Skip to content

MRTraining2.md

Latest commit

 

History

History
683 lines (439 loc) · 13.6 KB

MRTraining2.md

File metadata and controls

683 lines (439 loc) · 13.6 KB

footer: MR Methods for Faculty II, 10/6/17 slidenumbers: true

#MR Methods for Faculty II

Roeland Hancock

Associate Director, BIRC

Assistant Research Professor, Psychological Sciences

6 October 2017

#Internal Reproducibility

You should be able to effortlessly reproduce your own analysis

  • Version Control
  • Data provenance
  • Automated, reproducible computations
  • Documentation
  • Replicable computational environments

#Review

Internally reproducible research

  • Organize your data (BIDS)
  • Use scripts
  • Use version control (git)
  • Options for making reproducible software environments (singularity]

#Review

fMRI task design

  • Block vs event designs
  • Detection vs estimation efficiency
  • Design optimization is important

#Review

Resources for design optimization

fit,original

#Group vs Individual Activation

  • Group average data does not reflect individuals
  • Group analyses may not be reliable if the underlying activation is highly variable

#Low Within Subject Reliability

  • Test-retest fMRI reliability is poorly characterized
  • 10s of thousands of publised task fMRI studies
  • 63 with test-retest measures (Bennet et al., 2010)
  • Low test-retest reliability (mean ICC = .5; mean overlap = 29%)

#Limitations of Reliability

Poor reliability limits scientific value

  • Between group analyses
  • Brain-behavior correlations
  • Neurogenetics
  • Functional localizers
  • Databasing

#Improving Reliability

  • Increase SNR
  • Optimize design power
  • Minimize confounds (time of day, attention, practice)
  • But fMRI is fundamentally noisy
  • Select tasks with high reliability and/or do test-retest on your own data

fit,original

#Today

  • Data acquisition
  • Preprocessing for volumetric analysis
  • Single subject statistics (mass univariate)

#Implementing Your Task

  • Use your preferred stimulus presentation software
  • Check that timing is relatively consistent
  • Target output: event or block onset times, temporally aligned with BOLD data
  • Use the scanner TTL pulse to align your fMRI data

#fMRI Scan Sequence

Nominal setup at BIRC; other sequences are possible

  1. Researcher preps the experiment on stimulus PC
  2. Operator starts the BOLD sequence
  3. Dummy volume collection
  4. Scanner sends a trigger pulse
  5. Experiment starts

#Scanner Pulses

  • The scanner sends a 5 keystroke at the beginning of each volume that is kept
  • Wait for a 5 response to start the experiment
  • Offset event times based on the initial 5 response
  • More sophisticated timing setups are possible

#Testing your Experiment

Before scanning, check that

  • Your event timing is consistent between runs
  • You understand the timing variables
  • You understand how to align your events with fMRI data

#BOLD Display

right

See the MRI Display article on the BIRC wiki for visual angle calculations

#MRI Protocol

  • Develop a general idea of the scan parameters you want
  • Consider existing literature
  • Major projects: ADNI, HCP
  • Special considerations:
    • Specialized analysis methods
    • Unusually high temporal or spatial resolution
    • Challenging brain regions

#Parameters Spatial resolution:

  • 2-3 mm3 for typical fMRI and DWI
  • Spatial resolution vs SNR (2mm3 is a ~70% reduction in SNR from 3mm3)
  • Spatial resolution vs acquisition time
  • How big is your primary structure of interest?
  • Do you have a high movement population?

#Parameters Temporal resolution:

  • Typically one sample every few seconds
  • Faster is better...
  • Less motion between samples
  • Less physiological aliasing
  • More degrees of freedom
  • ... to a degree

#Parameters Temporal resolution:

Definitely choose the fastest unaccelerated sampling possible.

If you need more speed:

  • Partial imaging (iPAT/GRAPPA/ASSET/SENSE)
  • Fills in part of the MR signal
  • SNR cost (√2 or more)
  • Increased motion sensitivity

#Parameters Temporal resolution:

  • Simultaneous multislice (SMS/multiband)
  • Minimal SNR cost
  • Some increased motion sensitivity in some cases
  • <1s sampling possible
  • Also works for DWI

#Parameters

Other considerations:

  • Do you need whole brain coverage (partial coverage is faster)?
  • Are you interested in regions prone to signal loss (vOFC, amygdala, temporal regions)?

