarthy, Luba Benuskova, Liz Franz University of Otago, New Zealand

A Reduced-Dimension fmri! Shared Response Model

Multi-voxel pattern analysis: Decoding Mental States from fmri Activity Patterns

Classification of Subject Motion for Improved Reconstruction of Dynamic Magnetic Resonance Imaging

HST.583 Functional Magnetic Resonance Imaging: Data Acquisition and Analysis Fall 2008

Introduction to Neuroimaging Janaina Mourao-Miranda

MultiVariate Bayesian (MVB) decoding of brain images

Classification. Vladimir Curic. Centre for Image Analysis Swedish University of Agricultural Sciences Uppsala University

Statistical Analysis of Neuroimaging Data. Phebe Kemmer BIOS 516 Sept 24, 2015

Lecture 11: Classification

NEURO M203 & BIOMED M263 WINTER 2014

Supplementary Figure 1. Decoding results broken down for different ROIs

Spatial Regularization of Functional Connectivity Using High-Dimensional Markov Random Fields

White Pixel Artifact. Caused by a noise spike during acquisition Spike in K-space <--> sinusoid in image space

CS 229 Midterm Review

Last week. Multi-Frame Structure from Motion: Multi-View Stereo. Unknown camera viewpoints

Learning from High Dimensional fmri Data using Random Projections

Dimension Reduction CS534

Introductory Concepts for Voxel-Based Statistical Analysis

COSC160: Detection and Classification. Jeremy Bolton, PhD Assistant Teaching Professor

Supplementary Data. in residuals voxel time-series exhibiting high variance, for example, large sinuses.

An Introduction To Automatic Tissue Classification Of Brain MRI. Colm Elliott Mar 2014

22 October, 2012 MVA ENS Cachan. Lecture 5: Introduction to generative models Iasonas Kokkinos

Estimating Noise and Dimensionality in BCI Data Sets: Towards Illiteracy Comprehension

( ) =cov X Y = W PRINCIPAL COMPONENT ANALYSIS. Eigenvectors of the covariance matrix are the principal components

Announcements. Recognition I. Gradient Space (p,q) What is the reflectance map?

MR IMAGE SEGMENTATION

Bayesian Inference in fmri Will Penny

Marcel Worring Intelligent Sensory Information Systems

Introduction to digital image classification

Image Segmentation Techniques for Object-Based Coding

Facial Expression Recognition Using Non-negative Matrix Factorization

Image Processing. Image Features

Multivariate pattern classification

Reconstructing visual experiences from brain activity evoked by natural movies

CIE L*a*b* color model

Extracting Coactivated Features from Multiple Data Sets

10-701/15-781, Fall 2006, Final

MEDICAL IMAGE ANALYSIS

Multivariate Pattern Classification. Thomas Wolbers Space and Aging Laboratory Centre for Cognitive and Neural Systems

A Spectral-based Clustering Algorithm for Categorical Data Using Data Summaries (SCCADDS)

Robust Kernel Methods in Clustering and Dimensionality Reduction Problems

Mapping of Hierarchical Activation in the Visual Cortex Suman Chakravartula, Denise Jones, Guillaume Leseur CS229 Final Project Report. Autumn 2008.

CSE 6242 A / CS 4803 DVA. Feb 12, Dimension Reduction. Guest Lecturer: Jaegul Choo

First-level fmri modeling

Graphical Models, Bayesian Method, Sampling, and Variational Inference

Resting state network estimation in individual subjects

Pixels to Voxels: Modeling Visual Representation in the Human Brain

Random projection for non-gaussian mixture models

Slide 1. Technical Aspects of Quality Control in Magnetic Resonance Imaging. Slide 2. Annual Compliance Testing. of MRI Systems.

