Methods and systems for high-resolution EEG source localization using deep neural networks

Abstract

Methods and systems for estimating spatial locations of brain activity sources from EEG signals with fMRI-comparable resolution are disclosed. The method processes EEG signals with structural MRI data by computing a lead field matrix based on electrode coordinates and anatomical information. Brain activity sources are estimated using this matrix and a predefined source estimation algorithm. Energy levels for each source are calculated over a predefined time window and convolved with a hemodynamic response function. The resulting data undergoes spatial transformation into subject-specific space before mapping to standard brain template space. A deep neural network model comprising convolutional core and temporal combination components processes this mapped data to generate estimated BOLD signals. Application of a gray matter voxel mask during training and inference enhances spatial accuracy. The system implements this processing pipeline with at least one processor and memory configured to execute these operations.

Description

DescriptionTitle of Invention: METHODS AND SYSTEMS FOR HIGH- RESOLUTION EEG SOURCE LOCALIZATION USING DEEP NEURAL NETWORKS ITechnical Field

[0001] The present disclosure generally relates to the field of biomedical signal processing and brain imaging, particularly to high-resolution source localization of electroencephalography (EEG) signals. More specifically, it pertains to methods and systems for estimating the spatial locations and activity levels of underlying brain sources from EEG data with enhanced spatial resolution, comparable to that typically achieved with functional magnetic resonance imaging (fMRI).Background Art

[0002] Monitoring and understanding brain activity is fundamental to neuroscience research and clinical applications. FMRI is a widely adopted technique that generates detailed maps of brain activity with high spatial resolution, typically on the order of millimeters, by measuring changes in the Blood Oxygen Level Dependent (BOLD) signal associated with blood flow. Despite its strengths, fMRI is subject to several limitations. The technique relies on costly and stationary scanner equipment, incurs significant operational expenses, requires strict subject immobility during scans, and provides relatively low temporal resolution, limiting its ability to capture rapid brain dynamics.

[0003] In contrast, EEG offers a non-invasive, cost-effective, and portable alternative for measuring electrical brain activity directly, achieving high temporal resolution on the order of milliseconds. EEG imposes fewer restrictions on subjects, enabling recordings during a range of activities, and is more accessible than fMRI. However, EEG is hindered by poor spatial resolution due to volume conduction effects, which diffuse electrical signals through the head and obscure their precise origins within the brain.

[0004] To address these complementary strengths and weaknesses, prior efforts have sought to integrate the accessibility and temporal precision of EEG with the spatial detail of fMRI. Some approaches have employed computational techniques, including machine learning models and deep neural networks, to infer fMRI-likeBOLD signal distribution from EEG data, whether acquired simultaneously or independently.

[0005] Despite these developments, existing methods have exhibited several technical limitations. Many approaches have utilized subject-specific models, which often result in reduced accuracy and limited applicability across diverse subjects due to the complexity of individual variations. When evaluated with real experimental data, these methods have frequently struggled to reliably detect established activity patterns, often producing a high incidence of false positives and yielding only coarse spatial resolution.Summary of Invention

[0006] The following summary is a brief overview of the subject matter of the present invention and is not intended to be a comprehensive or definitive description thereof. It is not intended to identify essential features, nor is it intended to limit the scope of the invention. The proper scope of the present disclosure may be ascertained from the claims set forth below, given the detailed description and the drawings below.

[0007] Certain embodiments of the present invention may provide methods and systems for estimating spatial locations of brain activity sources corresponding to EEG signals with improved spatial resolution and robustness. In one preferred aspect, the method may involve acquiring a plurality of EEG signals and structural MRI data from a subject’s head. A lead field matrix can be computed using the structural MRI data and three-dimensional coordinates of a plurality of EEG electrodes. This may facilitate a more accurate initial estimation of brain activity sources via a predefined source estimation algorithm.

[0008] In the preferred aspect, the method may further comprise processing the initial estimated brain activity sources, by calculating a plurality of energy levels thereof over a predefined time window and generating brain activity signal data through convolving with a hemodynamic response function (HRF). This data may then be transformed into a subject-specific space and subsequently mapped onto a standard brain template space.

