A high-resolution reconstruction method and system for electrical impedance imaging based on deep learning

Abstract

The present invention discloses a method and system for high-resolution reconstruction of electrical impedance imaging based on deep learning. The present invention prepares an EIT data set by segmenting a CT or MRI data set and simulating electrode application to obtain the boundary voltage and conductivity distribution. The Gauss-Newton method is used to perform coarse conductivity reconstruction based on the measured boundary voltage, providing prior information for accurate conductivity reconstruction and improving the reconstruction quality. The present invention proposes a dual-branch collaborative fusion strategy network to achieve accurate reconstruction of conductivity. The network model includes two branches: a conductivity branch and a structural branch. The conductivity branch is used to learn the mapping from the measured boundary voltage to the conductivity; the structural branch is used to learn the mapping from the coarsely reconstructed conductivity to the accurate conductivity, and the results of the two branches are fused and trained to accurately reconstruct the conductivity. The present invention uses the above technology to accurately reconstruct the conductivity of human tissues and heterogeneous areas of interest.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a method and system for high-resolution reconstruction of electrical impedance imaging based on deep learning. Background Art

[0002] Electrical impedance tomography (EIT) is a non-invasive imaging technique in medical imaging. Traditional EIT reconstructs the electrical conductivity distribution of biological tissue by applying a small current and measuring the resulting voltage changes. However, due to the inherent pathological and ill-posed nature of the EIT method, the reconstructed conductivity images suffer from low spatial resolution and poor accuracy. Therefore, improving the resolution and accuracy of EIT imaging is a current research hotspot and challenge. Summary of the Invention

[0003] Purpose of the invention: In response to the above shortcomings, the present invention provides a high-resolution reconstruction method and system for electrical impedance imaging based on deep learning, with the aim of providing accurate reconstruction of the electrical conductivity of the biological tissue area of ​​interest.

[0004] Technical solution: To solve the above problems, the present invention discloses a high-resolution reconstruction method for electrical impedance imaging based on deep learning, comprising the following steps:

[0005] (1) Segment the acquired CT or MRI data set to separate normal tissue areas and foreign matter areas, and discretize the segmentation results into grids;

[0006] (2) The segmentation result after mesh discretization is used as a model, and simulated electrodes are applied to the model. The normal tissue area is imported into the model with the simulated electrodes applied, and the normal tissue area is assigned the conductivity value of the normal tissue area, and the other areas in the model except the normal tissue area are assigned the conductivity values ​​of other areas to obtain the first conductivity distribution value σ in the current model. Then, current is applied and the boundary voltage value v0 is solved using the finite element method, and the sensitivity matrix S is calculated based on the boundary voltage value v0. Then, the foreign matter area is imported into the model with the normal tissue area, and the corresponding foreign matter area conductivity value is assigned to the foreign matter area to obtain the second conductivity distribution value σ′ in the current model. Then, the final boundary voltage value v is calculated based on the given conductivity distribution value σ′ and the sensitivity matrix S.

[0007] (3) Use the Gauss-Newton method to perform rough reconstruction of the conductivity according to the boundary voltage value v to obtain the rough reconstructed conductivity σ k ;

[0008] (4) The boundary voltage v and the corresponding coarsely reconstructed conductivity are As a sample, iterate N times the conductivity That is, N samples are obtained; the N samples are partially divided into a training set and partially divided into a test set; the model includes a conductivity branch and a structure branch; the conductivity branch is used to learn the mapping from the boundary voltage value v measured in step (2) to the conductivity distribution value σ; the structure branch is used to learn the coarsely reconstructed conductivity σ k The two feature maps learned from the conductivity branch and the structure branch are then merged. The merged features are then processed by convolution operation and finally activated by Sigmoid function to generate the predicted conductivity σ pred as the final output.

[0009] Furthermore, in step (4), a loss function of the dual-branch collaborative fusion strategy network is constructed to guide the parameter adjustment of the collaborative fusion of the conductivity branch and the structure branch during the training process.

[0010] Furthermore, in step (4), the Adam optimizer is used to update the network model according to the loss function of the conductivity branch and the structure branch collaborative fusion strategy network.

