Electromagnetic imaging method, device, system, equipment and medium

Abstract

The invention relates to the technical field of biomedicine, and provides an electromagnetic imaging method, device, system and equipment and a medium, and the method comprises the steps: constructing an electrical impedance imaging model and an electromagnetic imaging model; generating a source domain data set by using the electrical impedance imaging model, and generating a target domain data set by using the electromagnetic imaging model; training a preset auxiliary model according to the source domain data set to obtain an improved auxiliary model; performing adaptation layer training and whole network fine tuning on the improved auxiliary model by using the target domain data set to obtain an electromagnetic imaging optimization model; and inputting to-be-tested data in the target domain data set into the electromagnetic imaging optimization model to output a reconstructed image of craniocerebral lesion signal distribution. By means of the scheme, the advantage that a large number of data sets are easy to generate during electrical impedance imaging model construction is utilized, under limited resources, the optimal balance between calculation time and imaging precision is achieved, and generalization of the model in similar problems is verified.

Description

Technical Field

[0001] This application relates to the field of biomedical technologies, and in particular, to an electromagnetic imaging method, device, system, equipment, and medium.

Background Art

[0002] In recent years, cerebrovascular diseases, as a common type of disease that seriously threatens human health, have attracted much attention. They mainly include cerebral atherosclerosis, cerebral artery injury, cerebral arteritis, etc. Especially for the middle-aged and elderly population, cerebrovascular diseases are characterized by high prevalence, high disability rate, and high mortality rate. Therefore, it is crucial to conduct real-time detection and imaging research on cerebral blood vessels. However, there are still difficulties in its real-time imaging technology at present.

[0003] Since existing imaging technologies face various problems such as high cost, potential harm to the human body during long-term use, and the need for complex operations, there is an urgent need to develop a new imaging technology that can meet the requirements of the brain for safety, rapidity, and accuracy. The unique electrical properties of human tissues can be used as a characteristic feature of tissues for imaging, which is expected to bring hope for brain imaging. With the in-depth research of various brain function detection technologies, medical detection means are gradually developing towards the direction of comprehensive, continuous, non-invasive, low-cost, and high-sensitivity detection capabilities. The bioimpedance detection technology can relate the impedance changes in the brain to the physiological and pathological activities of the brain by measuring the impedance of brain tissues and blood components. This technology has the advantages of high efficiency and convenience while meeting the above requirements. Therefore, imaging the cerebral blood vessels has become an important method in brain function monitoring. Although there are many mature algorithms available and in use currently, their drawbacks are also obvious. The biggest problem is that many algorithms require a large amount of labeled data, but obtaining a large amount of electromagnetic acquisition data clinically is extremely costly in terms of material and financial resources. Moreover, the electromagnetic imaging analysis method using the finite element method has problems such as large computational amount, long computational time, and inability to be reused, making it difficult to meet the actual needs.

Summary of the Invention

[0004] Embodiments of this application provide an electromagnetic imaging method, device, system, equipment, and medium, aiming to solve the technical problems such as limited measurement data volume, low measurement and imaging accuracy, and slow imaging speed existing in the related technologies.

[0005] In a first aspect, embodiments of this application provide an electromagnetic imaging method, including:

[0006] Constructing an electrical impedance tomography model and an electromagnetic imaging model;

[0007] Generate a source domain dataset using the electrical impedance tomography model, and generate a target domain dataset using the electromagnetic tomography model, where the source domain dataset includes first conductivity distribution data and first boundary voltage data obtained from a constant excitation current, and the target domain dataset includes second conductivity distribution data and second boundary voltage data obtained from an excitation magnetic field;

[0008] Train a preset auxiliary model according to the first conductivity distribution data and the first boundary voltage data in the source domain dataset to obtain an improved auxiliary model;

[0009] Use the first quantity of the second conductivity distribution data and the second boundary voltage data in the target domain dataset to perform adaptation layer training and full network fine-tuning on the improved auxiliary model to obtain an electromagnetic tomography optimization model;

[0010] Input the data to be tested in the target domain dataset into the electromagnetic tomography optimization model to output a reconstructed image of the signal distribution of cranial lesions; where the data to be tested is the second quantity of the second conductivity distribution data and the second boundary voltage data except for the first quantity.

[0011] In one embodiment, optionally, the electromagnetic tomography model includes a plurality of excitation coils, and each excitation coil corresponds to a TMR sensor, where the TMR sensor is located at the center of each excitation coil.

[0012] In one embodiment, optionally, the electrical impedance tomography model includes a plurality of electrodes and a test object with a preset shape, where the plurality of electrodes are equidistantly placed on the surface of the test object with a preset shape, and the first boundary voltage data corresponding to the source domain dataset is obtained by applying a constant excitation current and measuring the boundary voltage; and / or

[0013] The method for obtaining the first boundary voltage data by applying a constant excitation current and measuring the boundary voltage includes: performing simulation of the imaging forward problem based on a finite element target object and adjacent excitation modes under a constant excitation current to determine the corresponding first boundary voltage data.

[0014] In one embodiment, optionally, the method for generating the first conductivity distribution data corresponding to the source domain dataset using the electrical impedance tomography model includes:

[0015] Adjust the quantity, position, size, and shape of the finite element target object with a set conductivity to obtain the first conductivity distribution data; and / or

[0016] Generating the target domain dataset using the electromagnetic tomography model includes:

[0017] Add a target substance with a random shape and size consistent with the blood conductivity within the target area, and configure the second conductivity distribution data corresponding to the target substance that simulates the intracranial hemorrhage point;

[0018] Under the excitation magnetic field generated by the excitation coil, obtain the magnetic induction intensity at the boundary position by the TMR sensor to obtain the second boundary voltage data.

[0019] In one embodiment, optionally, train a preset auxiliary model according to the source domain dataset to obtain an improved auxiliary model, including:

[0020] Normalize the first conductivity distribution data and the first boundary voltage data in the source domain dataset respectively to obtain the processed conductivity distribution data and boundary voltage data;

[0021] Use the processed first boundary voltage data as the input of the preset auxiliary model, and the corresponding processed first conductivity distribution data as the training label to perform iterative training on the preset auxiliary model to obtain the improved auxiliary model.

[0022] In one embodiment, optionally, the improved auxiliary model is an improved neural network structure, including: an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first pooling layer, a second pooling layer, a fully connected layer, and an output layer;

[0023] The input layer deforms the input processed first boundary voltage data to obtain a first matrix;

[0024] The first convolutional layer uses multiple convolutional kernels with a first set size to extract features from the first matrix and outputs a feature vector of the first dimension;

[0025] The first pooling layer uses a maximum pooling window with a second set size and a set pooling stride to perform dimensionality reduction and important feature extraction on the feature vector of the first dimension to obtain a feature vector of the second dimension;

[0026] The second convolutional layer increases the number of output channels of the feature vector of the second dimension to obtain a feature vector of the third dimension;

[0027] The third convolutional layer and the fourth convolutional layer sequentially extract features from the feature vector of the third dimension to obtain a target feature vector;

[0028] The second pooling layer uses a maximum pooling window with a third set size and a set pooling stride to perform dimensionality reduction processing on the target feature vector to obtain a dimensionality-reduced target feature vector;

[0029] The fully connected layer integrates the dimensionality-reduced target feature vectors to obtain corresponding processed first conductivity distribution data.

