Lung EIT image reconstruction method based on graph structure

Abstract

The invention relates to a lung region electrical impedance tomography (EIT) image reconstruction method based on a graph structure, and establishes an EIT depth imaging model in which double branch channels of a high-frequency channel graph attention network and a low-frequency channel graph attention network are parallel, and the model is named as HLGAT. The method comprises the following steps: injecting current excitation into a finite element simulation model of a lung region to obtain boundary voltage data required by EIT image reconstruction, and performing non-uniform mesh generation on the lung region needing image reconstruction to obtain nodes required by a graph neural network and an adjacency relationship between the nodes, so as to obtain a lung region model; storing the data in a graph structure form suitable for network input; the HLGAT aggregates the low-frequency characteristic information and the high-frequency characteristic information through the two channels, node output characteristic information of the two channels is fused in a learnable mode, a new adjacency matrix is reconstructed according to the node characteristic information output by the network, then a new Laplace matrix is constructed, more accurate conductivity distribution can be obtained, and the accuracy of the conductivity distribution is improved. And the physical process of electrical impedance imaging can be described more accurately, so that the accuracy of model imaging is improved.

Description

Technical Field

[0001] The invention belongs to the field of electrical impedance tomography, in particular to a method for reconstructing a lung regional electrical impedance tomography image. Background Art

[0002] Electrical impedance tomography (EIT) is an imaging technology based on the internal conductivity distribution of an object. It is currently mainly used in the biomedical field. It applies current excitation to the surface of the object and measures the boundary voltage sequence on the surface of the object, thereby inferring the internal conductivity distribution of the object. EIT technology is suitable for real-time monitoring of the internal state of human organs and tissues. It has the advantages of non-invasiveness, no radiation, and strong real-time performance. In recent years, EIT has been widely used in the field of lung regional imaging to monitor the growth status of lung diseases in real time and evaluate the distribution of lung gas and ventilation status.

[0003] EIT imaging methods are usually divided into forward solving and reverse inverse problem solving: the forward problem refers to calculating the boundary voltage on the surface of an object when the current and conductivity distribution are known. Common methods include the finite element method and the boundary element method. The inverse problem solving refers to reconstructing the conductivity distribution inside the object based on the boundary voltage measured on the surface of the object. The inverse problem solving of EIT is pathological, so it needs to be calculated with the help of certain algorithms. Common methods include regularization, iterative optimization methods, and deep learning methods. In recent years, deep learning has been introduced into the inverse problem solving of EIT. Models such as convolutional neural networks are used to learn the conductivity distribution from the boundary voltage data, achieving good image reconstruction results.

[0004] Although EIT has very broad application prospects in many fields, it still faces some technical challenges in image reconstruction: the spatial resolution of EIT imaging is low, especially in areas with deeper internal structures. The measured surface boundary voltage data can only provide limited spatial information, so the imaging results are not as fine as CT images; EIT imaging is very sensitive to noise and artifacts, and body motion artifacts will cause the data accuracy to decrease, thereby affecting the accuracy of image reconstruction; EIT imaging technology based on deep learning requires a large amount of data sets and computing resources, the amount of calculation is huge, and the pattern interpretability is poor. It still faces problems such as insufficient imaging resolution, noise interference and model stability, so it is impossible to obtain an accurate EIT image. Summary of the invention

[0005] In view of this, the present invention extracts a graph structure EIT deep imaging model based on high-frequency and low-frequency dual-channel graph attention networks as the main body, named HLGAT. The attention mechanism is added to the graph neural network to construct a graph attention network. The high-frequency and low-frequency feature information of the nodes are aggregated by different graph structure networks respectively. The output of the characteristic information is fused in a learnable way to output the distribution of the conductivity of the lung area, and the EIT image is further reconstructed to improve the accuracy of EIT image reconstruction.

[0006] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0007] The present invention provides a lung region EIT image reconstruction method based on a graph structure, comprising the following steps:

[0008] Step S1: Design an EIT system, including a finite element model simulating the human lung region and an electrode array, which will be arranged at multiple locations in the human chest cavity to apply current to the lung region and record the corresponding boundary voltage data. These boundary voltage data are used to construct a data set for subsequent analysis and EIT image reconstruction.