[0011] Furthermore, in step (2), the expression of the sensitivity matrix S is: in, It represents the partial derivative of the boundary voltage with respect to the conductivity distribution, that is, the effect of a small change in conductivity on the voltage measurement.

[0012] Furthermore, in step (2), a new boundary voltage v1=Sσ' is calculated based on the solved sensitivity matrix S, and finally the boundary voltages v0 and v1 are preprocessed to obtain the final boundary voltage value v=(v1-v0) / v0 of the model.

[0013] Furthermore, step (3) includes the following sub-steps:

[0014] (3.1) Computing the Jacobian matrix: The Gauss-Newton method performs a linear search strategy in a specific descent direction, ignoring the second-order terms in the Taylor expansion approximation, as follows:

[0015]

[0016] Among them, f(σ k ) indicates that in σ k The objective function value, f(σ k +d) indicates that k The objective function value after a certain distance along the direction d, in addition, f(σ k )’s gradient and Hesser matrix are expressed as:

[0017]

[0018]

[0019] Where J(σ k ) represents r s (σ k )=F(σ k )-v’s Jacobian matrix, F(σ k ) is the observed value, r s is the residual function, s is the index of the voltage, and m is the number of voltages;

[0020] (3.2). Calculate the search direction: Search direction This can be approximated by solving the following equation:

[0021]

[0022] Where J(σ k ) T J(σ k ) is a diagonal matrix, λ is the regularization coefficient λ;

[0023] (3.3) Iterative calculation of conductivity distribution Iteratively calculates σ based on the previous initial estimate of the conductivity and the search direction k+1 , until the set number of iterations is met, otherwise return to step (3.1).

[0024] Furthermore, in step (4), in the conductivity branch, a cascaded fully connected U-Net network is used to learn the mapping of boundary voltage v to conductivity. The voltage value is used as the input of the network and initialized by a fully connected layer to obtain a matrix, and the matrix is ​​reshaped and converted. After the conversion, the matrix is ​​passed to the U-Net network for feature learning; the U-Net network includes several downsampling operations and several upsampling operations. Each downsampling operation will first pass through several convolution layers. After each convolution, the feature channel normalization method is used for normalization, and then the ReLU function is used for activation to finally obtain the conductivity branch feature matrix.

[0025] Furthermore, in the structure branch, the attention mechanism residual-U-Net network is used to learn the Gauss-Newton method to roughly reconstruct the conductivity To the actual conductivity mapping; this branch directly inputs the coarse reconstructed conductivity into the U-Net network for feature learning; the U-Net network includes several downsampling operations and several upsampling operations; each downsampling operation will first pass through several convolution layers, and after each convolution, it will be normalized using the feature channel normalization method, and then activated using the ReLU function; and an attention mechanism module is introduced in the structure branch to enhance the ability of the U-Net network to capture global information. The attention mechanism module includes a channel attention module and a spatial attention module, and the features obtained by the upsampling operation are concatenated to form a feature matrix, which is then passed into the attention mechanism module after the convolution layer to obtain a structure branch feature matrix.

[0026] Furthermore, the conductivity branch feature matrix and the structure branch feature matrix are synergistically integrated to achieve the final conductivity prediction: the conductivity branch feature matrix and the structure branch feature matrix are spliced ​​to form a 64*128*128 feature matrix X, and the feature matrix X is passed into three convolutional layers. After the first two convolutional layers, the BatchNorm (BN) normalization method is used for normalization and the ReLU function is used for activation; finally, the Sigmoid function is used to activate the output of the third convolutional layer to generate the final predicted conductivity value σ pred ,Right now

[0027]

[0028] The present invention also provides a deep learning-based electrical impedance imaging high-resolution reconstruction system, which is characterized by comprising:

[0029] The dataset creation module is used to segment the acquired CT or MRI dataset into normal tissue areas and foreign matter areas, and to perform grid discretization on the segmentation results.