[0009] In one general aspect, a deep neural network (DNN) model may be utilized to process the mapped signal data and generate an estimated blood oxygen leveldependent (BOLD) signal output that represents the locations of the brain activity sources with spatial resolution analogous to that of fMRI. In certain embodiments, the DNN model may comprise a convolutional core component and a temporal combination component. To enhance spatial precision, a mask corresponding to gray-matter voxels may be applied both during training, to restrict loss computation to gray-matter regions, and during inference, to filter the final BOLD output to graymatter locations only.

[0010] In another aspect, the method may include removing estimated sources associated with negative dipoles prior to energy calculation, improving signal quality. Furthermore, training the DNN model parameters may involve a multi-stage process, potentially starting with simulated data and followed by fine-tuning using simultaneously acquired real EEG and fMRI data from multiple subjects and potentially diverse cognitive tasks, thereby enhancing model generalizability across both subjects and tasks.

[0011] In another general aspect, the present disclosure may provide a system configured to perform such estimations. The system may comprise at least one processor and a memory storing executable instructions. These instructions may configure the at least one processor to perform operations including acquiring EEG signals and structural MRI data, computing the lead field matrix, performing initial source estimation, processing the estimates such as energy calculation, HRF convolution, transformations, standard space mapping, and applying the trained DNN model to generate the final high-resolution output.

[0012] In certain aspects, the system may support the utilization of the DNN architecture such as convolutional core and temporal combination components and the application of a gray matter voxels mask during both training and inference. Embodiments of the system may also be configured to perform negative dipole removal and support multi-stage training of the DNN model.Technical Problem

[0013] A fundamental challenge in functional neuroimaging is obtaining high spatial resolution maps of brain activity without incurring the significant costs, limited accessibility, and stringent operational constraints associated with fMRI. While EEG offers a more accessible, cost-effective, and temporally precise alternative, itinherently suffers from poor spatial resolution, limiting its ability to accurately localize brain activity sources.

[0014] Conventional prior art methods attempting this estimation demonstrate considerable deficiencies. These include insufficient spatial resolution, often significantly lower than fMRI capabilities, and poor accuracy, frequently characterized by identifying activity where none exists.

[0015] Additionally, many previous methods lack generalizability. Models developed often fail to perform reliably across different subjects, particularly those not included in the initial training data, partly due to the inadequate incorporation of individual-specific anatomical information.

[0016] Moreover, Generalizability across different cognitive tasks is also often limited, with models requiring task-specific training. These limitations, collectively prevent existing EEG-based estimation techniques form providing reliable and widely applicable solution for high-resolution functional brain mapping.Solution to Problem

[0017] To overcome the above shortcomings, the present disclosure provides systems and methods that may implement a multi-stage framework comprising subjectspecific biophysical modeling, advanced signal processing, and a deep neural network architecture. In various embodiments, the disclosed approach may enhance the accuracy and spatial resolution of brain activity estimation using non- invasive recordings such as EEG.

[0018] The system may be configured to incorporate subject-specific anatomical data to overcome insufficient spatial resolution and accuracy. A subject-specific lead field matrix may be computed using individual’s structural MRI data and EEG electrode coordinates, leading the electromagnetic forward model to the unique anatomy. This may allow for a more precise initial estimation of brain activity sources. Further refinements may be achieved by utilizing constraints related to physiologically plausible source locations, such as training to prioritize estimations within gray matter voxels or removing negative dipoles, thereby reducing spurious activations and enhancing effective spatial resolution.

[0019] Furthermore, the system may further utilize a two-stage processing strategy. While initial processing step may leverage subject-specific information for accuracy,via the lead field matrix, subsequent step may map the processed brain signal data onto a standardized brain template space. This standardization may provide consistent comparisons across individuals. The system may also employ a deep neural network model trained on datasets comprising simultaneously acquired EEG and fMRI data from a plurality of individuals. Such training may allow the model to learn generalizable spatio-temporal patterns of brain activity, thereby improving performance and reducing the risk of overfitting to individual-specific characteristics.