Supplementary Figure 1

INDEPENDENT COMPONENT ANALYSIS WITH FEATURE SELECTIVE FILTERING

Introduction to fmri. Pre-processing

Analysis of fmri data within Brainvisa Example with the Saccades database

Chapter 3 Set Redundancy in Magnetic Resonance Brain Images

Modern Medical Image Analysis 8DC00 Exam

SGN (4 cr) Chapter 11

Detecting Salient Contours Using Orientation Energy Distribution. Part I: Thresholding Based on. Response Distribution

Linear Methods for Regression and Shrinkage Methods

University of Florida CISE department Gator Engineering. Clustering Part 2

Function approximation using RBF network. 10 basis functions and 25 data points.

Biomagnetic inverse problems:

NA-MIC National Alliance for Medical Image Computing fmri Data Analysis

Segmentation of MR Images of a Beating Heart

Homework. Gaussian, Bishop 2.3 Non-parametric, Bishop 2.5 Linear regression Pod-cast lecture on-line. Next lectures:

Clustering and Visualisation of Data

Clustering Lecture 5: Mixture Model

Parametric Response Surface Models for Analysis of Multi-Site fmri Data

Basic Introduction to Data Analysis. Block Design Demonstration. Robert Savoy

Voxel selection algorithms for fmri

Adaptive Learning of an Accurate Skin-Color Model

CHAPTER 3 TUMOR DETECTION BASED ON NEURO-FUZZY TECHNIQUE

Spatial Variation of Sea-Level Sea level reconstruction

Functional MRI data preprocessing. Cyril Pernet, PhD

2. Data Preprocessing

Transcription:

Analysis of Functional MRI Timeseries Data Using Signal Processing Techniques Sea Chen Department of Biomedical Engineering Advisors: Dr. Charles A. Bouman and Dr. Mark J. Lowe S. Chen Final Exam October 7, 22 p.1/39

Introduction Overview Update: Supertemporal Resolution Analysis Review New simulations New data Clustered Components Analysis Motivation Theory Methods Results Conclusions S. Chen Final Exam October 7, 22 p.2/39

Goals We would like to aid in the understanding of the blood-oxygenation-level-dependent (BOLD) contrast mechanisms used in functional magnetic resonance imaging (fmri) through achieving a high signal-to-noise (SNR) estimate of the BOLD response. achieving a high temporal resolution estimate of the BOLD response. S. Chen Final Exam October 7, 22 p.3/39

fmri: The basic idea Experimental paradigm designed to activate neuronal metabolism Changes in blood oxygenation during activation parameters affecting MR signal changes in physical Contrast produced by difference between active and control states Data set is volume of pixels repeated over time One pixel of one slice through time Consider this pixel 8 16 24 32 4 Time Response Signal Stimulus Signal S. Chen Final Exam October 7, 22 p.4/39

Supertemporal Resolution: Motivation Short TR Better time resolution Lower SNR due to saturation effects BOLD signal is distorted by blood inflow effects Long TR Poorer time resolution BOLD effect more dominant in activation signal S. Chen Final Exam October 7, 22 p.5/39

Supertemporal Resolution: Review Assumption: Voxels exhibiting the same generating activation signal span different slices in a 2D acquisition Method exploits the timing characteristics of the 2D acquisition Bayesian prior used to implement temporal regularization S. Chen Final Exam October 7, 22 p.6/39

#! & % ) Supertemporal Resolution: Review MAP estimate for Supertemporal Resoution (STR) " where ' (' Optimization performed using conjugate gradient Regularization parameter strategy found by crossvalidation S. Chen Final Exam October 7, 22 p.7/39

Supertemporal Resolution: Updates Reduction in computation time Minor software revisions New hardware New simulations Introduced amplitude amplification factors (1x, 2x, 4x, 6x, 8x simulating increase in -field) Generated multiple (2 / ) datasets New human visual system data Three inch surface coil Multiple runs (3 TR = 2.s, 3 TR = ) of 3.5 cycles S. Chen Final Exam October 7, 22 p.8/39

Performance comparison Simple averaging (SA) method Alignment of slices into timeframe of first slice NO regularization Closed form solution s time resolution estimate from TR = s dataset Interpolation with regularization (IWR) method Alignment of slices into timeframe of first slice Regularization applied and chosen with crossvalidation Numerical optimization with conjugate gradient s time resolution estimate from TR = s dataset Supertemporal regularization (STR) method Slice timing considered in data model Regularization applied and chose with crossvalidation Numerical optimization with conjugate gradient s time resolution estimate from TR = 2. s dataset S. Chen Final Exam October 7, 22 p.9/39

