Data analysis

PIs: Dr Mohamed Bahri, PhD & Prof Christophe Phillips, Ir PhD

At the CRC, various types of data are commonly acquired and analysed for neuro-imaging experiments. These are mainly the following:

Positrons Emission Tomography, aka. PET, images

These are mainly images based on FDG (fluorodeoxyglucose) radio-tracer which maps cerebral energy metabolism.

(In the past images marked with H₂O¹⁵ were also used. These allowed activation experiments by mapping the variation of cerebral blood flow. This technique has now been superseded by functional MRI.)

Magnetic resonance imaging, aka. MRI

These are mainly functional MRI data (relying on the BOLD signal) but also structural MRI (T1 weighted images) and diffusion weighted MRI.

Electro-encephalographic, aka. EEG, signal

These data are recorded in any type of experiment: evoked potential, continuous signal at rest, sleep data (nap or whole night recording), and simultaneously with fMRI (with or without stimuli).

Magneto-encephalographic, aka. MEG, signal

Through our collaboration with the MEG unit at Hopital Erasme, we have access to MEG data, usually in combination with EEG data.

To process these data, the main software used at the CRC is SPM (http://www.fil.ion.ucl.ac.uk/spm/), which has been designed for the analysis of brain imaging data sequences (PET, IRMf and M/EEG). Thanks to our contacts with the “Wellcome Trust Centre for Neuroimaging”, we also take part in the constant development of SPM. An in-house toolbox, “fMRI Artefact rejection & Sleep Scoring Toolbox” or FASST (http://www.montefiore.ulg.ac.be/~phillips/FASST.html), compatible with SPM was developed for:

the handling (visualization, episode extraction, etc.) of long continuous multi-channel EEG/MEG recordings (typically sleep recordings);
the manual “scoring” of these continuous recordings (lasting up to several hours) and the extraction of some specific sleep features (for example slow wave sleep detection);
the automatic rejection of the artefacts induced in EEG data when they are jointly acquired with fMRI data.

Classical statistical analyses of neuroimaging data rely mainly on the “general linear model” and “statistical parametric mapping”, as implemented in the SPM software: the construction and assessment of spatially extended statistical processes used to test hypotheses about functional imaging data.

We also have some expertise with the following other methodological approaches:

“Dynamic causal modeling”, aka. DCM, for fMRI and EEG data. Generative models are built to provide an explanation for the observed data and infer the most probable cause of these signals
“Independent component analysis”, aka. ICA, for fMRI and EEG data. With fMRI, ICA allows the semi-automatic extraction of activation networks from resting state data. For the correction of the EEG acquired alongside fMRI, we use a “constrained ICA” approach.
“Supervised learning and classification”. Given a set of learning data, a “machine” is trained to categorize the data according to their label (for example healthy controls versus patients). Afterwards a new image can be automatically classified, by the trained machine, to one of the specific categories. Methods used presently at the CRC are « Support Vector Machine », « Relevance Vector Machine » and « Gaussian Processes ».

We also use other softwares as FSL, EEGlab, pMode, Camino, BrainVoyager, etc. depending on the researchers and projects need.