Adaptive multi-parameter feature fusion electrical impedance tomography method and related equipment

Abstract

The invention discloses a self-adaptive multi-parameter feature fusion electrical impedance tomography method and related equipment, and belongs to the field of electrical impedance tomography (EIT), an EIT voltage data set is obtained by constructing a finite element simulation model, an EIT initial image is obtained by solving an EIT image reconstruction inverse problem model, and the EIT voltage data set and the EIT initial image are obtained. And feature information related to the target is adaptively extracted from different regularization parameter imaging results through an adaptive multi-parameter feature fusion network, interference of irrelevant information is suppressed, and finally stable reconstruction of the EIT image is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of electrical impedance tomography, and specifically relates to an adaptive multi-parameter feature fusion electrical impedance tomography method and related equipment. Background Art

[0002] Electrical impedance tomography (EIT) is an advanced non-destructive electromagnetic imaging technology. It applies a safe current to the electrode array on the surface of the object being measured and collects the boundary voltage signal. It uses an image reconstruction algorithm to obtain the internal electrical characteristics distribution. With its unique advantages such as non-invasiveness, no radiation, low cost and continuous dynamic imaging, EIT has shown great application potential in medical and industrial fields such as lung function monitoring, stroke detection and multiphase flow measurement.

[0003] Image reconstruction of EIT is an inverse problem solving process of reconstructing the unknown electrical characteristic distribution based on the measured voltage. Unlike the solution of linear inverse problems such as CT and MRI, EIT faces greater challenges due to its inherent nonlinearity, ill-posedness and ill-posedness. A small perturbation in the measured voltage will cause a large deviation in the solution. Regularization algorithms stabilize the solution process by introducing additional constraints or information, and become an effective means to solve such ill-posed inverse problems. In the field of EIT, regularization methods such as Tikhonov regularization and total variation regularization (TV) have been widely used, effectively improving the quality of EIT imaging. However, the performance of regularization methods greatly depends on the selection of regularization parameters. If the regularization parameter is too large, it will lead to over-regularization of the solution, resulting in insufficient accuracy of the solution and a large error with the true solution. On the contrary, if the regularization parameter is set too small, the effect of regularization is small, resulting in an unstable solution. Although there are regularization parameter selection methods such as LCV and GCV, they take a long time to calculate and are difficult to converge in many cases. Therefore, most of the existing regularization parameters are still selected through empirical methods. How to select appropriate regularization parameters to achieve stable and high-quality imaging of EIT is still a key issue that needs to be solved urgently.

[0004] In recent years, deep learning-based EIT image reconstruction algorithms have been gradually proposed. Among them, deep learning EIT image post-processing algorithms that combine the interpretability of traditional EIT imaging algorithms with deep learning feature extraction capabilities have received extensive attention in the EIT field due to their excellent generalization and imaging quality. For example, an error-constrained network (EC-Net) designed by Jia et al. for image post-processing has achieved excellent image reconstruction performance and reconstruction speed. However, the imaging quality of the deep learning EIT image post-processing algorithm is highly dependent on the preliminary imaging results of the traditional algorithm. When the traditional algorithm fails to image due to incorrect selection of regularization parameters, the deep learning algorithm is also difficult to obtain good imaging results.

[0005] In summary, the imaging quality of existing EIT regularization imaging methods highly depends on the selection of regularization parameters. Currently, in the practical application of EIT, regularization parameters are still mostly selected by traditional methods, and the imaging quality is not stable. There is an urgent need for a new EIT imaging method that can achieve stable and high-quality imaging under different degrees of noise interference. Summary of the Invention

[0006] The present invention provides an adaptive multi-parameter feature fusion electrical impedance tomography method and related devices, which solves the problem that in the practical application of EIT, regularization parameters are still mostly selected by traditional methods and the imaging quality is not stable.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The adaptive multi-parameter feature fusion electrical impedance tomography method includes:

[0009] Construct a finite element simulation model, generate EIT measurement voltage data based on the finite element simulation model, and divide the EIT measurement voltage data into a training set, a validation set, and a test set;

[0010] Establish an inverse problem model for EIT image reconstruction using the relationship between the internal impedance change in the measured domain and the boundary measurement voltage, and solve the inverse problem model to obtain initial EIT imaging results with regularization parameters of different sizes;

[0011] Build an adaptive multi-parameter feature fusion network;

[0012] Use the training set and the validation set to train, optimize, and debug the constructed adaptive multi-parameter feature fusion network. During the training process, the adaptive multi-parameter feature fusion network extracts and fuses the feature information in multiple regularization parameter images from the initial EIT imaging results with regularization parameters of different sizes, and saves the optimal network parameters;

[0013] Deploy the network model according to the optimal network parameters, input the test set data into the adaptive multi-parameter feature fusion network, and through feature extraction and fusion, obtain a reconstructed image of the impedance distribution within the measured domain.

