Skip to content

Latest commit

 

History

History
executable file
·
269 lines (180 loc) · 7.64 KB

fmri_signals.rst

File metadata and controls

executable file
·
269 lines (180 loc) · 7.64 KB
Raw

FMRI signals

Here we will cover fMRI signals and their analysis.

This page content

The aim of this first part is to get familiar with fMRI data.

Warning

Carve in your mind that the only characters you can safely use in a file name are non-accented upper and lower case letters, numbers and two special characters: the underscore "_" and the dot "." . Don't cry if you run into problems with file names including accents and space.

Most of the problems one has to solve in any hands-on lab course are related to the students not knowing how to deal with the file structure of their own computer. If you are still unsure about concepts such as "home directory", "working directory", or "path", go back to your first lectures on python by Christophe Pallier and make sure you are able to explertly navigate in your file structure.

Before we start: installing the software

.. toctree::

 install.rst

Activation IRMf

Why do we do activation fMRI? Basically: localization Hypothesis: functional specialization Aim: given a target process, detect the brain areas involved Principle: the experimenter imposes on the subject a temporal variation related to the target process and looks for whichever part of the brain displays a signal in which the same variation could be found.

=> we need to get the time series of BOLD signal AND the time series of what happened to the subject AND they both have to be synchonous!

also

  • resting state fMRI
  • more than localization: connectivity

BOLD signal

BOLD signal: local vascular changes somehow related to increased brain activity can be detected by a MRI scanner (T2* signal increases). However the fluctuation of the signal induced by the BOLD effect is about few percents.

fMRI stimulation data

Same time reference than the MRI acquisitions. Describe the time series of conditions.

Reminder: we are looking for voxels in which part of the signal variance is explained by our manipulation, thus we need to know what is the shape of the temporal variation induces by our experimental paradigm.

Let's start with a basi example taken from a classical auditory stimulation dataset (exerpt from the SPM manual describing the paradigm):

This data set comprises whole brain BOLD/EPI images acquired on a
modified 2T Siemens MAGNETOM Vision system. Each acquisition consisted
of 64 contiguous slices (64×64×64 3×3mm×3 mm3 voxels). Acquisition took
6.05s, with the scan to scan repeat time (TR) set arbitrarily to 7s.
96 acquisitions were made (TR=7s) from a single subject, in blocks of 6,
giving 16 42s blocks. The condition for successive blocks alternated
between rest and auditory stimulation, starting with rest. Auditory
stimulation was bi-syllabic words presented binaurally at a rate of 60
per minute. The functional data starts at acquisition 4

To avoid T1 effects in the initial scans of an fMRI time series we
recommend discarding the first few scans. To make this example simple,
we’ll discard the first complete cycle (12 scans, 04-15), leaving 84
scans, image files 16-99.

Two different ways of defining a regressor

  1. manually:

    >>> reg=np.zeros(84)
>>> reg[6]=1
>>> reg[18]=1

  2. using a .csv file which defines in different columns the runs, the conditions, the onset time and the duration.

Convolution

>>> import nipy.modalities.fmri.hemodynamic_models as hm
>>> hrf = hm.spm_hrf(1.0, oversampling=1)
>>> plt.plot(np.convolve(reg, hrf))

Try to play a bit with the convolution using different regressors

fMRI scanner data

As usual, we would like to measure a variation => temporal resolution Localization => spatial resolution Trade-off : it takes more time to acquire more data points.

The MRI scanner allows the BOLD signal to be recorded in a given space where the experimenter puts the object to measure, in our case, a living brain (and around: the human that sustains it).

Scanner => sequential acquisition of data

  • 2D: one slice n * n voxels (a voxel is an element of volume, think 3D pixel)
  • 3D: many slices = one volume, which usually covers the whole brain
  • 4D: time series of sequentially acquired volumes == for each voxel, we get a time series

Let's look at the data:

Few words about DICOM files and nifti Conversion: Application Menu->NeuroDebian->Medical Imaging -> dcm2nii:

>>> import numpy as np
>>> from nipy import load_image
>>> import matplotlib.pyplot as plt
>>> i=load_image('host/python-cogstats/examples/wrf4d.nii.gz')
>>> i.shape
>>> i[0,0,0,0]

Now you can play with python in order to visualize

  1. a slice
  2. a volume
  3. a time series in a given voxel

Now, try to get some time series using Application Menu->NeuroDebian->Medical Imaging -> FSLview

Conclusions about the signal characteristics:

  • NOISY!!!!
  • Highly correlated, spatially, temporally!
  • spatial resolution
  • temporal resolution
  • different sources of variations

Remember, the fluctuation induced by brain activity is few percents => measurement of a small effect size in huge noise => need for a LOT of repetition! => substractive approach

Analysis principle

As said earlier, the idea is to find voxels in which the regressor describing a condition has some effect on the BOLD signal. Each regressor is actually explaining a part of the variance in a Linear Model

Analysis example

The first example file is 'host/examples/plot_auditory_analysis.py'.

Exercise

  • Run the example.
  • Change the .csv file to make it match the actual paradigm.
  • Compare with a random-generated regressor.
  • Remarks?

Contrasts for T- and F-tests

  • T-tests are used to compare the contributions of different regressors : one contrast vector
  • F-tests are used to compare a model to a submodel (same as anova(model0,model 1) in R) : matrix identifying regressors that will NOT be in the submodel

Another example

The second example file is 'host/examples/plot_localizer_analysis.py'

Exercise

  • Define some interesting contrasts prior to the analysis.
  • Run the example.

Multiple comparisons

Several strategies.

  1. Voxel-level
    • Uncorrected (silly)
    • Family-Wise: Bonferroni and others (only type I risk)
    • Estimation of independant elements of volume related to the spatial correlation: Random Field Theory
    • False Discovery Rate (type I and type II risks)
    • Reduce the number of voxels for the correction (ROI)

Number of voxels on which the model is estimated:

>>> fmri_glm.glms[0].get_beta().shape
>>> np.sum(fmri_glm.mask.get_data()!=0)
cumulative
>>> from scipy.stats import norm
>>> norm.ppf(1-0.001)

Exercise

  1. Caclulate a Bonferroni-correcter threshold
  2. Apply
  1. Cluster-level
    • Permutation test

Exercise

  1. From the Z-map of a contrast, retrieve cluster sizes.
  2. Sort the Cluster sizes.
>>> Z = z_map.get_data()
>>> cluster_map, n_clusters = ndimage.label(np.abs(Z) > 2.5)
>>> cluster_sizes_h0.append(np.bincount(cluster_map.ravel())[1:])
  1. What is the relathinship between the permutation and H0?
  2. What is permuted exactly?
  3. Run the file 'host/python-cogstats/examples/plot_permutation.py'

Second-level analysis

Is related to between-subject comparison. Fixed or Random effect analysis. Necessitate normalization.