An electrical impedance imaging method, apparatus, device, and medium

Abstract

The application discloses an electrical impedance imaging method, device, equipment and medium, and the electrical impedance imaging method comprises the following steps: cross-excitation is performed on a to-be-imaged object through a preset electrode array, current cross-excitation response data of the to-be-imaged object is acquired, and the electrode array comprises a plurality of electrodes arranged in a staggered manner; the to-be-imaged object is pre-reconstructed according to the current cross-excitation response data, and a current electrical impedance distribution map of the to-be-imaged object is obtained; feature extraction is performed on the current electrical impedance distribution map, current node adjacency relationships and current node features of the to-be-imaged object are acquired; and a target electrical impedance distribution map of the to-be-imaged object is generated based on the current node adjacency relationships and the current node features by using a preset image generation model. The method combines cross-excitation and feature extraction, effectively improves the reconstruction accuracy and spatial resolution of the electrical impedance distribution map, and can more accurately reflect real-time electrical characteristic changes in the to-be-imaged object.

Description

Technical Field

[0001] This invention relates to the field of electronic imaging technology, and in particular to a method, apparatus, device and medium for electrical impedance imaging. Background Technology

[0002] Electrical Impedance Tomography (EIT), as a non-invasive and radiation-free functional imaging technique, has shown great application potential in fields such as lung monitoring and brain functional imaging. Traditional three-dimensional electrical impedance tomography mainly uses multiple electrode strips to achieve spatial electric field coverage in order to acquire the boundary voltage signal of the subject, and reconstructs the internal conductivity distribution image by solving an inverse problem.

[0003] During dynamic monitoring, the subject's respiratory movements or changes in body position can cause real-time changes in the contact state between the electrode and the skin, leading to drastic fluctuations in the boundary voltage signal. Existing technologies struggle to effectively eliminate interference from changes in contact impedance, resulting in artifacts or distortions in the reconstructed images, severely impacting imaging accuracy and clinical diagnostic value. Summary of the Invention

[0004] This invention provides a method, apparatus, device, and medium for electrical impedance imaging to solve the problem that existing technologies cannot effectively eliminate interference caused by changes in contact impedance, resulting in artifacts or distortion in the reconstructed image.

[0005] In a first aspect, this application provides a method for electrical impedance imaging, comprising the steps of: performing cross-excitation on an object to be imaged through a preset electrode array to obtain current cross-excitation response data of the object to be imaged, wherein the electrode array includes multiple electrodes arranged in an alternating manner; pre-reconstructing the object to be imaged based on the current cross-excitation response data to obtain a current electrical impedance distribution map of the object to be imaged; extracting features from the current electrical impedance distribution map to obtain current node adjacency relationships and current node features of the object to be imaged; generating a target electrical impedance distribution map of the object to be imaged using a preset image generation model based on the current node adjacency relationships and current node features; wherein the image generation model is trained with sample node adjacency relationships and sample node features as input and a real electrical impedance distribution map as the desired output, wherein the real electrical impedance distribution map is obtained by simulation based on the real structural information of the sample object, and the sample node adjacency relationships and sample node features are obtained by feature extraction from the sample electrical impedance distribution map of the sample object; wherein the sample electrical impedance distribution map is obtained by pre-reconstructing the sample cross-excitation response data of the sample object, wherein the sample cross-excitation response data is obtained by performing cross-excitation on the sample object through the electrode array.

[0006] Secondly, this application provides a power impedance imaging device, comprising: a data acquisition module, configured to perform cross-excitation on an object to be imaged through a preset electrode array, and acquire current cross-excitation response data of the object to be imaged, wherein the electrode array includes a plurality of staggered electrodes; a pre-reconstruction module, configured to perform pre-reconstruction on the object to be imaged based on the current cross-excitation response data, and obtain a current power impedance distribution map of the object to be imaged; a feature extraction module, configured to perform feature extraction on the current power impedance distribution map, and obtain the current node adjacency relationship and current node features of the object to be imaged; and a target generation module, configured to generate a target based on the current node adjacency relationship and current node features. The system utilizes a preset image generation model to generate a target impedance distribution map of the object to be imaged. The image generation model is trained using sample node adjacency relationships and sample node features as input, and the true impedance distribution map as the desired output. The true impedance distribution map is obtained through simulation based on the actual structural information of the sample object. The sample node adjacency relationships and sample node features are extracted from the sample impedance distribution map of the sample object. The sample impedance distribution map is pre-reconstructed from the sample cross-excitation response data of the sample object, which is obtained by cross-exciting the sample object through the electrode array.

[0007] Thirdly, this application provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described electrical impedance imaging method.

[0008] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described electrical impedance imaging method.

[0009] In the above-mentioned technical solutions for electrical impedance imaging methods, devices, computer equipment, and storage media, the electrical impedance imaging method includes the following steps: cross-excitation is applied to the object to be imaged through a preset electrode array to obtain the current cross-excitation response data of the object to be imaged; the electrode array includes multiple staggered electrodes; pre-reconstruction is performed on the object to be imaged based on the current cross-excitation response data to obtain the current electrical impedance distribution map of the object to be imaged; feature extraction is performed on the current electrical impedance distribution map to obtain the current node adjacency relationship and current node features of the object to be imaged; based on the current node adjacency relationship and current node features, a preset image generation model is used to generate the target electrical impedance distribution map of the object to be imaged; the image generation model is trained with sample node adjacency relationship and sample node features as input and the real electrical impedance distribution map as the desired output; the real electrical impedance distribution map is obtained by simulation based on the real structural information of the sample object; the sample node adjacency relationship and sample node features are obtained by feature extraction from the sample electrical impedance distribution map of the sample object; the sample electrical impedance distribution map is pre-reconstructed from the sample cross-excitation response data of the sample object, which is obtained by cross-exciting the sample object through the electrode array. This method obtains richer boundary information through cross-excitation and extracts deep topological features by combining the pre-reconstruction results. It introduces an image generation model to deeply fit the nonlinear mapping relationship in the pre-reconstructed image, which can effectively utilize the topological structure and feature information between nodes, significantly improve the reconstruction accuracy and spatial resolution of the electrical impedance distribution map, and more accurately reflect the real-time changes in electrical properties inside the object to be imaged. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of a power impedance imaging method according to an embodiment of the present invention; Figure 2 This is an application scenario diagram of the electrical impedance imaging method in one embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electrode strip used in the electrical impedance imaging method in one embodiment of the present invention; Figure 4 This is a detailed flowchart of step S2 in the electrical impedance imaging method according to an embodiment of the present invention; Figure 5 This is a schematic diagram of step S22 in the electrical impedance imaging method according to an embodiment of the present invention; Figure 6 This is a specific flowchart of step S3 in the electrical impedance imaging method in one embodiment of the present invention; Figure 7 This is a schematic diagram of step S3 in the electrical impedance imaging method according to an embodiment of the present invention; Figure 8 This is a detailed flowchart of step S4 in the electrical impedance imaging method of one embodiment of the present invention; Figure 9 This is a schematic diagram of step S4 in the electrical impedance imaging method according to an embodiment of the present invention; Figure 10 This is a specific flowchart of step S41 in the electrical impedance imaging method of one embodiment of the present invention; Figure 11 This is a specific flowchart of step S42 in the electrical impedance imaging method of one embodiment of the present invention; Figure 12 This is a flowchart illustrating a specific process for model training in a resistive imaging method according to an embodiment of the present invention; Figure 13 This is a schematic diagram illustrating the principle of model training in a resistive imaging method according to an embodiment of the present invention; Figure 14 This is a schematic diagram of the discriminator model used in model training during electrical impedance imaging in one embodiment of the present invention; Figure 15 This is a schematic diagram of an electrical impedance imaging device according to an embodiment of the present invention; Figure 16 This is a schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] It is important to first clarify that during respiration, the shape of the thoracic cavity and the position and state of its internal tissues constantly change, leading to alterations in the electrical conductivity distribution within the thoracic cavity. At the end of inspiration, the lungs are fully expanded, and the electrical conductivity of the lung tissues is relatively high. Using 20% ​​of the highest electrical conductivity at the end of inspiration as a threshold, nodes are distinguished as pulmonary nodes and non-pulmonary nodes. During respiration, as the lungs expand and contract, the number and location of nodes exhibiting the electrical conductivity characteristics of pulmonary nodes change. Based on this, the following embodiments of the present invention, by dynamically tracking changes in pulmonary nodes, can more accurately reconstruct a real-time target impedance distribution image of the lungs, thereby achieving real-time monitoring of changes in the lung state during respiration.