[0030] The conductivity distribution and boundary voltage value data processing module is used to use the segmentation result after mesh discretization as a model, apply simulated electrodes on the model, import the normal tissue area into the model with the simulated electrodes applied, assign the normal tissue area conductivity value, assign other area conductivity values ​​to other areas in the model except the normal tissue area, to obtain the first conductivity distribution value σ in the current model, then apply current and use the finite element method to solve the boundary voltage value v0, and then calculate the sensitivity matrix S based on the boundary voltage value v0; then import the foreign matter area into the model that already has the normal tissue area, assign the corresponding foreign matter area conductivity value to the foreign matter area to obtain the second conductivity distribution value σ′ in the current model, and then calculate the final boundary voltage value v based on the given conductivity distribution value σ′ and the sensitivity matrix S;

[0031] The conductivity coarse reconstruction module is used to perform coarse conductivity reconstruction according to the boundary voltage value v through the Gauss-Newton method to obtain the coarse reconstructed conductivity σ k ;

[0032] Conductivity accurate reconstruction module, used to reconstruct the boundary voltage v and the corresponding coarse reconstructed conductivity of the iteration As a sample, iterate N times the conductivity That is, N samples are obtained; the N samples are partially divided into a training set and partially divided into a test set; the model includes a conductivity branch and a structure branch; the conductivity branch is used to learn the mapping from the boundary voltage value v measured in step (2) to the conductivity distribution value σ; the structure branch is used to learn the coarsely reconstructed conductivity σ k The two feature maps learned from the conductivity branch and the structure branch are then merged. The merged features are then processed by convolution operation and finally activated by Sigmoid function to generate the predicted conductivity σ pred as the final output.

[0033] Beneficial Effects: 1. High Resolution and Accuracy: This paper utilizes a novel deep learning-based reconstruction method to significantly improve the spatial resolution and accuracy of conductivity reconstruction in lesion areas. By applying deep learning technology, the low resolution and poor accuracy of traditional EIT methods, which suffer from pathological and ill-posed characteristics, are overcome, achieving more accurate conductivity distribution reconstruction.

[0034] 2. Reduced reliance on prior knowledge: Based on deep learning technology, this method eliminates the need for extensive prior knowledge and heuristic algorithms to improve image reconstruction quality. Deep learning models possess strong self-learning capabilities and can learn rich feature representations from large amounts of data, thus reducing reliance on prior knowledge. This makes the method more versatile and applicable. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of the dual-branch collaborative fusion strategy network construction method of the present invention.

[0036] Figure 2 Schematic diagram of the 16-electrode electrical impedance tomography measurement system of the present invention, wherein (a) is a schematic diagram of the chest EIT model with the addition of a lung grid, and (b) is a schematic diagram of the chest EIT model with the addition of a lung grid and a foreign body grid.

[0037] Figure 3 This is a schematic diagram of the overall training of the dual-branch collaborative fusion strategy network of the present invention.

[0038] Figure 4Schematic diagram of experimental results of the method of the present invention and several other classical conductivity reconstruction methods in the prior art. DETAILED DESCRIPTION

[0039] The technical solution of the present invention is further described below with reference to the accompanying drawings.

[0040] Example 1

[0041] In order to make the purpose, technical solutions and advantages of the present invention more clear, the following embodiments and appended Figure 1 、 2 3. Specifically, the present invention is further described in detail by taking the conductivity reconstruction of lung disease as an example. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention. The specific embodiment process of the present invention is as follows Figure 1 As shown, the steps are as follows:

[0042] Step 1: Create a lung EIT dataset;

[0043] Step 1.1: Select the human chest as the object, obtain a chest CT or MRI image for the object, and use the deep neural network U-Net model to segment the skin, lungs, and foreign matter areas of the chest. The segmentation results are discretized into a grid using Delaunay triangulation.