[0030] The output layer performs fully connected upsampling on the output of the processed first conductivity distribution data to obtain first conductivity distribution data that is consistent with the preset number of finite element divisions.

[0031] In one embodiment, optionally, the improved auxiliary model is trained with the adaptation layer and fine-tuned across the network using the target domain dataset to obtain an electromagnetic imaging optimization model, including:

[0032] Freeze the model parameters of the improved auxiliary model and add an adaptation layer to the improved auxiliary model.

[0033] Select a first number of target domain data from the target domain dataset to train the adaptation layer, and obtain a trained adaptation layer.

[0034] Select a second number of labeled target domain data from the target domain dataset, excluding the first number, and fine-tune the trained improved auxiliary model and the adaptation layer across the network to obtain the electromagnetic imaging optimization model.

[0035] In a second aspect, an embodiment of the present application provides an electromagnetic imaging device, including:

[0036] A construction module for constructing an electromagnetic imaging model and an electrical impedance tomography model.

[0037] A generation module for generating a source domain dataset using the electrical impedance tomography model and generating a target domain dataset using the electromagnetic imaging model, where the source domain dataset includes first conductivity distribution data and first boundary voltage data obtained from a constant excitation current, and the target domain dataset includes second conductivity distribution data and second boundary voltage data obtained from an excitation magnetic field.

[0038] A training module for training a preset auxiliary model according to the first conductivity distribution data and the first boundary voltage data in the source domain dataset to obtain an improved auxiliary model.

[0039] An adjustment module for performing adaptation layer training and full network fine-tuning on the improved auxiliary model using the first number of second conductivity distribution data and second boundary voltage data in the target domain dataset to obtain an electromagnetic imaging optimization model.

[0040] An output module, configured to input the data to be tested in the target domain dataset into the electromagnetic imaging optimization model to output a reconstructed image of the signal distribution of cranial lesions; wherein, the data to be tested is the second conductivity distribution data and the second boundary voltage data of a second quantity except for the first quantity.

[0041] In a third aspect, an electromagnetic imaging system is provided, including:

[0042] A plurality of excitation coils in a ring shape, a TMR sensor placed at a hollow position inside the plurality of excitation coils in the ring shape, and a target area formed on one side of the plurality of excitation coils;

[0043] A partitioning unit, configured to perform finite element partitioning on the target area to obtain a preset number of meshes of finite element partitioning;

[0044] A first acquisition unit, configured to measure the boundary voltage of a test object by using a plurality of electrodes and a test object with a preset shape in the target area to obtain the first boundary voltage data corresponding to the source domain dataset; before obtaining the first boundary voltage data, in the mesh, adjust the quantity, position, size, and shape of a finite element target object with a set conductivity to obtain the first conductivity distribution data corresponding to the source domain dataset;

[0045] A second acquisition unit, configured to apply a constant excitation current to a plurality of excitation coils, and obtain the second boundary voltage data of the magnetic induction intensity at the boundary position by the TMR sensor under the excitation magnetic field generated by the excitation coils; before obtaining the second boundary voltage data, add a target substance with a random shape and size and consistent with the blood conductivity in the target area, and determine the second conductivity distribution data of the target substance corresponding to the simulated intracranial hemorrhage point in the target domain dataset;

[0046] A training unit, configured to train a preset auxiliary model according to the first conductivity distribution data and the first boundary voltage data in the source domain dataset to obtain an improved auxiliary model;

[0047] An adjustment unit, configured to perform adaptation layer training and full network fine-tuning on the improved auxiliary model by using the first quantity of the second conductivity distribution data and the second boundary voltage data in the target domain dataset to obtain an electromagnetic imaging optimization model;

[0048] An output unit, configured to input the data to be tested in the target domain dataset into the electromagnetic imaging optimization model to output a reconstructed image of the signal distribution of cranial lesions; wherein, the data to be tested is the second conductivity distribution data and the second boundary voltage data of a second quantity except for the first quantity.

[0049] Fourthly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above electromagnetic imaging method are implemented.

[0050] Fifthly, a computer-readable storage medium is provided. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above electromagnetic imaging method are implemented.

[0051] In the solutions implemented by the above electromagnetic imaging method, device, equipment, and medium, an electrical impedance tomography (EIT) model and an electromagnetic imaging model are constructed; a source domain data set is generated using the EIT model, and a target domain data set is generated using the electromagnetic imaging model. A preset auxiliary model is trained according to the source domain data set to obtain an improved auxiliary model; the improved auxiliary model is subjected to an adaptation layer training and a full network fine-tuning using the target domain data set to obtain an electromagnetic imaging optimization model; the data to be tested in the target domain data set is input into the electromagnetic imaging optimization model to output a reconstructed image of the distribution of brain lesion signals. Through the above technical solutions of the present invention, the advantage of the EIT model in easily generating a large amount of data sets is utilized to obtain an improved auxiliary model based on this, and an optimization model is obtained through model migration, so that the small sample data obtained based on the electromagnetic imaging model can be accurately reconstructed. Under limited resources, an optimal trade-off between calculation time and imaging accuracy is achieved, and the generalization of the model in similar problems is verified.

BRIEF DESCRIPTION OF THE DRAWINGS

[0052] To more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0053] Figure 1 Shows a schematic flowchart of an electromagnetic imaging method according to an embodiment of the present application.

[0054] Figure 2 and Figure 3 Shows a schematic diagram of an electromagnetic imaging model according to an embodiment of the present application.

[0055] Figure 4 and Figure 5 Shows a schematic diagram of an electrical impedance tomography (EIT) model according to an embodiment of the present application.

[0056] Figure 6 Shows a schematic flowchart of using an auxiliary model to solve the imaging inverse problem according to an embodiment of the present application.

[0057] Figure 7 Shows a schematic diagram of the network structure of an improved auxiliary model according to an embodiment of the present application.

[0058] Figure 8 Shows a block diagram of an electromagnetic imaging device according to an embodiment of the present application.

[0059] Figure 9 Shows a block diagram of a computer device according to an embodiment of the present application.

Detailed implementation manners

[0060] For a better understanding of the technical solution of the present application, the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0061] It should be clear that the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts belong to the scope of protection of the present application.

[0062] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms of "a", "the" and "said" used in the embodiments of the present application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0063] Some embodiments of the present application will be described in detail below with reference to the accompanying drawings. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0064] Please refer to Figure 1 , Figure 1 Shows a schematic flowchart of an electromagnetic imaging method according to an embodiment of the present application.

[0065] As Figure 1 shown, the electromagnetic imaging method includes:

[0066] Step S101, constructing an electrical impedance tomography model and an electromagnetic imaging model;

[0067] Based on the model similarity between electromagnetic imaging and electrical impedance tomography, an electromagnetic imaging and electrical impedance tomography model is built.

[0068] In one embodiment, optionally, the electromagnetic imaging model includes a plurality of excitation coils, and a TMR sensor corresponds to each excitation coil, wherein the TMR sensor is located at the center of each excitation coil.

