Calculation method and system for conduction and scattering processes of low-frequency electromagnetic waves in brain

Abstract

The invention discloses a calculation method and system for conduction and scattering processes of low-frequency electromagnetic waves in the brain, and relates to the technical field of brain detection.The method has the advantages that brain tissues are simplified into homogeneous isotropic sphere models, accordingly, the calculation complexity can be obviously reduced, and foundations can be laid for efficient solution; an electromagnetic wave propagation and scattering physical model based on Maxwell equations is constructed, and the movement law of electromagnetic waves in brain tissue is accurately described; the physical model is deeply integrated into a deep learning algorithm, and three-dimensional intracerebral electromagnetic field distribution is efficiently solved; finally, the propagation path, energy distribution and scattered field characteristics of electromagnetic waves are analyzed, positioning and qualitative diagnosis of lesions such as encephaledema and tumors are achieved, key support is provided for brain tissue health assessment, the efficiency bottleneck of a traditional numerical method in brain electromagnetic calculation is overcome, the timeliness and accuracy of brain disease diagnosis are improved, and the method is suitable for popularization and application. And reliable technical support is provided for rapid bedside diagnosis of stroke and other emergencies.

Description

Technical Field

[0001] This application relates to the field of brain detection technology, and in particular to a calculation method and system for the conduction and scattering process of low-frequency electromagnetic waves in the brain. Background Technology

[0002] With the increasing urgency of early screening for brain diseases, non-invasive detection technology based on electromagnetic waves has become a key development direction in the field of neuromedicine. Low-frequency electromagnetic waves, due to their deep penetration ability and safety to biological tissues, have shown unique advantages in the diagnosis of lesions such as cerebral edema and brain tumors.

[0003] Current technologies primarily rely on traditional numerical methods, such as the finite element method, to simulate the behavior of electromagnetic waves in brain tissue. This approach faces significant challenges in clinical applications: due to the complex boundary conditions of the brain, whole-brain-scale electromagnetic calculations require processing a large number of mesh elements, resulting in single simulations taking hours or even days, failing to meet the needs of real-time clinical diagnosis; oversimplification of the brain model to improve computational speed leads to significant accuracy errors, substantially reducing the reliability of lesion detection; existing methods largely focus on calculating electromagnetic wave propagation paths, with insufficient quantitative analysis of scattered field characteristics, making it difficult to identify early, minute lesions; and, more importantly, these methods often rely on high-performance computing clusters, severely hindering the widespread adoption of this technology in primary healthcare settings.

[0004] While artificial intelligence methods have shown promise in accelerating computation, directly applying general neural network models would violate the fundamental physical laws of electromagnetic fields and lack the interpretability required for medical diagnosis.

[0005] Current technology has not yet effectively resolved the triple contradiction between accuracy, efficiency, and interpretability in intracranial electromagnetic computing, especially the lack of research on the correlation mechanism between scattering field characteristics and lesion nature. Summary of the Invention

[0006] The main objective of this application is to provide a calculation method and system for the conduction and scattering process of low-frequency electromagnetic waves in the brain, so as to solve the problem that the accuracy of traditional low-frequency electromagnetic waves for brain detection is not ideal in the prior art.

[0007] To achieve the above objectives, this application provides the following technical solution:

[0008] A calculation method for the conduction and scattering process of low-frequency electromagnetic waves in the brain, the calculation method being applied to low-frequency electromagnetic waves incident on and propagating in target brain tissue, the calculation method comprising:

[0009] Step S1: Obtain the dielectric constant and permeability parameters of the target brain tissue in a preset frequency band, and calculate the unified electromagnetic property parameters of the whole brain by weighted average.

[0010] Step S2: Construct a uniform spherical model with a diameter matching the anatomical dimensions of the human cranial cavity, and define the boundary conditions of the uniform spherical model from the inside to the outside as an open electromagnetic boundary from tissue to air.

[0011] Step S3: Using the uniform sphere model as the computational domain and the unified electromagnetic property parameters of the whole brain, construct the Maxwell differential control equations;

[0012] Step S4: By dynamically simulating external electromagnetic waves with different combinations of incident parameters, inputting them into the differential control equation, several datasets of electromagnetic field distribution in the brain are obtained.

[0013] Step S5: Discretize all brain electromagnetic field distribution datasets into three-dimensional grid datasets, and divide all three-dimensional grid datasets into sample sets and validation sets;

[0014] Step S6: Input all sample sets into a CNN / RNN hybrid deep neural network for training to establish a mapping relationship from incident parameters to the three-dimensional electromagnetic field distribution in the brain, and obtain the trained neural network model.

[0015] Step S7: Input all validation sets into the trained neural network model to obtain electromagnetic response prediction data through the mapping relationship.

[0016] As a further improvement to this application, step S7 involves inputting all validation sets into the trained neural network model to obtain electromagnetic response prediction data through the mapping relationship, followed by:

[0017] Step S10: Extract the electric field attenuation rate and distribution broadening from the electromagnetic response prediction data, and compare them with the preset healthy brain tissue benchmark database respectively.

[0018] Step S20: If the comparison result shows that the dielectric parameter is abnormally increased and the electric field spreads in concentric circles, then it is determined to be cerebral edema.

[0019] Step S30: If the comparison result shows a localized high absorption region accompanied by a loss of scattering directionality, it is determined to be a brain tumor.

[0020] As a further improvement to this application, step S1 involves obtaining the dielectric constant and permeability parameters of the target brain tissue at a preset frequency band, and calculating the unified electromagnetic property parameters of the whole brain through weighted averaging, including:

[0021] Step S11: The target brain tissue is divided into partitions based on a preset voxel size, and the dielectric constant and permeability parameters of each partition are collected in a preset frequency band.

[0022] Step S12, the unified permittivity and unified permeability parameters of the whole brain are obtained by weighted averaging using equation (1):

[0023] (1);

[0024] in, To achieve a unified dielectric constant for the entire brain, The first of the target brain tissue Each partition voxel For the first The dielectric constant of each voxel in the partition. To achieve a unified magnetic permeability parameter for the entire brain, For the first The permeability parameters of each voxel in the partition;

[0025] Step S13: Integrate the unified dielectric constant and unified magnetic permeability parameters of the whole brain for the same voxels into unified electromagnetic property parameters of the whole brain.

[0026] As a further improvement to this application, step S3, using the uniform sphere model as the computational domain and the unified electromagnetic property parameters of the whole brain as the construction of Maxwell's differential governing equations, includes:

[0027] Step S31, construct the Maxwell differential governing equation using equation (2):

[0028] (2);

[0029] in, For vector differential operators, For curl, Let be the electric field intensity vector of the computational domain. The imaginary unit, Let be the permeability parameter of the computational domain. Let be the magnetic field strength vector of the computational domain. The time-resonance oscillation frequency of the computational domain. The dielectric constant of the computational domain is given. The current density of the excitation source;

[0030] Step S32, according to equation (3), the constraint conditions of the Maxwell differential governing equations are defined to satisfy the energy dissipation characteristics of the external electromagnetic wave at the far-field boundary:

[0031] (3);

[0032] in, To measure the propagation distance of the stimulus source, The scattered field of the excitation source, The wave number of the external electromagnetic wave.

[0033] As a further improvement to this application, in step S4, external electromagnetic waves with different combinations of incident parameters are input into the differential control equation through dynamic simulation to obtain several brain electromagnetic field distribution datasets, including:

[0034] Step S41: Construct several multidimensional parameter matrices based on several combinations of incident parameters of external electromagnetic waves;

[0035] Step S42: The differential control equations are converted into a system of discrete algebraic equations using the finite difference time-domain method.

[0036] Step S43: Input each multidimensional parameter matrix into the discrete algebraic equation system and solve to obtain a brain electromagnetic field distribution dataset.