Simulation results: Performance.4 Mean square error of estimates versus synthetic amplitude amplification mean SA error mean IWR error mean STR error individual SA errors individual IWR errors individual STR errors e error e 2. error Mean Square Error (AU).3.2.1 1 2 3 4 5 6 7 8 Amplitude Amplification Factor Mean square error of simulation results for different analysis methods plotted against amplitude amplification factor S. Chen Final Exam October 7, 22 p.1/39

Simulation results: Examples.8 IWR estimate on synthetic dataset at 4x template amplitudes normalized IWR estimate injected BOLD signal.8 STR estimate on synthetic dataset at 4x template amplitudes normalized STR estimate injected BOLD signal.6.6.4.4.2.2 Intensity (AU) Intensity (AU).2.2.4.4.6.6.8 4 5 6 7 8 9 1 11 12 Time (s).8 4 5 6 7 8 9 1 11 12 Time (s) IWR estimate on TR=s data STR estimate on TR=2.s data Examples of second estimates at S. Chen Final Exam October 7, 22 p.11/39

Human data results: Simple averaging method.15.1 SA estimates for V (r#) data series V (r1) V (r2) V (r3).15.1 Statistics on SA estimates for V (r#) data series mean mean ± std Normalized intensity (AU).5.5 Normalized intensity (AU).5.5.1.1.15 4 5 6 7 8 9 1 11 12 Time (s).15 4 5 6 7 8 9 1 11 12 Time (s) SA estimates Mean and std. dev. of SA estimates Simple averaging estimates on the TR= second dataset (3 experiments) S. Chen Final Exam October 7, 22 p.12/39

Human data results: Interpolation with regularization method.15.1 IWR estimates for V (r#) data series V (r1) V (r2) V (r3).15.1 Statistics on IWR estimates for V (r#) data series mean mean ± std Normalized intensity (AU).5.5 Normalized intensity (AU).5.5.1.1.15 4 5 6 7 8 9 1 11 12 Time (s).15 4 5 6 7 8 9 1 11 12 Time (s) IWR estimates Mean and std. dev. of IWR estimates Interpolation with regularization estimates on the TR= second dataset (3 experiments) S. Chen Final Exam October 7, 22 p.13/39

Human data results: Supertemporal Resolution method.15.1 STR estimates for V (r#) data series 2. V (r1) 2. V (r2) 2. V (r3) 2..15.1 Statistics on STR estimates for V (r#) data series 2. mean mean ± std Normalized intensity (AU).5.5 Normalized intensity (AU).5.5.1.1.15 4 5 6 7 8 9 1 11 12 Time (s).15 4 5 6 7 8 9 1 11 12 Time (s) STR estimates Mean and std. dev. of STR estimates Supertemporal resolution estimates on the TR=2. second dataset (3 experiments) S. Chen Final Exam October 7, 22 p.14/39

Discussion and conclusions Simulated data Initially at low SNR, STR performs worse than IWR because small features masked by noise At increasing SNR, STR performs better than IWR/SA as inherent physical advantange becomes apparent In human data, STR estimates qualitatively different from IWR/SA estimates Conclusion: STR may be a valuable tool in characterizing small features in the BOLD signal at higher static field strengths or higher SNR S. Chen Final Exam October 7, 22 p.15/39

Clustered components analysis: Objectives Hypothesis: Activation by specific functional tasks responses in different parts of the brain Therefore, we propose the following goals: Distinct neural Design and run fmri experiment activating visual, auditory, and motor cortices. Estimate number of distinct neural responses (# of classes/clusters) Extract an estimate for each response Determine voxel memberships S. Chen Final Exam October 7, 22 p.16/39