[0069] In Figure 2 and Figure 3Among them, 1 is the excitation coil, 2 is the TMR sensor, and 3 is the target area. Among them, the electromagnetic imaging model includes: a plurality of excitation coils 1 in the shape of a ring and a TMR sensor 2 placed in the hollow position inside the plurality of excitation coils 1 (the TMR sensor 2 placed in the hollow position of the ring-shaped excitation coil 1); perform finite element division on the target area 3 formed on one side of the plurality of excitation coils 1 to obtain a preset number of meshes divided by finite elements. Among them, the preset number can be configured to 576.

[0070] The electromagnetic imaging model is a simulation model of tunneling magnetoresistance sensor (Tunnel ing Magnetoresistance, TMR)-electromagnetic tomography (Electromagnetic Tomography, EMT) technology for detecting brain lesion signals. As Figure 2 and Figure 3 shown, an EMT signal detection simulation structure of 8 excitation coils - 8 TMRs designed by using COMSOL Multiphysics (multiphysics simulation software) is constructed. Two cross-sections with excitation currents flowing in opposite directions are used to represent two-dimensional coils (excitation coils 1). The two cross-sections (rectangular cross-sections) corresponding to the same two-dimensional coil (excitation coil 1) are connected by a dotted line ( Figure 3 the cross-sections corresponding to taking 0 or a set value in the z direction of the plurality of excitation coils 1 in

[0071] In one embodiment, optionally, the electrical impedance imaging model includes a plurality of electrodes and a test object with a preset shape. Among them, the plurality of electrodes are equidistantly placed on the surface of the test object with a preset shape, and information (first boundary voltage data) is obtained by applying a constant excitation current and measuring the boundary voltage.

[0072] In the embodiments of the present disclosure, the electrical impedance imaging model includes: a plurality of electrode models with a preset shape. Among them, the conductivity of the test object and the background conductivity of the target area are set; under the background conductivity of the test object and the target area, by equidistantly placing a plurality of electrodes on the surface of the test object with a preset shape, applying a constant excitation current and measuring the boundary voltage to obtain the first boundary voltage data corresponding to the bioelectrical information.

[0073] For example, as Figure 4 and Figure 5As shown, the multiple electrode models with preset shapes are configured as a 16 - electrode circular model. By equidistantly placing 16 electrodes on the surface of a preset circular test object, applying a constant excitation current, and measuring the boundary voltage, bio - electrical information is obtained. Among them, a point - electrode circular finite - element model with 576 meshes containing 16 electrodes is selected. The radius of the circular target area is set to 1, and the background conductivity is 1 S / m.

[0074] Step S102: Generate a source - domain data set using the electrical impedance tomography model and generate a target - domain data set using the electromagnetic tomography model. Among them, the source - domain data set includes first conductivity distribution data and first boundary voltage data obtained from a constant excitation current, and the target - domain data set includes second conductivity distribution data and second boundary voltage data obtained from an excitation magnetic field.

[0075] Based on the finite - element and adjacent - excitation mode, use the EIDORS (Electrical Impedance Tomography and Diffuse Optical Tomography Reconstruction Software, an open - source software for electrical impedance tomography) software package in MATLAB to generate the source - domain data set. The excitation electrodes are adjacent - placed electrodes, and the potential differences between adjacent electrode pairs except the excitation electrodes are measured. As Figure 5 shown, electrodes 1 and 2 are used as excitation, and the potential differences of the remaining electrodes are measured (3 - 4, 4 - 5, 5 - 6, …, 15 - 16). One measurement with this configuration can collect 16×(16 - 3)=208 first - boundary - voltage data, which not only has rich bio - electrical information but also good sensitivity.

[0076] The specific method for setting the source - domain data set is as follows:

[0077] First, set the number, position, size, and shape of the target objects. Among them, the number of target objects ranges from 1 to 3. The position and size of the target objects are randomly determined within the target area. The conductivity range of the target objects is 0.1 S / m to 1.9 S / m. The target shapes of the target objects (shapes) include circles, triangles, and squares. The generation method is as follows:

[0078] (1) Circular target objects:

[0079] First, determine the center position of the circular target object in the target area as (xi, yi). In addition, after determining the center position, adding the radius size should ensure that it does not exceed the circular target area, that is, it is always inside the target area. The radius of the circle is selected between 0.25, 0.3, and 0.4, as long as the distance from the center to the origin plus the radius of the generated circle does not exceed 1.

[0080] (2) Square target object:

[0081] First, determine the center of the square target object within the target area. After determining the center of the square, each side of the square target is also within the circular target area, and the square can rotate at a certain angle. In this invention, only the cases of non-rotation and 45-degree rotation are selected, and the side length of the square is 0.4 or 0.55.

[0082] (3) Triangular target object:

[0083] The triangular targets are all equilateral triangles. The generation method is similar to that of the square. First, determine the center of the triangular target object within the target area, and then determine whether each vertex is within the target area. The triangular target can rotate at more angles than the square, including non-rotation and rotations of 30, 45, and 60 degrees, but the side length of the triangular target is constantly 0.6.

[0084] (4) Target object with a mixed shape:

[0085] The mixed target is generated based on the above three single-graphic targets (circular target object, square target object, triangular target object). Select non-overlapping images (images corresponding to the circular target object, square target object, and triangular target object) for finite element target design, and try to avoid selecting cases where the radius or side length is too large to avoid target overlap. The radius of the circle is selected as 0.25, the side length of the square is selected as 0.4, and the side length of the triangle is selected as 0.6.

[0086] After obtaining the finite element model, perform simulation of the forward imaging problem to obtain the boundary voltage information, and adopt the adjacent excitation mode as described for the excitation mode.

[0087] Step S103: Train a preset auxiliary model according to the first conductivity distribution data and the first boundary voltage data in the source domain dataset to obtain an improved auxiliary model.

[0088] In one embodiment, optionally, training a preset auxiliary model according to the source domain dataset to obtain an improved auxiliary model includes:

[0089] Perform normalization processing on the first boundary voltage data and the first conductivity distribution data in the source domain dataset respectively to obtain the processed first boundary voltage data and the first conductivity distribution data.

[0090] In this step, perform data preprocessing on the source domain dataset, and use the auxiliary model to perform inverse problem analysis on the processed source domain dataset. The auxiliary model parameters are trained based on the source domain dataset, and an improved auxiliary model is obtained based on the auxiliary model.

[0091] Before training the preset auxiliary model, the above two datasets (the first boundary voltage data and the first conductivity distribution data) need to be normalized and / or standardized. Normalize the first boundary voltage and the first conductivity distribution to obtain the processed first boundary voltage data and the first conductivity distribution data. The normalization is shown as follows:

[0092]

[0093] Among them, represents the data (the first boundary voltage data and the first conductivity distribution data) after normalization, x represents the unprocessed data (the first boundary voltage data and the first conductivity distribution data), x min and x max represent the minimum and maximum values in the dataset (the first boundary voltage data and the first conductivity distribution data), respectively. The boundary voltage in the dataset is standardized as the input data of the neural network, while the conductivity is normalized as the output data of the neural network, that is, the training label.