[0037] As a further improvement to this application, step S5 discretizes all brain electromagnetic field distribution datasets into three-dimensional mesh datasets, and divides all three-dimensional mesh datasets into sample sets and validation sets, including:

[0038] Step S51: Define a right-handed coordinate system with its origin located at the center of the uniform sphere model based on the uniform sphere model, and align the Z-axis of the right-handed coordinate system with the anatomical sagittal axis of the target brain tissue.

[0039] Step S52: Divide the right-handed coordinate system based on the preset voxel size to form several three-dimensional meshes;

[0040] Step S53: Map the current brain electromagnetic field distribution dataset to the three-dimensional grid of the right-hand coordinate system using trilinear interpolation.

[0041] Step S54: Zero-value filling is performed on the three-dimensional mesh with countless values. Based on all the three-dimensional meshes, a voxel dataset with physical labels is obtained, which is a three-dimensional mesh dataset.

[0042] Step S55: Divide the current voxel dataset into a sample set and a validation set.

[0043] As a further improvement to this application, step S6 involves inputting all sample sets into a CNN / RNN hybrid deep neural network for training to establish a mapping relationship from incident parameters to the three-dimensional electromagnetic field distribution within the brain, resulting in a trained neural network model, including:

[0044] Step S61: Organize all brain electromagnetic field distribution datasets into multi-channel three-dimensional tensors according to the spatial dimension of the computational domain;

[0045] Step S62: Organize all multidimensional parameter matrices into time variation feature matrices aligned with the time series of all multi-channel three-dimensional tensors;

[0046] Step S63: Define the number of input channels of the CNN branch of the CNN / RNN hybrid deep neural network to be consistent with the dimension of all multi-channel three-dimensional tensors, and define the dimension of the input features of the RNN branch to be consistent with the parameter dimension of all time-varying feature matrices;

[0047] Step S64: Extract the spatial feature vectors of all multi-channel three-dimensional tensors through the CNN branch, and extract the dynamic feature vectors of all time-varying feature matrices through the RNN branch;

[0048] Step S65: All spatial feature vectors and all dynamic feature vectors are merged in the splicing layer of the CNN / RNN hybrid deep neural network, and the resulting joint feature representation is the mapping relationship.

[0049] To achieve the above objectives, this application also provides the following technical solutions:

[0050] A computational system for the conduction and scattering process of low-frequency electromagnetic waves in the brain, the computational system being applied to the computational method described above, the computational system comprising:

[0051] The electromagnetic property parameter acquisition module is used to acquire the dielectric constant and magnetic permeability parameters of the target brain tissue in a preset frequency band, and to obtain the unified electromagnetic property parameters of the whole brain by weighted average calculation.

[0052] The uniform sphere model construction module is used to construct a uniform sphere model with a diameter matching the anatomical size of the human cranial cavity, and defines the boundary conditions of the uniform sphere model from the inside to the outside as an open electromagnetic boundary from tissue to air.

[0053] The differential control equation construction module is used to construct Maxwell's differential control equations using the uniform sphere model as the computational domain and the unified electromagnetic property parameters of the whole brain.

[0054] The electromagnetic wave dynamic simulation module is used to input external electromagnetic waves with different combinations of incident parameters into the differential control equation through dynamic simulation to obtain several brain electromagnetic field distribution datasets.

[0055] The electromagnetic field distribution discretization module is used to discretize all brain electromagnetic field distribution datasets into three-dimensional grid datasets, and to divide all three-dimensional grid datasets into sample sets and validation sets.

[0056] The hybrid deep neural network training module is used to input all sample sets into the CNN / RNN hybrid deep neural network for training, in order to establish the mapping relationship from the incident parameters to the three-dimensional electromagnetic field distribution in the brain, and obtain the trained neural network model.

[0057] The electromagnetic response data prediction module is used to input all validation sets into the trained neural network model to obtain electromagnetic response prediction data through the mapping relationship.

[0058] To achieve the above objectives, this application also provides the following technical solutions:

[0059] An electronic device includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the detection calculation method described above.

[0060] To achieve the above objectives, this application also provides the following technical solutions:

[0061] A computer-readable storage medium storing program instructions that, when executed by a processor, enable the detection calculation method described above.

[0062] Effective effects:

[0063] This application obtains the dielectric constant and permeability parameters of the target brain tissue at a preset frequency band, and calculates the unified electromagnetic property parameters of the whole brain through weighted averaging; constructs a uniform spherical model with a diameter matching the anatomical size of the human cranial cavity, and defines the boundary conditions of the uniform spherical model from the inside to the outside as an open electromagnetic boundary from the tissue to the air; constructs Maxwell's differential control equations using the uniform spherical model as the computational domain and the unified electromagnetic property parameters of the whole brain; inputs external electromagnetic waves with different combinations of incident parameters into the differential control equations through dynamic simulation to obtain several brain electromagnetic field distribution datasets; discretizes all brain electromagnetic field distribution datasets into three-dimensional mesh datasets, and divides all three-dimensional mesh datasets into sample sets and validation sets; inputs all sample sets into a CNN / RNN hybrid deep neural network for training to establish a mapping relationship from incident parameters to the three-dimensional electromagnetic field distribution in the brain, and obtains a trained neural network model; inputs all validation sets into the trained neural network model to obtain electromagnetic response prediction data through the mapping relationship. This application simplifies brain tissue into a homogeneous isotropic sphere model, significantly reducing computational complexity and laying the foundation for efficient solutions. It then constructs a physical model of electromagnetic wave propagation and scattering based on Maxwell's equations, accurately describing the motion of electromagnetic waves within brain tissue. Furthermore, by deeply integrating the physical model into a deep learning algorithm, it efficiently solves for the three-dimensional electromagnetic field distribution within the brain. Finally, it analyzes the propagation path, energy distribution, and scattering field characteristics of electromagnetic waves, enabling the localization and qualitative diagnosis of lesions such as cerebral edema and tumors, providing crucial support for brain tissue health assessment. This application overcomes the efficiency bottleneck of traditional numerical methods in brain electromagnetic computation, improves the timeliness and accuracy of brain disease diagnosis, and provides reliable technical support for rapid bedside diagnosis of acute conditions such as stroke. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating the steps of an embodiment of the calculation method for the conduction and scattering process of low-frequency electromagnetic waves in the brain according to this application.

[0065] Figure 2 This is a diagram showing the attenuation characteristics of electromagnetic waves in brain tissue, representing an embodiment of the calculation method for the conduction and scattering process of low-frequency electromagnetic waves in the brain according to this application.

[0066] Figure 3 A comparison diagram of the scattering energy ratio of one embodiment of the calculation method for the conduction and scattering process of low-frequency electromagnetic waves in the brain according to this application;

[0067] Figure 4 A comparison diagram of scattered energy distribution from one embodiment of the calculation method for the conduction and scattering process of low-frequency electromagnetic waves in the brain according to this application;

[0068] Figure 5A comparison diagram of forward scattering energy of one embodiment of the calculation method for the conduction and scattering process of low-frequency electromagnetic waves in the brain according to this application;

[0069] Figure 6 This is a comparison diagram of backscattered energy from one embodiment of the calculation method for the conduction and scattering process of low-frequency electromagnetic waves in the brain according to this application.

[0070] Figure 7 This is a functional module diagram of an embodiment of a computing system for the conduction and scattering process of low-frequency electromagnetic waves in the brain, as described in this application.

