A real-time reconstruction method for electrical impedance tomography based on 1D-UNETR and inverse finite element projection

Abstract

This invention relates to a real-time reconstruction method for electrical impedance tomography based on 1D-UNETR and inverse finite element projection, belonging to the field of electrical impedance tomography technology. This method obtains the time difference voltage sequence between the current moment and the background reference field through boundary electrodes, extracts multi-scale deep features of the voltage signal using the proposed 1D-UNETR architecture, and directly maps the high-dimensional feature manifold to the finite element mesh physical space by combining a depth inverse finite element projector, thereby directly reconstructing the conductivity perturbation distribution of foreign objects in the field domain, realizing high-precision real-time perception of the target's position, size, and shape, reducing the model size, and improving the image reconstruction speed.

Description

Technical Field

[0001] This invention belongs to the field of electrical impedance tomography technology and relates to a real-time reconstruction method for electrical impedance imaging based on 1D-UNETR and inverse finite element projection. Background Technology

[0002] Electrical impedance tomography (EIT) aims to non-invasively reconstruct the conductivity distribution of the internal medium by applying low-frequency AC excitation to the boundary of an object and measuring the response voltage. It has become an important development direction in modern medical imaging and industrial flaw detection.

[0003] In solving the EIT inverse problem, the accuracy and real-time performance of image reconstruction directly determine its application value. However, limited by the "soft field" effect and the limitations of boundary measurements, EIT imaging is a typical highly ill-conditioned and nonlinear inverse problem. Conventional model-driven algorithms, such as Tikhonov regularization and Landweber iteration, while having a physical basis, often face challenges such as low spatial resolution, severe artifacts, and excessively long iterative computation time when dealing with complex nonlinear relationships, making it difficult to meet the real-time monitoring requirements of high-speed dynamic processes.

[0004] In recent years, although deep learning technology has gradually become mainstream, existing mainstream methods generally suffer from the defects of "dimensional inflation" and "visual-physical mismatch." These methods forcibly map low-dimensional voltage signals to a high-dimensional pixel grid, resulting in redundant computational power and making it difficult to deploy on resource-constrained portable or edge devices. More seriously, pixel-based decoders tend to generate visually smooth images, often masking the true physical gradients of foreign object edges, resulting in reconstruction results lacking fidelity in physical values, which has significant limitations for real-time imaging requirements in high-speed dynamic application scenarios such as heartbeats and high-speed industrial fluids. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a real-time reconstruction method for electrical impedance imaging based on 1D-UNETR and inverse finite element projection. In the process of non-invasive monitoring of the measured field, the method relies on boundary electrodes to collect time-difference voltage sequences reflecting internal physical disturbances, uses a hybrid spectrum-space encoder to extract multi-scale features of the signal, and combines a depth inverse finite element projector to directly map the feature manifold to the finite element mesh physical space, thereby realizing end-to-end, high physical fidelity real-time positioning and reconstruction of the conductivity disturbance distribution of foreign objects in the field.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A real-time reconstruction method for electrical impedance imaging based on 1D-UNETR and inverse finite element projection includes the following steps: S1: Construct a dataset containing multiple sets of electrical impedance imaging samples using a finite element forward simulation solver. For each set of samples in the dataset, obtain the time-difference voltage sequence reflecting the physical disturbances within the field. The conductivity perturbation vector is obtained by considering the difference between the true conductivity distribution and the background conductivity distribution. Construct input and output data pairs It is used to supervise the training of reconstructed models to learn the nonlinear inverse mapping relationship from boundary differential voltage to the change in grid conductivity; S2: Construct a real-time reconstruction network for electrical impedance imaging based on 1D-UNETR and inverse finite element projection. The reconstruction network includes a spectral feature extractor, a global field coupling bottleneck layer, and feature reconstruction and inverse finite element projection. S3: Set a sparse adaptive physics loss function, utilizing input and output data to... Train and test the reconstructed network; S4: The trained reconstruction network is used to reconstruct electrical impedance imaging in real time to obtain the conductivity perturbation distribution of foreign objects in the field, thereby realizing real-time perception of the target's position, size, and shape.