Existing approaches to signal estimation Principle component analysis (PCA) Extracts orthogonal signals Disadvantage: Signals not usually orthogonal Independent component analysis (ICA) Extracts spatially independent signals Disadvantage: Signals may not be independent Conventional Clustering Groups signal vectors as spheres about a mean Disadvantage: Signals may not form spherical clusters General Comment: None of these methods start with an explicit model of the data. All go about estimating the distinct signals in an ad hoc way. S. Chen Final Exam October 7, 22 p.17/39

Analysis framework Dimensionality Reduction Signal subspace is orthogonal to noise subspace Noise can be accurately modeled in fmri Separate signal subspace (dim ) ) from noise subspace (dim Clustered Components Analysis Useful information is in shape of signal, amplitude unimportant Component direction is found instead of mean Amplitude can vary in cluster so long as shape preserved Clusters found in cylinders instead of spheres S. Chen Final Exam October 7, 22 p.18/39

Interpretation of Clustered Components Analysis Because amplitude of the voxel signal is not important, the method clusters around component directions, not component means. This means the clusters can be thought of as cylinders rather than the traditional spheres. S. Chen Final Exam October 7, 22 p.19/39

Dimensionality reduction: Harmonic decomposition Data model for harmonic decomposition : detrended voxel timecourse matrix ( =# of voxels) =# of timepoints, : matrix of sampled sines and cosines ( components) harmonic : harmonic image : maxtrix of residuals from the least squares fit S. Chen Final Exam October 7, 22 p.2/39

Dimensionality reduction: Signal subspace estimation Signal + noise covariance: Noise covariance: trace Signal covariance: Eigen decomposition Only the columns of the eigenvector matrix corresponding to the positive eigenvalues of yielding the modified eigenvector matrix are retained,. ( reduced dimensionality feature vector matrix: is a whitening vector matrix derived from ) S. Chen Final Exam October 7, 22 p.21/39

Data Model for Clustered Component Analysis Assumptions Only shape of the response important Amplitude is NOT important Noise independent in space and time (time-independence can be relaxed) Our Model is -dimensional column vector representation of the timecourse is the unknown scalar amplitude for pixel are the is class of the pixel is a Gaussian noise vector component directions, n e Xn + W n n e Xn where n = 1.5, X n = 1 e 1 e 2 e 3 S. Chen Final Exam October 7, 22 p.22/39

Clustered components approach Goal: Minimize minimum description length (MDL) criterion MDL loglikelihood # of parameters # of datapoints Unknown model parameters is the model order (number of clusters) is the amplitude of each pixel is the set of distinct neural responses are the prior probabilities for each class Use maximum likelihood (ML) estimate implicitly Find ML estimates and using the Expectation-Maximization (EM) algorithm for each model order Estimate model order MDL criterion by cluster merging and minimizing the S. Chen Final Exam October 7, 22 p.23/39

' ' Voxel likelihood function Likelihood for each voxel #! ML estimate of the scalar amplitude Voxel log-likelihood S. Chen Final Exam October 7, 22 p.24/39

#! ' '! ' ' ' ' % Maximum likelihood estimate Log-likelihood of the entire dataset # #! ML estimate of the parameters ( '& # "! S. Chen Final Exam October 7, 22 p.25/39

( % % ( % % & & % Expectation-maximization equations Posterior probability '& ( & & E-step ( & ( & M-step S. Chen Final Exam October 7, 22 p.26/39

! #!! # # #! '! # Order Estimation through Cluster Merging 1. Start with large number of clusters ( ) and initialize 2. Run EM algorithm to convergence 3. Choose the two clusters that minimize the distance function (which is the upper bound on the change in MDL) # #! 4. Merge clusters using 5. Decrement and initialize next iteration with new clusters 6. Repeat 2 through 5 until = 1 7. Choose number of components minimizing the MDL criterion! S. Chen Final Exam October 7, 22 p.27/39