[0014] Preferably, the adaptive multi-parameter feature fusion network includes a multi-parameter feature extraction module, a cross-parameter feature enhancement module, a multi-scale feature fusion module, and a decoder module.

[0015] Preferably, the multi-parameter feature extraction module is constructed by depthwise separable convolution, which is used to extract feature information from different regularized parameter images. The depthwise separable convolution decomposes the standard convolution into two parts: depth convolution and pointwise convolution. The depth convolution performs convolution operations independently on each input channel, while the pointwise convolution uses a 1×1 convolution kernel to combine the outputs of the depth convolution.

[0016] Preferably, the cross-parameter feature enhancement module includes a channel attention module, a spatial attention module, and a cross-parameter attention module, which are used to adaptively extract feature information related to the real target and suppress the interference of cluttered artifacts. The input feature map is sequentially weighted by the three attention features, so that it focuses on the feature information related to the target in the multi-parameter image.

[0017] Preferably, the channel attention module transforms the input feature map into channel attention features, and its formula is as follows:

[0018] M Channel (F) = Sig(MLP(MP(F)) + MLP(AP(F)))

[0019] where Sig represents the Sigmoid activation function, MLP represents the multi-layer perceptron, MP and AP respectively represent the max pooling and average pooling operations, where F is the input feature map, and M Channel (F) is the channel attention feature, and F is obtained by multiplying the input feature F and the channel attention feature 1。

[0020] The spatial attention module transforms the input feature map F1 into spatial attention features:

[0021] M Spatial (F1) = Sig(Conv[MP(F1); AP(F1)])

[0022] where Sig represents the Sigmoid activation function, Conv represents the convolution operation, F1 is, and M Spatial (F1) is the spatial attention feature.

[0023] Preferably, the cross-parameter attention module first performs max pooling and average pooling on the input feature map F2, then extracts features in the MLP, and sends them to the feature uncertainty module. In the feature uncertainty module, a low-rank Gaussian distribution N(μ, Σ) is used to simulate the correlation between parameter images. Here, μ is the mean, Σ is the covariance matrix, and F2 is obtained by multiplying the input feature F1 and the attention feature;

[0024] While approximating the Gaussian distribution, calculate the three components of the mean μ, covariance factor P, and covariance diagonal D that describe the distribution. The linear operation on the input T of the parameter feature uncertainty block is:

[0025] ω′ = W μ T, P′ = W P T, D′ = W D T

[0026] where W μ , W P , W D is the weight matrix for linear calculation, and reshaping the tensor ω′ gives the mean μ of N(μ, Σ), and the calculation formulas for the covariance matrix Σ are as follows:

[0027] ∑ = PP T + D

[0028] where P is the reshaped tensor of P′, D is the diagonal matrix with D′ as the diagonal, to obtain the mean and covariance matrix of the Gaussian distribution, then sampling a vector z from the Gaussian distribution N(μ, Σ), and further obtaining the cross - graph attention feature through Sigmoid activation and reshape:

[0029] M Cross_parameter = reshape(Sig(z)), z ∼ N(μ, ∑)

[0030] where M Cross_parameter is the cross - graph attention feature.

[0031] Preferably, the multi - scale feature fusion module extracts multi - scale spatial features using dilated convolutions with different dilation rates, and the formula for dilated convolution is as follows:

[0032]

[0033] where, in the case of two - dimensional signals, x is the input feature map, y is the output feature map, i is each position on the output feature map, r is the dilation rate, which determines the step size for sampling the input features, and w is the filter.

[0034] An adaptive multi - parameter feature fusion electrical impedance tomography system, comprising:

[0035] A dataset acquisition module: used to construct a finite - element simulation model, generate EIT measurement voltage data based on the finite - element simulation model, and divide the EIT measurement voltage data into a training set, a validation set, and a test set;

[0036] An imaging acquisition module: used to establish an inverse problem model for EIT image reconstruction using the relationship between the internal impedance change in the measured domain and the boundary measurement voltage, and solve the inverse problem model to obtain initial EIT imaging results with different magnitudes of regularization parameters;

[0037] A building module: used to build an adaptive multi - parameter feature fusion network;

[0038] Training module: It is used to train, optimize and debug the constructed adaptive multi-parameter feature fusion network by using the training set and the validation set. During the training process, the adaptive multi-parameter feature fusion network extracts and fuses the feature information in multiple regularized parameter images from the initial EIT imaging results with different sizes of regularized parameters, and saves the optimal network parameters.