[0020] Moreover, the system may utilize a spatio-temporal deep neural network architecture. This architecture may comprise a convolutional core component for extracting relevant spatial features from the mapped brain signal data and a temporal combination component designed to analyze the dynamics of these features over time. This architecture may enable the system to capture complex neural activity associated with a variety of cognitive states. The model may be trained using data obtained from a diverse set of cognitive tasks, allowing it to learn task-invariant representations of brain activity. As a result, the system may be applicable across multiple experimental conditions or clinical scenarios without requiring retraining for each specific task.Brief Description of Drawings

[0021] The present disclosure is best understood by reference to the following description and the accompanying figures. These figures are provided for illustrative purposes and do not limit the scope of the present disclosure. Of these figures:

[0022] FIG. 1 illustrates a flowchart of a method for estimating spatial locations of brain activity sources, in accordance with one or more embodiments of the present disclosure.

[0023] FIG. 2 illustrates a flowchart of a training process for a deep neural network model, in accordance with one or more embodiments of the present disclosure.

[0024] FIGS. 3A and 3B illustrate schematic diagrams of a convolutional core component and a temporal combination component of a deep neural network model, in accordance with one or more embodiments of the present disclosure.

[0025] FIGS. 4A-4D illustrate schematic diagrams of a plurality of tasks for which real EEG and fMRI data are simultaneously acquired, in accordance with one or more embodiments of the present disclosure.

[0026] FIG. 5 illustrates a diagram of an estimated BOLD signal output compared to a reference ground truth for one exemplary voxel, in accordance with one or more embodiments of the present disclosure.

[0027] FIG. 6 illustrates a schematic diagram of estimated brain activity sources compared to a reference ground truth, in accordance with one or more embodiments of the present disclosure.

[0028] FIG. 7A illustrates a group-level activation map from ERP-based deep neural network estimates of general visual stimulation, overlaid with corresponding fMRI activation contours, in accordance with one or more embodiments.

[0029] FIG. 7B illustrates a group-level contrast map from ERP-based deep neural network estimates contrasting object-image stimulation blocks with scrambled- image blocks, overlaid with corresponding fMRI contrast contours, in accordance with one or more embodiments.

[0030] FIG. 8 illustrates a block diagram of a system for estimating spatial locations of brain activity sources, in accordance with one or more embodiments of the present disclosure.Description of Embodiments

[0031] In the following detailed description, reference is made to the accompanying figures, which form a part hereof and are shown by way of illustration of example embodiments or aspects in which the disclosure may be practiced. The figures are not drawn to scale and are intended to illustrate the principles of the disclosure rather than to show actual dimensions or proportions. However, it will be evident to one skilled in the art that an example embodiment may be practiced without all of the disclosed details.

[0032] The present disclosure is not intended to be limited to the specific embodiments described herein, and it is contemplated that various changes, substitutions, and equivalents may be made without departing from the scope of the disclosure. The detailed description that follows is provided by way of elucidation, and is notintended to be limiting. Additionally, throughout this specification, the terms "a", "an", and "the" are intended to include plural references, and the term "in" is intended to include both "in" and "on".

[0033] It is to be understood that the use of phrases such as 'an embodiment,' 'one embodiment,' or 'an example embodiment' in this specification is intended to provide illustrative examples, and not to imply that the corresponding embodiment is limited to the specific features, structures, or characteristics described.

[0034] As used herein, the terms “brain activity sources” and “sources” are used interchangeably to refer to the underlying neural generators within the brain whose electrical activity is measured indirectly. The terms “standard brain template space” and “standard space” are used interchangeably to describe a normalized anatomical coordinate system such as MNI or Talairach space, allowing for comparison of brain data across different individuals.

[0035] The present disclosure relates to systems and methods for estimating spatial locations of brain activity sources from EEG signals. More particularly, embodiments of the disclosure may provide techniques for processing EEG signals in combination with structural MRI data to yield high-resolution brain activity estimations analogous to fMRI.

[0036] Referring to FIG. 1 , which may illustrate a flowchart of the method 100 for estimating spatial locations of brain activity sources according to one exemplary embodiment of the present disclosure. The method 100 may involve several steps, beginning with acquiring input data, including a plurality of EEG signals from a plurality of EEG electrodes positioned on a subject’s head, at step 101. The plurality of EEG signals may be recorded during one or more cognitive tasks or resting-state conditions.