[0009] Step S2: The original data cannot be directly used as the input of the graph neural network, so the EIT results need to be converted into the form of a graph structure. The present invention adopts non-uniform grid division technology to obtain the node information and adjacency matrix of the graph. Grid division not only affects the performance of the neural network model, but also directly affects the accuracy and computational efficiency of EIT imaging. According to the physical characteristics of the lung area and the complexity of the electric field distribution, the grid division will fully consider these factors to ensure the rationality of the grid division. Each pixel block is regarded as a node, and adjacent pixel units are connected by edges. The connection relationship only exists in the same area to form an undirected graph. The structural information of the graph is stored through the adjacency matrix. The adjacency matrix reflects the connection relationship between nodes and can characterize the propagation path of the current between different nodes. Based on the adjacency matrix, the Laplace matrix can be constructed to represent the propagation relationship between nodes, so that the model can more accurately reflect the propagation characteristics of the current. With the help of the Laplace matrix, the graph neural network can adjust the information transmission method between nodes, thereby improving the accuracy of EIT image reconstruction.

[0010] Step S3: Introduce the attention mechanism into the graph neural network and construct a graph attention network (GAT). The purpose of the graph neural network is to obtain the feature representation of the node by learning the connection relationship between the nodes. The introduced attention mechanism allows each node to assign different weights according to the importance of its neighbor nodes, thereby flexibly adjusting the influence of the neighbor nodes. The core idea of ​​GAT is to perform weighted averaging based on the neighbor relationship of the node. The weight of each neighbor node is calculated by the attention mechanism. The node feature update formula is as follows. The updated feature of node i is expressed by h i1 Indicates that N(i) is the set of neighbor nodes of node i, α ij is the attention weight between node i and node , W is the learnable weight matrix, h j is the feature representation of node j.

[0011]

[0012] For each pair of nodes i and j, an attention mechanism is used to calculate the relationship between them. The features between the nodes are calculated similarly to assign weights. The features of nodes i and j are mapped to a shared attention space to obtain an attention score e ij , the learned attention weight vector is denoted by α:

[0013] e ij =LeakyRelu(α T [Wh i ||Wh j ])

[0014] In order to ensure that the sum of the weights of all neighbor nodes is 1, the calculated attention coefficient is normalized using the Sofxmax function to obtain the attention weight of each neighbor node:

[0015]

[0016] In order to reconstruct the EIT image of the lung area more accurately, the graph convolution layer needs to be able to fully express the high-frequency and low-frequency information of adjacent nodes. The graph structure data composed of nodes and adjacency matrix is ​​input into the HLGAT network to aggregate the high-frequency and low-frequency information of neighbors. The calculation and extraction process of node feature information is as follows:

[0017]

[0018]

[0019]

[0020]

[0021]

[0022] In order to avoid mutual interference between filters and weight matrices, the fully connected layer is first used to perform nonlinear transformation on the input features. Then HLGAT learns positive attention values ​​and negative attention values ​​respectively to aggregate the high-frequency and low-frequency information of neighboring nodes and generate two feature representations of nodes. Finally, the two features are fused in an adaptive and learnable way. The high-frequency attention value and the low-frequency attention value are defined as follows:

[0023]

[0024]

[0025] In the above formula, and can be viewed as two shared convolution kernels. is set to a value greater than 0, It is set to a value less than 0, so that the feature information of high frequency and low frequency can be extracted respectively. Two activation functions NegRelu and PosRelu are designed. NegRelu learns positive attention value and PosRelu learns negative attention value, which are defined as follows:

[0026]

[0027]

[0028] NegRelu limits the high-frequency attention value to below 0, forcing the aggregation of high-frequency features; PosRelu limits the low-frequency attention value to above 0 to ensure further aggregation of low-frequency information. In this way, the balance of spatial features can be maintained in EIT image reconstruction while paying attention to overall and local multi-scale information. By extracting high-frequency and low-frequency features, the adjacency matrix is ​​further reconstructed to obtain a new Laplace matrix. During the iteration process, the structure of the graph is dynamically adjusted to better capture the complex relationship between node features, thereby significantly improving the accuracy and performance of HLGAT for image reconstruction.