[0039] Reconstruction module: It is used to deploy the network model according to the optimal network parameters, input the test set data into the adaptive multi-parameter feature fusion network, and obtain the reconstructed impedance distribution image within the measured domain through feature extraction and fusion.

[0040] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the adaptive multi-parameter feature fusion electrical impedance tomography method are implemented.

[0041] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the steps of the adaptive multi-parameter feature fusion electrical impedance tomography method are implemented.

[0042] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides an adaptive multi-parameter feature fusion electrical impedance tomography method. By constructing a finite element simulation model, an EIT voltage data set is obtained. By solving the EIT image reconstruction inverse problem model, an initial EIT image is obtained. Then, through the adaptive multi-parameter feature fusion network, the feature information related to the target is adaptively extracted from the imaging results of different regularized parameters, and the interference of irrelevant information is suppressed, and finally the stable reconstruction of the EIT image is realized. Description of the Drawings

[0043] Figure 1 It is a flow block diagram of the adaptive multi-parameter feature fusion electrical impedance tomography method of the present invention.

[0044] Figure 2 It is the three-dimensional finite element simulation model in Embodiment 1; in the figure: 1. Imaging target, 2. Background tissue, 3. Electrode array.

[0045] Figure 3 It is the specific network structure of the adaptive multi-parameter feature fusion network, including a multi-parameter feature extraction module, a cross-parameter feature enhancement module, a multi-scale feature fusion module, and a decoder module.

[0046] Figure 4 It is a schematic diagram of the cross-parameter feature enhancement module in the adaptive multi-parameter feature fusion network.

[0047] Figure 5It is the inverse problem imaging template used in the embodiments.

[0048] Figure 6 It is the data reconstruction result under a 90dB noise level, where the first column is the simulation model, the 2nd to 4th columns are the reconstruction results of the traditional Tikhonov regularization method, and the 5th column is the imaging result of the adaptive multi-parameter feature fusion network.

[0049] Figure 7 It is the data reconstruction result under a 30dB noise level, where the first column is the simulation model, the 2nd to 4th columns are the reconstruction results of the traditional Tikhonov regularization method, the 5th column is the imaging result of the EC-Net algorithm, and the 6th column is the imaging result of the adaptive multi-parameter feature fusion network.

[0050] Figure 8 It is the two-dimensional finite element simulation model in Embodiment 2; in the figure: 1. Imaging target, 2. Background tissue, 3. Electrode array.

[0051] Figure 9 It is the data reconstruction result of the two-dimensional simulation model and elliptical target, where the first column is the simulation model, the 2nd column is the reconstruction result of the total variation regularization method, the 3rd column is the imaging result of the EC-Net algorithm, and the 4th column is the imaging result of the adaptive multi-parameter feature fusion network.

[0052] Figure 10 It is the block diagram of the adaptive multi-parameter feature fusion electrical impedance tomography system of the present invention. Detailed implementation manners

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and illustrated herein can be arranged and designed in various different configurations.

[0054] Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0055] It should be noted that: similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0056] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0057] As Figure 1 shown, the present invention provides an adaptive multi-parameter feature fusion electrical impedance tomography method, including:

[0058] Construct a finite element simulation model, generate EIT measurement voltage data based on the finite element simulation model, and divide the EIT measurement voltage data into a training set, a validation set, and a test set;

[0059] Establish an inverse problem model for EIT image reconstruction by using the relationship between the internal impedance change in the measured domain and the boundary measurement voltage, and solve the inverse problem model to obtain initial EIT imaging results with different magnitudes of regularization parameters;

[0060] Build an adaptive multi-parameter feature fusion network;

[0061] Use the training set and the validation set to train, optimize, and debug the built adaptive multi-parameter feature fusion network. During the training process, the adaptive multi-parameter feature fusion network extracts and fuses the feature information in multiple regularization parameter images from the initial EIT imaging results with different magnitudes of regularization parameters, and saves the optimal network parameters;

[0062] Deploy the network model according to the optimal network parameters, input the test set data into the adaptive multi-parameter feature fusion network, and obtain a reconstructed image of the impedance distribution in the measured domain through feature extraction and fusion.