[0007] Furthermore, step S1 specifically includes the following steps: A dataset containing M sets of samples was constructed using a finite element forward simulation solver, and it was divided into a training set, a validation set, and a test set. For each set of samples in the dataset, a true conductivity distribution containing the location, size, and numerical randomization of foreign objects was first randomly generated within the field of the finite element mesh. And define the standard uniform background conductivity distribution when there are no foreign objects. Subsequently, based on the adjacent excitation-adjacent measurement protocol, the boundary measurement voltage vectors containing foreign objects are calculated using a forward solver. and the reference voltage vector under the background field By calculating the difference between the measured voltage and the reference voltage, a time-difference voltage sequence reflecting the physical disturbances within the field is obtained. The sequence contains several independent measurement data points, which are used as the initial input features for the reconstruction network. The conductivity perturbation vector is obtained by calculating the difference between the true conductivity distribution and the background conductivity distribution. The dimension of this vector is equal to the total number of elements in the finite element mesh. This is used as the training label for reconstructing the network, thereby constructing... Input and output data pairs are used to supervise the training of the reconstructed model to learn the nonlinear inverse mapping relationship from the boundary differential voltage to the change in grid conductivity.

[0008] Furthermore, the real-time reconstruction network for electrical impedance imaging based on 1D-UNETR and inverse finite element projection aims to establish a direct nonlinear mapping from the measurement manifold to the physical property manifold. The mathematical mapping definition for the entire process is as follows:

[0009] in, It is the input differential voltage sequence , It is the predicted differential conductivity distribution of the output. , The total number of triangular subdivision elements in the corresponding finite element mesh; function , and These represent the local gradient extractor, the global field coupling modeler, and the inverse finite element projector, respectively. The composite function formed by these three components constitutes a complete nonlinear inverse operator from the observation space to the physical parameter space.

[0010] Furthermore, the spectral feature extractor extracts the differential voltage sequence Input is a spectral feature extractor consisting of multiple layers of one-dimensional convolutions. ;No. Convolutional operations utilize a limited receptive field to extract local gradient features and compress data dimensionality through downsampling. The computation process is described as follows:

[0011] In the formula This represents a convolution with a stride of 2; this operation reduces the sequence length while doubling the channel dimension, thereby converting the specific signal waveform into abstract features; finally, it undergoes cascaded processing through k=3 stages.

[0012] Furthermore, the global coupling bottleneck layer utilizes a global field coupling modeler. Using the extracted local features as input, a self-attention mechanism is employed to simulate the global sensitivity matrix under soft field effects, calculating the correlation between query Q, key K, and value V. The core attention calculation formula is as follows:

[0013] in, The scaling factor is used to capture the electric field coupling effect between long-distance electrodes and fuse it into global context features via the output projection matrix.

[0014] Furthermore, feature reconstruction and inverse finite element projection are performed through a decoder. Global features are upsampled through progressively transposed convolutions and then fused with skip connection features from the encoder. To compensate for the loss of spatial information; The feature restoration process of each decoding layer is shown in the following equation:

[0015] Decoded Feature flattening generation Input depth inverse finite element projector By utilizing a two-stage wide-body fully connected layer and a physical mapping layer, the feature vectors are directly projected onto the finite element mesh physical space to calculate the predicted value of the conductivity change. :

[0016] Predicted values The output predictions are directly mapped to the unstructured grid. On each subdivided unit, end-to-end physical field reconstruction is completed.