#Hardware Options

BIRC has a 20-channel head coil and 64-channel head/neck coil

64-channel:

  • Better SNR, especially at the surface
  • Necessary for SMS or high acceleration
  • Smaller
  • More heterogeneity

#Flip Angle

  • Flip angle partially determines how much signal you get
  • Maximum signal at 90º for an unexcited sample
  • For partially excited samples, maximum at the Ernst angle
  • Maximum tSNR in BOLD at lower flip angles

#TE

  • Optimal signal when TE matches tissues T2*
  • Optimal TE varies across the brain, ~20-40ms
  • Regions with signal loss (temporal, vOFC) have shorter T2* (<30ms)
  • Regions with good signal (occipital cortex) have longer T2* (~40ms)
  • Slices take longer to acquire with longer TE

#Bias Correction

  • Receive coil channels have different spatial sensitivities
  • Can be corrected online (Prescan Normalization)
  • Or offline
  • Some sequences save corrected and uncorrected data

#Selecting Parameters

  • There is no set of universally best parameters
  • Consider
  • Your brain regions of interest
  • The expected level of subject movement
  • Your hypotheses and analysis needs
  • Pilot and use tSNR as a guide

#Scans (for fMRI)

Must have

  • Scout for positioning scans (~30s)
  • T1-weighted (~5-7 min)
  • Standard structural volume for normalization
  • Some type of field map (~10-120s)
  • Correct for EPI distortion
  • fMRI

right

#Time Management

  • Prep and setup: ~10 min
  • Scout, T1, fieldmap: ~6-8 min
  • Cleanup: ~5 min
  • Checkin/instructions/setup between scans: ~1 min
  • ~35-40 minutes for fMRI in a 1 hour booking

#After the Scan

  • Data appears in NiDB
  • Download DICOM data
  • Convert to BIDS
  • Preprocessing and analysis

#fMRI Processing and Analysis

  1. Quality control
  2. Minimal preprocessing
  3. Quality control
  4. Subject level statistics
  5. Group level statistics

#Quality Control 1

Make sure you have the expected data:

  • Files for each expected MRI series
  • Behavioral log files
  • Do data volumes have the correct dimensions?
  • Were the correct scan parameters used?
  • [BIDS-validator] (https://github.com/INCF/bids-validator) can help

#Quality Control 2

Check fMRI data for

  • Ghosts (signal outside the brain)
  • tSNR
  • Large initial intensity
  • Use QAP or mriqc
  • BIDS compatible!

#Minimal preprocessing (fMRI)

  • Slice time correction
  • Motion correction
  • Distortion correction
  • Co-registration

original,fit

original,fit

#Slice Time Correction

  • Interpolate the time series from each voxel to effectively align the data to a common time point
  • Alternative solutions
  • Ignore (for short TRs)
  • Model derivatives of the HRF

#Processing Order

There are different approaches to slice time correction:

  • Not at all
  • Before motion correction
  • Head movements can shift voxels in and out of the head-bad for interpolation
  • After motion correction
  • Motion correction can shift voxels into adjacent slices at different timepoints-particularly bad for interleaved acquisition

#Recommended Processing Order

  • Needed for effective connectivity
  • Always helps detection (Sladky et al, 2011)
  • but maybe not much if TR is short
  1. Denoise the data (spikes spread during interpolation)
  2. Correct for timing before motion correction if slices are interleaved

#Slice Order When were the slices acquired?

Varies by sequence and manufacturer:

  • ascending or descending
  • sequential or interleaved
  • starting from first or second slice

Or something more complicated, e.g. SMS

#Determining Slice Order

  • Check the raw data (e.g. DICOM images)
  • Always start with the raw data if you are unsure of scan parameters or don't trust processed headers
  • Inspect NIfTI headers (fslhd, nifti_tool) or .HEAD (3dinfo)
  • Usually alt+z2 (even # of slices) or alt+z (odd) here

Physics of Motion

Motion irreparably affects your data

  • Motion rotates the brain though regions of variable B0 inhomogeneity
  • Can alter correlations between brain regions
  • Introduce regions of signal loss
  • Moves regions of the brain between excitations
  • Introduces regions of changing signal intensity