[0063] The adaptive multi-parameter feature fusion network mainly consists of a multi-parameter feature extraction module, a cross-parameter feature enhancement module, a multi-scale feature fusion module, and a decoder module, specifically including:

[0064] The multi-parameter feature extraction module is built by depthwise separable convolution and is used to extract feature information from different regularization parameter images. The depthwise separable convolution decomposes the standard convolution into a depthwise convolution and a pointwise convolution. The depthwise convolution performs convolution operations independently on each input channel, while the pointwise convolution uses a 1×1 convolution kernel to combine the output of the depthwise convolution.

[0065] The cross-parameter feature enhancement module consists of a channel attention, a spatial attention, and a cross-parameter attention module, and is used to adaptively extract feature information related to the real target and suppress the interference of clutter artifacts. The input feature map F is sequentially weighted by three kinds of attention features, so that it focuses on the feature information related to the target in the multi-parameter image.

[0066] The channel attention module transforms the input feature map F into a channel attention feature, and its formula is as follows:

[0067] M Channel (F) = Sig(MLP(MP(F)) + MLP(AP(F)))

[0068] Among them, Sig represents the Sigmoid activation function, MLP represents the multi-layer perceptron, and MP and AP represent the max pooling and average pooling operations respectively.

[0069] The spatial attention module emphasizes the importance of position information in the image. This module transforms the input feature map F1 into spatial attention features:

[0070] M Spatial (F1) = Sig(Conv[MP(F1); AP(F1)])

[0071] Among them, Sig represents the Sigmoid activation function, and Conv represents the convolution operation.

[0072] The cross-parameter attention module first performs max pooling and average pooling on the input feature map, then extracts features in the MLP, and sends them to the feature uncertainty module. In this module, the low-rank Gaussian distribution N(μ, Σ) is used to simulate the correlation between parameter images. Among them, μ is the mean value, and Σ is the covariance matrix. While approximating the Gaussian distribution, the three components of the mean μ, covariance factor P, and covariance diagonal D that describe this distribution are calculated. The following linear operations are performed on the input T of the parameter feature uncertainty block:

[0073] μ′ = W μ T, P′ = W P T, D′ = W D T

[0074] Among them, W μ , W P , W D are the weight matrices for linear calculation. The reshaped tensor μ′ is used to obtain the mean μ of N(μ, Σ). The calculation formula for the covariance matrix Σ is as follows:

[0075] ∑ = PP T + D

[0076] Among them, P is the reshaped tensor of P′, and D is the diagonal matrix with D′ as the diagonal. Subsequently, a vector z is sampled from the Gaussian distribution N(μ, Σ), and further through Sigmoid activation and reshape, the cross-image attention features are obtained:

[0077] M Cross_parameter = reshape(Sig(z)), z ∼ N(μ, ∑)

[0078] Multi-scale feature fusion module:

[0079] The multi-scale feature fusion module uses dilated convolutions with different dilation rates to extract multi-scale spatial features. The formula for dilated convolution is as follows:

[0080]

[0081] Where, in the case of two-dimensional signals, x is the input feature map, y is the output feature map, i is each position on the output feature map, r is the dilation rate, which determines the step size of sampling the input features, and w is the filter. Compared with ordinary convolution, dilated convolution expands the receptive field from k to k+(k - 1)(r - 1), achieving control over the receptive field of the filter.

[0082] Decoder module:

[0083] The decoder module uses ordinary convolution to fuse multi-parameter feature information and reconstruct EIT images based on the information.

[0084] (1) Embodiment 1

[0085] Step 1: Construct a finite element simulation model. By changing parameters such as the target conductivity, size, and position in the simulation model, solve the EIT forward problem to generate a large amount of EIT measured voltage data, and divide it into a training set, a validation set, and a test set;

[0086] As Figure 2 shown, the established three-dimensional finite element simulation model includes an imaging target 1, a background tissue 2, and a 16-electrode array 3. The conductivity of the imaging target is set to 1 - 25 S / m, and the conductivity of the background tissue is set to 1 S / m. By changing the conductivity, position, and size of the imaging target and solving the EIT forward problem, a large number of EIT measured voltage data samples are obtained, and the data samples are divided into a training set, a validation set, and a test set. The EIT forward problem model is expressed as:

[0087] ΔV = SΔσ

[0088] Where, Δσ is the conductivity change inside the measured domain, ΔV is the boundary voltage data, and S represents the sensitivity matrix.