[0071] Figure 8 This is a schematic diagram of the structure of an embodiment of the electronic device of this application;

[0072] Figure 9 This is a schematic diagram of the structure of one embodiment of the storage medium of this application. Detailed Implementation

[0073] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0074] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0075] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0076] like Figure 1 As shown, this embodiment provides an example of a calculation method for the conduction and scattering process of low-frequency electromagnetic waves in the brain. In this embodiment, the calculation method is applied to low-frequency electromagnetic waves that are incident on and propagate in the target brain tissue.

[0077] Specifically, the calculation method includes the following steps:

[0078] Step S1: Obtain the dielectric constant and permeability parameters of the target brain tissue in a preset frequency band, and calculate the unified electromagnetic property parameters of the whole brain by weighted average.

[0079] Preferably, since the brain waves of human brain tissue are mainly concentrated in the delta wave (0.5-4Hz) band, the theta wave (4-8Hz) band, the alpha wave (8-13Hz), the beta wave (13-30Hz), and abnormal high frequencies (46-50Hz) that characterize pathology, the frequency band of the low-frequency electromagnetic wave in this embodiment can be set to 0.5Hz to 50Hz.

[0080] Preferably, the dielectric constant can be obtained using the parallel-plate capacitance method or the resonance method; the permeability parameter can be obtained by measuring the magnetization response through a magnetometer (VSM) or impedance analyzer by applying an alternating magnetic field, and calculating the relative permeability. In this embodiment, since it is a non-invasive probe and brain tissue cannot be sampled, the dielectric constant and permeability parameters can be obtained through publicly available resources, such as PhysioNet or biological tissue parameter databases (e.g., TissueFrequencyChart). By selecting the brain tissue type and setting the frequency range, the dielectric constant and permeability parameters at the corresponding frequency can be directly obtained.

[0081] Alternatively, the dielectric constant and permeability parameters can be derived by inversely analyzing the transmission loss and phase of brain tissue samples in a specific frequency band (e.g., 8 to 20 GHz) using the NRW (Nicholson-Ross-Weir) method, thus ensuring computational efficiency.

[0082] Preferably, this embodiment prioritizes obtaining the dielectric constant and permeability parameters through direct querying, which can be achieved by using a script to capture data from publicly available sources.

[0083] Step S2: Construct a uniform sphere model with a diameter matching the anatomical dimensions of the human cranial cavity, and define the boundary conditions of the uniform sphere model from the inside to the outside as an open electromagnetic boundary from tissue to air.

[0084] Preferably, since the anatomical dimensions of the human cranial cavity are known measurement data, the diameter of the uniform spherical model ranges from 18cm to 22cm, with the midpoint of 20cm taken as the diameter of the uniform spherical model. If higher accuracy is required, the individualized cranial cavity dimensions of the patient can be obtained through CT three-dimensional reconstruction.

[0085] Preferably, a uniform sphere model can be created using electromagnetic simulation software such as COMSOL Multiphysics and ANSYS HFSS, as well as the programming tool MATLAB.

[0086] Specifically, the internal medium parameters of the uniform sphere model are set to the unified electromagnetic property parameters of the whole brain in step S1, while the external environment of the uniform sphere model is set to air, that is, the dielectric constant and permeability parameters of the external environment are both 1.

[0087] Preferably, open electromagnetic boundary conditions are used to ensure that electromagnetic waves do not exhibit excessive reflection during dynamic simulation. Specifically, the normal component of the sphere's surface must be continuous, and the tangential components of the electric and magnetic fields must also be continuous.

[0088] Step S3: Using a uniform sphere model as the computational domain, Maxwell's differential control equations are constructed using the unified electromagnetic property parameters of the whole brain.

[0089] Preferably, since the model is a sphere, Maxwell's equations need to be converted to spherical coordinates (r, θ, ϕ), and the curl operator needs to be expanded according to spherical coordinates.

[0090] Step S4: By using dynamic simulation, external electromagnetic waves are input into the differential control equation with different combinations of incident parameters to obtain several datasets of electromagnetic field distribution in the brain.

[0091] Preferably, discrete frequency points can be divided according to logarithmic interval sampling, for example, 10 points per decade, while setting unit amplitude (1V / m) and zero phase as references, which can expand the ±30° phase difference into multiple amplitude-phase combinations. Under the condition of uniformly covering the entire space in spherical coordinates, a sampling point is taken every 15° in θ∈[0,π] and φ∈[0,2π].

[0092] Preferably, dynamic simulation can be performed using the electromagnetic simulation software COMSOL Multiphysics. This software has a built-in AC / DC module designed specifically for low-frequency electromagnetic fields, supports solving Maxwell's equations under the quasi-static approximation, and provides a biomedical materials library that can directly call up the dielectric parameters of brain tissue. It also supports parametric scanning and can perform batch simulations of different incident angle / frequency combinations.

[0093] Preferably, since COMSOL Multiphysics can directly derive the three-dimensional electric / magnetic field distribution matrix (real part + imaginary part), the three-dimensional electric / magnetic field distribution matrix (real part + imaginary part) can be directly defined as the brain electromagnetic field distribution dataset of this embodiment, that is, a combination of incident parameters yields a three-dimensional electric / magnetic field distribution matrix (real part + imaginary part).

[0094] Step S5: Discretize all brain electromagnetic field distribution datasets into three-dimensional grid datasets, and divide all three-dimensional grid datasets into sample sets and validation sets.

[0095] Preferably, since this embodiment uses a uniform sphere model, the three-dimensional network can use regular hexahedral elements to improve computational efficiency. When dividing the three-dimensional mesh, it is necessary to ensure that the voxel size is less than one-tenth of the electromagnetic wave wavelength.

[0096] Preferably, continuous electromagnetic field data E(x,y,z) and H(x,y,z) can be sampled according to grid node coordinates (X,Y,Z) and stored as a complex number format with real part + imaginary part, or as an amplitude / phase format.

[0097] Preferably, the ratio of the sample set to the validation set can be set to 6:4. If a test set needs to be added, the ratio of sample set:validation set:test set is 6:2:2. Simultaneously, the layers are stratified according to the incident parameters (frequency, angle) to ensure that each layer has a consistent distribution across the training / validation / test sets.

[0098] Preferably, the format of the three-dimensional mesh dataset is a three-dimensional tensor structure, i.e., the dimensions are Nr×Nθ×Nϕ×C, where C is the number of channels.

[0099] Preferably, the key pseudocode is as follows:

[0100] # 3D Mesh Discretization

[0101] def discretize_to_grid(field_data, grid_resolution=1e-3):

[0102] Nr, Nθ, Nϕ = int(0.2 / grid_resolution), int(0.2 / grid_resolution), int(0.2 / grid_resolution) # 20 cm diameter sphere

[0103] grid_data = np.zeros((Nr, Nθ, Nϕ, 6))

[0104] # 6 channels: Er, Eθ, Eϕ, Hr, Hθ, Hϕ

[0105] for i, j, k in product(range(nr), range(nθ), range(nϕ)):

[0106] grid_data[i, j, k] = interpolate_field(field_data, (i * dr, j * dθ, k * dϕ))

[0107] return grid_data

[0108] # Dataset splitting

[0109] def split_dataset(grid_datasets, train_ratio = 0.6, val_ratio = 0.2):

[0110] n_samples = len(grid_datasets)

[0111] indices = np.random.permutation(n_samples)

[0112] train_idx = indices[:int(n_samples * train_ratio)]

[0113] val_idx = indices[int(n_samples * train_ratio):int

[0114] :::ml-data{name=citationList}

[0115] ```json

[0116] Step S6: Input all sample sets into a CNN / RNN hybrid deep neural network for training to establish a mapping relationship from incident parameters to the three-dimensional electromagnetic field distribution in the brain, and obtain the trained neural network model.

[0117] Preferably, the 3D convolutional branch CNN of the CNN / RNN hybrid deep neural network is used to receive the discretized 3D grid dataset, and the 3D tensor structure Nr×Nθ×Nϕ×C mentioned above, where C corresponds to the number of channels of the CNN branch.

