chapter one
Introduction
The field of brain source localization using electroencephalography (EEG)/magnetoencephalography (MEG) signals emerged a few decades ago in neuroscience for a variety of clinical and research applications. To solve the EEG inverse problem, the forward problem needs to be solved first. The forward problem suggests the modeling of the head using advanced mathematical formulations [1]. Thus, head modeling for the solution of the forward problem is categorized as either analytical or numerical [2]. The numerical head modeling schemes have proved more efficient in terms of resolution provided, which leads to a better solution for the inverse problem. The commonly used techniques are the boundary element method (BEM), finite element method (FEM), and finite difference method (FDM). Among these methods, BEM is simpler and noniterative as compared with FEM and FDM. However, FEM has higher computational complexity as well as better resolution by covering more regions and allowing efficient computation of irregular grids [3,4].
Different neuroimaging techniques are used to localize active brain sources. However, when EEG is used to solve this problem, it is known as the EEG inverse problem. The EEG inverse problem is an ill-posed optimization problem as unknown (sources) outnumbers the known (sensors). Hence, to solve the EEG ill-posed problem, many techniques have been proposed to localize the active brain sources properlyāthat is, with better resolution [5]. The inverse techniques are generally categorized as either parametric approaches or imaging approaches [6]. The parametric methods assume the equivalent current dipole representation for brain sources. However, the imaging methods consider the sources as intracellular currents within the cortical pyramidal neurons. Hence, a current dipole is used to represent each of many tens of thousands of tessellation elements on the cortical surface. Thus, the source estimation in this case is linear in nature, as the only unknowns are the amplitudes of dipoles in the tessellation element. However, because the number of known quantities (i.e., electrodes) is significantly less than the number of unknowns (sources that are >10 K), the problem is underdetermined in nature. Hence, regularization techniques are used to control the degree of smoothing.
Mathematically, the EEG inverse problem is typically an optimization problem. Thus, there are a variety of minimization procedures, which include LevenbergāMarquardt and NelderāMead downhill simplex searches, global optimization algorithms, and simulated annealing [7]. However, all equivalent current dipole models apply principle component analysis (PCA) and singular value decomposition (SVD) to obtain a first evaluation related to the number and relative strength of field patterns existing in data before the application of a particular source model. According to the categorization discussed above, a number of techniques based on least squares principle, subspace factorization, Bayesian approaches, and constrained Laplacian have been developed for brain source localization. The methods are generally based on least squares estimation with minimum norm estimates (MNEs) [8ā10] and its modified form with Laplacian smoothness, such as low-resolution brain electromagnetic tomography (LORETA), standardized LORETA (sLORETA), and exact LORETA (eLORETA) [11ā14]; beamforming approaches [15]; and some parametric array signal processing-based subspace methods, such as multiple signal classification (MUSIC), recursive MUSIC, recursively applied and projected-MUSIC (RAP-MUSIC), and estimation of signal parameters via rotational invariance technique (ESPRIT) [16ā19]. The Bayesian framework-based methods are known as multiple sparse priors (MSP), which is the latest development in the field of brain source localization. This technique is discussed in detail in the literature [20ā23].
These methods are characterized according to certain parameters such as resolution, computational complexity, localization error, and validation. Some of these methods (LORETA, sLORETA, eLORETA) have low spatial resolution. However, array signal processing-based MUSIC and RAP-MUSIC offer better resolution but at the cost of high computational complexity. In addition, some other methods such as a focal underdetermined system solution (FOCUSS) provides a better solution with high resolution; however, due to heavy iterations in the weight matrix, FOCUSS has high computational complexity [24]. Some hybrid algorithms are also proposed for this purpose, such as weighted minimum normāLORETA (WMN-LORETA) [25], hybrid weighted minimum norm [26], recursive sLORETA-FOCUSS [27], shrinking LORETA-FOCUSS [28], and standardized shrinking LORETA-FOCUSS (SSLOFO) [29]. These hybrid algorithms provide better source localization but have a large number of iterations, thus resulting in heavy computational complexity. Besides, the techniques that are hybridized with LORETA and sLORETA suffer from low spatial resolution. Hence, the limitations of algorithms developed with any techniques include less stable solution, more computational burden, blurred solution, and slow processing.
After presenting a brief background on the field of brain source localization, we shall move on to unearth more about this dynamic and multidimensional research field. This research fieldāthat is, brain source localizationāinvolves brain signal/image processing, optimization algorithms, mathematical manipulations, and applied computational techniques. And a large variety of applications exist for this field, such as in brain disorder (epilepsy, brain tumor, schizophrenia, depression, etc.) applications, behavioral science applications, psychological studies, and traumatic brain injury applications. These topics are covered in detail later.
1.1Background
This section provides a detailed introduction for the human brain anatomy, modern neuroimaging techniques, which are used to capture image/signal data from the brain through different means, and overall economic burden due to brain disorders for which EEG source localization is used.
1.1.1Human brain anatomy and neurophysiology
The human brain is the most complex organ with 1012 neurons, which are interconnected via axons and dendrites, and 1015 synaptic connections. This complex structure allows it to release/absorb quintillion of neurotransmitter and neuromodulator molecules per second. The metabolism of the brain can be analyzed through radioactively labeled organic molecules or probes that are involved in processes of interest such as glucose metabolism or dopamine synthesis [30]. Brain development starts at a primary age of 17ā18 weeks of parental development and generates the electrical signals until death [31]. Neurons act as processing units for the brain activity to send or receive signals from/to various parts of the body to the brain. According to embryonic developments, the human brain can be divided into three regions anatomically: the forebrain (or prosencephalon), the midbrain (or mesencephalon), and the hindbrain (or rhombencephalon) [32]. However, broadly, the brain tissues are categorized as either gray matter or white matter. The brain surface is divided into four lobes: the frontal lobe, the parietal lobe, the temporal lobe, and the occipital lobe. A detailed discussion about brain anatomy and related functionality is provided in the following sections.
The brain is the most important and complex part of the central nervous system. Composed of various neurons that ac...