[0014] In one embodiment, such as Figure 1 As shown, a method for electrical impedance imaging is provided, comprising the following steps: Step S1: Cross-excite the object to be imaged through a preset electrode array to obtain the current cross-excitation response data of the object to be imaged. The electrode array includes multiple electrodes arranged in an alternating pattern.

[0015] It should be noted that the electrode array is a pre-designed imaging acquisition hardware, consisting of multiple electrodes arranged in an alternating pattern, which can be fitted to the surface of the object to be imaged. Cross-excitation is a dedicated electric field excitation acquisition method for electrical impedance imaging. By alternately injecting excitation current into the electrodes and acquiring the response voltage, the electric field information of the entire body surface is acquired. The current cross-excitation response data is the boundary voltage data acquired by the electrode array, which is the original foundational data for solving the subsequent inverse electrical impedance problem.

[0016] like Figure 2 As shown, in this embodiment, the object to be imaged is the human lung region. The electrode array is attached to the surface of the thoracic cavity, and the boundary voltage data of the lung region is collected through cross-excitation.

[0017] like Figure 3 As shown, the electrode array is integrated on a single electrode strip 100, which is made of flexible and stretchable silicone material. This allows it to adapt to the chest contours of patients of different body types, ensuring close contact between the electrode and the skin and reducing the impact of contact impedance on measurement accuracy. Furthermore, the surface of the electrode strip 100 is coated with a conductive gel layer, further improving the conductivity between the electrode and the skin, effectively reducing contact impedance, and thus improving the accuracy and stability of boundary voltage data acquisition.

[0018] Furthermore, the electrode strip 100 is approximately 0.2 mm thick and at least 10 cm wide, with 16 copper electrodes arranged in a crisscross pattern along its length. External electrode cables can be connected to the imaging device for data transmission. The inner layer of the electrode strip 100 uses graphene electrode sheets to cover the electrode contacts. These sheets are approximately 1.5 cm wide and 2 cm long. The center distance between the upper and lower electrodes is at least 5 cm. The lower electrodes are numbered sequentially from left to right as 1, 3, 5, 7, 9, 11, 13, and 15, while the upper electrodes are numbered sequentially from left to right as 2, 4, 6, 8, 10, 12, 14, and 16. Odd-numbered electrodes are located in the lower layer, and even-numbered electrodes are located in the upper layer, forming a staggered grid-like layout. This staggered arrangement effectively increases the spatial sampling density between electrodes, reduces the measurement blind zone, and improves the spatial resolution and reconstruction accuracy of impedance imaging.

[0019] It should be noted that the length and width of the electrode band 100 vary depending on the chest circumference. If a suitable size electrode band 100 is unavailable, 16 ECG electrode patches can be used as a substitute, with the application order as described above. Figure 3 The electrode arrangement should be consistent.

[0020] In this embodiment, an electrode array containing multiple staggered electrodes is first attached and fixed to the thoracic cavity surface area corresponding to the lungs using electrode strips 100, ensuring stable contact between the electrodes and the skin. Then, the electrode array is controlled to perform cross-excitation operations according to a preset excitation sequence, with different electrodes alternately performing current injection and voltage acquisition to fully acquire the electric field response signal of the thoracic cavity surface at the current moment. After integration and filtering, the current cross-excitation response data of the object to be imaged is obtained.

[0021] Specifically, adjacent or spaced-apart electrode pairs are selected as excitation and measurement electrodes, and weak AC excitation currents are injected sequentially while the response voltages between the remaining electrode pairs are acquired simultaneously. By traversing all possible excitation-measurement combinations, a complete boundary voltage dataset is constructed, which is the current cross-excitation response data.

[0022] like Figure 2As shown, when the excitation electrode pair is electrode 1 and electrode 2, an excitation current is applied between the two electrodes, and the voltages between the remaining electrode pairs 3-4, 4-5, ..., 14-15, and 15-16 are collected sequentially, yielding a total of 13 voltage values. Subsequently, an excitation is applied between electrodes 2 and 3, and the voltages between 4-5, 5-6, ..., 15-16, and 16-1 are collected sequentially, also yielding 13 voltage values. This process continues until an excitation is applied between adjacent electrode pairs and the corresponding data collection is completed, resulting in a complete set of 16 × 13 voltage data points, which constitute the current cross-excitation response data. This current cross-excitation response data comprehensively reflects the modulation effect of the intrathoracic conductivity distribution on the body surface electric field, providing a high signal-to-noise ratio input basis for subsequent solutions to the resistance inverse problem.

[0023] In this application, the staggered electrode array is adapted to the curved structure of the human chest cavity, and the hardware structure is simple and easy to wear. At the same time, the cross-excitation acquisition method can maximize the acquisition of boundary electric field information with a limited number of electrodes, make up for the deficiency of insufficient information acquisition by a single electrode, and provide complete and stable raw data support for subsequent image pre-reconstruction.

[0024] Step S2: Based on the current cross-excitation response data, pre-reconstruct the object to be imaged to obtain the current impedance distribution map of the object to be imaged.

[0025] It should be noted that the pre-reconstruction is a coarse imaging process based on solving the inverse electric field problem. Relying on the prior knowledge of the anatomical structure of the object to be imaged and regularization constraints, the boundary voltage data is transformed into an image of the conductivity distribution inside the object. The current impedance distribution map is the conductivity distribution image of the object to be imaged obtained from the preliminary reconstruction. It contains complete spatial structural information, but contains a small number of artifacts and blurring defects, providing basic image material for subsequent fine-tuning optimization.

[0026] In this embodiment, after obtaining the current cross-excitation response data, it is necessary to combine the prior information of the anatomical structure of the human lung region to perform three-dimensional mesh subdivision of the human lung region, and construct a three-dimensional mesh model including key anatomical structures such as the lungs, heart, and chest wall. Then, combining the three-dimensional mesh model and the cross-excitation response data, the inverse electric field problem of the human lung region is solved to obtain the current electrical impedance distribution map of the human lung region.