[0094] Similarly, the above method can be used to normalize and / or standardize the second boundary voltage data and the second conductivity distribution data corresponding to the target domain dataset. Normalize the second boundary voltage and the second conductivity distribution to obtain the processed second boundary voltage data and the second conductivity distribution data.

[0095] Such as Figure 6As shown, it is the process of using a convolutional neural network as a preset auxiliary model to solve the imaging inverse problem. The inverse problem is the image reconstruction process, which is to infer the conductivity distribution (the first conductivity distribution data) in the entire target area based on the potential (or current) data (the first boundary voltage data) measured at the boundary. The present invention uses an auxiliary model based on a neural network to initially complete the inverse problem calculation, which needs to be trained on a large dataset and adjust the weights and reduce errors by minimizing the objective function. Specifically, first, the collected first boundary voltage data is preprocessed (normalized and / or standardized) to meet the input requirements of the neural network corresponding to the preset auxiliary model. These processed first boundary voltage data are used as network inputs, and the first boundary conductivity distribution inside the target area of the corresponding test object is used as the training label. After multiple iterations, the neural network corresponding to the preset auxiliary model automatically learns and gradually optimizes its relevant weight and bias parameters to obtain an effective network model (the preset auxiliary model). The tested voltage data is input into the trained neural network, and the neural network corresponding to the preset auxiliary model obtains the corresponding conductivity values through inference operations. These conductivity values can then be used to generate the final reconstructed image, thereby realizing the visualization of the internal structure of the target area, that is, completing the image reconstruction stage. The specific steps are as follows:

[0096] (1) Convolutional layer: The convolutional layer extracts features from the input data (the first boundary voltage data normalized and / or standardized in the source domain dataset) through multiple convolutional kernels, which is the key part of feature extraction. Assume that the input matrix corresponding to the first boundary voltage data normalized and / or standardized in the source domain dataset is X = {x i,j |i = 1, 2,...I, j = 1, 2...J}; where i and j represent the number of rows and columns of the input matrix respectively, and x i,j is the element of the first boundary voltage data corresponding to the number of rows i and columns j of the input matrix after normalization and / or standardization. The size of the convolutional kernel is usually the same for height and width, set to F×F (F is configured as a positive integer greater than or equal to 1), then the convolutional kernel can be marked as W = {w m,n |m = 0, 1,...F - 1, n = 0, 1,...F - 1}, where m and n represent the number of rows and columns of the convolutional kernel respectively, and the convolutional kernel corresponding to the number of rows m and columns n is configured as w m,n . The calculation formula of the convolutional layer f cov is expressed as: where, a i,j is the element of the convolutional feature map output after the element of the first boundary voltage data corresponding to the number of rows i and columns j of the input matrix passes through the convolutional layer; b represents the bias of the convolutional layer.

[0097] (2) Pooling layer: The pooling layer reduces the data dimension of the convolutional feature map output by the convolutional layer to highlight the input features. Among them, the convolutional layer can be configured as multiple layers; the output of the l-th convolutional layer of the multiple layers is The output of the previous convolutional layer, i.e., the (l - 1)-th layer, is Then the expression of the pooling operation pool is:

[0098]

[0099] (3) Fully connected layer: The fully connected layer has two main functions in the network. First, the number and size of the fully connected layer can directly affect the performance of the network, and it enhances the function of the network by processing non-linear data. Secondly, the entire connected layer also acts as a classifier to integrate and map the features extracted from the previous layer into the discriminant space for classification. If the l-th layer in the above formula is the fully connected layer f, its output expression is:

[0100]

[0101] Among them, w l represents the weight of the fully connected layer f, and b l represents the bias of the fully connected layer f.

[0102] (4) Activation function layer: The activation function is usually located after the convolutional layer and the fully connected layer to introduce non-linear factors to enhance the expression ability of the network. The ReLU activation function used The expression is:

[0103]

[0104] Among them, represents the output of the fully connected layer.

[0105] As Figure 7 shown, for the inverse problem of image reconstruction of the signal distribution of cranial lesions, the present invention further designs an improved auxiliary model, that is, an improved CNN-LeNet-5 network structure. The following is the detailed configuration and function description of each layer of this network:

[0106] (1) Input layer (Input Layer): The input data of the neural network corresponding to the input layer is the measured boundary voltage signal (the first boundary voltage data). The first boundary voltage data is deformed to obtain a first matrix with a shape (size) of a set matrix size (for example, 16×13×1). The first matrix represents the simulation of the forward imaging problem based on the finite element target object and the adjacent excitation modes and adjacent measurement modes under a constant excitation current, and the corresponding first boundary voltage data is obtained.

[0107] (2) The first convolutional layer 1 (Conv Layer) uses 32 convolutional kernels of size 4×4 to extract the features of the first boundary voltage data in the first matrix and outputs a feature vector of the first dimension. Among them, the stride of each convolutional kernel moving each time is configured as 1. To ensure that the size of the feature matrix corresponding to the feature vector of the first dimension remains basically unchanged, padding is added on the basis of the feature matrix corresponding to the feature vector of the first dimension. A padding strategy of adding a layer of all 0s to the upper, left, and right boundaries and adding two layers to the lower boundary is adopted to fill the feature matrix corresponding to the feature vector of the first dimension, so that the output dimension is the first set matrix size (16×14). For example, 16×14×32.

[0108] (3) The first pooling layer 1 (Pool Layer): performs a first pooling operation on the feature vector of the first dimension output by the first convolutional layer 1 to obtain a feature vector of the second dimension. Among them, a maximum pooling window of 2×2 is adopted, and the pooling stride is 2. The first pooling operation is performed on the feature vector of the first dimension output by the first convolutional layer 1 to reduce the dimension of the feature vector of the first dimension while retaining the important features in the feature vector of the first dimension. Since pooling does not change the number of channels, the output size becomes 8×7×32.

[0109] (4) The second convolutional layer 2: increases the number of output channels of the feature vector of the second dimension to obtain a feature vector of the third dimension; among them, the setting of the second convolutional layer is similar to that of the first layer (32 convolutional kernels of size 4×4, and the stride of each convolutional kernel moving each time is configured as 1). Padding is added to the feature matrix after increasing the number of output channels of the feature vector of the second dimension. A padding strategy of adding a layer of all 0s to the upper, left, and right boundaries and adding two layers to the lower boundary is adopted to fill the feature matrix after increasing the number of output channels of the feature vector of the second dimension, so that the output dimension is the second set matrix size (8×8) smaller than the first set matrix size. For example, 8×8×64. Among them, the second convolutional layer 2 only increases the number of output channels to 64 to capture richer features, and the output dimension after the operation of the second convolutional layer 2 is 8×8×64.

