Section: New Results

Unveiling brain activity using M/EEG

Signals obtained from magneto- or electroencephalography (M/EEG) are very noisy and inherently multi-dimensional, i.e. provide a vector of measurements at each single time instant. To cope with noise, researchers traditionally acquire measurements over multiple repetitions (trials) and average them to classify various patterns of activity. This is not optimal because of trial-to-trial variability (waveform variations, jitters). The jitter-adaptive dictionary learning method (JADL) has been developed  [82] to better handle for this variability, with a particular emphasis on jitters. It was generalized to handle variability both in jitter and in duration, in a method called Adaptive Waveform Learning [8]. These methods  [83] are data-driven and learn a dictionary (prototype signals) from a set of signals, but are limited to a single channel, which restricts their capacity to work with very noisy multichannel data such as M/EEG. An extension to multidimensional signals has been developped in  [96] and [41].

• The skull conductivity is usually considered homogeneous, but the skull is actually made of several types of bone: hard (compacta) and soft (spongiosa) which may have different conductivity characteristics. By adapting a template to MR images of individual subjects, the influence of the spongiosa on source localization can be demonstrated  [97]. Estimating the conductivity values of the skull compartments is an important problem, for which theoretical results on uniqueness and robustness have been obtained  [64], [63], [26].

• Such studies show the need of easily obtaining EEG leadfields with various conductivity values. Recomputing a new leadfield for every different set of conductivities is expensive. We have thus developped a technique inspired by “reduced bases” which approximates the set of leadfields over a domain of conductivities using a low number of "base leadfields" [40]. The approach offers mathematical guarantees on the approximation level and provides an efficient methodological ground for attempting to compute both sources and conductivities in the EEG inverse problem.

Our expertise on building forward models in bioelectromagnetism has led to a collaboration with Oticon Medical, a cochlear implant manufacturer. Through Dang's PhD thesis [12], we developed computational models of cochlear implant stimulation, which can account for the anatomical shape of the inner ear, the shape of the implanted electrode, and the stimulation mode, for instance common ground or multi-mode grounding  [67], [66]. The OpenMEEG software was extended to cope with zero-conductivity regions (e.g. the silicon electrode holder). The cochlear implant Boundary Element model was coupled with a lumped capacitor and constant phase element model, allowing time-domain simulation. Thorough validation campaigns were conducted, in vitro (notably using a 3D printer) and in situ (in human specimens).