[0037] In one embodiment, at step 102, structural MRI data from the subject may be acquired to serve as the anatomical basis for subject-specific modeling. In alternative embodiments, computed tomography (CT) data may be acquired in lieu of MRI data, and, in the absence of both MRI and CT data, a standard brain template, such as MNI atlas, may be employed. Based on the anatomical information derived from the chosen data and the three-dimensional coordinates of the EEG electrodes, a lead field matrix may be computed at step 103. The leadfield matrix may represent a forward model that maps dipolar brain activity at various locations to the resulting EEG signals.

[0038] At step 104, a predefined source estimation algorithm may be used with the lead field matrix to estimate a plurality of brain activity sources, e.g., equivalent current dipoles, from the EEG signals. In one embodiment, the predefined source estimation algorithm may be selected from a group including, but not limited to, Weighted Minimum Norm Estimate (WMNE), Standardized Low-Resolution Electromagnetic Tomography (sLORETA), exact Low-Resolution Electromagnetic Tomography (eLORETA), beamforming techniques, dipole fitting methods, and other suitable inverse solution approaches. The resulting source distribution may represent an initial spatial localization of cortical activity over time.

[0039] In another embodiment, spurious negative dipole estimates on the wall opposite a true activation site of a cortical sulcus may be removed. Because dipole vectors on the two banks of a sulcus may be approximately symmetrical, true positive activation on one wall may produce a mirror-image negative estimate on the opposite wall. Discarding these spurious negative dipoles may enhance physiological plausibility and reduces artifacts.

[0040] At step 105, a plurality of energy levels may be computed for each estimated source. In one embodiment, the energy may be computed over a predefined temporal window corresponding substantially to the repetition time (TR) of a typical fMRI sequence, for example 1 or 2 seconds. This conversion may yield a signal more temporally aligned with fMRI dynamics and may facilitate subsequent comparison with BOLD signals.

[0041] In one embodiment, at step 106, the energy levels may be convolved with a hemodynamic response function (HRF). The HRF may be a canonical function such as the double-gamma model used in fMRI analysis, or an empirically estimated HRF specific to the subject or task. The convolution process may generate a new set of brain activity signal data that may reflect an approximate temporal transformation of neural energy into a pseudo-BOLD signal, forming the input to the next processing stages.

[0042] According to some implementations, at steps 107 and 108, further spatial processing may be performed to prepare the brain activity signal data for input intothe DNN model 300 and to facilitate standardization. At step 107, the generated brain activity signal data may be transformed into a subject-specific space derived directly from the acquired structural MRI data. This transformation may ensure that the activity signals are spatially aligned with the individual’s unique brain anatomy, thereby enhancing the precision of subsequent source estimation.

[0043] Subsequently, at step 108, the brain activity signal data residing in the subjectspecific space may be mapped onto a standard brain template space, such as the Montreal Neurological Institute (MNI) space or Talairach space according to one implementation. This standardization may be crucial for training a generalizable DNN model 300, as it allows the model to learn consistent patterns from anatomically corresponding locations across different individuals despite variations in brain size and shape, thereby reducing the risk of overfitting to individual-specific characteristics.

[0044] In one embodiment, at step 109, the mapped brain activity signal data may be processed utilizing a trained DNN model 300. The primary function of the DNN model 300 at this stage may be learning and applying a complex, non-linear transformation to the input data, thereby generating a final estimated output signal that represents the spatial locations of the underlying brain activity sources.

[0045] The DNN model 300 may be configured to generate an estimated BOLD signal output 502 with enhanced spatial resolution, potentially achieving resolutions on the order of millimeters similar to that of fMRI. In one embodiment, the output may be presented as an estimated BOLD signal time series for each voxel in the standard brain template space.

[0046] As further illustrated in FIGS. 3A and 3B, the architecture of the DNN model 300 may comprise components, including a convolutional core component 301 configured to extract relevant spatial features from the input data volumes and a temporal combination component 302 designed to integrate information across time points, thereby capturing the temporal dynamics of the estimated brain activity.

