Optimized design and analysis of Optimized design and analysis of - - PowerPoint PPT Presentation

Optimized design and analysis of Optimized design and analysis of sparse-sampling fMRI experiments sparse-sampling fMRI experiments

• f speech and hearing
• f speech and hearing

Tyler Perrachione John Gabrieli Satrajit Ghosh

Tyler Perrachione -- tkp@mit.edu May, 2011

2

Overview Overview

• What is sparse-sampling in fMRI and why is it

useful?

• What can we do to optimize sparse-sampling

design and analysis?

– Accounting for hemodynamics in sparse data analysis

• Computational modeling of sparse designs
• Human neuroimaging data

Tyler Perrachione -- tkp@mit.edu May, 2011

3

Sparse-sampling: What and why? Sparse-sampling: What and why?

• Acoustic noise from

switching MR scanner gradients

– Obfuscates auditory stimuli – Saturates activation of auditory cortices

• Head motion during vocal

responses adds noise, impedes registration

• Solution: Present stimuli

during silent delays between functional volume acquisition

Tyler Perrachione -- tkp@mit.edu May, 2011

4

What studies benefit from sparse? What studies benefit from sparse?

• Hearing & Language

– Children – Aging adults – Auditory impairment

auditory neuropathy, central auditory processing disorders

– Dyslexia

• Speech Production
• Neurosurgical mapping of

language regions

• Studies on older scanner

hardware

Schmidt et al., 2008

Tyler Perrachione -- tkp@mit.edu May, 2011

5

How can we optimize sparse fMRI? How can we optimize sparse fMRI?

• “Boxcar” models typically used

to represent sparse timeseries

– Ignores dynamic properties of hemodynamic response – Cannot distinguish the contribution of multiple event types to individual TAs

• Optimal delay between TRs?

– TR = 4? TR = 14? – Psychological and physiological factors

• Detection of activation in non-

auditory regions?

Tyler Perrachione -- tkp@mit.edu May, 2011

6

Simulating optimized sparse designs Simulating optimized sparse designs

• What delay is best?

– TR = 4, 8, 12s

• What model is best?

– Boxcar vs. HRF

• Generate a high temporal

resolution timeseries

• Convolve with canonical HRF
• Resample at points

corresponding to actual TA

• Add noise

– 5dB SNR – From actual resting fMRI

• Compare Models:

– 10,000 simulations in R

⊗

Tyler Perrachione -- tkp@mit.edu May, 2011

7

Simulating optimized sparse designs Simulating optimized sparse designs

Result of 10,000 simulations

Tyler Perrachione -- tkp@mit.edu May, 2011

8

Human neuroimaging data Human neuroimaging data

• Modified “Stroop” task

– Press button when color of text matches spoken color name.

• 14 control participants

– Right-handed, native- English, normal hearing

• Acquisition Parameters:

– TR = {4, 8, 12} sec – TA = 2 sec – 3.125 mm3 voxels – 5min {92, 47, 32 volumes}

• Analysis

– Functional data in SPM8 and NiPype

• Motion correction, volume

smoothing, within-subject model estimation

– Coregistration using Freesurfer 5.0.0 – Symmetric diffeomorphic nonlinear normalization in ANTS v1.5 red blue green

“green” “blue” “red” 

Tyler Perrachione -- tkp@mit.edu May, 2011

9

Human neuroimaging data Human neuroimaging data

Tyler Perrachione -- tkp@mit.edu May, 2011

10

Human neuroimaging data Human neuroimaging data

p < 0.001, FDR q = 0.05 L

Tyler Perrachione -- tkp@mit.edu May, 2011

11

Human neuroimaging data Human neuroimaging data

Convolved HRF Model Classic “Boxcar” Model

TR = 8 TR = 12 TR = 4

p < 0.001, FDR q = 0.05

L

Tyler Perrachione -- tkp@mit.edu May, 2011

12

Human neuroimaging data Human neuroimaging data

Model Comparison HRF > Boxcar TR = 4s

p < 0.001, FDR q = 0.05

L Simulations:

Tyler Perrachione -- tkp@mit.edu May, 2011

13

Human neuroimaging data Human neuroimaging data

TR = 4 TR = 8 TR = 12

Difference between task conditions Go > No-Go

p < 0.01, FDR q = 0.05 p < 0.01, FDR q = 0.05 p < 0.05, FDR q = 0.05

Tyler Perrachione -- tkp@mit.edu May, 2011

14

Conclusions Conclusions

• Optimizing sparse fMRI designs:

– HRF convolution significantly improves power of short (4s) TR sparse experiments – Short TRs improve detection of non-auditory activity – Very long TRs (12s) obscure activity in both auditory and non- auditory regions – HRF convolution can distinguish multiple conditions, allowing event-related sparse designs

• Increased power from optimized sparse designs:

– Reduces amount of scantime needed – Reduces length participants are in scanner – Reduces cost of fMRI experiments – Increases reliability of findings

Tyler Perrachione -- tkp@mit.edu May, 2011

15

Acknowledgments Acknowledgments

We thanks: Carlos Cardenas-Iniguez, Stephanie Del Tufo, John Lymberis, Steve Shannon, and Christina Triantafyllou. Imaging was performed at the Athinoula A. Martinos Imaging Center at McGovern Institute for Brain Research, MIT. This research was supported by the Ellison Medical Foundation, as well as NIH grant R03-EB008673 to S.G., and a NSF GRF to T.P.

Satrajit S. Ghosh John D.E. Gabrieli

tkp@mit.edu http://web.mit.edu/tkp/www/