Lecture 8: Issues in MRI and fMRI Analyses

1
Lecture 8Issues in MRI and fMRI Analyses
Child Psychiatry Research Methods Lecture Series

• Elizabeth Garrett
• esg_at_jhu.edu

2
Analysis of Imaging data

• New, exciting area of research growth (partly)
  due to increases in computer technology for
  handling data.
• ENORMOUS amount of data
• Two areas we will discuss today
• Volumetric MRI
• Functional MRI
• Todays talk will be tip of the iceberg, just
  to introduce some issues in statistical inference
  associated with imaging data. Very low on the
  quantitative scale.
• Other related topics
• MR Spectroscopy (MRSI)
• Diffusion Tensor Imaging (DTI)

3
Preface

• Two types of analysis that we hear about in
  MRI and fMRI studies
• (1) analysis that converts the readings from the
  MRI to data. This process is time intensive
  and takes a trained expert. There are multiple
  outputs from this process, including 3D pictures,
  volumes, intensities, etc.
• (2) analysis that uses the data from part (1)
  in a statistical way to answer scientific
  questions.

We will discuss ONLY (2). (1) is a whole other
can of worms.
4
MRI 3D image

• Intuitively, we are used to interpreting scans
  (like x-rays) qualitatively, not quantitatively.
• Example
• Consider a cube in the brain of only 10 by 10
  by 10 voxels.
• That is 1,000 data points!
• Data storage and data management gets costly
• Efficient methods are needed for dealing with
  this type of data.
• Statistical analysis requires data reduction.

5
Typical cross sections of MRI volume, 1mm3, 1 mm
1 pixel
6
Data representation
7
Side by side comparison
8
How do we quantify differences?

• Volumetric MRI
• Typically, we compare relative volumes of
  different areas of the brain
• Example Paula Lockharts comparison of children
  with Fetal Alcohol Syndrome (FAS) and controls
  (ADHD or normals).

9
(No Transcript)
10
How do we reduce dimensionality?

• Calculate volumes for each area of interest.
• Normalize to the overall volume (?)
• For each area, you have one number describing the
  area.
• Can compare areas in FAS kids to control kids.
• We can do this using simple statistical tools we
  already know.

11
Hypothetical results for cerebellar vermis volumes

• 20 FAS kids mean 400, s.d. 75
• 20 Control kids mean 450, s.d. 68
• t-test (degrees of freedom 20 20 - 2 38)

pvalue 0.033
12
Key Idea Data REDUCTION

• In Paulas example, a huge amount of data was
  reduced to one number for each child volume of
  cerebellar vermis.
• Other outcomes may be of interest
• Remember
• The statistical methods are not the hard part
  the hard part is summarizing an MRI.
• Cant say that things are different by visual
  inspection. Too subjective, although it might
  seem clear-cut.
• Converting from photos/scans to data takes a lot
  of space! Think about how much information you
  would need to try to reproduce a digital
  photograph and a photo is only 2 dimensions! A
  picture is really worth MORE than 1000 words.
• Try to choose measures that are as objective as
  possible.
• In embarking on an MRI study, you need someone
  who REALLY understands (1) (from preface) to be
  involved.

13
Analyzing Functional MRI (fMRI)

• All of the issues in MRI analysis are also issues
  in fMRI analysis
• But, fMRI is MUCH more complicated!
• Why?
• Functional shows how brain responds to stimulus
  over time.
• Not just a 3D picture of the brain, but a 3D
  image of how the brain is lighting up over
  multiple time points.

14
Example of lighting up
Moo and Hart, 2000
15
Easy Example of Modeling fMRI

• Design of experiment
• light turns on.
• 10 seconds off, 10 seconds on, 10 seconds off.
• Assumption
• areas that are activated will look like this

boxcar
16
How to figure out where boxcar occurs?

• At EACH voxel, plot intensity versus time

17
Assess statistically the boxcar model at each
voxel

• Perform a regression to see if activation is
  different when light is on versus off

Voxel i Time t
18
Regression Results

• Voxel 4235
• Coef Se t
  p
• on 0.024 0.014 1.70 0.09
• _cons -0.01 0.009 -1.11 0.27

• Voxel 947
• Coef Se t
  p
• on 0.99 0.014 70.39 lt0.001
• _cons -0.007 0.009 -0.78 0.44

19
Now what?

• Say you look at 2000 voxels.
• Then, you have 2000 t-tests.how can we summarize
  them?
• Consider image
• you can plot the significant voxels (see image)
• look at different areas of the brain
• look for patterns of activation
• for example,
• If you find that only voxel lights up in the in
  an area, that is likely not an interesting
  finding.
• If you find that a large proportion of voxels in
  an area in close proximity light up, then you
  have found something interesting.