[0047] In another embodiment, at step 110, a mask corresponding to gray matter voxels may be applied to the output signal generated by the DNN model 300. During a training phase, the mask may limit the loss function evaluation to gray matter regions. During an inference phase, the gray matter mask may be appliedto spatially filter the estimated BOLD signal output 502. As a result, a final BOLD signal characterization derived from the filtered output may exhibit improved spatial accuracy and / or functional specificity.

[0048] Referring to FIG. 2, a flowchart illustrating a training process of the DNN model 300 is illustrated, according to one embodiment of the present disclosure. The training process may be carried out in multiple stages to enhance model robustness and generalization.

[0049] At step 201 , a plurality of DNN model parameters may be initially trained using simulated EEG and fMRI data. Such simulated data may be generated via forward models and signal-synthesis techniques that approximate realistic neural activation patterns and hemodynamic responses.

[0050] At step 202, the pretrained model parameters may be fine-tuned using simultaneously acquired real EEG and fMRI data from one or more human subjects performing a plurality of cognitive tasks. In one implementation, the fine-tuning may first involve adjusting only a subset of the plurality of DNN model parameters, where these parameters may be associated with the input and output layers of the convolutional core component 301 and the temporal combination component 302, while holding all other parameters fixed. This initial phase may allow the model’s input / output mappings to adapt to real data distributions without disrupting the learned internal feature representations. Subsequently, in a second phase, all of the plurality of DNN model parameters may be jointly adjusted to refine internal weights and improve performance across different tasks and subjects.

[0051] Referring to FIG. 3A, the convolutional core component 301 of the DNN model 300 may implement an ll-Net architecture. In one embodiment, the initial layers may comprise two-dimensional convolutional operations with a stride of 2 and a kernel size of 2x2, enabling multiscale spatial feature extraction from the input brain activity volumes.

[0052] As illustrated in FIG. 3B, the temporal combination component 302 may integrate temporal continuity of the estimated BOLD signals. In one implementation, this component may perform a voxel-wise weighted combination of the current and immediately preceding outputs of the convolutional core component 301 according to the equation:where Y(n) and Y(n-1 ) are the output volumes at time steps n and n-1 , respectively, and Z(n) is a three-dimensional gating matrix matching those volumes. This formulation may improve signal-to-noise ratio by smoothing temporal fluctuations.

[0053] In alternative embodiments, the DNN model 300 may be implemented as a recurrent neural network (RNN) or long short-term memory (LSTM) network, wherein the convolutional core component 301 serves as a modular feature extractor within each recurrent block. Such architectures may capture longer-range temporal dependencies and account for regional variations in the hemodynamic response function across different brain areas.

[0054] Referring to FIGS. 4A-4D, exemplary tasks for acquiring simultaneous EEG and fMRI data from one or more human subjects are illustrated. Such concurrently acquired data may be employed during the fine-tuning stage, step 202, to adjust the plurality of deep neural network model parameters based on real task-evoked brain responses. In one embodiment, these data were acquired under approved institutional protocols from human subjects, and processed as described below.

[0055] In one embodiment, depicted in FIG. 4A, a visual recognition task may comprise four stimulation blocks 411 , 413, 415, and 417 of 12.5 seconds each, interleaved with three rest blocks 412 of 12.5 seconds each. During stimulation block 411 , an object image randomly selected from a set of one hundred images may be presented to elicit activation in the lateral occipital complex (LOC). In stimulation block 413, the same object image may be presented in scrambled form to control for low-level visual features. During stimulation block 415, a face images randomly selected from a set of one hundred twenty faces may be shown to engage the fusiform face area (FFA). Stimulation block 417 again may present the scrambled face image.

[0056] In another embodiment, illustrated in FIG. 4B, a motion perception task may comprise a first stimulation block 421 , a rest block 422, and a second stimulation block 423, each lasting 12.5 seconds. During stimulation block 421 , a field of white dots may move coherently in one of four directions (up, down, left, or right) with direction randomized between presentations to localize activity in motion-sensitive regions such as V5 / MT. During rest block 422, a static fixation screen may bedisplayed. In stimulation block 423, a field of static white dots with random spatial arrangement may be shown to control for low-level visual features.