[0017] Furthermore, the sparse adaptive physical loss function is expressed as:

[0018] In the formula, This represents the total number of subdivided elements in the finite element mesh (FEM Grid). Indicates the first The true conductivity distribution of each grid cell, This indicates that the network outputs the first... The predicted conductivity value for each grid cell. This represents the absolute physical error on that grid cell; This represents the mean square error term; This is a balancing coefficient used to adjust the contribution ratio of the two types of errors; The dynamic equilibrium weights, generated based on the real physical distribution, are expressed as:

[0019] in, and These represent the number of mesh cells in the foreign object region and the background region, respectively. A threshold for distinguishing between background and target.

[0020] The beneficial effects of this invention are as follows: This invention discloses a real-time reconstruction method for electrical impedance imaging based on 1D-UNETR and inverse finite element projection, which solves the problems of dimensional inflation and visual-physical mismatch in existing deep learning methods. Through architectural innovation, paradigm innovation and physical constraints, it achieves significant advantages such as lightweight model, fast inference speed and high physical inversion accuracy. It can effectively solve the real-time and high-fidelity imaging requirements in high-speed dynamic scenarios such as heartbeat and high-speed industrial fluids, and has broad application prospects and practical value.

[0021] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the real-time reconstruction method of electrical impedance tomography according to the present invention; Figure 2 The flowchart shows a real-time reconstruction method for electrical impedance imaging based on 1D-UNETR and inverse finite element projection according to the present invention. Figure 3 This is a structural diagram of the real-time reconstruction algorithm for electrical impedance imaging of the present invention; Figure 4 This is a structural diagram of the global field coupling bottleneck layer of the present invention; Figure 5 This is a graph showing the loss values ​​generated during the iterative model generation of this invention. Figure 6 This is a diagram illustrating the impedance tomography image reconstruction of the present invention. Detailed Implementation

[0023] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0024] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0025] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0026] Example 1: To address the shortcomings of existing deep learning EIT imaging technologies, such as dimensional inflation and visual-physical mismatch, this invention proposes a real-time electrical impedance tomography reconstruction method based on 1D-UNETR and inverse finite element projection. During non-invasive monitoring of the measured field, time-difference voltage sequences reflecting internal physical disturbances are acquired by boundary electrodes. Multi-scale features of the signal are extracted using a hybrid spectrum-space encoder. Combined with a deep inverse finite element projector, the feature manifold is directly mapped to the finite element mesh physical space, achieving end-to-end, high-physical-fidelity real-time localization and reconstruction of the conductivity disturbance distribution of foreign objects in the field.

[0027] This problem is highly nonlinear and ill-posed, and traditional pixelation methods face the challenge of redundant computational resources. This invention employs a direct nonlinear inversion strategy of "vector to mesh," performing calculations in two dimensions: feature extraction and manifold projection. Specifically, by eliminating the intermediate step of visual image generation, a direct mapping operator is constructed from the low-dimensional measurement space to the attribute space of the unstructured finite element mesh. This utilizes physical topology priors to constrain the solution space, thereby significantly reducing the number of model parameters and computational latency while fundamentally eliminating interpolation errors and visual illusions caused by mesh-pixel conversion, thus improving the accuracy and stability of the reconstructed solution in terms of physical numerical values.

[0028] Architecture and its mapping principles are as follows Figure 1 As shown. This invention discretizes the imaging field into an unstructured finite element mesh to reflect the time-difference voltage vector of internal foreign object disturbances. Using this as input, an end-to-end nonlinear inverse mapping relationship was constructed from the boundary measurement space to the finite element physical property space. This scheme directly inverts the change in conductivity value at the mesh topology level. It replaces the traditional pixel-level image decoding and ultimately obtains an unstructured grid reconstruction distribution that satisfies the physical field equation constraints and has numerical fidelity.