[0089] Step 2: Establish an inverse problem model for EIT image reconstruction using the relationship between the internal impedance change in the measured domain and the boundary measured voltage, and convert the measured voltage data into a reconstructed image. The inverse problem mathematical model is as follows:

[0090]

[0091] In the formula, ΔV is the voltage change value at two moments t1 and t2, For the reconstructed conductivity change, λ is the regularization parameter, R represents the regularization matrix, and the Tikhonov regularization method is used for image reconstruction. Here, R is the identity matrix. Images are reconstructed using three regularization parameters: 0.1, 0.01, and 0.001. The image reconstruction meshing template is as Figure 3 shown.

[0092] Step 3: As Figure 4 shown, first, the measured voltage data is converted into initial imaging results with different magnitudes of regularization parameters through inverse problem solving. Subsequently, the initial imaging results are fed into the constructed adaptive multi-parameter feature fusion network for adaptive extraction and fusion of image features with different regularization parameters. This network mainly consists of a multi-parameter feature extraction module, a cross-parameter feature enhancement module, a multi-scale feature fusion module, and a decoder module. The multi-parameter feature extraction module extracts spatial feature information from the initial imaging results with different magnitudes of regularization parameters; the cross-parameter feature enhancement module enables the network to focus on the characteristic information related to the imaging target in the imaging results with different parameters; the multi-scale feature fusion module enhances the multi-scale feature extraction ability of the network; and the decoder module realizes the fusion of features and image reconstruction.

[0093] Among them, the encoder of the multi-parameter feature extraction module is constructed using depthwise separable convolutions, and the decoder is constructed using ordinary convolutions.

[0094] As Figure 5 shown, the cross-parameter feature enhancement module consists of a channel attention module, a spatial attention module, and a cross-parameter attention module. The input feature map F passes through three types of attention feature weightings in sequence, enabling it to focus on the feature information related to the target in the multi-parameter images.

[0095] The channel attention module converts the input feature map F into channel attention features, and its formula is as follows:

[0096] M channel (F) = Sig(MLP(MP(F)) + MLP(AP(F)))

[0097] Among them, Sig represents the Sigmoid activation function, MLP represents the multi-layer perceptron, and MP and AP represent the max pooling and average pooling operations, respectively.

[0098] The spatial attention module emphasizes the importance of position information in the image. This module converts the input feature map F1 into spatial attention features:

[0099] M Spatial (F1) = Sig(Conv[MP(F1); AP(F1)])

[0100] Among them, Sig represents the Sigmoid activation function, and Conv represents the convolution operation.

[0101] The cross-parameter attention module first performs max pooling and average pooling on the input feature map, then extracts features in the MLP, and sends them to the feature uncertainty module. In this module, a low-rank Gaussian distribution N(μ,Σ) is used to simulate the correlation between parameter images, where μ is the mean and Σ is the covariance matrix. While approximating the Gaussian distribution, three components, namely the mean μ, covariance factor P, and covariance diagonal D, that describe the distribution are calculated. The following linear operations are performed on the input T of the parameter feature uncertainty block:

[0102] μ′ = W μ T, P′ = W P T, D′ = W D T

[0103] where W μ , W P , W D are the weight matrices for linear calculations, and reshaping the tensor μ′ gives the mean μ of N(μ,Σ). The calculation formula for the covariance matrix Σ is as follows:

[0104] ∑ = PP T + D

[0105] where P is the reshaped tensor of P′ and D is the diagonal matrix with D′ as the diagonal. Subsequently, a vector z is sampled from the Gaussian distribution N(μ,Σ), and through further Sigmoid activation and reshape, the cross-map attention feature is obtained:

[0106] M Cross_parameter = reshape(Sig(z)), z ∼ N(μ,∑)

[0107] The multi-scale feature fusion module uses dilated convolutions with different dilation rates to enhance the multi-scale feature extraction ability of the network; the formula for dilated convolution is as follows:

[0108]

[0109] where, in the case of two-dimensional signals, x is the input feature map, y is the output feature map, i is each position on the output feature map, r is the dilation rate, which determines the step size for sampling the input features, and w is the filter. Compared with ordinary convolution, dilated convolution expands the receptive field from k to k+(k - 1)(r - 1), achieving control over the receptive field of the filter.

[0110] Step 4: Use the training set and validation set data to train, optimize, and debug the constructed adaptive multi-parameter feature fusion network. Select the Adam algorithm as the optimization algorithm, and set the initial learning rate of the algorithm to 10 -4, the batch size is set to 8, and a total of 100 epochs are trained. We use the Focal loss as the loss function, and its calculation formula is as follows:

[0111]

[0112] In the formula, L Focal is the Focal loss function, n is the number of pixel points, and σ i is the label value of the i-th pixel point, is the corresponding predicted value of the pixel point, α is taken as 0.5, and γ is taken as 2.

