MEG and EEG data analysis with MNE-Python

Abstract

Magnetoencephalography and electroencephalography (M/EEG) measure the weak electromagnetic signals generated by neuronal activity in the brain. Using these signals to characterize and locate neural activation in the brain is a challenge that requires expertise in physics, signal processing, statistics, and numerical methods. As part of the MNE software suite, MNE-Python is an open-source software package that addresses this challenge by providing state-of-the-art algorithms implemented in Python that cover multiple methods of data preprocessing, source localization, statistical analysis, and estimation of functional connectivity between distributed brain regions. All algorithms and utility functions are implemented in a consistent manner with well-documented interfaces, enabling users to create M/EEG data analysis pipelines by writing Python scripts. Moreover, MNE-Python is tightly integrated with the core Python libraries for scientific comptutation (NumPy, SciPy) and visualization (matplotlib and Mayavi), as well as the greater neuroimaging ecosystem in Python via the Nibabel package. The code is provided under the new BSD license allowing code reuse, even in commercial products. Although MNE-Python has only been under heavy development for a couple of years, it has rapidly evolved with expanded analysis capabilities and pedagogical tutorials because multiple labs have collaborated during code development to help share best practices. MNE-Python also gives easy access to preprocessed datasets, helping users to get started quickly and facilitating reproducibility of methods by other researchers. Full documentation, including dozens of examples, is available at http://martinos.org/mne.

Keywords: electroencephalography (EEG); magnetoencephalography (MEG); neuroimaging; open-source; python; software.

Figure 1

Noisy raw MEG magnetometer signals corrupted by a) slow drifts, b) line noise (at 50 or 60 Hz), and c) heartbeats present across sensors. To clean signals data were filtered between 1 and 45 Hz. Subsequently, five signal space projection (SSP) vectors were applied (3 computed from empty room noise, 2 from ECG signals). The plots were generated using the plot method of the Raw class.

Figure 2

An evoked response (event-related fields in planar gradiometers of an Elekta-Neuromag Vectorview system) showing traces for individual channels (bad channels are colored in red). Epochs with large peak-to-peak signals as well as channels marked as bad can be discarded from further analyses. The figure was generated using the plot method of the Evoked class.

Figure 3

Cortical segmentation used for the source space in the distributed model with MNE. Left: The pial (red) and white matter (green) surfaces overlaid on an MRI slice. Right: The right-hemisphere part of the source space (yellow dots), represented on the inflated surface of the left hemisphere, was obtained by subdivision of an icosahedron leading to 10242 locations per hemisphere with an average nearest-neighbor distance of 3.1 mm. Left image was produced with FreeSurfer tksurfer tool and the right one with PySurfer (http://pysurfer.github.io) which internally depends on Mayavi (Ramachandran and Varoquaux, 2010).

Figure 4

Source localization of an auditory N100 component. Left: Results obtained using dSPM and a surface source space based on combined MEG and EEG data. The figure was generated using the plot method of the SourceEstimate class which internally calls PySurfer. Right: Results obtained using LCMV beamformer and a volume source space based on MEG channels only. The figure was generated using Freeview shipped with FreeSurfer.

Figure 5

Current estimates obtained from an individual subject can be remapped (morphed), i.e., normalized, to another cortical surface, such as that of the FreeSurfer average brain “fsaverage” shown here. The normalization is done separably for both hemispheres using a non-linear registration procedure defined on the sphere (Dale et al., ; Fischl et al., 1999). Here, the N100m auditory evoked response is localized using dSPM and then mapped to “fsaverage.” Images were produced with PySurfer.

Figure 6

Topographic and trellis plots of two automatically identified ICA components. The component #22 corresponds to the EOG artifact with a topography on the magnetometers showing frontal signals and a waveform typical of an eye blink. The component #6 on the right captures the ECG artifact with a waveform matching 3 heart beats.

Figure 7

Examples of clustering. (A) Time-frequency clustering showing a significant region of activation following an auditory stimulus. (B) A visualization of the significant spatio-temporal activations in a contrast between auditory stimulation and visual stimulation using the sample dataset. The red regions were more active after auditory than after visual stimulation, and vice-versa for blue regions. Image (B) was produced with PySurfer.

Figure 8

Sensor space decoding. At every time instant, a linear support vector machine (SVM) classifier is used with a cross-validation loop to test if one can distinguish data following a stimulus in the left ear or in the left visual field. One can observe that the two conditions start to be significantly differentiated as early as 50 ms and maximally at 100 ms which corresponds to the peak of the primary auditory response. Such a statistical procedure is a quick and easy way to see in which time window the effect of interest occurs.

Figure 9

Connectivity between brain regions of interests, also called labels, extracted from the automatic FreeSurfer parcellation visualized using plot_connectivity_circle. The image of the right presents these labels on the inflated cortical surface. The colors are in agreement between both figures. Left image was produced with matplotlib and right image with PySurfer.

Figure 10

Source localization with non-linear sparse solvers. The left plot shows results from TF-MxNE on raw unfiltered data (due to the built-in temporal smoothing), and the right plot shows results from γ-MAP on the same data but filtered below 40 Hz. One can observe the agreement between both methods on the sources in the primary (red) and secondary (yellow) visual cortices delineated by FreeSurfer using an atlas. The γ-MAP identifies two additional sources in the right fusiform gyrus along the visual ventral stream. These sources that would not be naturally expected from such simple visual stimuli are weak and peak later in time, which makes them nevertheless plausible.