[0110] (5) Third Convolutional Layer 3: Extract features from the feature vectors of the third dimension to obtain the convolutional feature vectors of the third dimension in the target feature vectors; the setting of the third convolutional layer is similar to that of the second layer (32 convolutional kernels of size 4×4, and the stride of each convolutional kernel moving each time is configured as 1). Padding is added to the feature matrix after increasing the number of output channels of the feature vectors of the third dimension. The padding strategy of adding a layer of all 0s to the upper, lower, left, and right boundaries is adopted for the feature matrix after increasing the number of output channels of the feature vectors of the third dimension. The dimension of the convolutional feature vectors of the third dimension corresponding to the third convolutional layer 3 is 8×8×64, which remains consistent with the size of the feature vectors of the third dimension.

[0111] (6) Fourth Convolutional Layer 4: Further extract features from the convolutional feature vectors of the third dimension in the target feature vectors to obtain the target convolutional feature vectors (target feature vectors) of the third dimension corresponding to the target feature vectors; among them, the fourth convolutional layer 4 is configured with 32 convolutional kernels of size 4×4, and the stride of each convolutional kernel moving each time is configured as 1. Padding is added to the feature matrix after increasing the number of output channels of the convolutional feature vectors of the third dimension. The padding strategy of adding a layer of all 0s to the upper, lower, left, and right boundaries is adopted for the feature matrix after increasing the number of output channels of the convolutional feature vectors of the third dimension. A new convolutional layer is added to improve the learning and expression ability of the model. The Padding strategy of using all 0 padding is also adopted so that the output dimension after operation is 8×8×64, which remains consistent with the size of the feature vectors of the third dimension or the convolutional feature vectors of the third dimension.

[0112] (7) Second Pooling Layer 2: Perform dimensionality reduction processing on the target feature vectors to obtain the dimensionality-reduced target feature vectors; among them, the second pooling layer 2 has the same parameters as the previous pooling layer (a maximum pooling window of 2×2, and the pooling stride is 2), and the output is 64 voltage matrices of 4×4.

[0113] (8) Fully Connected Layer (First Fully Connected Layer FCLayer): The main features of the voltage matrix after four rounds of convolution and two rounds of pooling processing have been extracted and are ready for the reconstruction of the brain lesion signal distribution image. Among them, the number of neurons in the first fully connected layer is configured as a first set number (for example, 256) to better integrate the features corresponding to the voltage matrix, and the size of the output matrix is a third set matrix size (1×1) corresponding to the first set number and smaller than the second set matrix size. For example, 1×1×256.

[0114] (9) Output Layer: The output layer is configured as a second fully-connected layer, and the second fully-connected layer directly produces the first conductivity distribution data corresponding to the first boundary voltage data required for reconstructing the brain lesion signal distribution image. Among them, the number of output neurons set in the second fully-connected layer is configured as a second set number greater than the first set number (for example, 576), and the size of the finally output conductivity distribution (conductivity) matrix is a fourth set matrix size (1×1) that is consistent with the third set matrix size corresponding to the second set number. For example, 1×1×576.

[0115] The activation function in the network all adopts the ReLU function to accelerate the convergence of network training and effectively avoid the problem of gradient disappearance. The Adam optimizer is used, the learning rate is set to 0.0005, the number of training epochs is 1200, the batch size is 256, and the mean squared error (MSE) is used as the loss function. By changing the input matrix size as described above, this network can effectively extract the key information for EIT (electrical impedance tomography) image reconstruction from the voltage data.

[0116] Using the electromagnetic imaging model, within the target area, the relationship between the vector boundary measurement value B (magnetic induction intensity) of the EMT system (electromagnetic imaging model) and the medium conductivity distribution σ(x, y) and permeability distribution μ(x, y) within the target area (physical field) can be expressed as:

[0117] B = f(σ(x, y), μ(x, y))

[0118] Among them, f represents the relationship function between the medium electrical property conductivity distribution σ(x, y) and permeability distribution μ(x, y) and the magnetic induction intensity B, and (x, y) represents the coordinates of the corresponding finite element within the target area (physical field). It can be seen from this formula that the vector boundary measurement value B (magnetic induction intensity) can be uniquely determined by the electrical property distribution within the target area (physical field).

[0119] The acquisition of the target domain dataset is to add substances with random shapes and sizes consistent with the blood conductivity within the target area, regarded as intracranial hemorrhage points, detect the magnetic induction intensity B of the boundary sensor (TMR sensor), and convert the magnetic induction intensity B to the axial characteristic (axial magnetic induction intensity) B′ = B x × cosθ + B y × sinθ, where B x and B y represent the components of the vector boundary measurement value B in the xy-axis directions respectively, and θ represents the angle between the vector boundary measurement value B and the positive x-axis direction. Based on the preset relationship between the axial magnetic induction intensity B′ and the boundary voltage data, the output voltage (second boundary voltage data) can be obtained.

[0120] For the electromagnetic imaging problem, the forward problems of EMT and EIT are described as the vector boundary measurement B = Sσ and the first boundary voltage data U = Fσ, respectively. Among them, S is the sensitivity matrix; F is the Jacobian matrix, which has the same effect as the sensitivity matrix S, and both are transfer coefficient matrices between the boundary data (vector boundary measurement B and the first boundary voltage data U) and the conductivity distribution inside the target area. The electromagnetic imaging optimization model and the improved auxiliary model have the same effect as the above transfer coefficient matrix. Therefore, based on the above-trained improved auxiliary model, an electromagnetic imaging optimization model will be obtained by combining a small amount of target domain data. Due to the certain distribution difference between the source domain dataset and the target domain dataset, simply copying the original model may result in poor reconstruction effects. How to adapt the improved auxiliary model to the target domain dataset and maximize the value of the target domain dataset is the key to obtaining the electromagnetic imaging optimization model.

[0121] Step S104: Use the first quantity of second conductivity distribution data and second boundary voltage data in the target domain dataset to perform adaptation layer training and full network fine-tuning on the improved auxiliary model to obtain the electromagnetic imaging optimization model.

[0122] In one embodiment, optionally, step S104 includes:

[0123] Freeze the model parameters of the improved auxiliary model and add an adaptation layer to the improved auxiliary model.

[0124] Select the first quantity of target domain data from the target domain dataset to train the adaptation layer to obtain the trained adaptation layer.

[0125] Select the second quantity of labeled target domain data other than the first quantity from the target domain dataset to perform full network fine-tuning on the trained auxiliary model and the adaptation layer to obtain the electromagnetic imaging optimization model.

[0126] Step S105: Input the data to be tested in the target domain dataset into the electromagnetic imaging optimization model to output a reconstructed image of the signal distribution of cranial lesions; wherein, the data to be tested is the second quantity of second conductivity distribution data and second boundary voltage data other than the first quantity.

[0127] Specifically, in combination with the Finetune idea, the present invention freezes the parameters of the pre-trained improved auxiliary model, adds an adaptation layer to obtain a new learning space, and then uses a small amount of target domain data to train the added adaptation layer. The added Batch Normalization (BN) layer can avoid the change in the distribution of the input data of the subsequent network layers caused by the change in the parameters of the forward network, and is beneficial to solving the problem of overfitting of the model due to the small amount of target domain data. The added Fully Connected Layer (FC) before the regressor can, on the one hand, enable the network to learn new knowledge, and on the other hand, by adjusting its weights, preserve or discard some features of the auxiliary model, so as to achieve the purpose of both preserving the valuable information of the source domain dataset and further learning the target domain data. The freezing is cancelled when the error of the test set converges stably during network training. Although the added adaptation layer can learn target domain information, this is based on the model trained in the source domain. The entire network is fine-tuned again using the labeled samples in a small amount of target domain datasets with a small learning rate, and finally an electromagnetic imaging optimization model that can effectively predict target domain data is obtained.