[0113] Step 5: Deploy the network model according to the optimal network parameters obtained by training, input the test set data into the trained adaptive multi-parameter feature fusion network, and obtain the impedance distribution reconstruction image in the measured domain through cross-parameter feature extraction and fusion of the reconstructed images under different regularization parameters.

[0114] (2) Comparative Example 1

[0115] Generate data according to the finite element simulation model described in (1), and compare the currently commonly used traditional Tikhonov regularization method, the EC-Net image post-processing method, and the adaptive multi-parameter feature fusion network imaging method proposed in the present invention. The imaging results are as Figure 6 、 Figure 7 shown. The first column is the real model setting, the 2-4th columns are the imaging results of the Tikhonov regularization method, the 5th column is the result of the EC-Net image post-processing method, and the 6th column is the imaging result of the adaptive multi-parameter feature fusion network imaging method.

[0116] First, as Figure 6As shown, in the weak noise data at a signal-to-noise ratio level of 90 dB, for the traditional Tikhonov regularization method, when the regularization parameter is relatively large at 0.1, the reconstructed image has serious artifacts. When the regularization parameter is 0.001, the accuracy of the reconstructed target is improved, but there is still a certain error. The EC-Net image post-processing method takes the imaging result with a regularization parameter of 0.001 as the input, effectively removes the image artifacts, and improves the image quality. However, under strong noise interference at a signal-to-noise ratio level of 30 dB, the Tikhonov regularization method is completely unable to image the target when the regularization parameter is small. When the Tikhonov method fails to image, the EC-Net image post-processing method is also unable to image the target because it cannot obtain effective information from a single image. When the regularization parameters are set to 0.1 and 0.01, although the Tikhonov regularization method can reconstruct the target, the reconstruction result has a large error and it is difficult to reflect the true target distribution. In contrast, the adaptive multi-parameter feature fusion network imaging method proposed by the present invention adaptively extracts and fuses the effective information in the imaging results of different regularization parameters, and obtains high-quality imaging results at different noise levels. The reconstructed target is closer to the real model.

[0117] (3) Example 2

[0118] Step 1: As shown in Figure 8 , construct a two-dimensional finite element simulation model, including an imaging target 1, a background tissue 2, and an electrode array 3. The target shape is set to an ellipse. By changing parameters such as the conductivity, size, and position of the target, solve the EIT forward problem to generate a large amount of EIT measured voltage data, and divide it into a training set, a validation set, and a test set;

[0119] Step 2: Use the relationship between the internal impedance change in the measured domain and the boundary measured voltage to establish an inverse problem model for EIT image reconstruction, and convert the measured voltage data into a reconstructed image. Here, we use the total variation regularization method for image reconstruction and construct the following mathematical model:

[0120]

[0121] In the formula, ΔV is the voltage change value at two time instants t1 and t2, is the reconstructed conductivity change, λ is the regularization parameter. Here, the L-curve method is used to select the regularization parameter, and three relatively optimal parameters are selected to reconstruct the image.

[0122] Step 3: Similar to Example 1, convert the measured voltage data into initial imaging results with different regularization parameters through inverse problem solving. Subsequently, send the initial imaging results into the constructed adaptive multi-parameter feature fusion network for adaptive extraction and fusion of image features with different regularization parameters.

[0123] Step 4: Use the training set and validation set data to train, optimize, and debug the constructed adaptive multi-parameter feature fusion network. Select the SGD algorithm as the optimization algorithm, and set the initial learning rate of the algorithm to 10 -4 , set the Batch size to 8, and train for a total of 100 epochs. Here, we use the Dice loss as the loss function, and its calculation formula is as follows:

[0124]

[0125] In the formula, L Dice is the Dice loss function, n is the number of pixel points, σ i is the label value of the i-th pixel point, is the corresponding pixel point prediction value, and β is 1.

[0126] Step 5: Deploy the network model according to the optimal network parameters obtained from training. Input the test set data into the trained adaptive multi-parameter feature fusion network, and through cross-parameter feature extraction and fusion of the reconstructed images under different regularization parameters, obtain the impedance distribution reconstruction image within the measured domain.