[0027] Specifically, such as Figure 4 As shown, step S2 includes the following sub-steps: Step S21: Based on the structural information of the object to be imaged, perform stereo subdivision of the object to be imaged to obtain a three-dimensional mesh model including multiple nodes.

[0028] It should be noted that the structural information refers to the prior anatomical structure information of the object to be imaged, including the spatial boundaries and distribution characteristics of each tissue. The stereo meshing adopts the Delaunay tetrahedral meshing method, which discretizes the continuous human thoracic cavity space into several tetrahedral elements and mesh nodes, forming a discrete geometric model that can be used for numerical calculation.

[0029] like Figure 5 As shown, standardized anatomical information of the human thoracic cavity and lungs is retrieved, and adaptive scaling is performed in conjunction with the body shape of the object to be imaged. The entire thoracic cavity space is then divided into three dimensions using Delaunay, generating a three-dimensional mesh model 20 containing a large number of uniformly distributed nodes that conforms to the actual anatomical boundaries of the human body. This three-dimensional mesh model 20 discretizes the continuous thoracic cavity space into a finite number of elements, clearly defining the geometric boundaries and topological connections of tissues such as the lungs, heart, and chest wall, providing an accurate spatial discretization basis for the subsequent numerical solution of the inverse electric field problem.

[0030] In this embodiment, the three-dimensional mesh model based on prior human anatomy can ensure that the three-dimensional mesh model fits the real physiological structure, improve the spatial rationality and accuracy of subsequent inverse problem solving, and provide a standardized and high-precision numerical calculation basis for solving electrical impedance distribution.

[0031] Step S22: Based on the three-dimensional mesh model and the current cross-excitation response data, perform regularized reconstruction of the object to be imaged to obtain the current electrical impedance distribution map.

[0032] It should be noted that the regularized reconstruction uses the Tikhonov regularization algorithm to constrain the ill-conditioned inverse problem of electrical impedance imaging, suppress solution divergence and noise amplification during the solution process, and achieve stable initial image reconstruction.

[0033] like Figure 5 As shown, the three-dimensional mesh model 20 obtained by subdivision is used as the calculation basis. The steady-state electric field Poisson equation is coupled, and the current cross-excitation response data 21 collected is used as the electric field boundary constraint condition. The Tikhonov regularization constraint term is introduced through the pre-reconstructor 30 to solve the inverse problem of EIT (Electrical Impedance Tomography). The electrical conductivity value inside the thoracic cavity is calculated by calculating the value of the electrical impedance of the object to be imaged through each mesh node, and the current electrical impedance distribution map 40 of the object to be imaged is reconstructed.

[0034] In this embodiment, the pre-reconstruction process employs regularization constraints, which can quickly output a complete coarse reconstruction image of the initial conductivity, preserving the core electrical features of lung ventilation while providing effective input for subsequent feature extraction and fine reconstruction.

[0035] Step S3: Extract features from the current impedance distribution map to obtain the current node adjacency relationship and current node features of the object to be imaged.

[0036] It should be noted that node adjacency relationships refer to the spatial topological connections between nodes in a 3D mesh, representing the fixed anatomical geometry of the object to be imaged. Node features are the standardized electrical feature vectors corresponding to each node, carrying real-time conductivity changes of the object to be imaged, and are the core input data for the image generation model.

[0037] Combination Figure 7 As shown, by extracting features from the current electrical impedance distribution map 40, the current node adjacency relationship and current node features 50 of the human lung region are obtained, providing structured data support for subsequent refined reconstruction based on graph neural networks. The current node adjacency relationship is the edge set of the three-dimensional mesh model, and the current node features are the conductivity information at each node. Together, they constitute the complete input tensor required by the image generation model.

[0038] Specifically, such as Figure 6 As shown, step S3 includes the following sub-steps: Step S31: Perform node identification on the current impedance distribution map and obtain the current node adjacency relationship between each node.

[0039] It should be noted that the node adjacency relationship is constructed based on the topological structure of the 3D mesh model and is usually represented in the form of an adjacency matrix or an edge list, which is used to describe the spatial connection state of each mesh node.

[0040] In this embodiment, based on the three-dimensional tetrahedral mesh model corresponding to the current impedance distribution map, all mesh nodes and tetrahedral elements are traversed, and two nodes belonging to the same tetrahedral element are determined to be adjacent nodes. Based on this, a node adjacency matrix is ​​constructed to obtain the current node adjacency relationship.

[0041] In this embodiment, by accurately reconstructing the three-dimensional anatomical topology of the human thoracic cavity, the spatial geometric prior information is fully preserved, providing reliable topological support for subsequent graph convolution feature extraction and ensuring that the spatial feature extraction conforms to the human physiological structure.

[0042] Step S32: Perform feature transformation on the current impedance distribution map to obtain node feature vectors with multiple dimensions.

[0043] It should be noted that the feature transformation is a linear dimension mapping operation. By using a preset weight matrix, the original conductivity values ​​are subjected to dimensionality increase and feature mapping to uncover the potential correlation features of the original data.

[0044] like Figure 7As shown, based on the current electrical impedance distribution map 40, the conductivity information of each node is obtained. Through the corresponding weight matrix, a linear mapping is performed to map the conductivity values ​​of each node to different feature spaces, resulting in multi-dimensional node feature vectors. This expands the single conductivity scalar into a high-dimensional feature representation that includes local gradient, spatial curvature, and local conductivity change trends, thereby enhancing the model's ability to capture subtle electrical changes during lung ventilation.

[0045] Furthermore, each feature dimension has a specific weight matrix to control the contribution of different feature dimensions. These weight matrices can be dynamically adjusted based on the training data to adapt to physiological differences and measurement noise distribution among individuals, thereby improving the robustness and generalization performance of feature representation. A fully connected linear operation is performed on the original conductance and weight matrix of each node to output the node feature vector of the corresponding dimension. This operation process can be represented as: ; in, Indicates the first Node feature vectors in each feature dimension =1,2,3; Indicates the first Weight matrices along each feature dimension are used to perform linear transformations in the feature space. Indicates the first The original conductivity values ​​of each node.

[0046] In this embodiment, multi-dimensional feature linear mapping breaks through the limitations of the original single-dimensional conductivity data feature representation, enriches the node feature dimensions, mines the potential correlation information of node electrical characteristics, and improves the richness of subsequent network feature learning.

[0047] Step S33: Activate the feature vectors of each node to obtain the corresponding activated feature vectors.

[0048] It should be noted that the activation feature vector introduces nonlinear factors through a nonlinear activation function, thereby enhancing the model's ability to express and fit complex electrical features.

[0049] In this embodiment, the LeakyReLU activation function is used for nonlinear transformation. The feature vectors of each node obtained by the feature transformation are uniformly processed using the LeakyReLU function for nonlinear activation, preserving weak negative feature information, and fitting the nonlinear correlation between thoracic conductivity and spatial topology to obtain the corresponding activated feature vectors.

[0050] Furthermore, the feature vectors of each node undergo a nonlinear transformation using the LeakyReLU activation function to obtain the corresponding activated feature vectors, which can be expressed as: ; in, Indicates the first Activation feature vectors in each feature dimension This represents the LeakyReLU activation function. Indicates the first Node feature vectors in each feature dimension =1,2,3.

