Presentazione di PowerPoint

1
Real-time Independent Component Analysis of
functional MRI time-series A new TBV (3.0) Plugin
for Real-Time ICA during fMRI
2
Real-time ICA of fMRI data Outline

• Data model and analysis tools in real-time fMRI
• Sliding-window vs Cumulative approaches
• Data-driven analysis tools in fMRI
• Component-based generative models for fMRI
• Spatial independent component analysis (s-ICA)
• Real-time (spatial) Independent Component
  Analysis
• Data model and implementation
• The Sliding-window FastICA algorithm
• Perfomances, operation and user interface
• Examples of applications
• Motor activity, Auditory and emotional activity
  during music listening
• A New plug-in for Turbo BrainVoyager 3.0
• Example of application for visual activity
  monitoring

3
Data Analysis Tools for Real-time fMRI (1)

• Real-time fMRI enables one to monitor a subjects
  brain activities during an ongoing session
• Results are to be delivered (and used) in/near
  real-time, i. e. within times in the order of one
  (or a few) TR(s) ...
• Trade-off between accuracy VS computational
  times
• gt Minimum batch of temporal observations time
  points to generate a reliable activation map
  (statistical power)
• gt Minimum time window size s to cover the
  essential dynamics of the activaiton
  (hemodynamics, stimulus changes, ...)

4
Data Analysis Tools for Real-time fMRI (2)

• Real-time fMRI enables one to monitor a subjects
  brain activities during an ongoing session
• Results are to be delivered (and used) in/near
  real-time, i. e. within times in the order of one
  (or a few) TR(s) ...
• Trade-off between accuracy VS computational
  times
• lt Maximum batch of temporal observation to
  generate the activation map in real-time
  (bottleneck computational load)
• lt Maximum time window size s to promptly detect
  transient (or temporally nonstationary) dynamic
  effects before these become irrelevant and
  sacrificed in favor of more repetitive and
  temporally stationary effects (Mitra and Pesaran,
  1999).

5
Data Analysis Tools for Real-time fMRI (3)

• Real-time fMRI utilizes two different approaches
• cumulative window (Cox et al., 1995)
• sliding window (Gembris et al., 2000 Posse et
  al., 2001)
• In the cumulative approach
• the entire partially measured fMRI time-series is
  analyzed in one step. One edge of the time window
  is fixed, whereas the other moves during the
  acquisition of new data.
• the specificity (wrt repetitive/stationary
  effects) increases over time (more data become
  available for averaging).
• The sensitivity (wrt transient/non-stationary
  effects) is reduced (more fluctuations become
  relevant)
• The computational load increases over time
  (unless spatial or temporal resolution is
  sacrificed!)

6
Data Analysis Tools for Real-time fMRI (4)

• Real-time fMRI utilizes two different approaches
• cumulative window (Cox et al., 1995)
• sliding window (Gembris et al., 2000 Posse et
  al., 2001)
• In the sliding-window approach
• The analysis is restricted to the most recently
  acquired data. Both edges of the window move
  during the acquisition.
• The accuracy is constant over time and the
  sensitivity to dynamic changes in brain activity
  can be maximized.
• The specificity is limited and critically
  dependent on SNR
• The computational load is constant

7
Esposito et al., Neuroimage 2003
8
Data-driven tools in Real-time fMRI (1)

• Off-line, data-driven tools nicely and usefully
  complement by hypethesis-driven analysis tools
• E. g., independent component analysis (ICA) can
  identify brain activity without a priori
  temporal assumptions on brain activity
• No info about experimental paradigm (stimulus)
• No detailed information about hemodynamics
• Rough knowledge of potentially relevant areas
• ...

9
Data-driven tools in Real-time fMRI (2)

• Real-time fMRI data analysis is traditionally
  based solely on hypothesis-driven tools (e. g.
  GLM) because data-driven tools (such as ICA) are
• computationally demanding (time consuming)
• difficult settings (options, contrains and
  constants)
• e. g. convergence problems (no result delivering)
• difficult selection of the results
• post-hoc (complex) interpretation
• ...