[0044] Step 1.2: Apply electrodes and simulate to construct an EIT dataset. This method uses a 16-electrode EIT sensor and adjacent excitation measurement mode. 16 electrodes are evenly attached around the chest skin to construct a chest model with 15,068 mesh nodes and 30,459 triangular facets. Specifically:

[0045] The lung mesh was imported into the chest mesh with electrodes attached to construct the model, and conductivity values ​​of 0.48 S / m and 0.0251 S / m were assigned to the chest region and lung region, respectively. Then, the boundary voltage value was measured using the adjacent excitation measurement mode. Figure 2As shown in (a), first, an excitation current of 1 mA is injected between electrode 1 and electrode 2, and the relative potential between the other 14 electrodes is measured, for a total of 13 voltage values. In the second step, the injection electrodes are converted from the original electrodes 1 and 2 to electrodes 2 and 3, and the measured voltages are the relative voltage values ​​of the 14 electrodes except electrodes 2 and 3, for a total of 13 voltage values. The measurement is repeated in this way, and finally 208 boundary voltage values ​​v0 can be measured. Next, the sensitivity matrix S is solved based on the given conductivity distribution σ in the constructed model (at this time, the conductivity value of 8969 grid nodes in the model is 0.48S / m, and the conductivity value of 6099 grid nodes is 0.0251S / m) and the boundary voltage value v0. The sensitivity matrix S describes the degree of influence of conductivity changes on voltage measurements. Specifically, the expression of the sensitivity matrix S is:

[0046]

[0047] in, represents the partial derivative of the boundary voltage with respect to the conductivity distribution, i.e., the effect of a small change in conductivity on the voltage measurement. This process can obtain a sensitivity matrix S of size 208*15068. Figure 2 As shown in (b), Figure 2 Based on the model (a), the heterogeneous mesh is imported and a conductivity value of 0.3 S / m is assigned to the heterogeneous region. The new boundary voltage v1 = Sσ' is calculated based on the model's conductivity distribution σ' (the conductivity value of 8969 grid nodes in the model is 0.48 S / m, the conductivity value of 5248 grid nodes is 0.0251 S / m, and the conductivity value of 851 grid nodes is 0.3 S / m) and the solved sensitivity matrix S. Finally, the measured boundary voltages v0 and v1 are preprocessed to obtain the final boundary voltage v = (v1 - v0) / v0 for the model, with the size of this variable being 1 * 208.

[0048] Step 2: Use the Gauss-Newton method to perform a rough reconstruction of the conductivity based on the measured boundary voltage;

[0049] Step 2.1. Calculate the Jacobian matrix. The Gauss-Newton method performs a linear search strategy in a specific descent direction, ignoring the second-order terms in the Taylor expansion approximation. The details are as follows:

[0050]

[0051] Among them, f(σ k ) indicates that in σ k The objective function value, f(σ k +d) indicates that k The objective function value after a certain distance along the direction d. In addition, f(σ k) can be expressed as:

[0052]

[0053]

[0054] Where J(σ k ) represents r s (σ k )=F(σ k )-v’s Jacobian matrix, F(σ k ) is the observed value, r s is the residual function, s is the index of the voltage, and m is the number of voltages and is set to 208.

[0055] Step 2.2, calculate the search direction. By ignoring the second-order term in formula (2) The search direction can be calculated However, due to the compact structure of the Jacobian matrix, the coefficient matrix is ​​an unbounded inverse matrix. Therefore, the Gauss-Newton method may lead to inaccurate solutions. Therefore, this patent uses the regularization in the Levenberg-Marquardt (LM) algorithm to solve σ k , the LM algorithm directly regularizes the ill-conditioned system by adding a diagonal matrix. Therefore, the search direction of nonlinear problems in EIT is This can be approximated by solving the following equation:

[0056]

[0057] Where J(σ k ) T J(σ k ) is a diagonal matrix, and the regularization coefficient λ is set to 0.5.

[0058] Step 2.3, iterative calculation of conductivity distribution Iteratively calculates σ based on the previous initial estimate of the conductivity and the search direction k+1 , Until the set number of iterations is met, otherwise return to step 2.1. Since the resolution of the conductivity image reconstructed in this patent is 128×128, the size of the conductivity coarsely reconstructed by the Gauss-Newton method is also 128×128.

[0059] Step 3: Use a dual-branch collaborative fusion strategy network to achieve accurate conductivity reconstruction.