[0127] (Four) Control Example 2

[0128] Generate data according to the two-dimensional finite element simulation model described in (Three). The imaging target is an elliptical target. Compare the total variation regularization method, the EC-Net image post-processing method, and the adaptive multi-parameter feature fusion network imaging method proposed by the present invention. The imaging results are as Figure 9 . The first column is the real model setting, the second column is the imaging result of the total variation regularization method, the third column is the result of the EC-Net image post-processing method, and the fourth column is the imaging result of the adaptive multi-parameter feature fusion network imaging method. From the imaging results, it can be seen that the adaptive multi-parameter feature fusion network imaging method proposed by the present invention has good generalization and is applicable to different regularization methods. In the case of a two-dimensional simulation model and an elliptical target, this imaging method also achieves high-quality target imaging.

[0129] As Figure 10 shown, the present invention provides an adaptive multi-parameter feature fusion electrical impedance tomography system, including:

[0130] Dataset acquisition module: used to construct a finite element simulation model, generate EIT measurement voltage data based on the finite element simulation model, and divide the EIT measurement voltage data into a training set, a validation set, and a test set;

[0131] Imaging acquisition module: It is used to establish an inverse problem model of EIT image reconstruction by using the relationship between the internal impedance change in the measured domain and the boundary measurement voltage, and solve the inverse problem model to obtain the initial EIT imaging results with different regularization parameters of different sizes;

[0132] Building module: It is used to build an adaptive multi-parameter feature fusion network;

[0133] Training module: It is used to train, optimize and debug the built adaptive multi-parameter feature fusion network by using the training set and the validation set. During the training process, the adaptive multi-parameter feature fusion network extracts and fuses the feature information in multiple regularization parameter images from the initial EIT imaging results with different regularization parameters of different sizes, and saves the optimal network parameters;

[0134] Reconstruction module: It is used to deploy the network model according to the optimal network parameters, input the test set data into the adaptive multi-parameter feature fusion network, and obtain the reconstructed impedance distribution image in the measured domain through feature extraction and fusion.

[0135] The terminal device provided by an embodiment of the present invention. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in each of the above method embodiments. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in each of the above device embodiments.

[0136] The computer program can be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor to complete the present invention.

[0137] The terminal device can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The terminal device may include, but is not limited to, a processor and a memory.

[0138] The processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0139] The memory can be used to store the computer program and / or module. By running or executing the computer program and / or module stored in the memory, and by invoking the data stored in the memory, the processor implements various functions of the terminal device.

[0140] If the module / unit integrated in the terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention, it can also be completed by a computer program instructing relevant hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0141] Although the embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the above specific embodiments and application fields. The above specific embodiments are merely illustrative and guiding, rather than restrictive. Those of ordinary skill in the art can also make many forms without departing from the scope protected by the claims of the present invention under the inspiration of the specification, and all of these fall within the scope of protection of the present invention.

Claims

1. An adaptive multi-parameter feature fusion electrical impedance tomography method, characterized in that: include: Construct a finite element simulation model, generate EIT measurement voltage data based on the finite element simulation model, and divide the EIT measurement voltage data into a training set, a validation set, and a test set; The inverse problem model of EIT image reconstruction is established by using the relationship between the internal impedance change of the measured domain and the boundary measurement voltage. The initial EIT imaging results with regularization parameters of different sizes are obtained by solving the inverse problem model. Build an adaptive multi-parameter feature fusion network; The adaptive multi-parameter feature fusion network is trained, optimized and debugged using the training set and the validation set. During the training process, the adaptive multi-parameter feature fusion network extracts and fuses feature information from multiple regularization parameter images from the initial EIT imaging results of regularization parameters of different sizes, and saves the optimal network parameters. The network model is deployed according to the optimal network parameters, and the test set data is input into the adaptive multi-parameter feature fusion network. Through feature extraction and fusion, the reconstructed image of the impedance distribution in the measured domain is obtained.

2. The adaptive multi-parameter feature fusion electrical impedance tomography method according to claim 1, characterized in that: The adaptive multi-parameter feature fusion network includes a multi-parameter feature extraction module, a cross-parameter feature enhancement module, a multi-scale feature fusion module and a decoder module.

3. The adaptive multi-parameter feature fusion electrical impedance tomography method according to claim 2, characterized in that: The multi-parameter feature extraction module is built by depthwise separable convolution, which is used to extract feature information from images with different regularization parameters. The depthwise separable convolution decomposes the standard convolution into two parts: depthwise convolution and pointwise convolution. The depthwise convolution performs convolution operations independently on each input channel, while the pointwise convolution uses a 1×1 convolution kernel to combine the outputs of the depthwise convolution.