[0051] It's important to note that the LeakyReLU activation function is similar to the ReLU activation function, with the main difference being how it handles negative inputs. For the ReLU activation function, when the input value is less than 0, its output value is directly set to 0. However, the LeakyReLU activation function does not set the output value to 0 when the input value is less than 0. Instead, it allows the output value to remain a small negative value and assigns a tiny gradient to this negative value. This helps avoid the vanishing gradient problem, thereby improving the model's training performance and generalization ability.

[0052] In this embodiment, the LeakyReLU activation function effectively introduces nonlinear expression capabilities, adapts to the complex dynamic changes in conductivity during lung ventilation, avoids feature loss in low conductivity regions, prevents gradient vanishing during model training, and improves the integrity of feature expression.

[0053] Step S34: Normalize each activated feature vector to obtain the current node features.

[0054] It should be noted that the normalization process uses the softmax function to uniformly map the feature values ​​to the 0-1 range, thereby standardizing the feature value range and adapting it to the input requirements of the subsequent image generation network.

[0055] like Figure 7 As shown, attention weights between nodes are calculated for each feature dimension, and global normalization is performed using the softmax function. This ensures that the features of each node form a relative weight distribution in the probability space, thereby constructing an attention distribution matrix with clear physical meaning. This provides a probabilistic weight basis for subsequent feature aggregation, ensuring that the sum of the relative weights is 1. The operation for the corresponding dimension can be represented as follows: ; in, Indicates the first The relative weight distribution across each feature dimension Indicates the first Activation feature vectors in each feature dimension =1,2,3 This represents the softmax function. This indicates the operation of the natural exponent.

[0056] Furthermore, by combining the relative weight distributions across each feature dimension, the current node's feature can be obtained. This feature integrates local topological structure and global attention information, enabling it to more accurately reflect the conductivity distribution characteristics of lung tissue. This feature not only preserves subtle changes in the original signal but also enhances the representation ability of key regions through an attention mechanism, laying a solid data foundation for the subsequent generation of high-resolution electrical impedance distribution maps.

[0057] In this embodiment, the softmax normalization mechanism ensures the consistency of feature vector distribution in the probability space, eliminating the interference of dimensional differences on model convergence. This mechanism smooths feature boundaries, suppresses the interference of outliers on feature aggregation, thereby improving the model's sensitivity to subtle lung lesions, enhancing its ability to capture weak pathological signals, and providing high signal-to-noise ratio feature inputs for subsequent image reconstruction.

[0058] Step S4: Based on the current node's adjacency relationship and current node features, generate a target impedance distribution map of the object to be imaged using a preset image generation model.

[0059] It should be noted that the image generation model is an offline-trained graph autoencoder model, including an encoder and a decoder, specifically adapted to 3D topological data of the thoracic cavity. It is used for fine-tuning, noise reduction, and edge optimization of coarsely reconstructed images, outputting a high-precision target electrical impedance distribution map. The image generation model is trained using sample node adjacency relationships and sample node features as input, and the true electrical impedance distribution map as the desired output. The true electrical impedance distribution map is obtained through simulation based on the actual structural information of the sample object. The sample node adjacency relationships and sample node features are extracted from the sample electrical impedance distribution map of the sample object. The sample electrical impedance distribution map is pre-reconstructed from the sample cross-excitation response data of the sample object, which is obtained by cross-exciting the sample object through an electrode array.

[0060] like Figure 9 and Figure 13 As shown, the image generation model 34 includes an encoder 31 and a decoder 32. The encoder 31 is used to extract high-dimensional abstract representations from node features, namely low-rank spatial features. The decoder 32 is responsible for mapping the low-rank spatial features back to the image space to reconstruct the electrical impedance distribution.

[0061] Specifically, such as Figure 8 As shown, step S4 includes the following sub-steps: Step S41: Perform spectral convolution on the adjacency relationship and features of the current node using the encoder to obtain low-rank spatial features.

[0062] It should be noted that the encoder employs a two-layer spectral convolution feature extraction structure. Spectral convolution is a graph convolution operation adapted to irregular 3D meshes, capable of combining topological relationships and nodal electrical features to complete feature extraction, noise reduction, and dimensionality compression, resulting in concise low-rank core features. Low-rank spatial features are a compact abstract representation of the original high-dimensional nodal features, removing redundant spatial details and noise interference while preserving the core topological patterns that determine the distribution of lung conductivity.

[0063] In this embodiment, the encoder is structured as a two-layer GCN (Graph Convolutional Network) to achieve two-layer spectral convolution feature extraction. The first spectral convolution layer primarily captures fine-grained conductivity variations within the local neighborhood, extracting high-frequency detail information. The second spectral convolution layer focuses on integrating the global topology, further compressing the feature dimension and suppressing noise interference, thereby extracting low-rank spatial features that characterize the overall conductivity distribution trend of the lungs. The LeakyReLU activation function and a linear activation function are applied to the first and second spectral convolution layers, respectively, to introduce nonlinear transformation capabilities and enhance the model's fitting effect on complex conductivity distributions.

[0064] Specifically, such as Figure 10 As shown, step S41 includes the following sub-steps: Step S411: Perform first spectral convolution on the current node's adjacency relationship and current node features to obtain the first spectral features.

[0065] It should be noted that the first spectral feature refers to the intermediate feature representation obtained after the first layer of spectral convolution processing. Essentially, it is a preliminary mapping and enhancement of the original node features in the frequency domain, aiming to capture the local nonlinear changes in conductivity within the thoracic cavity topology. The first spectral feature not only preserves the key electrical information in the original signal but also enhances the expressive power of the features through nonlinear transformation, laying the foundation for subsequent global feature integration.

[0066] like Figure 9 As shown, based on the current node's adjacency relationship For topological constraints and current node features As input, the encoder 31 performs the first layer of spectral convolution operation, and then performs nonlinear activation using the LeakyReLU activation function to extract the nonlinear electrical features of the thoracic cavity space, thus obtaining the first spectral features. The corresponding process can be represented as follows: ; in, Indicates the first spectral feature, Indicates the adjacency relationship of the current node. Indicates the characteristics of the current node. This represents the first weight matrix of the first layer of spectral convolution. This represents the convolution operation function for the first layer of spectral convolution. This represents the LeakyReLU activation function for the first layer of spectral convolution.

[0067] In this embodiment, the complex spatial topology and nonlinear conductivity distribution of the human thoracic cavity are fully fitted by the first spectral convolution processing, so as to comprehensively extract the effective features of lung ventilation and filter out shallow noise and artifact information.

[0068] Step S412: Perform second spectral convolution on the first spectral features to obtain low-rank spatial features.

[0069] It should be noted that the second spectral convolution processing refers to further performing deep frequency domain filtering and dimensionality compression operations based on the first spectral features.

[0070] like Figure 9 As shown, encoder 31 performs a second-layer linear spectral convolution on the first spectral features, combined with a linear activation function, to achieve feature dimension compression and redundant information removal, thus refining the low-rank spatial features that carry the core imaging information. The corresponding process can be represented as follows: ; in, Represents the characteristics of low-rank spaces. Indicates the first spectral feature, Indicates the characteristics of the current node. This represents the second weight matrix of the second-layer spectral convolution. This represents the convolution operation function for the second-layer spectral convolution. This represents the linear activation function for the second-layer spectral convolution.