[0060] Step 3.1, divide the data into training set and test set, 80% of all data as training set, 20% of the data as test set, the training set is used to train the model, and the test set is used to test the performance of the model. A single sample of each part of the data contains the boundary voltage v measured in step 1 and the corresponding conductivity roughly reconstructed in step 2.

[0061] Step 3.2: Construct a dual-branch collaborative fusion strategy network. The framework of the dual-branch collaborative fusion strategy network is as follows: Figure 3 As shown in the figure, the network model consists of two branches: conductivity branch and structure branch. The conductivity branch uses a cascaded fully connected U-Net network to learn the mapping from the boundary voltage measured in step 1 to the conductivity; the structure branch uses the attention mechanism residual U-Net network to learn the mapping from the coarsely reconstructed conductivity in step 2 to the precise conductivity. The two feature maps learned from the conductivity branch and the structure branch are then merged. Subsequently, the merged features are processed by three layers of convolution operations and finally activated using the Sigmoid function to generate the predicted conductivity σ pred As the final output. Specifically including:

[0062] In the conductivity branch, a cascaded fully connected U-Net network is used to learn the mapping from boundary voltage v to conductivity. A voltage value of size 1*208 is used as the network input and initialized by a 16384-dimensional fully connected layer. This results in a 1*16384 matrix, which is then reshaped to a 128*128 matrix and passed to the U-Net network for feature learning. The U-Net network includes four downsampling operations and four upsampling operations. Each downsampling operation first passes through three 3*3 convolutional layers with a stride of 1, followed by one 3*3 convolutional layer with a stride of 2. After each convolution, the feature channel normalization method is used and then activated using the ReLU function, resulting in a feature matrix of size 32*128*128.

[0063] In the structure branch, the attention mechanism residual-U-Net network is used to learn the Gauss-Newton method to roughly reconstruct the conductivity The architecture branch maps the reconstructed conductivity to the actual conductivity. This branch directly inputs the 128*128 coarse reconstructed conductivity into the U-Net network for feature learning. The U-Net network consists of four downsampling operations and four upsampling operations. Each downsampling operation first passes through three 3*3 convolutional layers with a stride of 1, followed by one 3*3 convolutional layer with a stride of 2. Each convolution is normalized using the feature channel normalization method and then activated using the ReLU function. In addition, an attention mechanism module is introduced in the architecture branch to enhance the U-Net network's ability to capture global information. This attention mechanism module includes a channel attention module and a spatial attention module. The features obtained from the four upsampling operations are concatenated to form a feature matrix of size 960*128*128. This feature matrix is ​​then passed through a 3*3 convolutional layer with a stride of 1 to obtain a feature matrix of size 256*128*128. This feature matrix is ​​then passed to the attention mechanism module, resulting in a feature matrix of size 32*128*128.

[0064] The feature matrices learned by the conductivity branch and the structure branch are synergistically integrated to achieve the final conductivity prediction. The two feature matrices of size 32*128*128 obtained by the two branches are concatenated to form a 64*128*128 feature matrix X, and this feature matrix is ​​passed to three 3*3 convolutional layers with a stride of 1. After the first two convolutional layers, the BatchNorm (BN) normalization method is used for normalization and the ReLU function is used for activation. Finally, the Sigmoid function is used to activate the output of the third convolutional layer to generate the final predicted conductivity value σ pred ,Right now:

[0065]

[0066] In addition, in the above model, the Adam optimizer can also be used to update the model according to the loss function of the dual-branch collaborative fusion strategy network. The initial learning rate is 0.0003, and a total of 300 epochs are trained. A batch of boundary voltage values ​​and the corresponding coarsely reconstructed conductivity are taken as the input of the dual-branch collaborative fusion strategy network to generate the target conductivity distribution image. During the training process, the structural similarity index of the predicted conductivity distribution image and the true conductivity distribution image is monitored, and the epoch with the best performance on the training data is recorded. In this way, the model parameters with the best performance can be selected and saved for the test set data. Among them, the loss function of the dual-branch collaborative fusion strategy network is constructed to guide the parameter adjustment of the dual-branch collaborative fusion strategy network during the training process. The mean square error is used as the loss function of the dual-branch collaborative fusion strategy network model. In addition, in order to improve the generalization ability of the model, a regularization term is added to the loss function. The loss function formula is as follows:

[0067]

[0068] Among them, the regularization coefficient ω is 0.01, ‖θ‖ 2 represents the L2 norm of the model parameters, and N represents the actual conductivity distribution σ act The number of N is 16384, σ pred represents the conductivity value predicted by the dual-branch collaborative fusion strategy network, σ act and σ pred The size of the image is 128*128. In addition, in order to prove the effectiveness of the method of the present invention, three classic conductivity reconstruction methods in the prior art, namely Convolutional Neural Networks (CNN), U-Net Neural Network and Error-Constraint Network (Ec-Net), are introduced to compare with the reconstruction method of this embodiment. The experimental results are shown in Figure 2. Figure 4 As shown in the figure, the results show that compared with the CNN, U-Net, and Ec-Net methods, the patented method significantly improves the quality of reconstructed images. It not only reduces reconstruction artifacts but also accurately reconstructs the location and size of foreign objects. This demonstrates the excellent reconstruction capabilities of the patented method.

[0069] An embodiment of the present invention provides a deep learning-based high-resolution reconstruction method for electrical impedance tomography, which uses a two-branch collaborative fusion strategy network to achieve accurate conductivity reconstruction. A collaborative fusion training strategy is adopted for the conductivity branch and the structure branch. The conductivity branch uses a cascaded FC-U-Net network to learn the mapping from boundary voltage to conductivity. The structure branch uses the results of Gauss-Newton reconstruction as input, and an attention mechanism is introduced in the structure branch to simultaneously express the dependencies between spatial and channel dimensions, thereby enhancing the model's ability to capture global information. Finally, the features of the two branches are fused for coordinated end-to-end training.

[0070] Example 2

[0071] Corresponding to the above method, the present invention also provides a high-resolution reconstruction system for electrical impedance imaging based on deep learning, comprising:

[0072] The dataset creation module is used to segment the acquired CT or MRI dataset into normal tissue areas and foreign matter areas, and to perform grid discretization on the segmentation results.

[0073] The conductivity distribution and boundary voltage value data processing module is used to use the segmentation result after mesh discretization as a model, apply simulated electrodes on the model, import the normal tissue area into the model with the simulated electrodes applied, assign the normal tissue area conductivity value, assign other area conductivity values ​​to other areas in the model except the normal tissue area, to obtain the first conductivity distribution value σ in the current model, then apply current and use the finite element method to solve the boundary voltage value v0, and then calculate the sensitivity matrix S based on the boundary voltage value v0; then import the foreign matter area into the model that already has the normal tissue area, assign the corresponding foreign matter area conductivity value to the foreign matter area to obtain the second conductivity distribution value σ′ in the current model, and then calculate the final boundary voltage value v based on the given conductivity distribution value σ′ and the sensitivity matrix S;

[0074] The conductivity coarse reconstruction module is used to perform coarse conductivity reconstruction according to the boundary voltage value v through the Gauss-Newton method to obtain the coarse reconstructed conductivity σ k ;

[0075] Conductivity accurate reconstruction module, used to reconstruct the boundary voltage v and the corresponding coarse reconstructed conductivity of the iteration As a sample, iterate N times the conductivity That is, N samples are obtained; the N samples are partially divided into a training set and partially divided into a test set; the model includes a conductivity branch and a structure branch; the conductivity branch is used to learn the mapping from the boundary voltage value v measured in step (2) to the conductivity distribution value σ; the structure branch is used to learn the coarsely reconstructed conductivity σ k The two feature maps learned from the conductivity branch and the structure branch are then merged. The merged features are then processed by convolution operation and finally activated by Sigmoid function to generate the predicted conductivity σ pred as the final output.