[0057] In a further embodiment, shown in FIG. 4C, an auditory language task may comprise a first stimulation block 431 , a rest block 432, and a second stimulation block 433. During stimulation block 431 , five meaningful words, randomly selected from a set of fifty, may be played to activate language-related regions such as Broca’s and Wernicke’s areas. During rest block 432, silence or fixation may be presented. In stimulation block 433, five pseudowords, randomly selected from a set of fifty pronounceable non-words, may be played.

[0058] Referring to FIG. 4D, a working-memory task may be presented using a 3x4 grid block 441. In one embodiment, the task may comprise alternating easy and hard stimulation blocks. During each easy block, three individual cells of the grid block 441 may be illuminated sequentially. During each hard block, three pairs of grid cells may be illuminated sequentially, with each pair illuminated simultaneously in a single step. Upon completion of each block, two candidate illumination patterns 443 may be presented to the subject, who may be prompted to select the pattern corresponding to the sequence just observed.

[0059] Data acquired from each of these cognitive tasks may be processed according to the method steps, including lead field matrix computation, source estimation, energy calculation and HRF convolution, spatial transformation into a subject-specific space and standard brain template space, and DNN processing. The resulting task-evoked EEG-fMRI data may serve to fine-tune the plurality of DNN model parameters, thereby improving the model’s generalization across diverse cognitive conditions.

[0060] Referring to FIG. 5, there may be shown a comparison between an estimated BOLD signal output 502 generated by the DNN model 300 and a corresponding ground truth BOLD signal 501 acquired via fMRI for one exemplary voxel during the visual recognition task. In this exemplary embodiment, the DNN model 300 may not be trained on data from the subject under evaluation, thereby demonstrating the model’s ability to generalize to unseen subjects.

[0061] Referring to Table 1 , precision and recall metrics may be reported for the DNN model 300 across the four cognitive tasks of FIGS. 4A-4D. Precision and recallmay be calculated based on the overlap between predicted and ground-truth clustered activation regions in a test subject whose data were not used during training. As shown, the model may achieve high precision and recall across all tasks, demonstrating its ability to generalize to unseen subjects and reliably localize task-evoked brain activity sources.Table 1Precision and Recall of the deep neural network model 300 in four cognitive tasks

[0062] Referring to FIG. 6, a comparison of activation maps may be illustrated for a human subject whose data were not included in the training of the DNN model 300. The top row may show the ground truth spatial distribution of brain activity sources as measured by fMRI, while the bottom row may show the corresponding activation estimated by the DNN model 300.

[0063] Referring to FIG. 7A, a group-level activation map derived from ERP-based deep neural network estimates is illustrated. In one exemplary embodiment, ERPs elicited by general visual stimulation may be provided, subject-by-subject, to the deep neural network model 300 in a leave-one-out cross-validation across nine subjects. For each held-out subject, the DNN model 300 may generate a voxelwise estimated BOLD signal time series, which may then be subjected to a second- level, random -effects group analysis. FIG. 7A may depict regions of significant activation identified from the DNN estimates (shaded areas, p < 0.05) overlaid with corresponding fMRI activation contours (p < 0.05) obtained from conventional fMRI analysis of the same stimulus.

[0064] Referring to FIG. 7B, a group-level contrast map derived from ERP-based deep neural network estimates is illustrated. In one exemplary embodiment, event- related potentials elicited by object-image stimulation block 411 and by scrambled- image stimulation block 413 may be provided, subject-by-subject, to compute respective energy time series. A contrast signal may be generated by subtracting the scrambled-image energy series from the object-image energy series, and this contrast signal data may then be processed by the DNN model 300 to produce an estimated contrast BOLD time series for each subject. In another exemplary embodiment, leave-one-out cross-validation may be performed across nine subjects, and the resulting subject-level contrast estimates may be entered into a second-level, random -effects group analysis. FIG. 7B illustrates regions of significant contrast activation identified from the DNN estimates (shaded areas, p < 0.05) overlaid with corresponding fMRI contrast contours (p < 0.05) obtained from conventional group-level fMRI analysis of the same contrast.