[0029] Step S4: Using HLGAT to generate new Laplce and node feature information, the conductivity distribution of the lung area can be solved. Based on these conductivity data, the EIT image of the lung area can be reconstructed. By using relevant evaluation indicators, the performance of the model and the quality of the reconstructed image are quantitatively evaluated to obtain more accurate EIT image reconstruction results.

[0030] The graph-structure-based lung EIT image reconstruction method of the present invention has the following advantages:

[0031] A dual-channel graph attention network model of HLGAT was constructed to reconstruct the EIT image of the lung region. HLGAT aggregates low-frequency and high-frequency information through two channels respectively, and fuses the outputs of the two channels in a learnable way. The dual-channel model can improve the performance of node classification. The graph attention network learns to connect the positive and negative attention values ​​between nodes, performs weighted summation of adjacent node features, and updates the embedding of the current node; the graph attention network pays attention to the neighborhood of the node, follows the self-attention strategy, calculates the hidden representation of each node in the graph, improves the over-smoothing phenomenon to a certain extent, and effectively adjusts the combination of high-frequency and low-frequency signals. The adjacency matrix and Laplace matrix are reconstructed through the feature information extracted by HLGAT, which can provide stronger adaptability and expression ability in the HLGAT network, help improve the process of node feature aggregation, more effectively aggregate node information, improve the model's ability to capture complex relationships in node feature information, significantly improve the performance of model image reconstruction, and obtain a more accurate EIT image of the lung region.

[0032] The HLGAT model has very obvious advantages in the application of lung regional EIT imaging compared with other EIT imaging methods. Multi-scale feature fusion can provide richer feature expression, can process local and global information at the same time, and effectively aggregate the feature information of neighboring nodes. Traditional CNN or GNN models often cannot flexibly balance features of different scales, and cannot adaptively optimize the structure of the network model according to the characteristics of the data. In addition, traditional EIT imaging methods lack refined processing of feature information of different frequencies, and therefore cannot obtain a high-resolution EIT image. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0034] Figure 1 The non-Euclidean space undirected graph is obtained by meshing and mapping the lung area finite element simulation model of the present invention;

[0035] Figure 2 A schematic diagram of the graph attention network model (GAT) used in the present invention;

[0036] Figure 3 A schematic diagram of the HLGAT network model constructed by the present invention;

[0037] Figure 4 It is a principle diagram of the lung region EIT image reconstruction method based on graph structure of the present invention;

[0038] Figure 5Schematic diagram of the process of lung area EIT image reconstruction according to the present invention. DETAILED DESCRIPTION

[0039] The present invention is further described in detail below through specific examples. The following examples are only descriptive and not restrictive, and the protection scope of the present invention cannot be limited by them. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0040] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0041] First, an EIT system is designed. The system includes a finite element model of the human lung area and an electrode array. The electrodes are placed at different positions in the human chest cavity to inject current into the lung area and record the corresponding voltage data to form a voltage measurement data set.

[0042] Secondly, the non-uniform grid partitioning method is used to obtain the node information and adjacency matrix of the graph structure, such as Figure 1 As shown. Grid division not only affects the performance of the network model, but also directly affects the accuracy and computational efficiency of EIT imaging. The grid is divided according to the physical characteristics of the lung area space and the complexity of the electric field distribution. Pixels are connected by adjacent pixels. Each pixel block can be regarded as a node. The connection relationship only exists in the same area. It is mapped into an undirected graph in non-Euclidean space, and the adjacency matrix is ​​used to store the structural information of the graph. The adjacency matrix represents the mutual connection relationship between nodes and reflects the propagation path of the current between different nodes. The construction of the adjacency matrix is ​​based on the interaction relationship between nodes. The Laplace matrix can be obtained through the adjacency matrix and the degree matrix. It is used to represent the propagation relationship between nodes, which helps the model to accurately reflect the characteristics of current propagation. Through the Laplace matrix, the network model can further adjust the information transfer between nodes, thereby improving the accuracy of EIT image reconstruction;