[0118] The CNN branch can use a 5×5×5 kernel with a stride of 2, 64 channels, and ReLU activation in the first layer; followed by three layers of 3×3×3 convolutions, with batch normalization and max pooling after each layer.

[0119] The time-varying feature matrix described below is input into the bidirectional LSTM with 128 hidden units to capture frequency domain correlation.

[0120] The next step is to concatenate the flattened features (approximately 1024 dimensions) output by the CNN with the temporal features (256 dimensions) output by the RNN, and then perform non-linear fusion through a fully connected layer (512 nodes) to obtain the mapping relationship.

[0121] Preferably, a composite loss function can be designed: the first term measures the deviation between the predicted electric field value and the standard numerical solution, and the second term forces the network prediction result to satisfy the basic conservation law of the electromagnetic field, as shown in the following formula:

[0122]

[0123] in Represents the loss function. This indicates that the network predicts the electric field. Indicates a reference solution. Indicates the weighting coefficient. It is a 2-norm operator.

[0124] Preferably, the key pseudocode is as follows:

[0125] class HybridModel(nn.Module):

[0126] def __init__(self):

[0127] super().__init__()

[0128] # CNN branch

[0129] self.cnn = nn.Sequential(

[0130] nn.Conv3d(6, 64, kernel_size=5, stride=2),

[0131] nn.BatchNorm3d(64),

[0132] nn.ReLU(),

[0133] nn.MaxPool3d(2),

[0134] # ...more convolutional layers )

[0136] # RNN branch

[0137] self.rnn = nn.LSTM(input_size=6, hidden_size=128,bidirectional=True)

[0138] # Fusion Layer

[0139] self.fc = nn.Sequential(

[0140] nn.Linear(1024+256, 512),

[0141] nn.SiLU(),

[0142] nn.Linear(512, Nr*Nθ*Nϕ*3) # Outputs a three-dimensional field )

[0144] def forward(self, x_grid, x_params):

[0145] cnn_feat = self.cnn(x_grid).flatten(1)

[0146] rnn_out, _ = self.rnn(x_params)

[0147] rnn_feat = rnn_out.mean(dim=1)

[0148] fused = torch.cat([cnn_feat, rnn_feat], dim=1)

[0149] return self.fc(fused).view(-1, Nr,Nθ,Nϕ,3)

[0150] Step S7: Input all validation sets into the trained neural network model to obtain electromagnetic response prediction data through mapping relationships.

[0151] Further, in step S7, all validation sets are input into the trained neural network model to obtain electromagnetic response prediction data through mapping relationships. This is followed by the following steps:

[0152] Step S10: Extract the electric field attenuation rate and distribution broadening from the electromagnetic response prediction data, and compare them with the preset healthy brain tissue benchmark database.

[0153] Step S20: If the comparison result shows an abnormally high dielectric parameter and a concentrically spreading electric field, then it is determined to be cerebral edema.

[0154] Step S30: If the comparison result shows a localized high absorption region accompanied by a loss of scattering directionality, it is determined to be a brain tumor.

[0155] Three diagnostic features were extracted from the comparison results: the percentage decrease in local electric field strength relative to the healthy baseline; the change in the ratio of forward and backward scattering of electromagnetic waves; and the broadening of the scattering energy distribution within a specific angular range.

[0156] The diagnostic criteria for cerebral edema are as follows: when the target area simultaneously meets the conditions of abnormally elevated mediator parameters, concentric diffusion of electric field attenuation, and no shadowing effect, it is determined to be an edema lesion.

[0157] Among them, the diagnostic criteria for brain tumors are: when a localized high absorption area appears in the target region, accompanied by a distortion of the electromagnetic wave propagation path, and the scattering energy distribution shows a lack of directionality, it is determined to be a tumorous lesion.

[0158] Preferably, see Figures 2 to 6 The images show comparisons of various electromagnetic information results between a normal brain and brain edema and brain tumors. Figure 2 This is a diagram showing the attenuation characteristics of electromagnetic waves in brain tissue. Figure 3 A comparison chart of scattering energy ratios; Figure 4 A comparison diagram of scattered energy distribution; Figure 5 A comparison diagram of forward scattering energy; Figure 6 This is a comparison diagram of backscattered energy.

[0159] Further, in step S1, the dielectric constant and permeability parameters of the target brain tissue in a preset frequency band are obtained, and the unified electromagnetic property parameters of the whole brain are calculated by weighted averaging. This specifically includes the following steps:

[0160] Step S11: Divide the target brain tissue into regions based on the preset voxel size, and collect the dielectric constant and permeability parameters of each region in the preset frequency band.

[0161] Preferably, the voxel size needs to be less than one-tenth of the electromagnetic wave wavelength. If the 20cm diameter model of this embodiment is used, the voxel size is generally between 1mm and 4mm. The frequency band of the low-frequency electromagnetic wave can be set to 0.5Hz to 50Hz according to the above discussion.

[0162] Preferably, the dielectric constant and permeability parameters of this embodiment can be obtained through public channels, such as PhysioNet and biological tissue parameter databases (such as TissueFrequencyChart). By selecting the brain tissue type and setting the frequency range, the dielectric constant and permeability parameters at the corresponding frequency can be obtained directly.

[0163] Step S12, the uniform dielectric constant and uniform magnetic permeability parameters of the whole brain are obtained by weighted averaging using equation (1):

[0164] (1).

[0165] in, To achieve a unified dielectric constant for the entire brain, The first target brain tissue Each partition voxel For the first The dielectric constant of each voxel in the partition. To achieve a unified magnetic permeability parameter for the entire brain, For the first The permeability parameters of each voxel in the partition.

[0166] Step S13: Integrate the unified dielectric constant and unified magnetic permeability parameters of the whole brain for the same voxels into unified electromagnetic property parameters of the whole brain.

[0167] Further, step S3, using a uniform sphere model as the computational domain, constructs Maxwell's differential governing equations through unified electromagnetic property parameters of the whole brain, specifically including the following steps:

[0168] Step S31, construct Maxwell's differential governing equations using equation (2):

[0169] (2).

[0170] in, For vector differential operators, For curl, The electric field intensity vector in the computational domain. The imaginary unit, For the magnetic permeability parameter of the computational domain, The magnetic field strength vector in the computational domain, To calculate the time-harmonic oscillation frequency in the computational domain, Let be the dielectric constant of the computational domain. The current density is the excitation source current density.

[0171] Step S32, according to equation (3), the constraint conditions of Maxwell's differential governing equations are defined to satisfy the energy dissipation characteristics of the external electromagnetic wave at the far-field boundary:

[0172] (3).

[0173] in, To measure the propagation distance of the stimulus source, The scattered field of the excitation source, The wave number of the external electromagnetic wave.

[0174] Further, in step S4, external electromagnetic waves are input into the differential control equations with different combinations of incident parameters through dynamic simulation to obtain several datasets of brain electromagnetic field distributions, including:

[0175] Step S41: Construct several multidimensional parameter matrices based on several combinations of incident parameters of external electromagnetic waves.

[0176] Preferably, the incident parameters can be combined according to the frequency range, amplitude and phase, and incident direction.

[0177] Among them, frequency range: the discrete frequency points are divided according to the preset frequency band (such as 1Hz-1MHz), usually sampled at logarithmic intervals (such as 10 points / decade); amplitude and phase: the unit amplitude (1V / m) and zero phase are set as the reference, which can be expanded into multiple amplitude and phase combinations (such as ±30° phase difference); incident direction: uniformly covers the entire space in the spherical coordinate system, and takes discrete angles θ∈[0,π] and φ∈[0,2π] (such as one sampling point every 15°).

