Computer login: labuser2 BR@ins08 PREPROCESSING Removing - - PowerPoint PPT Presentation

Computer login: labuser2 BR@ins08

preprocessed func,onal func,onal ,meseries MODEL FITTING

General linear model using:

• Canonical HRF
• FIR ‘sFck’ funcFons
• Other basis funcFons

STATISTICAL INFERENCE

Thresholding: FWE, FDR, GRF

parameter es,mates & sta,s,cal maps ‘ac,va,on’ maps Plots

contrast vector design matrix

PREPROCESSING

CondiFoning Data:

• rescaling / inten. norm.
• temporal filtering
• spaFal smoothing

Removing ArFfacts:

• spike detecFon
• slice Fming
• moFon

correcFon

• B0 unwarping

Single‐subject analysis

preprocessed func,onal func,onal ,meseries ‘raw’ k‐space data MODEL FITTING

General linear model using:

• Canonical HRF
• FIR ‘sFck’ funcFons
• Other basis funcFons

STATISTICAL INFERENCE

Thresholding: FWE, FDR, GRF

REGISTRATION / NORMALIZATION

parameter es,mates & sta,s,cal maps

RECONSTRUCTION

high res. anatomical standard template ‘ac,va,on’ maps Plots

contrast vector

ROI extracted ,mecourse

design matrix

PREPROCESSING Single‐subject analysis

CondiFoning Data:

• rescaling / inten. norm.
• temporal filtering
• spaFal smoothing

Removing ArFfacts:

• spike detecFon
• slice Fming
• moFon

correcFon

• B0 unwarping

Slice‐Timing

WHY

• Different slices are collected at

different Fmes in TR WHEN

• debatable: some say not

necessary, as effects are small – may be more helpful if averaging across slices

• BEST: with moFon correcFon
• therwise: some say before

moFon correcFon, some say a\er

HOW

• FSL ‘sliceFmer’ or GUI, SPM GUI

MoFon CorrecFon

WHY

• If a subject moves during (or

between) scans, then a chunk of Fssue may move to a different voxel WHEN

• typically always, unless using a bite‐

bar

• before filtering, smoothing, a\er

spike correcFon HOW

• notoriously difficult
• FSL: mcflirt or GUI, SPM: realignment, mrTools

SpaFal Filtering (smoothing)b

WHY

• increase SNR by averaging noise
• f nearby voxels
• increase likelihood of overlap in a

standard space

• Gaussian random fields and other

correcFons require ‘smooth’ data

WHEN HOW

• convoluFon w/ a Gaussian kernel
• FSL fslmaths or GUI, SPM smoothing
• if expected acFvaFons are

relaFvely large

• when averaging across subjects in

a volumetric space

• NO: ROI or individual subject

(debatable)

• Before temporal filtering

Temporal Filtering / Detrending

WHY

• Increase model fit by removing low‐

frequency changes not related to protocol WHEN

• almost always
• as a last step
• except perhaps for ‘resFng state’
• r other scans which look at

slowly varying signals, or if doing a frequency analysis

HOW

• high‐pass filtering: gaussian

convoluFon in Fme

• FSL fslmaths –bpd, GUI
• SPM: in model specificaFon
• MODEL is also filtered
• vs. Detrending
• matlab scripts?
• inclusion in model

Model Fifng

• General Linear Model (GLM):

Y = Xβ + ε

data explanatory variables weights error Worsley, Keith J. (2001) Ch. 14: StaFsFcal analysis of acFvaFon images. In Func%onal MRI: An Introduc%on to Methods. Jezzard, Mathews, Smith ed. Oxford. ε ~ Nn(0, Vσ2) = * +

= * + X

O

sFmulaFon hemodynamic response funcFon (HRF)

ROI Analysis

• ExtracFng data from an ROI:

– increases signal‐to‐noise (assuming the voxels are actually doing the same or similar thing) – generally avoids the mulFple comparisons problem – also avoids the need for normalizaFon across subjects – results in a massive data reducFon

• write simple scripts! plot it!