4. The adaptive multi-parameter feature fusion electrical impedance tomography method according to claim 3, characterized in that: The cross-parameter feature enhancement module includes channel attention, spatial attention and cross-parameter attention modules, which are used to adaptively extract feature information related to the real target and suppress the interference of cluttered artifacts. The input feature map is weighted by three types of attention features in turn to focus on the feature information related to the target in the multi-parameter image.

5. The adaptive multi-parameter feature fusion electrical impedance tomography method according to claim 4, characterized in that: The channel attention module converts the input feature map into channel attention features, and its formula is as follows: M Channel (F)=Sig(MLP(MP(F))+MLP(AP(F))) Among them, Sig represents the Sigmoid activation function, MLP represents the multi-layer perceptron, MP and AP represent the maximum pooling and average pooling operations respectively, where F is the input feature map, M Channel (F) is the channel attention feature, and F1 is obtained by multiplying the input feature F and the channel attention feature; The spatial attention module transforms the input feature map F1 into spatial attention features: M Spatial (F1)=Sig(Conv[MP(F1);AP(F1)]) Among them, Sig represents the Sigmoid activation function, Conv represents the convolution operation, F1 is, M Spatial (F1) is the spatial attention feature.

6. The adaptive multi-parameter feature fusion electrical impedance tomography method according to claim 5, characterized in that: The cross-parameter attention module first performs maximum pooling and average pooling on the input feature map F2, then extracts features in the MLP and sends them to the feature uncertainty module. In the feature uncertainty module, a low-rank Gaussian distribution N(μ,Σ) is used to simulate the correlation between parameter images, where μ is the mean and Σ is the covariance matrix. F2 is obtained by multiplying the input feature F1 and the attention feature. While approximating the Gaussian distribution, the mean μ, covariance factor P, and covariance diagonal D describing the distribution are calculated, and the linear operation is performed on the input T of the parameter feature uncertainty block as follows: μ′=W μ T,P′=W P T,D′=W D T Among them, W μ ,W P ,W D is the weight matrix for linear calculation, reshape the tensor μ′ to get the mean μ of N(μ,Σ), and the calculation formula of the covariance matrix Σ is as follows: ∑=PP T +D Among them, P is the reshaped tensor of P′, D is the diagonal matrix with D′ as the diagonal, and the mean and covariance matrix of the Gaussian distribution are obtained. Then, the vector z is sampled from the Gaussian distribution N(μ,Σ), and the cross-graph attention feature is further obtained through Sigmoid activation and reshape: M Cross_parameter =reshape(Sig(z)),z~N(μ,∑) Among them, M Cross_parameter is the cross-graph attention feature.

7. The adaptive multi-parameter feature fusion electrical impedance tomography method according to claim 3, characterized in that: The multi-scale feature fusion module uses dilated convolutions with different dilation rates to extract multi-scale spatial features. The formula for dilated convolution is as follows: Among them, in the case of a two-dimensional signal, x is the input feature map, y is the output feature map, i is each position on the output feature map, r is the dilation rate, which determines the step size of our sampling of the input features, and w is the filter.

8. Adaptive multi-parameter feature fusion electrical impedance tomography system, characterized in that: include: Dataset acquisition module: used to build a finite element simulation model, generate EIT measurement voltage data based on the finite element simulation model, and divide the EIT measurement voltage data into a training set, a validation set, and a test set; Imaging acquisition module: used to establish an inverse problem model of EIT image reconstruction by using the relationship between the internal impedance change of the measured domain and the boundary measurement voltage, and solve the inverse problem model to obtain initial EIT imaging results with regularization parameters of different sizes; Building module: used to build an adaptive multi-parameter feature fusion network; Training module: used to train, optimize and debug the built adaptive multi-parameter feature fusion network using the training set and the validation set. During the training process, the adaptive multi-parameter feature fusion network extracts and fuses feature information from multiple regularization parameter images from the initial EIT imaging results of regularization parameters of different sizes, and saves the optimal network parameters. Reconstruction module: It is used to deploy the network model according to the optimal network parameters, input the test set data into the adaptive multi-parameter feature fusion network, and obtain the impedance distribution reconstructed image in the measured domain through feature extraction and fusion.

9. 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, the steps of the adaptive multi-parameter feature fusion electrical impedance tomography method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the adaptive multi-parameter feature fusion electrical impedance tomography method according to any one of claims 1 to 7 are implemented.

Images