[Homer-users] problem

[thuppert]

Baseline data is used in the principle component filter to remove systemic
fluctuations.

To do this:
1)  load multiple experimental files (including functional and a blank
(baseline) file).  Do this by loading one at a time or using the multiple
select option on "open file"

2)  Mark the baseline file as such.  Near the pulldown menu displaying the
open files, there is a check box labeled "baseline data?".  Switch to the
file(s) which represent baseline data and check this box  (leave the box
unchecked for all functional runs).

3)  After selecting the filtering parameters, click the "Update All" button
to have these settings applied to all files.  The program will process the
baseline file(s) first, calculate the principle components from this file,
and then apply the PCA filter to the functional runs.


Documentation for this filter is described in:
Zhang Y, Brooks DH, Franceschini MA and Boas DA. Eigenvector-based spatial
filtering for reduction of physiological interference in diffuse optical
imaging. Journal Biomed Optics 10(1): 011014. 2005.


For the second part of the question- how many components to remove?
Unfortunately, there is no magic answer to this.  PCA works by breaking the
spatial covariance matrix (calculated from the baseline data) into its
"characteristic" components.  The strongest component represents the most
spatially covariant "signal" in the data.  In general, this can be used to
approximate systemic fluctuations as these are pulsing across the entire
head.  This component is then projected from the functional run, which
assumes that this spatial covariance (if due to physiology) will persist in
the functional data.  The two PCA filters #2 and #3 differ only by whether
the covariance is calculated from the optical density (i.e. per wavelength)
or from hemoglobin concentration changes (HbO and HbR).  If there were no
instrument noise (i.e. the only source of error in the data were from
spatially covariant physiology), both of these would work the same  (and the
3rd filter would be the better choice).  In real data, since one wavelength
(in our case 690nm) is generally nosier (lower SNR) the 2nd filter is
sometimes the better choice.  It sometimes depends on the data.

Instrument noise is random and therefore is NOT spatially covariant.  If the
dominant source of noise in the data is instrumental (on either the baseline
or functional), than the PCA filter will not work very well.  If you  prune
low SNR channels first (deselect them on the probe geometry- which discludes
them from the calculation of spatial covariance and the effects of the PCA
filter), this will help.  PCA works best with a large probe, such that the
spatial covariance due to systemic fluctuations is much larger then the
functional activation area.

As a general rule for selecting how many components (once you have
understood the math which you are actually doing) is to look at the
singular-value spectrum (the button "view SV spectrum")   This is a plot of
the eigen-values for the components.  Ideally you would like to see the
majority (80-90%) of the spatial covariance stored in the first one or two
components  (i.e. a sharply saturating curve).  You do not want a slowly
increasing curve, as this is the case where the PCA will fail.  We find it
useful to remove the first 1-3 components (up to the 80%) point.

Also note, the PCA filtered time-course are the ones labeled "-covariance
reduced".  I.e. to view the effects of the PCA filter (#2) compare the time
course for dOD vs dOD-cov reduced.


Ted Huppert, M.Sc.

PhD student-Harvard Univ.
Dept of Biophysics
Photon Migration Imaging lab
Mass General Hospital/CNY

Tele: (617)726-9338

thuppert at nmr.mgh.harvard.edu



  -----Original Message-----
  From: yzhu [mailto:yzhu at nlpr.ia.ac.cn]
  Sent: Thursday, March 30, 2006 4:16 AM
  To: thuppert at nmr.mgh.harvard.edu
  Subject: [Homer-users] problem


  Dear All,

     I am not clear how to use the baseline data. Could I use the data which
was collected during subjects resting as the baseline?  And how to choose
the number of singular-vectors when I use the PCA filter #2&#3?

  Thanks for your help!
  Regards,
  Zhu Ye
  -----Original Message-----
  From: yzhu [mailto:yzhu at nlpr.ia.ac.cn]
  Sent: Thursday, March 30, 2006 4:16 AM
  To: thuppert at nmr.mgh.harvard.edu
  Subject: [Homer-users] problem


  Dear All,

     I am not clear how to use the baseline data. Could I use the data which
was collected during subjects resting as the baseline?  And how to choose
the number of singular-vectors when I use the PCA filter #2&#3?

  Thanks for your help!
  Regards,
  Zhu Ye
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/homer-users/attachments/20060330/3888197d/attachment.html