[0029] like Figure 2 As shown, this method includes the following steps: 1. Dataset Construction: The first step is the data generation and preprocessing stage, which utilizes a finite element forward simulation solver to construct data containing... A dedicated dataset of samples was created and strictly divided into training, validation, and test sets in an 8:1:1 ratio. For each sample in the dataset, a true conductivity distribution containing the location, size, and numerical randomization of foreign objects was first generated randomly within the finite element mesh domain. And define the standard uniform background conductivity distribution when there are no foreign objects. Subsequently, based on the adjacent excitation-adjacent measurement protocol, the boundary measurement voltage vectors containing foreign objects were calculated using a forward solver. and the reference voltage vector under the background field Based on this, by calculating the difference between the measured voltage and the reference voltage, a time-difference voltage sequence reflecting the physical disturbances within the field is obtained. The sequence contains a total of Each independent measurement data point is used as the initial input feature for the reconstructed network; simultaneously, the difference between the true conductivity distribution and the background conductivity distribution is calculated to obtain the conductivity perturbation vector. The dimension of this vector is equal to the total number of elements in the finite element mesh. This is used as the training label for reconstructing the network, thereby constructing... Input and output data pairs are used to supervise the training of the reconstructed model to learn the nonlinear inverse mapping relationship from the boundary differential voltage to the change in grid conductivity.

[0030] 2. Construction of 1D-UNETR and Inverse Finite Element Projection Network 2.1 Physics-Driven Architecture Design The inverse problem of impedance imaging is essentially the process of inferring internal dielectric parameters from boundary measurements. Its physical field exhibits a significant duality: on the one hand, measurements from adjacent electrodes show strong local correlation; on the other hand, due to soft-field effects, small perturbations in local conductivity instantaneously couple to the entire boundary, exhibiting significant global nonlocality. Traditional pure convolutional networks, limited by their local receptive field, struggle to capture this long-distance coupling; while pure Transformer networks, though adept at global modeling, lack inductive bias for local high-frequency boundary signals. To effectively model this physical duality, this invention proposes a hybrid 1D-UNETR architecture, such as... Figure 3 As shown.

[0031] Unlike previous methods based on the 2D image domain, this architecture aims to establish a direct nonlinear mapping from the measurement manifold to the physical property manifold.

[0032] The mathematical mapping for the entire process is defined as follows:

[0033] in, It is the input differential voltage sequence , It is the predicted differential conductivity distribution of the output. ,in The total number of triangular subdivision elements in the corresponding finite element mesh. (Function) , and These represent the local gradient extractor, the global field coupling modeler, and the inverse finite element projector, respectively. This composite function constitutes a complete nonlinear inverse operator from the observation space to the physical parameter space.

[0034] 2.2 Multi-scale spectral feature extraction encoder The invention aims to extract spectral features reflecting local abrupt changes at the boundary. Since high-frequency fluctuations in the EIT boundary differential voltage sequence typically correspond to the location and size information of internal foreign objects, a 1D convolutional network with a limited receptive field is employed as a learnable local gradient extractor. To ensure the model has sufficient feature capacity to fit highly nonlinear inverse mappings, the base channel width is set to C=96, a design that doubles the feature representation capability compared to traditional architectures.

[0035] like Figure 3 As shown, this module contains k=3 levels. For the k-th level, the processing flow is formulated into the following two steps: (1) Local gradient feature extraction: For the input of stage k This module uses stacked convolutional blocks to mine high-order gradient features. First, a preliminary feature transformation is performed as shown in the equation. Then, deep feature refinement is performed to obtain the feature sequence for this stage, as shown in the equation.

[0036]

[0037]

[0038] In the formula, This represents a one-dimensional convolution operator with a kernel size of 3. For batch normalization, The ReLU activation function is used. The first convolutional layer captures the basic gradient pattern, and the second convolutional layer further refines the feature response, focusing on the weak voltage perturbations caused by abrupt changes in internal conductivity.

[0039] (2) Spectral domain downsampling: In order to expand the receptive field and extract abstract semantics, the invention abandons the max pooling operation that may cause loss of high-frequency information, and instead uses stride convolution for downsampling, as shown in the following mathematical expression.