[0071] In this embodiment, by employing a linear activation function in the second spectral convolution process, the feature dimensions can be simplified, core effective information can be preserved, and imaging distortion caused by feature redundancy can be avoided, providing high-quality latent feature support for accurate reconstruction by the subsequent decoder.

[0072] It should be noted that GCN can use the spectral convolution function of each layer. An embedded graph structure is used, and the target nodes and edges (corresponding to node adjacency relationships) are trained. The encoding process of a certain layer of GCN can be represented as: ; in, and These represent the input and output of the convolutional layer, respectively. , This is the filter parameter matrix. Spectral convolution function. Represented as: ; in, express The corresponding diagonal matrix, diagonal elements , representing the The total number of connections to each node; , For identity matrix, and The number of rows and columns are the same; This is the activation function for the corresponding layer.

[0073] Step S42: Decode the low-rank spatial features using a decoder to obtain the target impedance distribution map.

[0074] It should be noted that the decoder aims to map the compressed low-rank spatial features back to the high-dimensional image space in order to restore the spatial resolution and detailed texture of the original image, thereby generating a high-fidelity target impedance distribution map.

[0075] In this embodiment, the decoder adopts a multi-layer transposed convolutional structure to progressively upsample low-rank spatial features and combines skip connections to fuse detailed information from different levels of the encoder in order to compensate for spatial loss during downsampling and improve the clarity and structural integrity of the reconstructed image.

[0076] Specifically, such as Figure 11 As shown, step S42 includes the following sub-steps: Step S421: Transpose the low-rank spatial features to obtain the transposed spatial features.

[0077] It should be noted that the transpose operation is a matrix dimension transformation operation used to match the dimension of the feature inner product operation, thereby enabling the calculation of feature associations between nodes. The transposed spatial feature is an intermediate representation of the original low-rank spatial feature after dimensional transformation, with its row and column dimensions interchanged with the original feature matrix, laying the mathematical foundation for subsequent calculation of the global association between nodes.

[0078] Combination Figure 9 The low-rank spatial features output by encoder 31 Perform matrix transpose operation to obtain the transpose space features with dimension matching. .

[0079] Step S422: Perform inner product processing on the transposed spatial features and the low-rank spatial features to obtain the node similarity matrix.

[0080] It should be noted that the inner product operation is used to calculate the feature similarity between any two grid nodes, characterizing the spatial correlation strength and electrical distribution consistency between nodes. The node similarity matrix is ​​a mathematical expression used to quantify the interdependence between nodes. Its element values ​​reflect the degree of similarity in electrical properties among different grid nodes, providing crucial structural prior information for subsequent probability mapping and image reconstruction.

[0081] Combination Figure 9 Transpose spatial features Features of low-rank spaces Perform matrix inner product operations, traverse all node combinations, and generate a node similarity matrix that can characterize the spatial association and electrical similarity of nodes.

[0082] In this embodiment, by accurately reconstructing the spatial topological association and electrical distribution association of each node in the thoracic cavity, the true anatomical structure of the human body and the lung ventilation distribution characteristics are restored.

[0083] Step S423: Perform probability mapping on the node similarity matrix to obtain the target electrical impedance distribution map.

[0084] It should be noted that the probability mapping uses the sigmoid activation function to map the similarity values ​​to probability values ​​in the 0-1 range, thereby achieving accurate reconstruction of conductivity distribution and topology.

[0085] Combination Figure 9 via the sigmoid function A global probability mapping is performed on the node similarity matrix, and the probability of mapping the node similarity matrix to [0,1] represents the confidence that there is a spatial adjacency between two nodes. This completes the accurate reconstruction of the topology and conductivity values, and outputs a target impedance distribution map with clear edges, artifact elimination, and accurate values. The corresponding operation process can be represented as follows: ; in, This represents the joint probability distribution corresponding to each node in the target impedance distribution diagram. Indicates the first The node feature vector of each node. Indicates the first The node feature vector of each node. No. The node and the first The node connection relationships of each node. Represents the node similarity matrix. The node and the first The probability of connections between nodes, and , Indicates the first The transpose of the node feature vectors of each node. This represents the sigmoid function.

[0086] In this embodiment, the above-mentioned decoding process of the decoder can repair the edge blurring, artifact distortion and lack of detail defects of the pre-reconstructed image, greatly improve the spatial resolution and numerical accuracy of three-dimensional electrical impedance imaging of the lungs, and accurately reflect the dynamic ventilation changes of the lungs.

[0087] Furthermore, before step S1, model training is required to obtain a preset image generation model.

[0088] like Figure 13 and Figure 14 As shown, firstly, an adversarial generative network needs to be constructed, which includes a generator model 34 and a discriminator model 35. The generator model 34 uses the Graph Auto-Encoder (GAE) framework in graph neural networks. GAE is a special autoencoder technique that combines graph embedding and autoencoder ideas, including... Figure 9 The encoder 31, decoder 32, and latent spatial mapping module connecting them are used to generate sample electrical impedance distribution maps with high resolution and low artifacts for sample objects. The discriminator model 35 is based on a standard multi-layer perceptron (MLP), including an input layer 51, a hidden layer 52, and an output layer 53. It is used to extract high-dimensional features from the real electrical impedance distribution map and the generated sample electrical impedance distribution map, and output the discrimination result. Based on the discrimination result, the adversarial loss is calculated, and then the model parameters of the generator model 34 are optimized by backpropagation, so that the sample electrical impedance distribution map generated by the generator model 34 approximates the real data distribution in statistical distribution.

[0089] Furthermore, based on the adversarial generative network, the discriminator model 35 and the generator model 34 are adversarially trained to obtain an image generation model.

[0090] In this embodiment, the core of adversarial training lies in providing feedback on the authenticity of the generated results through the discriminator model 35, thereby guiding the optimization direction of the entire adversarial generative network and continuously improving the structural fidelity and physical consistency of the generated images until the generator model 34 can output a high-quality electrical impedance distribution map that is difficult to distinguish from real data in terms of visual features and numerical distribution, thus obtaining the desired image generation model.

[0091] It should be noted that before training the generator model, a positive sample dataset needs to be constructed for training the discriminator model. Initial training of the discriminator model involves obtaining the true electrical impedance distribution map and its true node adjacency relationships corresponding to the sample objects. The true node adjacency relationships refer to the topological information extracted from the true electrical impedance distribution map of the sample objects, reflecting the actual spatial connectivity of the internal medium of the sample objects. These true node adjacency relationships and the generated sample electrical impedance distribution map constitute the inputs to different training stages of the discriminator model, used to evaluate the physical rationality of the generated sample electrical impedance distribution map under topological constraints. The true electrical impedance distribution map is a distortion-free standard ground truth image obtained by performing a forward problem simulation on the sample objects, serving as the training positive samples.

[0092] like Figure 13 As shown, the actual electrical impedance distribution obtained through simulation is... The input is processed by the second processor 42 for feature extraction. The feature extraction process is the same as that in the first processor 41, and the true electrical impedance distribution map is obtained. Real node adjacency relationships and real node features This will then establish the actual node adjacency relationships. Input discriminator model 35 to train it so that it can output a discrimination result of the first probability value 1, which corresponds to the true electrical impedance distribution. It has a true label, indicating that it is a real image. The discriminator model 35 learns the true electrical impedance distribution by minimizing the classification error. The complex topological patterns and physical laws contained in the data can be accurately identified to determine the topological structure and physical feature distribution of real data, thereby providing effective gradient feedback for the generator model 34 in adversarial games.

