experiment design
Post on 03-Jan-2016
38 Views
Preview:
DESCRIPTION
TRANSCRIPT
Novel Tools for (Functional) Magnetic
Resonance Image Analysis
Development and Implementation in the Scientific and Statistical
Computing Core
Robert W CoxRobert W Coxand a cast of several
Raw dataMR-scanner BOLD EPI
AnatomyFunction
Func. & Anat.
BOLD signal
ScannerSubjectStimulus Delivery
Group analysis
Experiment design
ReconstructionDistortion correction
Co-registration
Statistical modelsInference
Scientific & Statistical Computing CoreScientific & Statistical Computing Core• Develop and implement new methodologies to meet
user needs• Consult with IRP users/groups regarding
– Experimental design– Processing methods and tools– Statistical inferences
• Conduct classes on designing and processing FMRI experiments
• Answer FMRI / MRI questions on message board • Distribute & maintain our open-source software tools• Facilitate cross-talk between different FMRI tools:
– AFNI, FSL, FMRIstat, FreeSurfer, Caret, SPM, …
Rich sourceof ideas fornovel tools
AFNI + SUMA
• AFNI = collection of programs for FMRI analysis– Visualization
• 2D, 3D, time-series, cortical surface (SUMA)– Time Series Analysis
• Linear & nonlinear regression– Statistics on 3D Image Collections
• 1-5 way ANOVA; non-parametrics; SEM– Data editing tools
• Spatial and temporal filtering• 3D image registration• Clustering; ROI drawing & Atlas-based ROIs
The AFNIAFNI / SSCC Philosophy• Enable users to stay close to their data
– Save intermediate results– Look at images and data in connected ways
• User controls processing steps and parameters– Everyone has an opinion– Special problems need special solutions
• Efficient (fast) implementations– Things that are easy and fast to do will get done
more often• Help the users
– Until our patience runs out
Features Added to AFNIAFNI and SUMASUMA in Response to User Requests and / or
Problems / Complaints(at least in part)
Next Set of Slides
Feature: Atlases• Problem: Navigating in a complicated folded up 3D
object (i.e., the brain) with few easily recognized landmarks
• Solution: Coordinate-based brain atlases– Accepting the 5-10 mm uncertainty of brain coordinates
• Atlas #1: Talairach-Tournoux atlas– As parsed by Peter Fox’s group at UT San Antonio
• Atlas #2: Cytoarchitectonic atlases from Karl Zilles’ group at Forschungszentrum Jülich– 10 brains being sliced & diced & stained & scanned– About 40% complete at this time
• Where Am I? + Jump To + Colorization + ROIs• Plans: keep up with Zilles; Animal atlases? …
Example: Where Am I?
• Problem: other skull stripping software (e.g., BET in
FSL) didn’t work reliably enough• Solution was to re-visit problem from scratch, and
build on BET’s surface growing algorithm• Then add new features: special knowledge about
where the eyes are likely to be; 3D edges; etc.• Then test it on the hard cases from NIH (ab)users• Extra feature: extend it to monkey images• Plans: continue testing and improvements
Feature: Skull Stripping A
Feature: De-Spiking• Problem: occasional big spikes in echo planar
images gathered for functional MRI– Problem eventually traced to gradient coil– In the meantime: can studies be saved?
• Wrecks the standard time series analysis
A
Feature: Amplitude Modulated FMRI• Situation: Each stimulus event comes with an
auxiliary parameter– May be measured (GSR, reaction time, …) or may
be determined by experimenter• Want to determine if FMRI response magnitude is
proportional to this auxiliary parameter• Solution was to add amplitude modulated
regressors to AFNI’s 3dDeconvolve program– Two regressors per condition– First is: each stimulus response identical– Second is: each stimulus response proportional
to auxiliary parameter for that stimulus
• Plans: 2-3 params/event; event-wise amplitudes
Feature: Nonlinear Regression Models• Pharmacological models for time series analysis
– AFNI’s nonlinear regression program 3dNLfim• Michaelis-Menton dynamics for BOLD FMRI with
psychoactive drugs (e.g., ethanol)
• Dynamic Contrast Enhanced MRI for quantifying Gd contrast leakage through blood-brain barrier
Feature: Smart Blurring• FMRI time series datasets are often smoothed
(blurred) in space to– Reduce noise (by averaging)
– Increase intra-subject activation “blob” overlap• Blurring brain & non-brain signals together is silly• When combining data from different scanners (i.e.,
multi-center studies), image smoothness varies– Should blur images until they have the same
level of smoothness so that inter-scanner combinations make statistical sense
• Developed a method for blurring inside a mask that stops when image noise reaches specified level of smoothness: ut =∇⋅[D(x,t)∇u(x,t)]
Feature: Structural Equation Modeling• SEM is a form of connectivity analysis• Input: correlations between activated ROIs
– Regions where the activations fluctuate in strength together will be more highly correlated
• Input: connectivity diagram between ROIs• Output: strength of connections• Can also search for “better” fitting connections
Feature: All-in-One Analysis Program• Common complaint: “AFNI is tooooooo hard to use”• Analysis of single subject data involves several
steps, each instantiated in separate programs– Registration, smoothing, normalizing, model
analysis• Solution is a program afni_proc.py that will run all
these programs in a coherent sequence– Intermediate results are saved to make it
possible to track backwards when results are confusing
• This script is not intended to let the user avoid understanding the data analysis process!