#Motion Artifact

inline

#Minimize motion

  • Train subjects in a mock scanner
  • Use padding or restraints
  • Emphasize importance of staying still
  • Monitor compliance

#Realignment

Goal: put voxels in the same place throught the scan

  • Spatially correct for movements from volume to volume in an EPI time series
  • 6 degrees of freedom (DOF)
  • rotation (roll, pitch, yaw)
  • translation (x, y, z) shifts
  • Interpolate voxels to a fixed grid

original

#Choices

  • Another b0 volume that will be used with anatomical alignment
  • For example the reference image from a multi band time series
  • First volume-maybe higher signal
  • Third volume-maybe a better match to the rest of the time series
  • Middle volume-minimize interpolation distance
  • Min outlier volume-a volume with minimal artifacts

#EPI distortion

  • Field inhomogeneity distorts EPI along the phase encode axis
  • Particularly problematic for DWI
  • Ideally correct for this using a fieldmap and/or nonlinear alignment

A/P

left

right

L/R

left

right

#Approaches

  • Collect a 'real' fieldmap
  • Collect multiple scans with flipped PE direction
  • Nonlinear warping

#Distortion correction

  1. Calculate a field map
  2. Align the fieldmap and EPI
  3. Unwarp the EPI distortion
  4. Best done with motion correction

FSL tools: fugue and topup

#Co-registration

Goal: Align the fMRI data to another dataset

  • Typically a T1-weighted anatomical volume
  • The T1 (and aligned fMRI) can then be aligned to a template

#Co-registration options

  • Rigid body
  • Not ideal between modalities
  • Affine transformation
  • Also adjusts shearing and scaling
  • Doesn't address geometric distortion
  • Non-linear
  • Addresses distortion
  • Recently recommended (Friston et al., 2017)

right

#Preprocessing Pipelines

  • AFNI and FSL do some or all steps
  • fmriprep
  • Also generates images for review

#Quality Control

  • Establish acceptable criteria before analysis
  • Mix of visual inspection and quantitative mettics
  • Visual
  • Inspect registration
  • Quantitative
  • Motion
  • tSNR

#Quality Control Tools

#Statistics

  • Smoothing to increase SNR
  • High pass filtering (.01-.02 Hz) to remove slow drifts
  • Filtering needs to be considered during design
  • Scaling
  • Single subject GLM

#Motion Regressors

  • Effort to account for motion-relate signals
  • Realignment produces 6 (translation and rotation on 3 axes) motion parameters
  • Motion effects extend over time
  • Include derivatives or other expanded parameter sets

#Motion Metrics

  • RMS: absolute volume displacement
  • FD (framewise displacement): volume to volume displacement
  • DVARS: volume to volume changes in intensity

#Censoring

  • aka scrubbing
  • Remove motion contaminated volumes from analysis
  • Possibly some successive or prior volumes
    • Delete data (affects filtering and model)
    • Regressors
    • Interpolation (affects df)

#Statistics

  • BOLD response is modeled in a GLM
  • Stimulus onsets * HRF or basis
  • GLM includes motion regressors and other confounds

original,fit

#Basis Functions

  • 'Canonical HRF': a double gamma response
    • Assumes the HRF has a particular shape
  • FIR, tent, sine, spline
  • Estimates HRF

#Stimulus Functions

  • Delta function
  • Each event is instantaneous
  • Block function
  • Events extend in time
  • HRF reaches a set maximum after time

#Design Matrix

  • Specify a basis and stimulus function for each condition
  • Convolve the basis and stimulus functions
  • Add motion and other non-task regressors

#Statistics

  • GLM (OLS or REML)
  • Linear constrasts between conditions
  • and/or F tests
  • Result: t/z/F values and beta values

#Multiple Comparisons

  • Mass univariate statistics are over 10-100s of thousands of tests
  • Statistics need to be corrected for number of comparisons
  • Corrections should account for spatial structure

#Correction Options

  • Familywise Error (FWE)
  • False Discovery Rate (FDR)
  • Random field cluster-based correction
  • Threshold based statistics (TBSS)
  • Permutation cluster-based correction
  • Permutation