[0043] Next, we build the HLGAT network model and introduce the attention mechanism into the graph neural network (GNN) to construct the graph attention network model (GAT), as shown in Figure 2 As shown in the figure. GNN aims to learn the feature expression of nodes through the connection relationship between nodes. The attention mechanism enables each node to assign different weights to its neighbor nodes and perform weighted calculations on neighbor nodes, so as to learn the relative importance of nodes more flexibly. The core idea of ​​the GAT network is to perform weighted averaging based on the neighbor relationship of nodes. In this process, the weight of each neighbor node is calculated through the attention mechanism. The node feature update formula is shown below. The updated feature of node i is expressed by h i1 Indicates that N(i) is the set of neighbor nodes of node i, α i, is the attention weight between node i and node j, W is the learnable weight matrix, h j is the feature representation of node j.

[0044]

[0045] For each pair of nodes i and j, an attention mechanism is used to calculate the relationship between them. The features between the nodes are calculated similarly to assign weights. The features of nodes i and j are mapped to a shared attention space to obtain an attention score e ij , the learned attention weight vector is denoted by α:

[0046] e ij =LeakyRelu(α T [Wh i ||Wh j ])

[0047] In order to ensure that the sum of the weights of all neighbor nodes is 1, the calculated attention coefficient is normalized using the Sofxmax function to obtain the attention weight of each neighbor node:

[0048]

[0049] In order to reconstruct the EIT image of the lung area more accurately, the graph convolution layer needs to be able to fully express the high-frequency and low-frequency information of adjacent nodes. The graph structure data composed of nodes and adjacency matrix is ​​input into the HLGAT network to aggregate the high-frequency and low-frequency information of neighbors. The calculation and extraction process of node feature information is as follows:

[0050]

[0051]

[0052]

[0053]

[0054]

[0055] The learnable matrix is ​​denoted by W F , W H and W L Indicates that W H With W LAdjust the relationship between the high frequency and low frequency of the current node, and input the node feature information extracted by HLGAT into the fully connected layer FC to output the distribution of conductivity and reconstruct the EIT image of the lung area. In order to avoid the entanglement of the filter and the weight matrix, the FC layer is first used to perform a nonlinear transformation on the original feature information. Then HLGAT learns negative attention values ​​and positive attention values ​​to aggregate the high frequency and low frequency information of neighboring nodes, and obtains two feature information expressions of the node. Finally, the expression of the two feature information is adaptively fused in a learnable way. The high frequency attention value and the low frequency attention value are defined as follows:

[0056]

[0057]

[0058] In the above formula, and can be viewed as two shared convolution kernels. is set to a value greater than 0, It is set to a value less than 0, so that the feature information of high frequency and low frequency can be extracted respectively. Two activation functions NegRelu and PosRelu are designed. NegRelu learns positive attention value and PosRelu learns negative attention value, which are defined as follows:

[0059]

[0060]

[0061] NegRelu can limit the high-frequency attention value to below 0, thereby forcing the high-frequency feature information to be aggregated; on the contrary, PosRelu can limit the low-frequency information to above 0, forcing the low-frequency information to be further aggregated, such as Figure 3 As shown. Aggregating high-frequency and low-frequency feature information can improve the accuracy of EIT image reconstruction, maintain the balance of spatial features, and provide a multi-scale representation method, so that the image reconstruction algorithm can focus on both overall and local features. By extracting the high-frequency and low-frequency feature vector information of the node, the adjacency matrix is ​​reconstructed, and a new Laplace matrix is ​​constructed using the new adjacency matrix. The structure of the graph is dynamically adjusted during the iteration process to better capture the relationship between node features or weight changes. The node feature information extracted by HLGAT forms a node feature matrix, and the node features are used to calculate the Euclidean distance between nodes. The distance is converted into a weighted form, and the attention weights calculated by HLGAT are combined with the adjacency matrix weights based on spatial distance to obtain a new adjacency matrix, thereby further obtaining a new Laplace matrix, which can better capture the complex and changing relationships between nodes, more effectively aggregate the feature information of nodes, and significantly improve the image reconstruction performance of the model.