[0178] Step S42: The differential control equations are transformed into a system of discrete algebraic equations using the finite difference time-domain method.

[0179] Preferably, an interlaced grid arrangement of electromagnetic field components can be adopted, with the electric field component located at the center of the grid edge and the magnetic field component located at the center of the grid surface, forming a spatial ring relationship. At the same time, a leapfrog approach is adopted in time to realize the alternating update of electric and magnetic field components, with a time step interval of Δt / 2.

[0180] Preferably, converting Maxwell's equations to component form in spherical coordinates (r, θ, φ) requires handling the radial and angular derivatives:

[0181] .

[0182] In particular, the pole region (θ=0,π) needs to be protected from singularities by local mesh refinement or coordinate transformation. The symbol for this formula has already appeared above, and its meaning will not be repeated here.

[0183] Preferably, since this embodiment has an open boundary, a perfectly matched layer (PML) is added outside the computational domain as an absorption layer, and non-reflection is achieved through complex coordinate stretching.

[0184] Preferably, the core Python code is as follows:

[0185] # Pseudocode Example: Core Iteration of FDTD in Spherical Coordinates

[0186] for n in time_steps:

[0187] # Update magnetic field components

[0188] H_phi += (dt / mu) * ( (E_r[theta+1] - E_r[theta]) / (r*dtheta) - (r*E_theta[r+1] - r*E_theta[r]) / (r*dr) )

[0189] # Update electric field components (including PML absorption terms)

[0190] E_r *= (1 - sigma_r*dt / eps) / (1 + sigma_r*dt / eps)

[0191] E_r += (dt / eps) / (1 + sigma_r*dt / eps) * ( (H_phi / (r*sin_theta) -H_phi_prev) / (r*dphi) - (H_theta[r] - H_theta[r-1]) / dr )

[0192] # Boundary handling

[0193] apply_pml_boundary(E, H)

[0194] Step S43: Input each multidimensional parameter matrix into the discrete algebraic equation system and solve to obtain a dataset of electromagnetic field distribution in the brain.

[0195] Further, in step S5, all brain electromagnetic field distribution datasets are discretized into three-dimensional mesh datasets, and all three-dimensional mesh datasets are divided into sample sets and validation sets, specifically including the following steps:

[0196] Step S51: Define a right-handed coordinate system with its origin located at the center of the uniform sphere model, based on the uniform sphere model. The Z-axis of the right-handed coordinate system is aligned with the anatomical sagittal axis of the target brain tissue.

[0197] Preferably, since the three-dimensional mesh is a regular hexahedron, spherical coordinates are not applicable. In this embodiment, a right-handed coordinate system is established to conform to the orientation of the three-dimensional mesh.

[0198] Preferably, the sagittal axis is one of the three basic axes in human anatomy, extending in the anteroposterior direction and parallel to the ground, forming a spatial positioning reference together with the coronal axis (left-right direction) and the vertical axis (up-down direction). In brain models, the sagittal axis usually corresponds to the midline in the anteroposterior direction, passing through key anatomical landmarks such as the anterior commissure (AC) and posterior commissure (PC).

[0199] Preferably, the anterior commissure (AC) is located anterior to the junction of the left and right hemispheres below the corpus callosum, and the posterior commissure (PC) is located dorsally to the aqueduct of the midbrain. The direction of the sagittal axis can be precisely defined by the AC-PC line, which coincides with the anatomical sagittal axis in a standard brain model.

[0200] Preferably, after determining the anatomical sagittal axis, the origin of the right-hand coordinate system is the geometric center of the uniform sphere model, the Z-axis extends from front to back along the AC-PC line (sagittal axis), the X-axis is defined as the left-right direction (i.e., the coronal axis) according to the right-hand rule, and the Y-axis is determined by X×Z=Y and points in the up-down direction (i.e., the vertical axis).

[0201] Step S52: Divide the right-handed coordinate system based on the preset voxel size to form several three-dimensional meshes.

[0202] Step S53: Map the current brain electromagnetic field distribution dataset to a three-dimensional grid in a right-handed coordinate system using trilinear interpolation.

[0203] Preferably, the trilinear interpolation method is a method for linearly interpolating discrete sampled data on a three-dimensional tensor product grid. The grid allows for non-overlapping arrangement in each dimension, but it is not a triangular finite element analysis grid. This method obtains the target point value by selecting eight adjacent data points around the interpolation point and performing weighted linear interpolation along the x, y, and z axes. Since the formula is existing technology, it is not provided in this embodiment.

[0204] Preferably, the core Python code is as follows:

[0205] import numpy as np

[0206] def trilinear_interp(volume, points):

[0207] """Core function for 3D mesh data interpolation"

[0208] Args:

[0209] volume: Original 3D array (shape: DxHxW)

[0210] points: An array of target point coordinates (shape: Nx3)

[0211] Returns:

[0212] An array of interpolated values ​​(shape: N,)

[0213] """

[0214] # Coordinate normalization

[0215] points = np.asarray(points)

[0216] dims = np.array(volume.shape) - 1

[0217] grid_coords = points * dims

[0218] # Get the indexes of the 8 corner points

[0219] coords = np.floor(grid_coords).astype(int)

[0220] delta = grid_coords - coords

[0221] x0, y0, z0 = coords.T

[0222] x1, y1, z1 = np.clip(coords + 1, 0, dims).T

[0223] # Trilinear Interpolation Calculation

[0224] c000 = volume[x0, y0, z0]

[0225] c100 = volume[x1, y0, z0]

[0226] c010 = volume[x0, y1, z0]

[0227] # ...the other 5 corner points

[0228] c00 = c000*(1-delta[:,0]) + c100*delta[:,0]

[0229] c01 = c010*(1-delta[:,0]) + c110*delta[:,0]

[0230] # ...other planar interpolations

[0231] return c0*(1-delta[:,2]) + c1*delta[:,2]

[0232] Step S54: Zero-value filling is performed on the three-dimensional mesh with countless values. Based on all the three-dimensional meshes, a voxel dataset with physical labels is obtained, which is a three-dimensional mesh dataset.

[0233] Step S55: Divide the current voxel dataset into a sample set and a validation set.

[0234] Preferably, the ratio of the sample set to the validation set can be set to 6:4. If a test set needs to be added, the ratio of sample set:validation set:test set is 6:2:2. Simultaneously, the layers are stratified according to the incident parameters (frequency, angle) to ensure that each layer has a consistent distribution across the training / validation / test sets.

[0235] Further, in step S6, all sample sets are input into a CNN / RNN hybrid deep neural network for training to establish a mapping relationship from the incident parameters to the three-dimensional electromagnetic field distribution in the brain, thereby obtaining the trained neural network model. This specifically includes the following steps:

[0236] Step S61: Organize all brain electromagnetic field distribution datasets into multi-channel three-dimensional tensors according to the spatial dimension of the computational domain.

[0237] Preferably, the 3D convolutional branch CNN of the CNN / RNN hybrid deep neural network is used to receive the discretized 3D grid dataset, and the 3D tensor structure Nr×Nθ×Nϕ×C mentioned above, where C corresponds to the number of channels of the CNN branch.

[0238] Step S62: Organize all multidimensional parameter matrices into time variation feature matrices aligned with the time series of all multi-channel three-dimensional tensors.

[0239] Step S63: Define the number of input channels of the CNN branch of the CNN / RNN hybrid deep neural network to be consistent with the dimension of all multi-channel three-dimensional tensors, and the dimension of the input features of the RNN branch to be consistent with the parameter dimension of all time-varying feature matrices.

[0240] Step S64: Extract the spatial feature vectors of all multi-channel 3D tensors through the CNN branch, and extract the dynamic feature vectors of all time-varying feature matrices through the RNN branch.