[0040]

[0041] In the formula This represents a convolution with a stride of 2. This operation reduces the sequence length while doubling the channel dimension, thereby converting the specific signal waveform into abstract "spectral" features. Finally, after k=3 stages of cascaded processing, the encoder outputs the deep features as shown in the following equation.

[0042]

[0043] 2.3 Global Field Coupling Bottleneck Layer To overcome the inherent limitation of convolution operators in capturing EIT soft field effects due to their local receptive field, a global field coupling bottleneck layer is introduced at the end of the encoder. Its structure is as follows Figure 4 As shown. This module aims to explicitly simulate the global nonlocality of the potential distribution across sequence distances. The module consists of an N=2 stack. For the... l The layer calculation process is formalized in this invention into the following four key steps: (1) Spatial location coding: Since the self-attention mechanism is inherently permutation invariant, a learnable position encoding parameter is introduced to annotate the physical geometry information of the electrodes. For the encoder's output characteristics Initial sequence input It is defined as shown in the following formula.

[0044]

[0045] In the formula, the position encoding parameter , The introduction of this method breaks the symmetry of the sequence, enabling the model to distinguish between the relative and absolute positions of the electrodes, thus correctly preserving the toroidal spatial topology of the boundary measurements.

[0046] (2) Multi-head self-attention: MHSA is the core mechanism for capturing global coupling. To capture the dependencies between electrodes in parallel from different representation subspaces, the invention sets the number of attention heads to [value missing]. For the first l Layer input First, it is generated through linear projection. The group query, key, and value matrix is ​​shown in the following formula.

[0047]

[0048] In the formula , For the first iThe learnable projective weight matrix corresponds to each head. The formula for calculating single-head attention is shown below.

[0049]

[0050] In the formula, For each head, there are feature dimensions, and Softmax is the normalization matrix.

[0051] Subsequently, the multi-head mechanism concatenates the outputs of all heads along the channel dimension and outputs them through an output matrix. The linear fusion is performed as shown in the following equation.

[0052]

[0053] In this invention, the attention matrix has a clear physical meaning; it acts as a data-driven global impedance coupling map, quantizing the first... The measurement point for the first... The nonlocal effects of these potential features enable the network to explicitly simulate the global sensitivity characteristics of the Jacobian matrix in analytical algorithms.

[0054] (3) Bit-by-bit feedforward network: To enhance the model's nonlinear fitting capability, features at each location are independently passed through a feedforward network. This network comprises two linear transformations and one nonlinear activation, as shown in the following equation.

[0055]

[0056] in, For GELU activation function, The feature dimension is expanded to 4 times to increase the latent space capacity, and This allows for the nonlinear recombination of features in a higher-dimensional manifold space.

[0057] (4) Residual learning and layer normalization: To ensure the stability of gradient propagation in deep networks, each sublayer is connected using a Post-Norm structure. The calculation process is shown in the following equation.

[0058]

[0059]

[0060] in Representation layer normalization. By standardizing the feature dimensions and applying learnable affine transformations, it effectively alleviates the gradient vanishing and covariate shift problems in deep network training.

[0061] After N=2 layers of stacking, the bottleneck layer outputs the final global feature sequence as shown in the following formula.

[0062]

[0063] 2.4 Feature Recombination and Inverse Finite Element Projection decoder The aim is to map latent features that aggregate global and local information back to the physical space. Its core innovation lies in constructing a physically-aware projection mechanism, which enables direct regression from feature manifolds to unstructured finite element mesh manifolds.

[0064] (1) Multi-scale feature fusion based on splicing To preserve the location information of foreign objects while restoring the boundary shape, the invention employs a U-shaped structure and fuses multi-scale features through skip connections. Unlike the summation operation of residual networks, the invention uses channel splicing to maximize the preservation of high-frequency gradient information.

[0065] For the decoding stage k, the fusion process is defined as two steps: first, upsampling is performed through transposed convolution as shown in the formula; then, features from the same level as the encoder are fused. The splicing and fusion are performed as shown in the following formula.