[0062] Finally, the new Laplace matrix and node information reconstructed by HLGAT are used to solve the distribution of EIT conductivity in the lung area, as shown in Figure 4 As shown in the figure, the EIT image is reconstructed according to the conductivity distribution, and the performance of the model and the effect of EIT image reconstruction are evaluated using relevant measurement indicators to obtain a more accurate EIT image. The flowchart of the graph-based lung region EIT image reconstruction method is shown in the figure. Figure 5 shown.

[0063] The above description is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept, which all belong to the protection scope of the present invention.

Claims

1. A graph-based method for reconstructing electrical impedance tomography images of the lung region has been established. A dual-branch channel of a high-frequency graph attention network and a low-frequency graph attention network is used as the main body of an EIT deep imaging model named HLGAT. The method mainly includes the measurement of the voltage sequence at the boundary of the lung region, the acquisition of node information and node adjacency relationships by non-uniform grid division of the lung imaging region, the construction of the HLGAT network model and the reconstruction of a new adjacency matrix and Laplace matrix based on the node feature information output by the HLGAT, the solution of the conductivity distribution of the lung imaging region according to the new Laplace matrix, and the reconstruction of the EIT image. It is characterized by: The steps include: Step S1: Measuring boundary voltage. Design an EIT system, which includes a finite element model of the human lung region and an electrode array. The electrodes are placed at different positions in the human chest cavity and are used to inject current into the lung region, record corresponding voltage data, and form a voltage measurement data set. Step S2: Non-uniform grid division and acquisition of nodes and adjacency matrix. The original EIT data information cannot be directly used as the input of the network, so the EIT data results need to be constructed into a graph structure and stored in a form suitable for the input of the graph neural network. The present invention adopts a non-uniform grid division method to obtain the node information and adjacency matrix of the graph structure. The grid division not only affects the performance of the network model, but also directly affects the accuracy and calculation efficiency of EIT imaging. The grid is divided according to the physical characteristics of the lung area space and the complexity of the electric field distribution. Pixels are connected by adjacent pixels. Each pixel block can be regarded as a node. The connection relationship only exists in the same area. It is mapped into an undirected graph in non-Euclidean space. The adjacency matrix is ​​used to store the structural information of the graph. The adjacency matrix represents the mutual connection relationship between nodes and reflects the propagation path of the current between different nodes. The construction of the adjacency matrix is ​​based on the interaction relationship between nodes. The Laplace matrix can be obtained through the adjacency matrix and the degree matrix. It is used to represent the propagation relationship between nodes, which helps the model to accurately reflect the characteristics of current propagation. Through the Laplace matrix, the network model can further adjust the information transfer between nodes, thereby improving the accuracy of EIT image reconstruction; Step S3: Construction of HLGAT network. The attention mechanism is introduced into the graph neural network (GNN) to construct the graph attention network model (GAT). GNN aims to learn the feature expression of nodes through the connection relationship between nodes. The attention mechanism enables each node to assign different weights to its neighbor nodes and perform weighted calculations on neighbor nodes, so as to learn the relative importance of nodes more flexibly. The core idea of ​​the GAT network is to perform weighted averaging based on the neighbor relationship of the nodes. In this process, the weight of each neighbor node is calculated through the attention mechanism. The node feature update formula is shown below. The updated feature of node i is expressed by h i1 Indicates that N(i) is the set of neighbor nodes of node i, α ij is the attention weight between node i and node j, W is a learnable weight matrix, and h is the feature representation of node j. For each pair of nodes i and j, an attention mechanism is used to calculate the relationship between them. The features between the nodes are calculated similarly to assign weights. The features of nodes i and j are mapped to a shared attention space to obtain an attention score e ij , the learned attention weight vector is denoted by α: and ij =LeakyRelu(α T [Wh i ||Wh j ]) In order to ensure that the sum of the weights of all neighbor nodes is 1, the calculated attention coefficient is normalized using the Sofxmax function to obtain the attention weight of each neighbor node: In order to reconstruct the EIT image of the lung area more accurately, the graph convolution layer needs to be able to fully express the high-frequency and low-frequency