Jump to: navigation, search

[edit] Kendrick's fMRI tests (2011/02/02)

On Feb 2, 2011, Bob and Kendrick scanned Keith Main at CNI. The data included a simple fMRI task and also some anatomical data. The data are at ~knk/multiclass/KM20110202. The data included: [3 repetitions of a BRAVO T1 FSPGR whole brain scan], [3 fMRI EPI runs], [T2 inplane FSE], [T2 3D whole brain], and [T2 3D whole brain at high-res].

Visual display:

  • We used a large LCD monitor placed in the control room, facing the scanner room.
  • Subject viewed the monitor through mirrors mounted on the coil. Approximate visual field-of-view was 5-6 deg in the vertical direction.

Details on the visual stimulus:

  • Trials were 8 s (3 s stimulation, 5 s rest).
  • There were 13 distinct stimuli (various noise patterns presented in a circular aperture).
  • Each stimulus was presented once in each run and in random order.
  • Each run was about 2.5 min long. 114 TRs.
  • 16 s of rest at beginning and end of each run. Stimuli were presented in blocks of 5 trials with 1 rest trial in-between.
  • There were a total of 13 stimuli x 3 runs = 39 stimulus trials.

EPI details:

  • 24 slices @ 2.5 mm, 64 x 64 matrix size, 160 mm x 160 mm FOV, 1336.437 ms TR, 29.7 ms TE
  • Slices were near-axial, tilted downwards to go through the temporal lobe.
  • Frequency-encode direction was A/P to avoid aliasing from the nose.
  • 2x acceleration was on. (Perhaps this caused SNR decrease; however, it did allow 24 slices to be acquired (vs. 21 slices at the Lucas Center)).
  • The scan and the stimulus were manually started simultaneously (or pretty close to simultaneously).


  • Discard first five volumes, slice time correct, motion correct.
  • Fit GLM model. This model used an "average" HRF measured from different subjects (for a 3-s visual stimulus) and included 13 beta weights, one for each stimulus. The model also included polynomials to estimate baseline and drift. The GLM model was fit using n-fold cross-validation. Model accuracy was quantified using R^2 (which was calculated after projecting polynomials out from both the data and the model prediction).


  • Here's the inplane (downsampled to the resolution of the functional data):


  • Here's a sample functional volume:


  • The distortion of the functional volumes appears to be very low. (However, it's hard to show this via a wiki page). Also, the dropout seems quite low, except for the expected large dropout near the temporal lobe. Also, the stability of the distortion characteristics is quite good from run to run (again, this is not evident from this wiki page). All of these observations imply that either the field is quite well shimmed and stable and/or the 2x accelerated EPI sequence is quite good at handling distortion.
  • Here's a map of the cross-validated R^2 of the GLM:

Knk-20110202pcr r05.png

Black corresponds to R^2 values that are 0 or less. So, everything that has a color is almost certainly a true "activation". These results are encouraging that BOLD activity can be detected reliably. However, unfortunately I don't have a dataset that these results can be directly compared to (so I don't know if the number of activated voxels is high or low given the amount of data we acquired).

  • Instead of looking at the R^2, we can use a different metric in order to compare to existing data that I've acquired at the Lucas Center. Specifically, consider the figure put at the bottom of this page. In that figure, notice that the CNI EPI 2x is not very red (unlike most of the data at the Lucas Center). This suggests that the temporal SNR of the data that we acquired was low. Perhaps this can be remedied by downgrading to 1x acceleration or by changing RF coils.

To do:

  • Use Freesurfer to process the whole-brain anatomical data?
  • Do a direct comparison of Atsushi's spiral and EPI.
  • Evaluate different coils with respect to temporal SNR.
  • Try 1x EPI acceleration. Play with the EPI phase-encode direction.
  • Keep 21 slices, but reduce the TR (to get more data points and therefore better effective SNR).
  • Need to be able to trigger the scan automatically.

[edit] Kendrick's fMRI tests (2011/02/07)

On Feb 7, 2011, We scanned Mehdi Senoussi at CNI. The data included a simple fMRI task and also some anatomical data. The data are at ~knk/multiclass/MS20110207. The data included: [2 repetitions of a BRAVO T1 FSPGR whole brain scan], [3+1+3+1 fMRI runs], [T2 inplane FSE], and [T2 3D whole brain at high-res].

Same visual display and visual stimulus as the previous tests. There were four datasets:

D1. 3 stimulus fMRI runs: EPI 2x, 21 slices, 2.5 mm ^ 3, 1145.517 ms TR, 28.7 ms TE, flip 65
D2. 1 rest fMRI run: same as above, except EPI 1x, 1336.437 ms TR, flip 68
D3. 3 stimulus fMRI runs: spiral, 1336.437 ms TR
D4. 1 rest fMRI run: same as above


  • Here's a sample functional volume from D1:

Knk-20110207 2xraw.png

  • Here's a sample functional volume from D2:

Knk-20110207 1xraw.png

  • Here's a map of the cross-validated R^2 of the GLM from D1:

Knk-20110207 2xr2.png

This is really poor CNR.

  • Here's the temporal SNR (stability) for D1:

Knk-20110207 2xcoeffalt.png

  • Here's the temporal SNR (stability) for D2:

Knk-20110207 1xcoeffalt.png

It appears that the reason for the poor CNR is the poor temporal SNR.

[edit] Temporal SNR summary from different datasets


Personal tools