[0066]

[0067]

[0068] in, This is the output of the previous decoder. These are the features after upsampling. This indicates a splicing operation along the channel dimension. for Convolution is used to halve the number of feature channels in the concatenated array and integrate cross-channel information. This fusion strategy ensures that the global semantic information from the Transformer used for localization and the low-level gradient details of the encoder used to define the shape are fully preserved in the channel dimension, avoiding the blurring of details that may be caused by simple element-wise addition.

[0069] (2) Depth Inverse Finite Element Projector: To avoid redundant computations in the pixel domain, this invention proposes a "feature-physical manifold mapping" mechanism, namely a depth inverse finite element projector. Unlike traditional decoders that aim to reconstruct visual images, this module consists of two wide-body fully connected layers, designed to directly project the feature space onto an unstructured mesh space. This projection transformation is defined as shown in the following equation.

[0070]

[0071] in, For the final decoded sequence Flattened vector, It is the ReLU activation function. and These are the weight matrix and bias of the first hidden layer, which is designed to map primary features to an extremely high-dimensional nonlinear representation space. and The weight matrix and bias vector of the second projective hidden layer are maintained by a 4096-dimensional wide-body structure. This layer aims to provide ample parameter capacity to fit the complex electric field coupling relationship in the EIT inverse problem and prevent bottleneck loss of features during projection. and The parameters of the physical mapping output layer are, where , This layer eliminates nonlinear activation and directly maps deep latent features to grid conductivity values ​​with explicit physical meaning through linear combination, directly mapping hidden features to a grid containing... The FEM mesh space of each cell.

[0072] The deep inverse finite element projector forces the network to learn a nonlinear basis function mapping from the latent feature space to the FEM cell space. This enables direct regression of the physical manifold, fundamentally avoiding the interpolation errors and visual illusions caused by traditional pixelation methods, and ensuring the physical fidelity of the reconstruction results.

[0073] 3. Sparse Adaptive Physical Loss Function Traditional image generation loss functions often fail due to the extremely sparse physical characteristics of foreign object regions (ROIs) in EIT imaging. SSIM, which focuses on structural similarity perceived by human vision, is easily masked by dominant background smoothing, thus ignoring the conductivity deviations of minute lesions. To return to the essence of EIT as a quantification of physical fields, this invention abandons visual perception metrics and proposes a sparsity-adaptive physical loss function. It is shown in the following formula.

[0074]

[0075] In the formula, This represents the total number of subdivided elements in the finite element grid (FEM Grid). Indicates the first The true conductivity distribution of each grid cell, This indicates that the network outputs the first... The predicted conductivity value for each grid cell. This represents the absolute physical error on the grid cell. This metric directly reflects the inversion accuracy of the physical parameters. It finely corrects the gradient of the foreign object edge in the later stages of training to prevent the loss of details caused by excessive smoothing. This represents the mean squared error term. This metric accelerates model convergence in the early stages of training and strongly penalizes large-scale conductivity deviations. The balancing coefficient is used to adjust the contribution ratio of the two types of errors. The invention is configured as follows: . The weights are dynamically balanced based on the real physical distribution. The core design idea is inverse class weighting, which forces the network to focus on high-frequency physical boundaries by increasing the weights of sparse regions, as shown in the following equation.

[0076]

[0077] in, and These represent the number of mesh cells in the foreign object region and the background region, respectively. A threshold is used to distinguish between the background and the target. This weighting mechanism mathematically ensures that the contributions of the foreign object region and the background region to the total loss are normalized to the same order of magnitude. This design fundamentally eliminates the risk that features will be ignored by the model due to the small ROI area, ensuring the accuracy of the imaging results in terms of physical values, thereby eliminating the risk of features being ignored due to target sparsity.