[0093] Furthermore, the update gradient of discriminator model 35 can be expressed as: ; in, Represents the true electrical impedance distribution map The first in A real node edge, Represents the sample electrical impedance distribution map The first in Generate node edges, This represents the discrimination probability of discriminator model 35 against the edges of the input samples, used to distinguish between real and generated images. A true impedance distribution diagram / Sample electrical impedance distribution map The total number of edges in the middle node.

[0094] After the discriminator model 35 has completed its initial training, it is jointly trained with the generator model 34 to construct an adversarial generative network, entering an alternating iterative training phase. In this phase, the generator model 34 strives to generate more realistic electrical impedance distribution maps to deceive the discriminator model 35, while the discriminator model 35 continuously improves its ability to distinguish between real and fake samples. The two work together to optimize in the game, ultimately enabling the generator model 34 to reconstruct high-fidelity electrical impedance images from the topological structure.

[0095] Specifically, such as Figure 12 As shown, the process of training an adversarial generative network includes the following steps: Step S01: Generate the sample impedance distribution map using the generator model.

[0096] It should be noted that the sample impedance distribution map is a fake image reconstructed by the generator model based on the sample topology and sample features, used as a negative sample input for adversarial training.

[0097] like Figure 13 As shown, the sample impedance distribution diagram The generation process follows the same procedure as described in steps S1-S4 above, which involves the adjacency relationship of sample nodes of multiple sample objects. Features of sample nodes The initialized generator model 34 is input, and the encoder 31 extracts high-dimensional features and latent space mapping. The decoder 31 then reconstructs the model, outputting sample impedance distribution maps of multiple sample objects. The only difference is that the object to be imaged is replaced with a sample object. The corresponding image generation model is now the initialized generator model 34. The rest of the processing logic remains unchanged; the difference lies in the input data source, which utilizes the adjacency relationships of multiple sample object nodes. Features of sample nodes As input, rather than real-time measurement data of the object to be imaged.

[0098] Among them, the sample impedance distribution map required for training can be generated in batches using generator model 34. This provides sufficient data support for subsequent training of adversarial games using real and fake samples.

[0099] Step S02: Input the sample impedance distribution map or the actual node adjacency relationship into the input layer and perform one-dimensional flattening to obtain a one-dimensional feature vector.

[0100] It should be noted that the one-dimensional flattening process is a feature format preprocessing operation that converts the two-dimensional feature matrix into a one-dimensional feature vector to adapt to the fully connected input format of the discriminator model.

[0101] like Figure 14 As shown, the sample impedance distribution diagram output by generator model 34 is presented. Or actual node adjacency relationships The input layer 51 of the input discriminator model 35 is subjected to one-dimensional flattening and normalization to obtain a one-dimensional feature vector with uniform dimension.

[0102] Among them, by analyzing the sample impedance distribution map Or actual node adjacency relationships Flattening the data into a one-dimensional feature vector can unify the input data format of the discriminator model 35, eliminate the dimension mismatch problem, and ensure the stable and effective feature extraction of the subsequent hidden layer.

[0103] Step S03: Input the one-dimensional feature vector into the hidden layer for hidden layer processing to obtain hidden discriminative features.

[0104] like Figure 14 As shown, the hidden layer 52 consists of three layers with neuron dimensions of 32, 16, and 16 respectively. Each layer uses the tanh activation function to extract subtle differences in hidden discriminative features between real and fake samples.

[0105] In this embodiment, the normalized one-dimensional feature vector is sequentially input into three hidden layers 52, performing linear transformation and tanh nonlinear activation layer by layer to progressively uncover subtle distribution differences in features between real and fake images, ultimately extracting highly discriminative hidden features. The one-dimensional feature vector output from the input layer 51 is denoted by R0. The output of the first hidden layer 521 is then f1(W1R0+t1), where W1 is the first hidden weight and t1 is the first hidden bias term. The outputs of the second and third hidden layers 522 follow the same pattern. Here, f(·) is the tanh function.

[0106] In this application, the discriminator model accurately captures subtle artifacts, numerical and contour differences in reconstructed or ground truth images through multi-layer nonlinear feature mapping, significantly improving the discriminator model's resolution capability.

[0107] Step S04: Input the hidden discriminative features into the output layer and process the output to obtain the discrimination result.

[0108] It should be noted that the output layer is configured with a softmax activation function, which is used to map the hidden features to classification probabilities in the (0,1) interval, and output the discrimination result of whether the corresponding image is a real sample or a fake sample.

[0109] like Figure 14As shown, the hidden discriminative features are input into the output layer 53 of the discriminator model 35. Through linear transformation combined with softmax probabilistic regression, the output layer yields a discrimination result with a first probability value close to 1 (corresponding to the current input sample being the true electrical impedance distribution map) or a second probability value close to 0 (corresponding to the current input sample being the reconstructed sample electrical impedance distribution map). The process from hidden layer 52 to output layer 53 can be viewed as a multi-class logistic regression, i.e., softmax regression. Therefore, the output of output layer 53 is softmax(W4R3+t4), where R3 represents the output of hidden layer 52, f3(W3R2+t2). Here, W4 is the output layer weight, t4 is the output layer bias term, and R3 represents the output of the third hidden layer 523, where W3 is the third hidden weight and t3 is the third hidden bias term.

[0110] In this embodiment, the output layer outputs either a first probability value that the corresponding sample is a true electrical impedance distribution map or a second probability value that the sample electrical impedance distribution map is a true image. The closer the second probability value is to 1, the higher the probability that the sample electrical impedance distribution map generated by the generator model is a true image; conversely, the closer the second probability value is to 0, the higher the probability that the sample electrical impedance distribution map generated by the generator model is a fake image. This second probability value not only intuitively reflects the discriminator model's confidence in the authenticity of the sample, but also provides a quantitative basis for the subsequent calculation of the loss function and the backpropagation of network parameters.

[0111] Step S05: Update the parameters of the generator model or discriminator model based on the discrimination result to obtain the image generation model.

[0112] It should be noted that the model parameter update is a backpropagation iterative optimization process. Relying on the loss function constraint, the model parameters of the discriminator model and the generator model are updated alternately to achieve adversarial game optimization.

[0113] In this embodiment, a fixed number of iterations is used, and an alternating freeze optimization strategy is employed. The loss value is calculated based on the discrimination results, and the model parameters of the discriminator model 35 and the generator model 34 are updated sequentially through backpropagation. This iterative training is repeated until the loss converges, ultimately yielding the trained image generation model. Specifically, this can be represented as follows: ; in, and These represent the generator model and the discriminator model, respectively. This represents the first probability value by which the discriminator model determines that the true electrical impedance distribution map of the input is the true image. This represents the second probability value by which the discriminator model determines that the impedance distribution map of the sample generated by the generator model is a true image. The discriminator model tends to make the first probability value... Approaching 1 means identifying real samples as true as possible; at the same time, it tends to make the second probability value... The probability should be close to 0, meaning the generated sample should be identified as fake as accurately as possible. The generator model, on the other hand, tends to make the second probability value... The goal is to generate a realistic sample impedance distribution map that is as close to 1 as possible, which can deceive the discriminator model.