[0128] The comparative experiment on the accuracy of the model reconstruction results is quantified as a series of comparative experiments involving two intelligent algorithms (electromagnetic imaging optimization model and convolutional neural network, CNN) and two traditional non-intelligent algorithms (Gauss-Newton method, GN and Tikhonov regularization method, TK). In order to quantitatively compare the performance of these algorithms, various quantitative evaluation indicators such as Root Mean Square Error (RMSE), Structural Similarity Index (SSIM), Mean Absolute Error (MAE) and Correlation coefficient (CC) are used to evaluate the effect of image reconstruction. The calculation formula of each evaluation indicator is shown as follows:

[0129]

[0130]

[0131] c1 = (k1L) 2 , c2 = (k2L) 2

[0132]

[0133] In the above formula, σ represents the conductivity value calculated by the reconstruction algorithm, represents the conductivity value of the original model, λ σ is the variance of the calculated conductivity, is the variance of the model conductivity, μ σ Calculate the average value of the conductivity, is the average value of the model conductivity, represents the covariance of the conductivities of the two, c1 and c2 are two constants of the SSIM evaluation index, where k1 = 0.01, k2 = 0.03, and L is the range of size changes determined by the preset conductivity value.

[0134] The final results show that the electromagnetic imaging optimization model exhibits the best performance in all evaluation indexes. Especially in the SSIM and CC indexes, their values are close to 1, indicating that the reconstructed image has a very high structural similarity and correlation with the real image. The performance of CNN is better than that of GN and TK, but there is still a certain gap compared with the electromagnetic imaging optimization model.

[0135] Figure 8 shows a block diagram of an electromagnetic imaging device according to an embodiment of the present application.

[0136] As Figure 8 shown, in a second aspect, an embodiment of the present application provides an electromagnetic imaging device 80, including:

[0137] A construction module 81 for constructing an electromagnetic imaging model and an electrical impedance tomography model;

[0138] A generation module 82 for generating a source domain data set by using the electrical impedance tomography model and generating a target domain data set by using the electromagnetic imaging model, where the source domain data set includes first conductivity distribution data and first boundary voltage data obtained by a constant excitation current, and the target domain data set includes second conductivity distribution data and second boundary voltage data obtained by an excitation magnetic field;

[0139] A training module 83 for training a preset auxiliary model according to the source domain data set to obtain an improved auxiliary model;

[0140] An adjustment module 84 for performing adaptation layer training and full network fine-tuning on the improved auxiliary model by using the target domain data set to obtain an electromagnetic imaging optimization model;

[0141] An output module 85 for inputting the data to be tested in the target domain data set into the electromagnetic imaging optimization model to output a reconstructed image of the signal distribution of cranial lesions.

[0142] In one embodiment, optionally, the construction module 81 is used for:

[0143] The electromagnetic imaging model includes a plurality of excitation coils, and a TMR sensor corresponds to each excitation coil, where the TMR sensor is located at the center of each excitation coil.

[0144] The electrical impedance tomography model includes a plurality of electrodes and a test object with a preset shape. Among them, the plurality of electrodes are equidistantly placed on the surface of the test object with a preset shape, and the first boundary voltage data corresponding to the source domain data set is obtained by applying a constant excitation current and measuring the boundary voltage; and / or,

[0145] The method for obtaining the first boundary voltage data by applying a constant excitation current and measuring the boundary voltage includes: performing simulation of the forward imaging problem based on a finite element target object and adjacent excitation modes under a constant excitation current to determine the corresponding first boundary voltage data.

[0146] In one embodiment, optionally, the generation module 82 is used for:

[0147] Adjust the number, position, size and shape of the finite element target objects with a set conductivity in the target area, and set the background conductivity in the target area to obtain the first conductivity distribution data; and / or,

[0148] Add target substances with random shapes and sizes consistent with the blood conductivity in the target area, and configure the second conductivity distribution data corresponding to the target substances simulating intracranial hemorrhage points;

[0149] Use the electrical impedance tomography model, perform simulation of the forward imaging problem based on the finite element target object and adjacent excitation modes under a constant excitation current to determine the corresponding first boundary voltage data.

[0150] Use the electromagnetic imaging model to obtain the second boundary voltage data by acquiring the magnetic induction intensity at the boundary position by the TMR sensor under the excitation magnetic field generated by the excitation coil.

[0151] In one embodiment, optionally, the training module is used for:

[0152] Perform normalization processing on the first conductivity distribution data and the first boundary voltage data in the source domain data set respectively to obtain the processed first conductivity distribution data and the first boundary voltage data;

[0153] Use the processed first boundary voltage data as the input of the preset auxiliary model, and its corresponding processed first conductivity distribution data as the training label to perform iterative training on the preset auxiliary model to obtain the improved auxiliary model.

[0154] The improved auxiliary model is an improved neural network structure, including: an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first pooling layer, a second pooling layer, a fully connected layer, and an output layer; the input layer deforms the processed first boundary voltage data to obtain a first matrix; the first convolutional layer uses multiple convolutional kernels of a first set size to extract features from the first matrix and outputs a feature vector of a first dimension; the first pooling layer uses a maximum pooling window of a second set size and a set pooling stride to reduce the dimension and extract important features from the feature vector of the first dimension to obtain a feature vector of a second dimension; the second convolutional layer increases the number of output channels of the feature vector of the second dimension to obtain a feature vector of a third dimension; the third convolutional layer and the fourth convolutional layer extract features from the feature vector of the third dimension to obtain a target feature vector; the second pooling layer uses a maximum pooling window of a third set size and a set pooling stride to perform dimensionality reduction on the target feature vector to obtain a dimensionally reduced target feature vector; the fully connected layer integrates according to the dimensionally reduced target feature vector to obtain corresponding processed first conductivity distribution data; the output layer performs fully connected upsampling processing on the output processed first conductivity distribution data to obtain first conductivity distribution data consistent with a preset number of finite element divisions.

[0155] In this embodiment and other possible embodiments, the first pooling layer uses a 2×2 maximum pooling window and a pooling stride of 2 to reduce the dimension and extract important features from the feature vector of the first dimension to obtain a feature vector of a second dimension; the second convolutional layer increases the number of output channels of the feature vector of the second dimension to obtain a feature vector of a third dimension; the third convolutional layer and the fourth convolutional layer extract features from the feature vector of the third dimension to obtain a target feature vector; the second pooling layer uses a 2×2 maximum pooling window and a pooling stride of 2 to perform dimensionality reduction on the target feature vector to obtain a dimensionally reduced target feature vector.