[0078] 4. Model Generation and Evaluation Following the above-described dataset and network structure, iterative training is performed using the input training and test sets. The iteration count is set to 200, the batch size to 32, and the learning rate to 0.001. After each iteration, the model is generated. During the iteration process, the changes in the loss value generated by the loss function are recorded. Figure 5 As shown. The input validation set is used to validate the model's imaging speed and loss rate. A comparison between the standard image and the reconstructed image is shown below. Figure 6 As shown, the average imaging speed of the reconstructed images on the validation set is 0.32s, and the average loss is 0.142.

[0079] The loss value exhibits a steep decreasing trend in the early stages of training. This indicates that the multi-scale spectral feature extraction stage in the 1D-UNETR architecture can quickly capture the low-frequency principal component features of the differential voltage sequence regarding the location of the foreign object, achieving initial separation of the background field and the target field. As iterations deepen, the loss curve enters a stable decreasing range without significant oscillations or overfitting. This verifies the effectiveness of the sparse adaptive hybrid physical loss function proposed in this invention, effectively avoiding the imaging blurring problem caused by conventional loss trapping in local minima. The average single-frame imaging time on the validation set remains in the millisecond range, an order of magnitude improvement compared to traditional iterative algorithms and current deep learning pixel imaging algorithms. This is mainly attributed to the design of the inverse finite element projector, which compresses the complex iterative inversion process into a single forward matrix operation, eliminating the computational bottleneck of repeated updates and inversions of the Jacobian matrix in traditional methods. The average test loss remains at an extremely low level, and the reconstructed image closely matches the standard image, proving that the vector-to-grid paradigm successfully avoids pixelation interpolation errors and directly recovers the true distribution of the field at the physical topology level.

[0080] Example 2: An electronic device, comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method described in Embodiment 1 when executing the computer program.

[0081] Example 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0082] Example 4: A computer program product includes a computer program that, when executed by a processor, implements the method described in Example 1.

[0083] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily refer to the same embodiment.

[0084] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0085] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0086] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0087] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0088] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be 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, or discrete hardware components.

[0089] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0090] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A real-time reconstruction method of electrical impedance tomography based on 1D-UNETR and inverse finite element projection, characterized in that: Includes the following steps: S1: Construct a dataset containing multiple sets of electrical impedance imaging samples using a finite element forward simulation solver, for each set of samples in the dataset, obtain a time-difference voltage sequence reflecting the physical disturbance inside the field domain , and the difference between the true conductivity distribution and the background conductivity distribution is the conductivity disturbance vector , construct the input and output data pairs , for supervised training of the reconstruction model to learn the nonlinear inverse mapping relationship from the boundary difference voltage to the grid conductivity change S2: Construct a real-time reconstruction network for electrical impedance imaging based on 1D-UNETR and inverse finite element projection. The reconstruction network includes a spectral feature extractor, a global field coupling bottleneck layer, and feature reconstruction and inverse finite element projection. S3: Set a sparse adaptive physics loss function, utilizing input and output data to... Train and test the reconstructed network; S4: The trained reconstruction network is used to reconstruct electrical impedance imaging in real time to obtain the conductivity perturbation distribution of foreign objects in the field, thereby realizing real-time perception of the target's position, size, and shape.

2. The real-time reconstruction method for electrical impedance imaging based on 1D-UNETR and inverse finite element projection according to claim 1, characterized in that: Step S1 specifically includes the following steps: A dataset containing M sets of samples was constructed using a finite element forward simulation solver, and it was divided into a training set, a validation set, and a test set. For each set of samples in the dataset, a true conductivity distribution containing the location, size, and numerical randomization of foreign objects was first randomly generated within the field of the finite element mesh. And define the standard uniform background conductivity distribution when there are no foreign objects. Subsequently, based on the adjacent excitation-adjacent measurement protocol, the boundary measurement voltage vectors containing foreign objects are calculated using a forward solver. and the reference voltage vector under the background field By calculating the difference between the measured voltage and the reference voltage, a time-difference voltage sequence reflecting the physical disturbances within the field is obtained. The sequence contains several independent measurement data points, which are used as the initial input features for the reconstruction network. The conductivity perturbation vector is obtained by calculating the difference between the true conductivity distribution and the background conductivity distribution. The dimension of this vector is equal to the total number of elements in the finite element mesh. This is used as the training label for reconstructing the network, thereby constructing... Input and output data pairs are used to supervise the training of the reconstructed model to learn the nonlinear inverse mapping relationship from the boundary differential voltage to the change in grid conductivity.