[0241] Preferably, the CNN branch can use a first-layer convolutional kernel of 5×5×5, stride of 2, number of channels of 64, and ReLU activation; followed by three layers of 3×3×3 convolutions, each layer is followed by batch normalization and max pooling.

[0242] The time-varying feature matrix described below is input into the bidirectional LSTM with 128 hidden units to capture frequency domain correlation.

[0243] Step S65: All spatial feature vectors and all dynamic feature vectors are merged in the concatenation layer of the CNN / RNN hybrid deep neural network, and the resulting joint feature representation is the mapping relationship.

[0244] Preferably, the flattened features (approximately 1024 dimensions) output by the CNN are concatenated with the temporal features (256 dimensions) output by the RNN, and then nonlinearly fused through a fully connected layer (512 nodes) to obtain the mapping relationship.

[0245] This embodiment obtains the dielectric constant and permeability parameters of the target brain tissue at a preset frequency band, and calculates the unified electromagnetic property parameters of the whole brain through weighted averaging. A uniform spherical model with a diameter matching the anatomical dimensions of the human cranial cavity is constructed, and the boundary conditions of the uniform spherical model from the inside to the outside are defined as an open electromagnetic boundary from the tissue to the air. Using the uniform spherical model as the computational domain, Maxwell's differential control equations are constructed based on the unified electromagnetic property parameters of the whole brain. External electromagnetic waves with different combinations of incident parameters are input into the differential control equations through dynamic simulation to obtain several brain electromagnetic field distribution datasets. All brain electromagnetic field distribution datasets are discretized into three-dimensional mesh datasets, and all three-dimensional mesh datasets are divided into sample sets and validation sets. All sample sets are input into a CNN / RNN hybrid deep neural network for training to establish a mapping relationship from incident parameters to the three-dimensional electromagnetic field distribution in the brain, resulting in a trained neural network model. All validation sets are input into the trained neural network model to obtain electromagnetic response prediction data through the mapping relationship. This embodiment simplifies brain tissue into a homogeneous isotropic sphere model, significantly reducing computational complexity and laying the foundation for efficient solutions. Furthermore, it constructs a physical model of electromagnetic wave propagation and scattering based on Maxwell's equations, accurately describing the motion of electromagnetic waves within brain tissue. By deeply integrating the physical model into a deep learning algorithm, it efficiently solves for the three-dimensional electromagnetic field distribution within the brain. Finally, it analyzes the propagation path, energy distribution, and scattering field characteristics of electromagnetic waves, enabling the localization and qualitative diagnosis of lesions such as cerebral edema and tumors, providing crucial support for brain tissue health assessment. This embodiment overcomes the efficiency bottleneck of traditional numerical methods in brain electromagnetic computation, improving the timeliness and accuracy of brain disease diagnosis, and providing reliable technical support for rapid bedside diagnosis of acute conditions such as stroke.

[0246] like Figure 7 As shown, this embodiment provides an example of a computational system for the conduction and scattering process of low-frequency electromagnetic waves in the brain. In this embodiment, the computational system is applied to the computational method as described in the above embodiment.

[0247] Specifically, the computing system includes an electromagnetic characteristic parameter acquisition module 1, a uniform sphere model construction module 2, a differential control equation construction module 3, an electromagnetic wave dynamic simulation module 4, an electromagnetic field distribution discretization module 5, a hybrid deep neural network training module 6, and an electromagnetic response data prediction module 7.

[0248] The module comprises four sub-modules: Electromagnetic property parameter acquisition module 1, which acquires the dielectric constant and permeability parameters of the target brain tissue at a preset frequency band and calculates the unified electromagnetic property parameters of the whole brain through weighted average calculation; Uniform sphere model construction module 2, which constructs a uniform sphere model with a diameter matching the anatomical dimensions of the human cranial cavity and defines the boundary conditions of the uniform sphere model from the inside to the outside as an open electromagnetic boundary from the tissue to the air; Differential control equation construction module 3, which uses the uniform sphere model as the computational domain and constructs Maxwell's differential control equations based on the unified electromagnetic property parameters of the whole brain; and Electromagnetic wave dynamic simulation module 4, which simulates external electromagnetic waves with different incident parameters through dynamic simulation. The input differential control equations are combined to obtain several brain electromagnetic field distribution datasets; the electromagnetic field distribution discretization module 5 is used to discretize all brain electromagnetic field distribution datasets into three-dimensional mesh datasets, and divide all three-dimensional mesh datasets into sample sets and validation sets; the hybrid deep neural network training module 6 is used to input all sample sets into a CNN / RNN hybrid deep neural network for training, in order to establish a mapping relationship from incident parameters to the three-dimensional electromagnetic field distribution in the brain, and obtain the trained neural network model; the electromagnetic response data prediction module 7 is used to input all validation sets into the trained neural network model, so as to obtain electromagnetic response prediction data through the mapping relationship.

[0249] Furthermore, the computing system includes an electromagnetic response data comparison module, a cerebral edema determination module, and a brain tumor determination module that are electrically or signal-connected in sequence; the electromagnetic response data comparison module and the electromagnetic response data prediction module 7 are electrically or signal-connected.

[0250] Among them, the electromagnetic response data comparison module is used to extract the electric field attenuation rate and distribution broadening degree from the electromagnetic response prediction data, and compare them with the preset healthy brain tissue benchmark database respectively; the cerebral edema determination module is used to determine cerebral edema if the comparison result shows an abnormal increase in dielectric parameters and the electric field spreads in concentric circles; the brain tumor determination module is used to determine brain tumor if the comparison result shows the presence of a local high absorption area accompanied by a loss of scattering direction.

[0251] Furthermore, the electromagnetic characteristic parameter acquisition module 1 specifically includes a first electromagnetic characteristic parameter acquisition unit, a second electromagnetic characteristic parameter acquisition unit, and a third electromagnetic characteristic parameter acquisition unit that are electrically or signal-connected in sequence; the third electromagnetic characteristic parameter acquisition unit is electrically or signal-connected to the uniform sphere model construction module 2.

[0252] The first electromagnetic property parameter acquisition unit is used to partition the target brain tissue based on a preset voxel size and collect the dielectric constant and permeability parameters of each partition in a preset frequency band.

[0253] The second electromagnetic property parameter acquisition unit is used to obtain the unified permittivity and unified permeability parameters of the whole brain by weighted averaging using equation (1):

[0254] (1).

[0255] in, To achieve a unified dielectric constant for the entire brain, The first target brain tissue Each partition voxel For the first The dielectric constant of each voxel in the partition. To achieve a unified magnetic permeability parameter for the entire brain, For the first The permeability parameters of each voxel in the partition.

[0256] The third electromagnetic property parameter acquisition unit is used to integrate the unified permittivity and unified permeability parameters of the whole brain voxels of the same partition into unified electromagnetic property parameters of the whole brain.

[0257] Furthermore, the differential control equation construction module 3 specifically includes a first differential control equation construction unit and a second differential control equation construction unit that are electrically or signal-connected in sequence; the first differential control equation construction unit is electrically or signal-connected to the uniform sphere model construction module 2, and the second differential control equation construction unit is electrically or signal-connected to the electromagnetic wave dynamic simulation module 4.

[0258] The first differential control equation building unit is used to construct Maxwell's differential control equations through equation (2):

[0259] (2).

[0260] in, For vector differential operators, For curl, The electric field intensity vector in the computational domain. The imaginary unit, For the magnetic permeability parameter of the computational domain, The magnetic field strength vector in the computational domain, To calculate the time-harmonic oscillation frequency in the computational domain, Let be the dielectric constant of the computational domain. The current density is the excitation source current density.