[0156] In one embodiment, optionally, the improved auxiliary model is an improved neural network structure, including: an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first pooling layer, a second pooling layer, a fully connected layer, and an output layer;

[0157] The input layer deforms the processed first boundary voltage data to obtain a first matrix;

[0158] The first convolutional layer uses 32 4×4 convolutional kernels to extract features from the first matrix and outputs a feature vector of a first dimension;

[0159] The first pooling layer uses a 2×2 maximum pooling window with a pooling stride of 2 to reduce the dimension and extract important features of the feature vector in the first dimension, obtaining a feature vector in the second dimension;

[0160] The second convolutional layer increases the number of output channels of the feature vector in the second dimension, obtaining a feature vector in the third dimension;

[0161] The third convolutional layer and the fourth convolutional layer perform feature extraction on the feature vector in the third dimension to obtain a target feature vector;

[0162] The second pooling layer uses a 2×2 maximum pooling window with a pooling stride of 2 to perform dimensionality reduction on the target feature vector, obtaining a dimensionality-reduced target feature vector;

[0163] The fully connected layer integrates according to the dimensionality-reduced target feature vector to obtain corresponding processed first conductivity distribution data;

[0164] The output layer performs fully connected upsampling processing on the output of the processed first conductivity distribution data to obtain first conductivity distribution data that is consistent with the preset number of finite element divisions.

[0165] In one embodiment, optionally, the adjustment module is used for:

[0166] Freezing the model parameters of the improved auxiliary model and adding an adaptation layer to the improved auxiliary model;

[0167] Selecting a first quantity of target domain data from the target domain dataset to train the adaptation layer, obtaining a trained adaptation layer;

[0168] Selecting a second quantity of labeled target domain data other than the first quantity from the target domain dataset to perform full network fine-tuning on the trained auxiliary model and the adaptation layer, obtaining the electromagnetic imaging optimization model.

[0169] In a third aspect, an electromagnetic imaging system is provided, including:

[0170] A plurality of excitation coils 1 in a circular ring shape, a TMR sensor 2 placed in the hollow position inside the plurality of excitation coils 1 in the circular ring shape, and a target area formed on one side of the plurality of excitation coils 1;

[0171] A partitioning unit for performing finite element partitioning on the target area to obtain a preset number of finite element partitioned grids;

[0172] The first acquisition unit is configured to measure the boundary voltage of the test object by using a plurality of electrodes and a test object with a preset shape in the target area to obtain first boundary voltage data corresponding to the source domain dataset; before obtaining the first boundary voltage data, adjust the number, position, size and shape of the finite element target objects with a set conductivity in the grid to obtain first conductivity distribution data corresponding to the source domain dataset;

[0173] The second acquisition unit is configured to apply a constant excitation current to a plurality of excitation coils 1, and obtain second boundary voltage data by acquiring the magnetic induction intensity at the boundary position by the TMR sensor 2 under the excitation magnetic field generated by the excitation coils; before obtaining the first boundary voltage data, add target substances with a random shape and size and consistent with the blood conductivity in the target area, and determine second conductivity distribution data corresponding to the target substances simulating intracranial hemorrhage points in the target domain dataset;

[0174] The training unit is configured to train a preset auxiliary model according to the first conductivity distribution data and the first boundary voltage data in the source domain dataset to obtain an improved auxiliary model;

[0175] The adjustment unit is configured to perform adaptation layer training and full network fine-tuning on the improved auxiliary model by using the first quantity of second conductivity distribution data and second boundary voltage data in the target domain dataset to obtain an electromagnetic imaging optimization model;

[0176] The output unit is configured to input the data to be tested in the target domain dataset into the electromagnetic imaging optimization model to output a reconstructed image of the signal distribution of the cranial lesion; wherein, the data to be tested is the second quantity of second conductivity distribution data and second boundary voltage data except for the first quantity.

[0177] In a fourth aspect, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above electromagnetic imaging method are implemented.

[0178] In a fifth aspect, a computer-readable storage medium is provided. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above electromagnetic imaging method are implemented.

[0179] It should be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described electromagnetic imaging device and each module can refer to the corresponding processes in the foregoing embodiments of the electromagnetic imaging method, and will not be described herein again.

[0180] It should be noted that those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the above-described model training device and each module can refer to the corresponding processes in the foregoing embodiments of the electromagnetic imaging method, and will not be elaborated herein.

[0181] The above electromagnetic imaging device can be implemented in the form of a computer program, and this computer program can run on a computer device as shown in Figure 9 the figure.

[0182] Figure 9 The block diagram of a computer device according to an embodiment of the present application is shown.

[0183] Referring to Figure 9 , this computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the memory can include a storage medium and an internal memory.

[0184] The storage medium can store an operating system and a computer program. This computer program includes program instructions, and when the program instructions are executed, the processor can be made to execute any one of the electromagnetic imaging methods for multi-source data provided in the embodiments of the present application.

[0185] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0186] The internal memory provides an environment for the operation of the computer program in the storage medium. When this computer program is executed by the processor, the processor can be made to execute any one of the electromagnetic imaging methods for multi-source data. The storage medium can be non-volatile or volatile.

[0187] This network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 9 the structure shown in is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0188] It should be understood that the processor may be a Central Processing Unit (CPU), and the processor may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0189] In addition, an embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions for executing the steps of the method in the embodiment of the first aspect.

[0190] It should be noted that for the functions or steps that the above computer-readable storage medium or electronic device can achieve, reference may be made to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described in detail here.

[0191] It should be understood that the term "and / or" used herein is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " herein generally represents an "or" relationship between the associated objects before and after.

[0192] It should be understood that although the terms first, second, etc. may be used in the embodiments of the present application to describe the setting units, these setting units should not be limited to these terms. These terms are only used to distinguish the setting units from each other. For example, without departing from the scope of the embodiments of the present application, the first setting unit may also be referred to as the second setting unit, and similarly, the second setting unit may also be referred to as the first setting unit.

[0193] Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrase "if determined" or "if detected (stated condition or event)" may be interpreted as "when determined" or "in response to determining" or "when detected (stated condition or event)" or "in response to detecting (stated condition or event)".

[0194] In several embodiments provided by the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection to each other can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be in electrical, mechanical, or other forms.

[0195] In addition, in each embodiment of the present application, the functional units can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated units can be implemented in the form of hardware or in the form of a combination of hardware and software functional units.

[0196] Those of ordinary skill in the art can understand that all or part of the processes of implementing the above method embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the method embodiments as described above. Among them, any reference to the memory, storage, database, or other media used in the various embodiments provided by the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0197] The above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. An electromagnetic imaging method, characterized in that: The method comprises: Construct electrical impedance imaging models and electromagnetic imaging models; Generate a source domain data set using the electrical impedance imaging model, and generate a target domain data set using the electromagnetic imaging model; wherein the source domain data set includes first conductivity distribution data and first boundary voltage data obtained by a constant excitation current, and the target domain data set includes second conductivity distribution data and second boundary voltage data obtained by an excitation magnetic field; Training a preset auxiliary model according to the first conductivity distribution data and the first boundary voltage data in the source domain data set to obtain an improved auxiliary model; Using a first quantity of second conductivity distribution data and second boundary voltage data in the target domain data set to perform adaptation layer training and full network fine-tuning on the improved auxiliary model to obtain an electromagnetic imaging optimization model; The data to be tested in the target domain data set is input into the electromagnetic imaging optimization model to output a reconstructed image of the distribution of craniocerebral lesion signals; wherein the data to be tested is a second quantity of second conductivity distribution data and second boundary voltage data in addition to the first quantity.