10
Component-based Generative Models (1)
Measured fMRI time-series
C1
C3
Time (scans)
C2
Cn
11
Component-based Generative Models (2)
Mixing Unmixing
voxels
voxels
C1j C2j ... Cnj
COMPONENTS (C)
time
time
W-1(A)
DATA (Y)
Yj
Ai
Al
12
Principal Component Analysis

• Orthogonality Principle (simple linear
  decorrelation)

• Maximum variance principle (VARIMAX)
• (1) time-courses must be also orthogonal
  (uncorrelated)
• (2) components ordered by relative contribution
  to variance

13
Independent Component Analysis (1)

• Independency Principle (non-linear
  decorrelation)

• Information Theory Minimization of mutual
  information
• Maximize entropy flow of a neural network H(C)
  -gt max (Infomax)
• Maximize Non-gaussianity of components N(C) -gt
  max (Fastica)
• Statistical dependency is removed along one
  dimension (e.g. space)
• (1) time-courses can be correlated (spatial
  ICA)
• (2) components not ordered by relative
  contribution to variance

14
ICA vs PCA
Formisano, et al., Magnetic Resonance Imaging 2004
15
Independent Component Analysis (2)

• (Like PCA) ICA requires the computation of the
  data covariance matrix of the voxels time
  courses included in the analysis
• (Unlike PCA) spatial ICA only models the spatial
  distributions of brain activities (and builds
  accordingly the output maps)
• What ICA offers in addition to PCA does not
  depend on the covariance but only the spatial
  statistics
• While the statistical power of covariance
  estimation depends on the temporal window of
  observation (and the number of time points), the
  power of the spatial distribution estimation only
  depends on the voxel space

16
The power of spatial statisistics (1)
Signal
Features
Noise (pure)
17
The power of spatial statisistics (2)
Z-score (activation parameter)
18
Real-time ICA (1)

• The computational load of spatial ICA algorithms
  grows much more with the temporal dimension than
  with the number of voxels included in the
  analysis
• If we fix the temporal window the power of
  spatial statistics is constant. If the temporal
  window is large enough to ensure enough accuracy
  of the maps, the computation load can be held
  constant in a sliding-window approach
• In order to deliver components as fast as
  possible a deflation scheme can be used to
  extract ICA components one by one (FastICA
  algorithm by Hivarinen 1999). This renders the
  ICA component maps immediately available even in
  the presence of convergence problems.

19
The FastICA algorithm
one-unit function
deflation
multi-unit function
symmetric
20
Real-time ICA (2)

• Rt-ICA -gt sliding-window approach FastICA
• The window is chosen to solve the trade-off
  between accuracy and computational load.
• This approach works and can be useful if
• FastICA delivers useful and accurate components
  among the first extracted ICs in a relatively
  low number of iteration per run. If not, we
  cannot assume no activity
• The selection can be aided and supported by
  (rough) prior knowledge about where activity of
  interest takes place but selectivity should be
  unambigous
• Cumulative maps about a process of interest can
  be obtained by adequately tracking over time (and
  combining) subsequent sliding-window ICA
  components

21
Esposito et al., Neuroimage 2003
22
Esposito et al., Neuroimage 2003
23
Esposito et al., Neuroimage 2003
24
Esposito et al., Neuroimage 2003
25
Esposito et al., Neuroimage 2003
26
ICA in real-time fMRI during visual
stimulation A new plugin for Turbo Brain Voyager
3.0
27
ICA in real-time fMRI during visual
stimulation A new plugin for Turbo Brain Voyager
3.0
28
ICA in real-time fMRI during visual
stimulation A new plugin for Turbo Brain Voyager
3.0
29
ICA in real-time fMRI during visual
stimulation A new plugin for Turbo Brain Voyager
3.0
TBV LOG
Incoming Data
Real time ROI Selection
Data Pointer
ICA Component Rankings Spatial correlations and/or
other relevant parameters
RTICA PLUGIN
MAP VIEWER NeuroFeedback (MAP ANALYZER)
Ranked ICA Component Maps
30
ICA in real-time fMRI during visual
stimulation A new plugin for Turbo Brain Voyager
3.0
31
ICA in real-time fMRI during visual
stimulation A new plugin for Turbo Brain Voyager
3.0
32
Real-time ICA of fMRI data Conclusions

• Real-time ICA during fMRI is feasible in many
  circumanstances and has some potentials in
  monitoring brain activity under typical real-time
  fMRi settings
• The Sliding-window fastICA algorithm has
  comparable performances to GLM under highly
  controlled situations but requires no timing
  information and no critical settings
• This opens the possibility of monitoring
  non-triggered, non-repetitive and non-stationary
  neural activity with only mininal spatial prior
  on the networks involved
• Integration of rt-ICA generated maps in
  neurofeedback experiments now possible with the
  new Plugin for TurboBrainVoyager 3.0

33
Thank You!support_at_brainvoyager.com