information of adjacent nodes. The graph structure data composed of nodes and adjacency matrix is ​​input into the HLGAT network to aggregate the high-frequency and low-frequency information of neighbors. The calculation and extraction process of node feature information is as follows: The learnable matrix is ​​denoted by W F , W H and W L Indicates that W H With W L Adjust the relationship between the high frequency and low frequency of the current node, and input the node feature information extracted by HLGAT into the fully connected layer FC to output the distribution of conductivity and reconstruct the EIT image of the lung area. In order to avoid the entanglement of the filter and the weight matrix, the FC layer is first used to perform a nonlinear transformation on the original feature information. Then HLGAT learns negative attention values ​​and positive attention values ​​to aggregate the high frequency and low frequency information of neighboring nodes, and obtains two feature information expressions of the node. Finally, the expression of the two feature information is adaptively fused in a learnable way. The high frequency attention value and the low frequency attention value are defined as follows: In the above formula, and can be viewed as two shared convolution kernels. is set to a value greater than 0, It is set to a value less than 0, so that the feature information of high frequency and low frequency can be extracted respectively. Two activation functions NegRelu and PosRelu are designed. NegRelu learns positive attention value and PosRelu learns negative attention value, which are defined as follows: NegRelu can limit the high-frequency attention value to below 0, thereby forcing the high-frequency feature information to be aggregated; on the contrary, PosRelu can limit the low-frequency information to above 0, forcing the low-frequency information to be further aggregated. Aggregating high-frequency and low-frequency feature information can improve the accuracy of EIT image reconstruction, maintain the balance of spatial features, and provide a multi-scale representation method so that the image reconstruction algorithm can focus on both overall and local features at the same time. By extracting the high-frequency and low-frequency feature vector information of the node, the adjacency matrix is ​​reconstructed, and a new Laplace matrix is ​​constructed using the new adjacency matrix. The structure of the graph is dynamically adjusted during the iteration process to better capture the relationship or weight change between node features. The node feature information extracted by HLGAT forms a node feature matrix, and the node features are used to calculate the Euclidean distance between nodes. The distance is converted into a weighted form. The attention weights calculated by HLGAT are combined with the adjacency matrix weights based on spatial distance to obtain a new adjacency matrix, thereby further obtaining a new Laplace matrix, which can better capture the complex and changing relationships between nodes, more effectively aggregate the feature information of nodes, and significantly improve the image reconstruction performance of the model. Step S4: using the new Laplace matrix and node information reconstructed by HLGAT to solve the EIT conductivity distribution of the lung area, reconstructing the EIT image according to the conductivity distribution, and using relevant measurement indicators to evaluate the performance of the model and the effect of EIT image reconstruction, thereby obtaining a more accurate EIT image; 2. When measuring the boundary voltage in step S1 according to claim 1, current is applied through the electrode array and voltage is measured at the electrodes, satisfying the boundary conditions of electrode measurement: The current density is represented by J, n is the normal vector of the electrode surface, I k is the current applied to the kth electrode, is the position of the kth electrode, is the voltage measured on the electrode, which satisfies the current boundary condition and the voltage boundary condition.

3. According to claim 1, when step S2 adopts non-uniform grid partitioning to obtain nodes and adjacency matrix, the characteristic of non-uniform grid is that the spatial distribution of nodes is adjusted according to the geometric characteristics or conductivity distribution of the region, and the adjacency matrix of the graph structure describes the connection relationship between nodes. The element A of the adjacency matrix ij Usually the interaction element K between nodes ij and the geometric distance d between nodes ij Decided:

4. According to step S3 of claim 1, when reconstructing the adjacency matrix and the Laplace matrix, the adjacency matrix is ​​reconstructed by using the method of calculating the spatial distance, and the connection weights between nodes can be adjusted by the spatial distance to construct a spatially aware adjacency matrix; the connection weights are adjusted in combination with the distribution of conductivity, so that the adjacency matrix can better reflect the propagation characteristics of the electric field, and further enable the HLGAT model to more effectively combine the physical spatial characteristics and conductivity information, thereby improving the accuracy of EIT image reconstruction in the lung area.

Images