2. The method according to claim 1, characterized in that The electromagnetic imaging model includes a plurality of excitation coils, and each excitation coil corresponds to a TMR sensor.

3. The method according to any one of claims 1 to 2, characterized in that The electrical impedance imaging model comprises a plurality of electrodes and a test object of a preset shape, wherein the plurality of electrodes are equidistantly placed on a surface of the test object of the preset shape, and first boundary voltage data corresponding to the source domain data set is obtained by applying a constant excitation current and measuring the boundary voltage; and / or, The method for obtaining first boundary voltage data by applying a constant excitation current and measuring the boundary voltage includes: simulating an imaging forward problem based on a finite element target object and adjacent excitation modes under a constant excitation current to determine corresponding first boundary voltage data.

4. The method according to any one of claims 1 to 3, characterized in that The method of generating first conductivity distribution data corresponding to the source domain data set by using the electrical impedance imaging model comprises: adjusting the number, position, size and shape of finite element target objects having set conductivity to obtain first conductivity distribution data; and / or, Generating a target domain dataset using the electromagnetic imaging model includes: Adding a target substance with a random shape and size consistent with the blood conductivity in the target area, and configuring second conductivity distribution data corresponding to the target substance to simulate an intracranial hemorrhage point; Under the excitation magnetic field generated by the excitation coil, the magnetic induction intensity at the boundary position is acquired by the TMR sensor to obtain the second boundary voltage data.

5. The method according to any one of claims 1 to 4, characterized in that The preset auxiliary model is trained according to the source domain data set to obtain an improved auxiliary model, including: Normalizing the first conductivity distribution data and the first boundary voltage data in the source domain data set to obtain processed first conductivity distribution data and first boundary voltage data; Using the processed first boundary voltage data as the input of the preset auxiliary model and the corresponding processed first conductivity distribution data as training labels, iteratively training the preset auxiliary model to obtain the improved auxiliary model; and / or, The improved auxiliary model is an improved neural network structure, including: an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first pooling layer, a second pooling layer, a fully connected layer and an output layer; The input layer deforms the first boundary voltage data after input processing to obtain a first matrix; The first convolution layer uses a plurality of convolution kernels of a first set size to perform feature extraction on the first matrix and output a feature vector of a first dimension; The first pooling layer uses a maximum pooling window of a second set size and a set pooling step size to reduce the dimension of the feature vector of the first dimension and extract important features to obtain a feature vector of the second dimension; The second convolutional layer increases the number of output channels of the feature vector of the second dimension to obtain a feature vector of the third dimension; The third convolution layer and the fourth convolution layer sequentially extract features from the feature vector of the third dimension to obtain a target feature vector; The second pooling layer uses a maximum pooling window of a third set size and a set pooling step size to perform dimensionality reduction processing on the target feature vector to obtain a target feature vector after dimensionality reduction; The fully connected layer integrates the target feature vector after dimension reduction to obtain the corresponding processed first conductivity distribution data; The output layer performs a fully connected upsampling process on the processed first conductivity distribution data to obtain first conductivity distribution data consistent with a preset number of finite element divisions.

6. The method according to any one of claims 1 to 5, characterized in that Using the target domain dataset to perform adaptation layer training and full network fine-tuning on the improved auxiliary model to obtain an electromagnetic imaging optimization model, including: Freezing model parameters of the improved auxiliary model and adding an adaptation layer to the improved auxiliary model; Selecting a first amount of target domain data from the target domain data set to train the adaptation layer to obtain a trained adaptation layer; A second quantity of labeled target domain data other than the first quantity is selected from the target domain data set, and the trained improved auxiliary model and adaptation layer are fine-tuned in the entire network to obtain the electromagnetic imaging optimization model.

7. An electromagnetic imaging device, characterized in that: include: A building module, used for building an electromagnetic imaging model and an electrical impedance imaging model; A generating module, used to generate a source domain data set using the electrical impedance imaging model, and to generate a target domain data set using the electromagnetic imaging model, wherein the source domain data set includes first conductivity distribution data and first boundary voltage data obtained by a constant excitation current, and the target domain data set includes second conductivity distribution data and second boundary voltage data obtained by an excitation magnetic field; A training module, used for training a preset auxiliary model according to the first conductivity distribution data and the first boundary voltage data in the source domain data set to obtain an improved auxiliary model; An adjustment module, configured to perform adaptation layer training and full network fine-tuning on the improved auxiliary model using a first quantity of second conductivity distribution data and second boundary voltage data in the target domain data set, so as to obtain an electromagnetic imaging optimization model; An output module is used to input the test data in the target domain data set into the electromagnetic imaging optimization model to output a reconstructed image of the distribution of craniocerebral lesion signals; wherein the test data is a second quantity of second conductivity distribution data and second boundary voltage data in addition to the first quantity.

8. An electromagnetic imaging system, characterized in that: include: A plurality of ring-shaped excitation coils (1), a TMR sensor (2) placed in a hollow position inside the plurality of ring-shaped excitation coils (1), and one side of the plurality of excitation coils (1) form a target area; A division unit, used for performing finite element division on the target area to obtain a preset number of finite element divided grids; A first acquisition unit, configured to acquire first boundary voltage data corresponding to a source domain data set by measuring a boundary voltage of the test object using a plurality of electrodes and a test object of a preset shape in the target area; Before acquiring the first boundary voltage data, in the grid, adjusting the number, position, size and shape of finite element target objects with set conductivity to obtain first conductivity distribution data corresponding to the source domain data set; The second acquisition unit is used to obtain second boundary voltage data by applying a constant excitation current to the multiple excitation coils (1) and obtaining the magnetic induction intensity at the boundary position by the TMR sensor (2) under the excitation magnetic field generated by the excitation coils; before obtaining the second boundary voltage data, adding a target substance with a random shape and size consistent with the blood conductivity in the target area, and determining the second conductivity distribution data corresponding to the target substance of the simulated intracranial hemorrhage point corresponding to the target domain data set; A training unit, configured to train a preset auxiliary model according to the first conductivity distribution data and the first boundary voltage data in the source domain data set to obtain an improved auxiliary model; An adjustment unit, configured to perform adaptation layer training and full network fine-tuning on the improved auxiliary model using a first quantity of second conductivity distribution data and second boundary voltage data in the target domain data set, so as to obtain an electromagnetic imaging optimization model; An output unit is used to input the test data in the target domain data set into the electromagnetic imaging optimization model to output a reconstructed image of the distribution of craniocerebral lesion signals; wherein the test data is a second quantity of second conductivity distribution data and second boundary voltage data in addition to the first quantity.

9. A computer device, characterized in that: include: at least one processor; and, a memory communicatively coupled to the at least one processor; The memory stores instructions executable by the at least one processor, wherein the instructions are configured to execute the method according to any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: Computer executable instructions are stored, and the computer executable instructions are used to execute the method according to any one of claims 1 to 6.

Images