3. The real-time reconstruction method for electrical impedance imaging based on 1D-UNETR and inverse finite element projection according to claim 1, characterized in that: The real-time reconstruction network for electrical impedance imaging based on 1D-UNETR and inverse finite element projection aims to establish a direct nonlinear mapping from the measurement manifold to the physical property manifold. The mathematical mapping definition for the entire process is as follows: in, It is the input differential voltage sequence , It is the predicted differential conductivity distribution of the output. , The total number of triangular subdivision elements in the corresponding finite element mesh; function , and These represent the local gradient extractor, the global field coupling modeler, and the inverse finite element projector, respectively. The composite function formed by these three components constitutes a complete nonlinear inverse operator from the observation space to the physical parameter space.

4. The real-time reconstruction method for electrical impedance imaging based on 1D-UNETR and inverse finite element projection according to claim 1, characterized in that: The spectral feature extractor extracts the differential voltage sequence. Input is a spectral feature extractor consisting of multiple layers of one-dimensional convolutions. ;No. Convolutional operations utilize a limited receptive field to extract local gradient features and compress data dimensionality through downsampling. The computation process is described as follows: In the formula This represents a convolution with a stride of 2; this operation reduces the sequence length while doubling the channel dimension, thereby converting the specific signal waveform into abstract features; finally, it undergoes cascaded processing through k=3 stages.

5. The real-time reconstruction method for electrical impedance imaging based on 1D-UNETR and inverse finite element projection according to claim 1, characterized in that: The global coupling bottleneck layer utilizes a global field coupling modeler Using the extracted local features as input, a self-attention mechanism is employed to simulate the global sensitivity matrix under soft field effects, calculating the correlation between query Q, key K, and value V. The core attention calculation formula is as follows: in, The scaling factor is used to capture the electric field coupling effect between long-distance electrodes and fuse it into global context features via the output projection matrix.

6. The real-time reconstruction method for electrical impedance imaging based on 1D-UNETR and inverse finite element projection according to claim 1, characterized in that: Feature Recombination With inverse finite element projection through decoder Global features are upsampled through progressively transposed convolutions and then fused with skip connection features from the encoder. To compensate for the loss of spatial information; The feature restoration process of each decoding layer is shown in the following equation: Decoded Feature flattening generation Input depth inverse finite element projector By utilizing a two-stage wide-body fully connected layer and a physical mapping layer, the feature vectors are directly projected onto the finite element mesh physical space to calculate the predicted value of the conductivity change. : Predicted values The output predictions are directly mapped to the unstructured grid. On each subdivided unit, end-to-end physical field reconstruction is completed.

7. The real-time reconstruction method for electrical impedance imaging based on 1D-UNETR and inverse finite element projection according to claim 1, characterized in that: The sparse adaptive physical loss function is expressed as: In the formula, This represents the total number of subdivided elements in the finite element mesh (FEM Grid). Indicates the first The true conductivity distribution of each grid cell, This indicates that the network outputs the first... The predicted conductivity value for each grid cell. This represents the absolute physical error on that grid cell; This represents the mean square error term; This is a balancing coefficient used to adjust the contribution ratio of the two types of errors; The dynamic equilibrium weights, generated based on the real physical distribution, are expressed as: in, and These represent the number of mesh cells in the foreign object region and the background region, respectively. A threshold for distinguishing between background and target.

Images