[0114] In this application, the imaging capability of the generator model is continuously optimized through adversarial game theory, so that the output image of the generator model can infinitely approach the real simulation image, thus completely making up for the inherent defects of single-electrode band imaging.

[0115] Furthermore, model training also includes constructing a loss function, which comprises adversarial loss and regression loss. These two types of losses are complementary constraints and work together to optimize the model. The adversarial loss is the sum of the cross-entropy loss between the true electrical impedance distribution map and the true label described in the discrimination result, and the cross-entropy loss between the sample electrical impedance distribution map and the false label described in the discrimination result; the regression loss is the node distribution difference loss between the sample electrical impedance distribution map and the true electrical impedance distribution map. The loss function can be expressed as: ; in, and These represent the generator model and the discriminator model, respectively. Indicating resistance to loss, Indicates regression loss, and , This represents the conductivity value of the corresponding node in the sample impedance distribution diagram. This represents the conductivity value of the corresponding node in the actual electrical impedance distribution diagram, where N is the total number of nodes. This represents the squared operation of the L2 norm. The regression loss ranges from [0, +∞), and is equal to 0 when the predicted result perfectly matches the true conductivity distribution. The larger the error, the larger this value becomes. The optimization objective is to minimize the regression loss. .

[0116] In this embodiment, the adversarial loss is calculated by separately calculating the cross-entropy between the real electrical impedance distribution map and the real label output by the discriminator model, and the cross-entropy between the electrical impedance distribution map of the fake sample and the fake label. The sum of these two is used as the final adversarial loss term to ensure that the discriminator has sufficient gradient signal when distinguishing between real and fake samples, thereby effectively driving the generator to learn the distribution pattern of real data. The regression loss focuses on the pixel-level or node-level numerical reconstruction accuracy. By calculating the mean square error (MSE) of the conductivity distribution between the generated image and the real image, it quantifies the deviation of the generated image from the real value in physical properties, ensuring that the reconstruction result is not only visually realistic, but also strictly consistent in physical magnitude, thereby guaranteeing the quantitative reliability of the reconstructed image in clinical diagnosis or industrial testing. The two types of losses are superimposed to form a loss function, which is used for iterative updating of generator parameters.

[0117] In other embodiments, the weight coefficients of the loss function can be dynamically adjusted according to the actual application scenario. For example, in scenarios that emphasize image visual quality, the weight of adversarial loss can be appropriately increased to enhance the texture details and structural integrity of the generated image; while in scenarios that emphasize quantitative analysis accuracy, the weight of regression loss should be increased to strengthen the ability to restore physical quantities and ensure that the model's prediction accuracy on key indicators meets stringent engineering or medical standards. Regression loss can also employ other robust loss function forms, such as structural similarity (SSIM) loss or perceptual loss, to further constrain the consistency of the generated image with the true distribution in local texture and global structure, thereby improving the model's generalization ability and reconstruction stability in complex noisy environments.

[0118] In this application, adversarial loss constrains the overall contour and feature distribution of the image, eliminating large-area artifacts and distortions; regression loss constrains the node-by-node electrical numerical accuracy, avoiding numerical offset. The dual constraints significantly improve the clarity and accuracy of the model reconstruction.

[0119] Furthermore, before formal iterative training, network parameter initialization and training hyperparameter configuration were first completed: based on network training experience, the base learning rate of the generator model and the learning rate of the discriminator model were uniformly initialized to 1e-3. After training started, the learning rate dynamically decreased in each round with a decay rate of 0.99, achieving adaptive fine-tuning of the learning rate. The Adam optimizer, which has adaptive gradient adjustment characteristics, was selected to complete gradient descent and network parameter update calculations. During training, the Adam optimizer accurately completed iterative optimization of model parameters by calculating the gradient update step size of each network parameter in real time. Simultaneously, based on the pre-training test results, the total number of iterative training iterations was set to 200.

[0120] like Figure 13As shown, the specific training steps are as follows: After 200 rounds of complete iterative training, the impedance data samples generated by the generator model 34 can infinitely approximate the real conductivity distribution samples, the discriminator model 35 cannot effectively distinguish the authenticity of the input samples, and the model adversarial game reaches the Nash equilibrium state, marking the completion of the overall algorithm training.

[0121] In summary, this application discloses a power impedance imaging method, comprising the following steps: performing cross-excitation on the object to be imaged through a preset electrode array to obtain the current cross-excitation response data of the object to be imaged, wherein the electrode array includes multiple staggered electrodes; pre-reconstructing the object to be imaged based on the current cross-excitation response data to obtain the current power impedance distribution map of the object to be imaged; extracting features from the current power impedance distribution map to obtain the current node adjacency relationship and current node features of the object to be imaged; generating the target power impedance distribution map of the object to be imaged using a preset image generation model based on the current node adjacency relationship and current node features; the image generation model is trained with sample node adjacency relationship and sample node features as input and the real power impedance distribution map as the desired output, wherein the real power impedance distribution map is obtained by simulation based on the real structural information of the sample object, and the sample node adjacency relationship and sample node features are obtained by feature extraction from the sample power impedance distribution map of the sample object; the sample power impedance distribution map is pre-reconstructed from the sample cross-excitation response data of the sample object, wherein the sample cross-excitation response data is obtained by cross-exciting the sample object through the electrode array. This method obtains richer boundary information through cross-excitation and extracts deep topological features by combining the pre-reconstruction results. It introduces an image generation model to deeply fit the nonlinear mapping relationship in the pre-reconstructed image, which can effectively utilize the topological structure and feature information between nodes, significantly improve the reconstruction accuracy and spatial resolution of the electrical impedance distribution map, and more accurately reflect the real-time changes in electrical properties inside the object to be imaged.

[0122] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0123] In one embodiment, an electrical impedance imaging device is provided, which corresponds one-to-one with the electrical impedance imaging method described in the above embodiments. For example... Figure 15 As shown, the electrical impedance tomography (EIT) device includes a data acquisition module 101, a pre-reconstruction module 102, a feature extraction module 103, and a target generation module 104. Detailed descriptions of each functional module are as follows: The data acquisition module 101 is used to perform cross-excitation on the object to be imaged through a preset electrode array and acquire the current cross-excitation response data of the object to be imaged. The electrode array includes multiple electrodes arranged in an alternating manner.

[0124] The pre-reconstruction module 102 is used to pre-reconstruct the object to be imaged based on the current cross-excitation response data to obtain the current electrical impedance distribution map of the object to be imaged.

[0125] The feature extraction module 103 is used to extract features from the current impedance distribution map and obtain the current node adjacency relationship and current node features of the object to be imaged.

[0126] The target generation module 104 is used to generate a target impedance distribution map of the object to be imaged based on the current node adjacency relationship and current node features using a preset image generation model. The image generation model is trained with the sample node adjacency relationship and sample node features as input and the real impedance distribution map as the expected output. The real impedance distribution map is obtained by simulation based on the real structural information of the sample object. The sample node adjacency relationship and sample node features are obtained by feature extraction from the sample impedance distribution map of the sample object. The sample impedance distribution map is pre-reconstructed from the sample cross-excitation response data of the sample object. The sample cross-excitation response data is obtained by cross-exciting the sample object through an electrode array.