[0065] Referring now to FIG. 8, a block diagram illustrating a system 800 for estimating spatial locations of brain activity sources from EEG signals, according to one exemplary embodiment, is provided. The system 800 may comprise EEG electrodes 802 configured to acquire a plurality of EEG signals from a subject’s head, and a structural MRI scanner 804 configured to acquire structural MRI data from the subject’s head. These EEG and MRI data may be provided as inputs to an integrated processing pipeline for further processing as described below.

[0066] In one embodiment, the system 800 may include at least one processor and associated memory, configured to execute instructions corresponding to the various processing steps of the pipeline. In particular, the processor may be configured to perform input data acquisition 811 , receiving EEG signals from electrodes 802 and structural MRI data from MRI scanner 804. At step 813, the processor may compute a lead field matrix based on the acquired structural MRI data and three-dimensional coordinates of the EEG electrodes 802. This lead field matrix computation may involve subject-specific anatomical modeling to accurately represent the electrical conductivity and geometry of brain and head tissues.

[0067] At step 815, the processor may utilize the computed lead field matrix together with a predefined source estimation algorithm to estimate a plurality of brain activity sources from the EEG signals according to one implementation. In oneembodiment, the source estimation algorithm may be selected from a group including, but not limited to, WMNE, sLORETA, eLORETA, beamforming, dipole fitting, or other suitable inverse solution methods. In another embodiment, at step 817, the processor may further improve source estimation accuracy by removing estimated brain activity sources associated with negative dipoles, which may represent physiological artifacts arising from inverse model ambiguities near cortical sulci.

[0068] At step 819, the processor may calculate a plurality of energy levels for each of the estimated brain activity sources over a predefined time interval substantially matching a typical fMRI repetition time, TR, according to one implementation. The processor may then, at step 821 , convolve the calculated energy levels with a HRF to generate brain activity signal data that approximates hemodynamic changes detectable by fMRI.

[0069] In a subsequent step 823, the processor may spatially map the generated brain activity signal data into a standard brain template space, such as the MNI or Talairach space. This mapping can facilitate inter-subject comparisons and provides standardized anatomical references, improving the generalizability of the system’s subsequent processing and outputs.

[0070] In one implementation, at step 825, the mapped brain activity signal data may be processed using a deep neural network model 825 configured to generate an estimation of a BOLD signal. The deep neural network inference step 825 may comprise two distinct computational components: a convolutional core component 825a configured to extract relevant spatial features from the mapped signal data volumes, and a temporal combination component 825b configured to integrate and enhance temporal coherence across successive time points. The generated BOLD signal output of the deep neural network model 825 thereby may have spatial resolution comparable to fMRI.

[0071] In one embodiment, the processor may apply a gray matter mask 827 to the estimated BOLD signal output generated by the deep neural network model 825. During a training phase, the mask may limit the loss function evaluation to gray matter regions. During an inference phase, the mask may be applied to spatially filter the estimated BOLD signal output 831. As a result, the final BOLD signalcharacterization derived from the filtered output may exhibit improved spatial accuracy and functional specificity.

[0072] The final output of the system 800, may be the estimated BOLD signal, representing the spatial locations of brain activity sources in a standardized brain template space with improved spatial resolution, accuracy, and generalizability compared to conventional EEG source localization methods. The estimated BOLD signal output 831 can subsequently be used for clinical diagnostics, cognitive neuroscience research, or other applications that traditionally depend upon high- resolution fMRI data.Industrial Applicability

[0073] According to the aforementioned various embodiments, the present invention may find industrial applicability in the fields of non-invasive neuroimaging and clinical diagnostics of neurological disorders. The system and method may enable precise localization of seizure foci with high spatial resolution for diagnosing epilepsy, and may support cognitive neuroscience research by providing accurate mapping of task-evoked brain activation across diverse cognitive tasks. Additionally, the invention may facilitate brain-computer interface applications, allowing real-time interaction and neurofeedback for therapeutic or assistive purposes.