Claims

1. A high-resolution reconstruction method for electrical impedance imaging based on deep learning, characterized in that: The following steps are involved: (1) Segment the acquired CT or MRI data set to separate normal tissue areas and foreign matter areas, and discretize the segmentation results into grids; (2) The segmentation result after mesh discretization is used as a model, and simulated electrodes are applied to the model. The normal tissue area is imported into the model with the simulated electrodes applied, and the normal tissue area is assigned the conductivity value of the normal tissue area, and the other areas in the model except the normal tissue area are assigned the conductivity values ​​of other areas to obtain the first conductivity distribution value σ in the current model. Then, current is applied and the boundary voltage value v0 is solved using the finite element method. Then, the sensitivity matrix S is calculated based on the boundary voltage value v0; Then, the foreign matter region is introduced into the model that already has a normal tissue region, and the foreign matter region is assigned a corresponding foreign matter region conductivity value to obtain a second conductivity distribution value σ′ in the current model. Then, the final boundary voltage value v is calculated based on the given conductivity distribution value σ′ and the sensitivity matrix S. (3) Use the Gauss-Newton method to perform rough reconstruction of the conductivity according to the boundary voltage value v to obtain the rough reconstructed conductivity σ k ; (4) The boundary voltage v and the corresponding coarsely reconstructed conductivity are As a sample, iterate N times the conductivity That is, N samples are obtained; the N samples are partially divided into a training set and partially divided into a test set; the model includes a conductivity branch and a structure branch; the conductivity branch is used to learn the mapping from the boundary voltage value v measured in step (2) to the conductivity distribution value σ; the structure branch is used to learn the coarsely reconstructed conductivity σ k Mapping to precise conductivity; The two feature maps learned from the conductivity branch and the structure branch are then merged; the merged features are then processed by convolution operation and finally activated using the Sigmoid function to generate the predicted conductivity σ pred as the final output.

2. The high-resolution reconstruction method for electrical impedance imaging based on deep learning according to claim 1, characterized in that: In step (4), a loss function of the dual-branch collaborative fusion strategy network is also constructed to guide the parameter adjustment of the collaborative fusion of the conductivity branch and the structure branch during the training process.

3. The high-resolution reconstruction method for electrical impedance imaging based on deep learning according to claim 1 or 2, characterized in that: In step (4), the Adam optimizer is used to update the network model according to the loss function of the conductivity branch and the structure branch collaborative fusion strategy network.

4. The high-resolution reconstruction method for electrical impedance imaging based on deep learning according to claim 1, characterized in that: In step (2), the expression of the sensitivity matrix S is: in, It represents the partial derivative of the boundary voltage with respect to the conductivity distribution, that is, the effect of a small change in conductivity on the voltage measurement.

5. The high-resolution reconstruction method for electrical impedance imaging based on deep learning according to claim 4, characterized in that: In step (2), a new boundary voltage v1=Sσ' is calculated based on the solved sensitivity matrix S, and finally the boundary voltages v0 and v1 are preprocessed to obtain the final boundary voltage value v=(v1-v0) / v0 of the model.

6. The high-resolution reconstruction method for electrical impedance imaging based on deep learning according to claim 1, characterized in that: Step (3) includes the following sub-steps: (3.1) Computing the Jacobian matrix: The Gauss-Newton method performs a linear search strategy in a specific descent direction, ignoring the second-order terms in the Taylor expansion approximation, as follows: Among them, f(σ k ) indicates that in σ k The objective function value, f(σ k +d) indicates that k The objective function value after a certain distance along the direction d, in addition, f(σ k )’s gradient and Hesser matrix are expressed as: Where J(σ k ) represents r s (σ k )=F(σ k )-v’s Jacobian matrix, F(σ k ) is the observed value, r s is the residual function, s is the index of the voltage, and m is the number of voltages; (3.2). Calculate the search direction: Search direction This can be approximated by solving the following equation: Where J(σ k ) T J(σ k ) is a diagonal matrix, λ is the regularization coefficient λ; (3.3) Iterative calculation of conductivity distribution Iteratively calculates σ based on the previous initial estimate of the conductivity and the search direction k+1 , until the set number of iterations is met, otherwise return to step (3.1).