[0261] The second differential governing equation building unit is used to define the constraint conditions of Maxwell's differential governing equation according to equation (3), which satisfy the energy dissipation characteristics of external electromagnetic waves at the far-field boundary:

[0262] (3).

[0263] in, To measure the propagation distance of the stimulus source, The scattered field of the excitation source, The wave number of the external electromagnetic wave.

[0264] Furthermore, the electromagnetic wave dynamic simulation module 4 specifically includes a first electromagnetic wave dynamic simulation unit, a second electromagnetic wave dynamic simulation unit, and a third electromagnetic wave dynamic simulation unit that are electrically connected in sequence; the first electromagnetic wave dynamic simulation unit is electrically or signal-connected to the second differential control equation construction unit, and the third electromagnetic wave dynamic simulation unit is electrically or signal-connected to the electromagnetic field distribution discretization module 5.

[0265] The first electromagnetic wave dynamic simulation unit is used to construct several multidimensional parameter matrices based on several combinations of incident parameters of external electromagnetic waves; the second electromagnetic wave dynamic simulation unit is used to convert the differential control equations into a system of discrete algebraic equations using the finite difference time-domain method; and the third electromagnetic wave dynamic simulation unit is used to input each multidimensional parameter matrix into the system of discrete algebraic equations and solve for a dataset of electromagnetic field distribution in the brain.

[0266] Furthermore, the electromagnetic field distribution discretization module 5 specifically includes a first electromagnetic field distribution discretization unit, a second electromagnetic field distribution discretization unit, a third electromagnetic field distribution discretization unit, a fourth electromagnetic field distribution discretization unit, and a fifth electromagnetic field distribution discretization unit that are electrically or signal-connected in sequence; the first electromagnetic field distribution discretization unit is electrically or signal-connected to the third electromagnetic wave dynamic simulation unit, and the fifth electromagnetic field distribution discretization unit is electrically or signal-connected to the hybrid deep neural network training module 6.

[0267] The first electromagnetic field distribution discretization unit is used to define a right-handed coordinate system with its origin located at the center of the uniform sphere model, based on the uniform sphere model. The Z-axis of the right-handed coordinate system is aligned with the anatomical sagittal axis of the target brain tissue. The second electromagnetic field distribution discretization unit is used to divide the right-handed coordinate system based on a preset voxel size, forming several three-dimensional grids. The third electromagnetic field distribution discretization unit is used to map the current brain electromagnetic field distribution dataset to the three-dimensional grid of the right-handed coordinate system using trilinear interpolation. The fourth electromagnetic field distribution discretization unit is used to fill the three-dimensional grid with zero values, and obtain a voxel dataset with physical labels based on all the three-dimensional grids, which is a three-dimensional grid dataset. The fifth electromagnetic field distribution discretization unit is used to divide the current voxel dataset into a sample set and a validation set.

[0268] Furthermore, the hybrid deep neural network training module 6 specifically includes a first hybrid deep neural network training unit, a second hybrid deep neural network training unit, a third hybrid deep neural network training unit, a fourth hybrid deep neural network training unit, and a fifth hybrid deep neural network training unit that are electrically or signal-connected in sequence; the first hybrid deep neural network training unit and the fifth electromagnetic field distribution discretization unit are electrically or signal-connected.

[0269] The training unit comprises five components: a first hybrid deep neural network (HNN) training unit, a second hybrid deep neural network training unit, and a third hybrid deep neural network training unit. The first unit organizes all brain electromagnetic field distribution datasets into multi-channel three-dimensional tensors according to the spatial dimension of the computational domain. The second unit organizes all multi-dimensional parameter matrices into time-varying feature matrices aligned with the temporal sequence of all multi-channel three-dimensional tensors. The third unit defines the number of input channels of the CNN branch of the CNN / RNN hybrid deep neural network as consistent with the dimension of all multi-channel three-dimensional tensors, and the input feature dimension of the RNN branch as consistent with the parameter dimension of all time-varying feature matrices. The fourth unit extracts spatial feature vectors from all multi-channel three-dimensional tensors through the CNN branch and extracts dynamic feature vectors from all time-varying feature matrices through the RNN branch. The fifth unit merges all spatial feature vectors and all dynamic feature vectors at the concatenation layer of the CNN / RNN hybrid deep neural network, and the resulting joint feature representation is the mapping relationship.

[0270] It should be noted that this embodiment is a functional module embodiment based on the above method embodiment. For the preferred, extended, limited, exemplified and principle explanation parts of this embodiment, please refer to the above embodiment. This embodiment will not repeat them.

[0271] This embodiment obtains the dielectric constant and permeability parameters of the target brain tissue at a preset frequency band, and calculates the unified electromagnetic property parameters of the whole brain through weighted averaging. A uniform spherical model with a diameter matching the anatomical dimensions of the human cranial cavity is constructed, and the boundary conditions of the uniform spherical model from the inside to the outside are defined as an open electromagnetic boundary from the tissue to the air. Using the uniform spherical model as the computational domain, Maxwell's differential control equations are constructed based on the unified electromagnetic property parameters of the whole brain. External electromagnetic waves with different combinations of incident parameters are input into the differential control equations through dynamic simulation to obtain several brain electromagnetic field distribution datasets. All brain electromagnetic field distribution datasets are discretized into three-dimensional mesh datasets, and all three-dimensional mesh datasets are divided into sample sets and validation sets. All sample sets are input into a CNN / RNN hybrid deep neural network for training to establish a mapping relationship from incident parameters to the three-dimensional electromagnetic field distribution in the brain, resulting in a trained neural network model. All validation sets are input into the trained neural network model to obtain electromagnetic response prediction data through the mapping relationship. This embodiment simplifies brain tissue into a homogeneous isotropic sphere model, significantly reducing computational complexity and laying the foundation for efficient solutions. Furthermore, it constructs a physical model of electromagnetic wave propagation and scattering based on Maxwell's equations, accurately describing the motion of electromagnetic waves within brain tissue. By deeply integrating the physical model into a deep learning algorithm, it efficiently solves for the three-dimensional electromagnetic field distribution within the brain. Finally, it analyzes the propagation path, energy distribution, and scattering field characteristics of electromagnetic waves, enabling the localization and qualitative diagnosis of lesions such as cerebral edema and tumors, providing crucial support for brain tissue health assessment. This embodiment overcomes the efficiency bottleneck of traditional numerical methods in brain electromagnetic computation, improving the timeliness and accuracy of brain disease diagnosis, and providing reliable technical support for rapid bedside diagnosis of acute conditions such as stroke.

[0272] Figure 8 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Figure 8 As shown, the electronic device 8 includes a processor 81 and a memory 82 coupled to the processor 81.

[0273] The memory 82 stores program instructions for implementing the federated learning-based collaborative energy-saving method for government data clusters in any of the above embodiments.

[0274] The processor 81 is used to execute program instructions stored in the memory 82 for collaborative energy saving of government data clusters based on federated learning.

[0275] The processor 81 can also be referred to as a CPU (Central Processing Unit). The processor 81 may be an integrated circuit chip with signal processing capabilities. The processor 81 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0276] Furthermore, Figure 9 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. See also: Figure 9 In this embodiment of the application, the storage medium 9 stores program instructions 91 capable of implementing all the above methods. These program instructions 91 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0277] In the several embodiments provided in this application, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, signal, or other forms.