Feature: Diffusion Tensor Analysis• Goal: Compute the Diffusion Tensor (etc.) from
Diffusion Weighted image collections– Problem #1: log+linear method is inaccurate in
highly anisotropic locations (the cool places to be)– Problem #2: published nonlinear solution
methods not available in open-source software• Solution was to create and implement an efficient
robust nonlinear method for finding the diffusion tensor D in each voxel– Also, a optional nonlinear image smoother (2D
and 3D) to reduce noise in homogenous areas• Our code now incorporated into DTI Query, an
open-source tractography program from Stanford
Feature: Inter-Modality Registration• Goal: Efficiently align 3D volumes acquired with
different imaging contrasts• Solution is a general program using histogram-
based measurements of image matching (e.g., mutual information)
• This one is still very much a work-in-progress– Works pretty well on “simple” cases (e.g., whole-
brain to whole-brain)
– Dealing with partial-brain to whole-brain and with brain images that have holes in them is less reliable right now
– Also want to add non-affine warping capabilities
Example: Inter-Modality Registration A
Skull Stripped MRI
… masked CT
… CT overlaid on MRI in color - unaligned
… CT overlaid on MRI in color - aligned
Feature: Analysis of Mn Contrast MRI
• Mn is an MRI contrast agent and a calcium analog• Goal: time-dependent in vivo tract tracing in
monkeys• Problems abound:
– Like FMRI, signal changes are small– Other artifacts from day-to-day scanning are
larger– Simple image subtraction isn’t reliable
• Next 3 slides: some data and results …
Mn Data: Different Days A
Mn Data: Subtract & t-Test A
Mn Data: Cleverer t-Test A
Features Added to AFNIAFNI and SUMASUMA in Response
to Our Own Crazy Thoughts
(mostly)
Next Set of Slides
Functional Functional activationactivation & Motion estimation in realtime
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
AFNI
Dimon
MR Scanner, Image Files
Realtime FMRI
Feedback Receiver
Surface-Based Analyses• Create cortical surface models, project 3D data to
these surfaces, analyze in that space– Respects geometry and topology of cortex
• Most AFNI statistical tools now work with image data defined over surfaces as well as over 3D volumes
QuickTime™ and aYUV420 codec decompressor
are needed to see this picture.
• Movie capture from SUMA• Activation map projected from AFNI
Visualization & Links Between Modes
NIfTINeuroimaging Informatics Technology Initiative
• Goal: facilitate inter-operability of FMRI data analysis software
• First fruit: NIfTI-1.1 standard for storing datasets defined over 3D volumes (plus time axis)– Works with AFNI, FSL, SPM, BrainVoyager, …
• Agreement is not a one-time thing– Ongoing process is needed to deal with compatibility, extensions,
new ideas along the same line, …
• Efforts underway:– NIfTI-G: standard for storing cortical surface models (and
associated data)– NIfTI-W: standard for storing non-affine spatial warps
• Programs “talk” to each other (esp. AFNI & SUMA)• Exchange data• Issue commands - you can script many parts of the AFNI & SUMA graphical interfaces
3dSkullStrip
SUMA
AFNI
Closely Linked Communication
Developer-friendlinessRealtime physiological monitoring using AFNI:Jerzy Bodurka, FIM/LBC/NIMH
• Train Support Vector Machine (SVM) classifier on a collection of pre-categorized 3D brain images
• e.g., “looking at house” and “looking at face”
• Classifies new 3D images into the categories
Brain State Classification
R L
From Stephen LaConte;Emory, transitioning to Rice
• Much of our most fruitful and satisfying work comes from close and ongoing interactions with investigators that have interesting problems– Derived from studies that are pushing the
envelope of deriving information from MRI
• We are here to provide solutions to problems (of image analysis)– Your current short-term problems (lots of these!)
– Your actual longer-term problems– What we think your future needs will be
Penultimate Slide
Ultimate Slide
top related