Synthetic fmri Images Synthetic data Baseline control images created at each sample point During periods of activation, 3 different realistic signals with varying amplitudes were injected Gaussian white noise added to each voxel at each timepoint time Verification and comparison using different analysis methods applied before and after signal subspace estimation (SSE) PCA, using 3 components corresponding to the 3 largest variances Spatial ICA constrained to yield 3 components Spatial ICA unconstrained, using 3 best components Fuzzy C-means (FCM) clustering constrained to yield 3 clusters CCA S. Chen Final Exam October 7, 22 p.28/39

Paradigm Design For our dimensionality reduction, activation must be periodic Block activation scheme 1 cycle = 32 seconds control (rest state), 32 seconds 1 scan = 16 seconds lead in - 4.5 paradigm cycles - 16 seconds lead out (only use samples during paradigm) Hz sample rate (TR = 2 seconds) To illustrate the power of the clustering method, many different types of functional cortex must be activated Visual: Flashing 8Hz checkerboard Auditory: Forward vs. Backward speech (backwards is the control) Motor: Self paced finger tapping Lead-in Lead-out Off On Off On Off On Off On Off : :16 :48 1:2 1:52 2:24 2:56 3:28 4: 4:32 5:4 5:2 S. Chen Final Exam October 7, 22 p.29/39

Simulated data: Hard classifications Results for CCA applied to synthetic data S. Chen Final Exam October 7, 22 p.3/39

Simulated data: Qualitative results 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 PCA FCM constrained ICA 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 1 2 3 4 5 6 7 8 9 1 11 unconstrained ICA CCA Estimates after application of SSE S. Chen Final Exam October 7, 22 p.31/39

Simulated data: Quantitative results Mean squared error for analyses on synthetic data before and after signal subspace estimation (SSE) Before SSE After SSE PCA FCM ICA (c) ICA (u) CCA Number of voxels classified correctly on synthetic data before and after signal subspace estimation (SSE) out of 192 total voxels PCA ICA (c) ICA (u) FCM CCA Before SSE 61 113 38 95 167 After SSE 111 162 77 151 169 S. Chen Final Exam October 7, 22 p.32/39

Human data: Timesequence realizations 2 1.5 Class 1 Class 2 Class 3 Class 4 Class 5 1 1 1.5 2 2 4 6 8 1 12 14 16 18 2 22 First 5 clusters S. Chen Final Exam October 7, 22 p.33/39

Human data: Timesequence realizations 2 1.5 Class 6 Class 7 Class 8 Class 9 1 1 1.5 2 2 4 6 8 1 12 14 16 18 2 22 Clusters 6-9 S. Chen Final Exam October 7, 22 p.34/39

Human data: Hard classifications 5 5 1 1 15 15 2 2 25 25 5 1 15 2 25 5 1 15 2 25 Motor cortex (first 5 clusters): (L) Upper slice, (R) Lower slice S. Chen Final Exam October 7, 22 p.35/39

Human data: Hard classifications 5 5 1 1 15 15 2 2 25 25 5 1 15 2 25 5 1 15 2 25 Auditory cortex (first 5 clusters): (L) Upper slice, (R) Lower slice S. Chen Final Exam October 7, 22 p.36/39

Human data: Hard classifications 5 5 1 1 15 15 2 2 25 25 5 1 15 2 25 5 1 15 2 25 Visual cortex (first 5 clusters): (L) Upper slice, (R) Lower slice S. Chen Final Exam October 7, 22 p.37/39

Conclusions Clustered component analysis is a new method of extracting signals where only shape, not amplitude, is important CCA has been shown to perform well on simulated data The experimental results show the following: The distinct neuronal signals do not correlate strongly with the known functional cortices The clusters tend to lie along sulcal-gyral boundaries, possibly correlated with vasculature CCA can be used with dimensionality reduction strategies other than the ones used in our experiments CCA may also be adapted for use with applications other than fmri S. Chen Final Exam October 7, 22 p.38/39

Acknowledgements Major Professors: Dr. Mark J. Lowe and Dr. Charles A. Bouman Committee Members: Dr. Peter C. Doerschuk and Dr. Edward J. Delp Department of Biomedical Engineering and Division of Imaging Sciences S. Chen Final Exam October 7, 22 p.39/39