[0278] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for calculating the conduction and scattering process of low-frequency electromagnetic waves in the brain, the method being applied to low-frequency electromagnetic waves incident on and propagating in target brain tissue, characterized in that... The calculation method includes: Step S1: Obtain the dielectric constant and permeability parameters of the target brain tissue in a preset frequency band, and calculate the unified electromagnetic property parameters of the whole brain by weighted average. Step S2: Construct a uniform spherical model with a diameter matching the anatomical dimensions of the human cranial cavity, and define the boundary conditions of the uniform spherical model from the inside to the outside as an open electromagnetic boundary from tissue to air. Step S3: Construct Maxwell's differential control equations using the uniform sphere model as the computational domain and the unified electromagnetic property parameters of the whole brain; Step S4: By dynamically simulating external electromagnetic waves with different combinations of incident parameters, inputting them into the differential control equation, several datasets of electromagnetic field distribution in the brain are obtained. Step S5: Discretize all brain electromagnetic field distribution datasets into three-dimensional grid datasets, and divide all three-dimensional grid datasets into sample sets and validation sets; Step S6: Input all sample sets into a CNN / RNN hybrid deep neural network for training to establish a mapping relationship from incident parameters to the three-dimensional electromagnetic field distribution in the brain, and obtain the trained neural network model. Step S7: Input all validation sets into the trained neural network model to obtain electromagnetic response prediction data through the mapping relationship.

2. The calculation method according to claim 1, characterized in that, Step S7: Input all validation sets into the trained neural network model to obtain electromagnetic response prediction data through the mapping relationship, followed by: Step S10: Extract the electric field attenuation rate and distribution broadening from the electromagnetic response prediction data, and compare them with the preset healthy brain tissue benchmark database respectively. Step S20: If the comparison result shows that the dielectric parameter is abnormally high and the electric field spreads in concentric circles, then it is determined to be cerebral edema. Step S30: If the comparison result shows a localized high absorption region accompanied by a loss of scattering directionality, it is determined to be a brain tumor.

3. The calculation method according to claim 1, characterized in that, Step S1: Obtain the dielectric constant and permeability parameters of the target brain tissue in a preset frequency band, and calculate the unified electromagnetic property parameters of the whole brain by weighted average, including: Step S11: The target brain tissue is divided into partitions based on a preset voxel size, and the dielectric constant and permeability parameters of each partition are collected in a preset frequency band. Step S12, the unified permittivity and unified permeability parameters of the whole brain are obtained by weighted averaging using equation (1): （1）； in, To achieve a unified dielectric constant for the entire brain, The first of the target brain tissue Each partition voxel For the first The dielectric constant of each voxel in the partition. To achieve a unified magnetic permeability parameter for the entire brain, For the first The permeability parameters of each voxel in the partition; Step S13: Integrate the unified dielectric constant and unified magnetic permeability parameters of the whole brain for the same voxels into unified electromagnetic property parameters of the whole brain.

4. The calculation method according to claim 1, characterized in that, Step S3, using the uniform sphere model as the computational domain and the unified electromagnetic property parameters of the whole brain as the construction domain, constructs the Maxwell differential governing equations, including: Step S31, construct the Maxwell differential governing equation using equation (2): （2）； in, For vector differential operators, For curl, Let be the electric field intensity vector of the computational domain. The imaginary unit, Let be the permeability parameter of the computational domain. Let be the magnetic field strength vector of the computational domain. The time-resonance oscillation frequency of the computational domain. The dielectric constant of the computational domain is given. The current density of the excitation source; Step S32, according to equation (3), the constraint conditions of the Maxwell differential governing equations are defined to satisfy the energy dissipation characteristics of the external electromagnetic wave at the far-field boundary: （3）； in, To measure the propagation distance of the stimulus source, The scattered field of the excitation source, The wave number of the external electromagnetic wave.

5. The calculation method according to claim 1, characterized in that, Step S4: By dynamically simulating external electromagnetic waves with different combinations of incident parameters, inputting them into the differential control equation, several datasets of brain electromagnetic field distributions are obtained, including: Step S41: Construct several multidimensional parameter matrices based on several combinations of incident parameters of external electromagnetic waves; Step S42: The differential control equations are converted into a system of discrete algebraic equations using the finite difference time-domain method. Step S43: Input each multidimensional parameter matrix into the discrete algebraic equation system and solve to obtain a brain electromagnetic field distribution dataset.

6. The calculation method according to claim 3, characterized in that, Step S5: Discretize all brain electromagnetic field distribution datasets into three-dimensional mesh datasets, and divide all three-dimensional mesh datasets into sample sets and validation sets, including: Step S51: Define a right-handed coordinate system with its origin located at the center of the uniform sphere model based on the uniform sphere model, and align the Z-axis of the right-handed coordinate system with the anatomical sagittal axis of the target brain tissue. Step S52: Divide the right-handed coordinate system based on the preset voxel size to form several three-dimensional meshes; Step S53: Map the current brain electromagnetic field distribution dataset to the three-dimensional grid of the right-hand coordinate system using trilinear interpolation. Step S54: Zero-value filling is performed on the three-dimensional mesh with countless values. Based on all the three-dimensional meshes, a voxel dataset with physical labels is obtained, which is a three-dimensional mesh dataset. Step S55: Divide the current voxel dataset into a sample set and a validation set.

7. The calculation method according to claim 1, characterized in that, Step S6: Input all sample sets into a CNN / RNN hybrid deep neural network for training to establish a mapping relationship from incident parameters to the three-dimensional electromagnetic field distribution in the brain, resulting in a trained neural network model, including: Step S61: Organize all brain electromagnetic field distribution datasets into multi-channel three-dimensional tensors according to the spatial dimension of the computational domain; Step S62: Organize all multidimensional parameter matrices into time variation feature matrices aligned with the time series of all multi-channel three-dimensional tensors; Step S63: Define the number of input channels of the CNN branch of the CNN / RNN hybrid deep neural network to be consistent with the dimension of all multi-channel three-dimensional tensors, and define the dimension of the input features of the RNN branch to be consistent with the parameter dimension of all time-varying feature matrices; Step S64: Extract the spatial feature vectors of all multi-channel three-dimensional tensors through the CNN branch, and extract the dynamic feature vectors of all time-varying feature matrices through the RNN branch; Step S65: All spatial feature vectors and all dynamic feature vectors are merged in the splicing layer of the CNN / RNN hybrid deep neural network, and the resulting joint feature representation is the mapping relationship.

8. A computational system for the conduction and scattering process of low-frequency electromagnetic waves in the brain, said computational system being applied to the computational method as described in any one of claims 1 to 7, characterized in that, The computing system includes: The electromagnetic property parameter acquisition module is used to acquire the dielectric constant and magnetic permeability parameters of the target brain tissue in a preset frequency band, and to obtain the unified electromagnetic property parameters of the whole brain by weighted average calculation. The uniform sphere model construction module is used to construct a uniform sphere model with a diameter matching the anatomical size of the human cranial cavity, and defines the boundary conditions of the uniform sphere model from the inside to the outside as an open electromagnetic boundary from tissue to air. The differential control equation construction module is used to construct Maxwell's differential control equations using the uniform sphere model as the computational domain and the unified electromagnetic property parameters of the whole brain. The electromagnetic wave dynamic simulation module is used to input external electromagnetic waves with different combinations of incident parameters into the differential control equation through dynamic simulation to obtain several brain electromagnetic field distribution datasets. The electromagnetic field distribution discretization module is used to discretize all brain electromagnetic field distribution datasets into three-dimensional grid datasets, and divide all three-dimensional grid datasets into sample sets and validation sets. The hybrid deep neural network training module is used to input all sample sets into the CNN / RNN hybrid deep neural network for training, in order to establish the mapping relationship from the incident parameters to the three-dimensional electromagnetic field distribution in the brain, and obtain the trained neural network model. The electromagnetic response data prediction module is used to input all validation sets into the trained neural network model to obtain electromagnetic response prediction data through the mapping relationship.

9. An electronic device, characterized in that, The method includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the computation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed by a processor, enable the computation method as described in any one of claims 1 to 7.

Images