Trainee Project
Title:
Brain source localization based on iterative regression methods
Dates:
2024/05/02 - 2024/09/30
Supervisor(s):
Description:
The localization and characterisation of neuronal electrical generators are fundamental for the understanding of healthy or pathological brain functions. Thiese studies are mainly based on electrophysiological recordings, a type of measurement with very high temporal resolution. Based on a set of M measurements and the knowledge of a distributed source model (a dictionary of N source projections covering the brain volume), we need to deduce the set of sources responsible for these measurements. The solution to this problem is not unique, as different source different source configurations can lead to the same propagation result on the sensors. Numerous algorithms have been developed in order to identify the desired solution from the set of admissible solutions.
The aim of this subject is to evaluate methods enforcing sparsity based on iterative regression approaches. A well-known family of approaches for solving this type of problem are Matching Pursuit-type methods (with more elaborate variants such as Orthogonal MP or Orthogonal Least Squares), although they are little used in the context of brain source localisation. These so-called bottom-up methods consist of adding a new source among N at each iteration, thereby reducing a criterion, generally a quadratic criterion of adequacy to the data, penalised by the number of previously selected sources. Some approaches also make it possible to reconsider the sources chosen during the iterations and to be able to remove them if the criterion to be optimised is improved. One example is the SBR method (for Single Best Replacement [1]), which we have already considered in our work [2]. Another family of related approaches are so-called top-down regression approaches, where all N sources making up the propagation model are recruited from the outset to reconstruct the measurements, and where the aim is to iteratively eliminate the sources that do not contribute significantly to these measurements.
This training work will start by testing the bottom-up regression methods already developed by the team members on simulated data sets, then on real data acquired during intracerebral electrical stimulation protocols for which the positions of the sources are known. Secondly, the top-down methods will be developed and tested on the same data sets. Finally, all these methods will be tested on real data from recordings collected during face recognition protocols in humans, with the aim of identifying the brain regions that support this cognitive function.
The candidate is required to have a strong background in statistical signal processing and good skills in Matlab programming.
The aim of this subject is to evaluate methods enforcing sparsity based on iterative regression approaches. A well-known family of approaches for solving this type of problem are Matching Pursuit-type methods (with more elaborate variants such as Orthogonal MP or Orthogonal Least Squares), although they are little used in the context of brain source localisation. These so-called bottom-up methods consist of adding a new source among N at each iteration, thereby reducing a criterion, generally a quadratic criterion of adequacy to the data, penalised by the number of previously selected sources. Some approaches also make it possible to reconsider the sources chosen during the iterations and to be able to remove them if the criterion to be optimised is improved. One example is the SBR method (for Single Best Replacement [1]), which we have already considered in our work [2]. Another family of related approaches are so-called top-down regression approaches, where all N sources making up the propagation model are recruited from the outset to reconstruct the measurements, and where the aim is to iteratively eliminate the sources that do not contribute significantly to these measurements.
This training work will start by testing the bottom-up regression methods already developed by the team members on simulated data sets, then on real data acquired during intracerebral electrical stimulation protocols for which the positions of the sources are known. Secondly, the top-down methods will be developed and tested on the same data sets. Finally, all these methods will be tested on real data from recordings collected during face recognition protocols in humans, with the aim of identifying the brain regions that support this cognitive function.
The candidate is required to have a strong background in statistical signal processing and good skills in Matlab programming.
Department(s):
| Biology, Signals and Systems in Cancer and Neuroscience |