20
Focus in the previous example

• Area of activation
• If t-test for ?1 is significant, then we say
  that the voxel is lit up.
• We ignore HOW lit up it is, just that it is lit
  up.
• This example has focused on trying to assess AREA
  of activation.

21
Alternative approach

• Consider intensity of activation

p lt 0.001
22
Intensity of Activation

• In both voxels, activation was significant.
• But, in voxel 947, the level of activation was
  much higher.
• We may be more interested in voxels that show
  greater intensities.
• Consider image see high intensity in the center
  of an activated region, tapering as move away
  from center.

23
Using data from multiple subjects

• Add a random intercept/effect for each
  individual.
• For each voxel, you are averaging across many of
  the intensity plots like we saw in previous
  slides.

Voxel i Time t Subject n
24
Longitudinal fMRI Studies

• How does function change in degenerative disease,
  such as multiple sclerosis?
• Can use linear regression again.
• Consider previous example
• what if they arent too different voxels, but two
  different time points?
• How do we assess if function, within a voxel, is
  deteriorating over time, quantitatively?

25
(No Transcript)
26
Longitudinal Model with One Subject
Voxel i Time t Year j
27
Linear Regression Model
28
Longitudinal Model with Multiple Subjects
Voxel i Time t Subject n Year j
29
Things to consider in fMRI studies

• Area of activation and intensity of activation
  are the usual outcomes of interest. Must
  specify!
• Complicated tasks are hard to analyze
• already established finger tapping, light on/off
• still under investigation Go - No Go
• Multiple subjects
• Method (i.e. linear regression) translates nicely
  to multiple subjects
• recall repeated measures and longitudinal
  analysis
• include a subject-specific adjustment and data
  from across individuals
• can be analyzed simulataneously.
• Longitudinal Analyses
• Not that much more difficult conceptually than
  the multiple subjects example
• Data management issues are the limiting,
  complicating factor
• When longitundinal data AND multiple subjects,
  need to consider variance structure (not
  addressed today).
• Study design is critical in these experiments
  number of subjects, tasks chosen

30
Biostatistical Resources and Further Studies

• Biostatistics Consulting Center
• In the biostatistics department, SPH.
• http//www.jhsph.edu/biostats/consulting.html
• fMRI journal club (for serious fMRI stuff)
• fmri_jc_at_yahoogroups.com
• Psychiatric Neuro-Imaging people

31
Biostatistics Courses

• Biostatistics 611-612
• Provides the basic tools for reading medical
  literature.
• Focuses on understanding the terminology (e.g.
  pvalue, odds ratio, regression), and very little
  on equations.
• 2 terms (one semester)
• summer epi institute (3 weeks intensive)
• 1st and 2nd terms
• If you want to understand statistics at a
  reading level, but never want to do data
  analysis, this would be the course for you.

32
Biostatistics Courses

• Biostatistics 621-624
• Provides tools for performing data analysis.
  More in depth and hands-on than 611-612
• Data analysis using Stata
• 4 terms (can take fewer)
• 2 sections, both beginning in 1st term.
• If you want to be able to do your own statistics,
  sample size calculations, etc. but do not want to
  get overly involved in the mathy aspects of
  statistics, this is the course for you!

33
Biostatistics Courses

• Biostatistics 651-654
• Calculus-based statistics
• Only for those who want to REALLY understand
  statistics
• 4 terms, 1 section, beginning 1st term.
• Learn everything from 621-624, but in more depth.
• Only take this course if you are a strong math
  student and calculus is fresh in your mind.

34
Other related courses offered at JHMI

• Clinical Research Methods
• 2 weeks intensive (9-5 every day)
• Epi and biostats
• This full time course is intended for clinical
  post-doctoral fellows and junior faculty of the
  School of Medicine.
• www.jhsph.edu/gtpci/icc.html
• Science of Clinical Investigation
• 4 courses taught in sequence throughout the
  academic year.
• Each course consists of approximately eight
  three-hour sessions held on consecutive Monday
  evenings from 530-830 PM.
• Goal To enhance course participants' theoretical
  understanding of and practical skills in the
  design, implementation, analysis, and
  interpretation of data from clinical
  investigations.
• http//www.jhsph.edu/gtpci/cis.html