[0127] Specific limitations regarding the electrical impedance imaging device can be found in the limitations of the electrical impedance imaging method described above, and will not be repeated here. Each module in the aforementioned electrical impedance imaging device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the operations corresponding to each module.

[0128] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 16 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a power impedance imaging method.

[0129] In one embodiment, 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, it implements the electrical impedance tomography method described in the above embodiments, for example... Figure 1S1-S4, as shown, will not be described again here to avoid repetition. Alternatively, the processor executes the computer program to implement the functions of each module / unit in this embodiment of the electrical impedance imaging device, for example... Figure 15 The functions of the data acquisition module 101, pre-reconstruction module 102, feature extraction module 103, and target generation module 104 shown are not described again here to avoid repetition.

[0130] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program implements the electrical impedance imaging method described above, for example... Figure 1 S1-S4, as shown, will not be described again here to avoid repetition. Alternatively, when this computer program is executed by a processor, it implements the functions of each module / unit in this embodiment of the electrical impedance imaging device, for example... Figure 15 The functions of the data acquisition module 101, pre-reconstruction module 102, feature extraction module 103, and target generation module 104 shown are not described again here to avoid repetition. The computer-readable storage medium can be non-volatile or volatile.

[0131] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0132] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0133] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0134] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions 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 within the protection scope of the present invention.

Claims

1. A method for electrical impedance imaging, characterized in that, Including the following steps: Cross-excitation is applied to the object to be imaged using a preset electrode array to obtain the current cross-excitation response data of the object to be imaged. The electrode array includes multiple electrodes arranged in an alternating pattern. Based on the current cross-excitation response data, the object to be imaged is pre-reconstructed to obtain the current electrical impedance distribution map of the object to be imaged. Feature extraction is performed on the current electrical impedance distribution map to obtain the current node adjacency relationship and current node features of the object to be imaged; Based on the current node's adjacency relationship and current node features, a target impedance distribution map of the object to be imaged is generated using a preset image generation model; The image generation model is trained by taking the adjacency relationship and features of sample nodes as input and the real electrical impedance distribution map as the desired output. The real electrical impedance distribution map is obtained by simulation based on the real structural information of the sample object. The adjacency relationship and features of sample nodes are obtained by feature extraction from the sample electrical impedance distribution map of the sample object. The sample impedance distribution map is obtained by pre-reconstructing the sample cross-excitation response data of the sample object, which is obtained by cross-exciting the sample object through the electrode array.

2. The electrical impedance imaging method according to claim 1, characterized in that, The step of pre-reconstructing the object to be imaged based on the current cross-excitation response data to obtain the current electrical impedance distribution map of the object to be imaged includes: Based on the structural information of the object to be imaged, the object to be imaged is stereotactically segmented to obtain a three-dimensional mesh model including multiple nodes. Based on the three-dimensional mesh model and the current cross-excitation response data, the object to be imaged is reconstructed using regularization to obtain the current electrical impedance distribution map.

3. The electrical impedance imaging method according to claim 2, characterized in that, The step of extracting features from the current impedance distribution map to obtain the current node adjacency relationship and current node features of the object to be imaged includes: Node identification is performed on the current impedance distribution map to obtain the current node adjacency relationship between each node; The current electrical impedance distribution map is subjected to feature transformation to obtain node feature vectors with multiple dimensions; The feature vectors of each node are activated to obtain the corresponding activated feature vectors. The activation feature vectors are normalized to obtain the current node features.

4. The electrical impedance imaging method according to claim 1, characterized in that, The image generation model includes an encoder and a decoder; the step of generating a target impedance distribution map of the object to be imaged using a preset image generation model based on the current node's adjacency relationship and current node features includes: The encoder performs spectral convolution on the current node's adjacency relationship and current node features to obtain low-rank spatial features. The low-rank spatial features are decoded by the decoder to obtain the target electrical impedance distribution map.

5. The electrical impedance imaging method according to claim 4, characterized in that, The step of performing spectral convolution processing on the current node's adjacency relationship and current node features through the encoder to obtain low-rank spatial features includes: The adjacency relationship and features of the current node are subjected to a first spectral convolution to obtain the first spectral features; The first spectral feature is subjected to a second spectral convolution to obtain the low-rank spatial feature.

6. The electrical impedance imaging method according to claim 4, characterized in that, The step of decoding the low-rank spatial features using the decoder to obtain the target electrical impedance distribution map includes: The low-rank spatial features are transposed to obtain transposed features; The transpose feature and the low-rank spatial feature are processed by inner product to obtain the node similarity matrix; By performing probability mapping on the node similarity matrix, the target electrical impedance distribution map is obtained.

7. The electrical impedance imaging method according to claim 1, characterized in that, The electrical impedance imaging method further includes: Construct an adversarial generative network, which includes a generator model and a discriminator model; The discriminator model and the generator model are subjected to adversarial training to obtain the image generation model.

8. The electrical impedance imaging method according to claim 7, characterized in that, The discriminator model includes an input layer, a hidden layer, and an output layer; the adversarial training of the discriminator model and the generator model to obtain the image generation model includes: The sample electrical impedance distribution map is generated using the generator model; The sample impedance distribution map or the real node adjacency relationship of the real impedance distribution map is input into the input layer for one-dimensional flattening to obtain a one-dimensional feature vector. The one-dimensional feature vector is input into the hidden layer for hidden layer processing to obtain hidden discriminative features; The hidden discriminative features are input into the output layer for output processing to obtain the discriminative result; The parameters of the generator model or the discriminator model are updated based on the discrimination result to obtain the image generation model.

9. The electrical impedance imaging method according to claim 8, characterized in that, The step of performing adversarial training on the generator model using the discriminator model to obtain the image generation model further includes: A loss function is constructed, which includes adversarial loss and regression loss; the adversarial loss is the sum of the cross-entropy loss between the true electrical impedance distribution map and the true label described by the discrimination result, and the cross-entropy loss between the sample electrical impedance distribution map and the false label described by the discrimination result; the regression loss is the distribution difference between the sample electrical impedance distribution map and the true electrical impedance distribution map.

10. A power impedance imaging device, characterized in that, include: The data acquisition module is used to perform cross-excitation on the object to be imaged through a preset electrode array, and acquire the current cross-excitation response data of the object to be imaged. The electrode array includes multiple electrodes arranged in an alternating pattern. The pre-reconstruction module is used to pre-reconstruct the object to be imaged based on the current cross-excitation response data to obtain the current electrical impedance distribution map of the object to be imaged. The feature extraction module is used to extract features from the current electrical impedance distribution map to obtain the current node adjacency relationship and current node features of the object to be imaged; The target generation module is used to generate a target impedance distribution map of the object to be imaged based on the current node's adjacency relationship and current node features, using a preset image generation model. The image generation model is trained by taking the adjacency relationship and features of sample nodes as input and the real electrical impedance distribution map as the desired output. The real electrical impedance distribution map is obtained by simulation based on the real structural information of the sample object. The adjacency relationship and features of sample nodes are obtained by feature extraction from the sample electrical impedance distribution map of the sample object. The sample impedance distribution map is obtained by pre-reconstructing the sample cross-excitation response data of the sample object, which is obtained by cross-exciting the sample object through the electrode array.

11. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the electrical impedance imaging method as described in any one of claims 1 to 9.

12. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the electrical impedance imaging method as described in any one of claims 1 to 9.

Images