[0074] The invention may also contribute to pharmaceutical and neuromodulation studies by enabling evaluation of treatment effects on brain function, and may assist in mental health assessments by mapping neural activity related to conditions such as anxiety or depression. Furthermore, the invention may support personalized medicine through subject-specific brain mapping, and may offer a cost-effective, portable alternative to conventional fMRI by integrating EEG acquisition hardware with structural MRI modeling and deep neural network inference. These capabilities may make the invention suitable for use in hospital diagnostic suites, outpatient clinics, research laboratories, and wearable neurofeedback devices. I

Claims

Claims

1. A method for estimating spatial locations of brain activity sources corresponding to EEG signals, comprising: a. acquiring a plurality of EEG signals from a plurality of EEG electrodes positioned on a subject head; b. acquiring structural magnetic resonance imaging (MRI) data from the subject head; c. computing a lead field matrix based on the acquired structural MRI data and three-dimensional coordinates of the plurality of EEG electrodes; d. estimating a plurality of brain activity sources corresponding to the acquired plurality of EEG signals using the lead field matrix and a predefined source estimation algorithm; e. calculating a plurality of energy levels for each of the estimated plurality of brain activity sources over a predefined time window; f. generating brain activity signal data by convolving the calculated plurality of energy levels with a hemodynamic response function; g. transforming the generated brain activity signal data into a subjectspecific space derived from the acquired structural MRI data; h. mapping the transformed brain activity signal data from the subjectspecific space onto a standard brain template space; i. processing the mapped brain activity signal data utilizing a deep neural network model configured to generate an estimated blood oxygen level dependent signal output representing spatial locations of the plurality of brain activity sources with a spatial resolution analogous to fMRI, wherein the deep neural network model comprises a convolutional core component and a temporal combination component; and j. applying a mask corresponding to gray matter voxels to the estimated blood oxygen level dependent signal output by the deep neural network model.

2. The method according to claim 1 , further comprising:a. removing the estimated plurality of brain activity sources associated with negative dipoles prior to calculating the plurality of energy levels.

3. The method according to claim 1 , wherein training the deep neural network model comprises: a. training a plurality of deep neural network model parameters using simulated EEG and fMRI data; and b. fine-tuning the plurality of deep neural network model parameters using simultaneously acquired real EEG and fMRI data from one or more human subjects, the fine-tuning step comprising:- adjusting a subset of the plurality of deep neural network model parameters associated with input and output layers of the convolutional core component and the temporal combination component; and- adjusting all of the plurality of deep neural network model parameters.

4. A system for estimating spatial locations of brain activity sources corresponding to EEG signals, the system comprising: a. at least one processor; and b. a memory communicatively coupled to the at least one processor, wherein the at least one processor is configured to:- acquire a plurality of EEG signals from a plurality of EEG electrodes positioned on a subject head;- acquire structural magnetic resonance imaging (MRI) data from the subject head;- compute a lead field matrix based on the acquired structural MRI data and three-dimensional coordinates of the plurality of EEG electrodes;- estimate a plurality of brain activity sources corresponding to the acquired plurality of EEG signals using the lead field matrix and a predefined source estimation algorithm;calculate a plurality of energy levels for each of the estimated plurality of brain activity sources over a predefined time window;- generate brain activity signal data by convolving the calculated plurality of energy levels with a hemodynamic response function;- transform the generated brain activity signal data into a subject-specific space derived from acquired structural MRI data;- map the transformed brain activity signal data from the subject-specific space onto a standard brain template space;- process the mapped brain activity signal data utilizing a deep neural network model configured to generate an estimated blood oxygen level dependent signal output representing spatial locations of the plurality of brain activity sources with a spatial resolution analogous to fMRI, wherein the deep neural network model comprises a convolutional core component and a temporal combination component; and- applying a mask corresponding to gray matter voxels to the estimated blood oxygen level dependent signal output by the deep neural network model.

5. The system according to claim 4, wherein the processor further configured to remove the estimated plurality of brain activity sources associated with negative dipoles prior to calculating the plurality of energy levels.

6. The system according to claim 4, wherein training of the deep neural network model comprises: a. training a plurality of deep neural network model parameters using simulated EEG and fMRI data; and b. fine-tuning the plurality of deep neural network model parameters using simultaneously acquired real EEG and fMRI data from one or more human subjects, the fine-tuning step comprising:- adjusting a subset of the plurality of deep neural network model parameters associated with input and output layers of the convolutional core component and the temporal combination component; and- adjusting all of the plurality of deep neural network model parameters. ;

Images