7. The high-resolution reconstruction method for electrical impedance imaging based on deep learning according to claim 1, characterized in that: In step (4), in the conductivity branch, a cascaded fully connected U-Net network is used to learn the mapping from boundary voltage v to conductivity. The voltage value is used as the input of the network and initialized by a fully connected layer to obtain a matrix. The matrix is ​​then reshaped and converted. After conversion, the matrix is ​​passed to the U-Net network for feature learning. The U-Net network includes several downsampling operations and several upsampling operations. Each downsampling operation will first pass through several convolution layers. After each convolution, the feature channel normalization method is used for normalization, and then the ReLU function is used for activation to finally obtain the conductivity branch feature matrix.

8. The high-resolution reconstruction method for electrical impedance imaging based on deep learning according to claim 7, characterized in that: In the structure branch, the attention mechanism residual-U-Net network is used to learn the Gauss-Newton method to roughly reconstruct the conductivity Mapping to actual conductivity; this branch directly inputs the coarse reconstruction conductivity into the U-Net network for feature learning; The U-Net network includes several downsampling operations and several upsampling operations; each downsampling operation will first pass through several convolution layers, and after each convolution, it will be normalized using the feature channel normalization method, and then activated using the ReLU function; and an attention mechanism module is introduced in the structural branch to enhance the U-Net network's ability to capture global information. The attention mechanism module includes a channel attention module and a spatial attention module, and the features obtained by the upsampling operation are concatenated to form a feature matrix. The feature matrix is ​​then passed into the attention mechanism module after passing through the convolution layer to obtain a structural branch feature matrix.

9. The high-resolution reconstruction method for electrical impedance imaging based on deep learning according to claim 8, characterized in that: The conductivity branch feature matrix and the structure branch feature matrix are synergistically integrated to achieve the final conductivity prediction: the conductivity branch feature matrix and the structure branch feature matrix are spliced ​​to form a 64*128*128 feature matrix X, and the feature matrix X is passed to three convolutional layers. After the first two convolutional layers, the BatchNorm (BN) normalization method is used for normalization and the ReLU function is used for activation; finally, the Sigmoid function is used to activate the output of the third convolutional layer to generate the final predicted conductivity value σ pred ,Right now σ pred = Sigmoid(Conv(ReLU(BN(Conv(ReLU(BN(Conv(X)))))))).

10. A high-resolution reconstruction system for electrical impedance imaging based on deep learning, characterized in that: include: The dataset creation module is used to segment the acquired CT or MRI dataset into normal tissue areas and foreign matter areas, and to perform grid discretization on the segmentation results. The conductivity distribution and boundary voltage value data processing module is used to use the segmentation result after mesh discretization as a model, apply simulated electrodes to the model, import the normal tissue area into the model with the simulated electrodes applied, assign the normal tissue area's conductivity value, and assign other areas in the model other than the normal tissue area's conductivity values ​​to obtain a first conductivity distribution value σ in the current model. Then, current is applied and the finite element method is used to solve the boundary voltage value v0, and the sensitivity matrix S is calculated based on the boundary voltage value v0. Then, the foreign matter region is introduced into the model that already has a normal tissue region, and the foreign matter region is assigned a corresponding foreign matter region conductivity value to obtain a second conductivity distribution value σ′ in the current model. Then, the final boundary voltage value v is calculated based on the given conductivity distribution value σ′ and the sensitivity matrix S. The conductivity coarse reconstruction module is used to perform coarse conductivity reconstruction according to the boundary voltage value v through the Gauss-Newton method to obtain the coarse reconstructed conductivity σ k ; Conductivity accurate reconstruction module, used to reconstruct the boundary voltage v and the corresponding coarse reconstructed conductivity of the iteration As a sample, iterate N times the conductivity That is, N samples are obtained; the N samples are partially divided into a training set and partially divided into a test set; the model includes a conductivity branch and a structure branch; the conductivity branch is used to learn the mapping from the boundary voltage value v measured in step (2) to the conductivity distribution value σ; the structure branch is used to learn the coarsely reconstructed conductivity σ k Mapping to precise conductivity; The two feature maps learned from the conductivity branch and the structure branch are then merged; the merged features are then processed by convolution operation and finally activated using the Sigmoid function to generate the predicted conductivity σ pred as the final output.

Images