Ph. D. Project
Title:
Analysis of large-scale brain networks underlying human face perception
Dates:
2019/05/10 - 2022/02/14
Supervisor(s): 
Description:
Human face perception is arguably the most important visual function for human social communication. Understanding the neural basis of this function is of primary importance, starting by identifying the involved neural structures and their relationships.

In this PhD thesis we wish to rely on direct measurements of the visual responses to apply and develop signal processing tools to characterize the underlying face recognition network, using mainly intracranial Stereo-EEG (SEEG) recordings. The CHRU of Nancy has a solid experience in SEEG recordings. SEEG records the activity with a temporal resolution of the order of the millisecond, at the temporal scale of the studied phenomena. The sensors are directly implanted in the brain volume, avoiding the skull propagation barrier and providing high SNR measurements of the brain activity. During their period of SEEG investigation (usually five days), and with their consent, implanted patients can be involved in stimulation protocols. The functional activations of the brain to a given stimuli are then recorded. In the context of face recognition, our research team develops specific protocols based on fast periodic visual stimulation (FPVS) : the human brain is stimulated at a specific rate leading to periodic responses that can be captured on the measurements and can be and quantified in the frequency domain. Such methodology has led to direct proofs of the involvement of brain areas in the occipito-temporal visual pathways in response to face stimuli, confirming previous studies carried out with fMRI.

The analysis of large-scale networks (i.e. neural systems distributed across the entire extent of the brain) supporting cognitive function is a current active field of research. At the macroscopic scale, it is well accepted that synchronized neural mass populations activated by a stimuli can be assimilated to nodes of the network. The identification of functional network edges between these nodes comes from the analysis of time series data in the time (e.g. cross-correlation function) or frequency (e.g., spectral coherence or phase synchrony measures) domain. In either domain, the used relation measure can be symmetric (e.g. linear cross-correlation), characterizing undirected edge network, or asymmetric revealing directed links between pair of nodes (e.g. Granger Causality analysis). The combination of SEEG recordings with FPVS protocols provide valuable data for functional analysis of evoked responses : the instant onset and duration of the stimuli are precisely known, as well as the frequency components of interest reflecting the specific activation of the brain structures to the visual stimuli. Based on classical relation measures cited above, we expect to identify the edge of the nodes with high confidence.

Although implanted inside the brain and thus being in the immediate neighbourhood of neuronal sources, SEEG signals are nevertheless a mixing of close and distant contributions from several macroscopic populations, because of volume conduction. The proper analysis of functional brain networks then require to separate and localize these various contributions. Our research team has gained strong experience on SEEG source reconstruction in the past few years. In particular, sparse approaches retained our attention, explaining the measurements with a few number of sources. Sparsity is relevant in this context where the components related to the stimuli are well localized in the time-frequency plane, and where several trials are available to average and emphasize the targeted brain activations. Another aspect is the sparse implantation of the SEEG contacts, making of SEEG inverse problem a particularly ill-posed problem. Careful examination of the source space that might be identified with a given SEEG implantation will be required, and techniques currenty developed in our team need to be adapted for the SEEG setup.

From the reconstructed sources, we will also explore the use of dynamic causal modeling (DCM), which infers the causal architecture of distributed dynamical systems modeling interactions between neural mass populations. This methodology has already been explored in the context of steady-state responses, and offers the possibility to compare the statistical relevance of competing models under Bayesian optimization. Various hypotheses can be tested on the structure of the network and/or the strength of connectivity, the best hypothesis (model) can be selected from its likelihood when confronted to the data.

It is also important to note that the analysis of relationships between different sources and/or measurements at different brain locations must be completed by the relationships over frequency and time scales. It is widely accepted that cross-frequency couplings (within single or multiple signals) are important features characterizing normal or pathological brain functions (theta-
gamma coupling in working memory, beta-gamma abnormal coupling in Parkinson). Several techniques exist for estimating cross-frequency coupling but they need to be adapted or developed for the specific answers of the brain during FPVS. Beyond multi-site and multi-frequency band couplings and connectivity, an extremely interesting approach explores multi-scale/ multi-modality approaches, involving signals with different origins but physiologically related, such as microscopic and macroscopic recordings. Indeed, an alternative opportunity we wish to explore during this PhD thesis is the investigation of brain electrical activities using microscopic electrode setup, also implanted at the CHRU of Nancy. At this level, the activity is a mixture of LFP (electric potential generated by a local population of neurons) and of action potentials (AP) of single neurons close to the electrode. Methods for the identification and separation of these two main contributions are currently developed by our research team, making it it possible to study the relationships between the components and to determine the causal links between the explored structures across both scales. Promising results on real data have already been achieved, and should be extended and generalized to several patients and protocols during this thesis.

For further details and references, see http://w3.cran.univ-lorraine.fr/perso/radu.ranta/sujet_word_fin.pdf
Keywords:
SEEG, connectivity analysis, cross-frequency coupling, brain source imaging, dynamic causal modeling
Department(s): 
Biology, Signals and Systems in Cancer and Neuroscience