Organ shape solving system, conductivity reconstruction solving system, EIT imaging system and method

Abstract

The invention provides an organ shape solving system, a conductivity reconstruction solving system and an EIT imaging system and method, and belongs to the field of medical imaging, and the technical key points are as follows: a storage module is used for storing a boundary voltage matrix delta V of a patient and a Jacobian matrix J obtained by solving tissue contour information based on idealized prior setting; the initial EIT imaging module is used for obtaining an initial EIT image; the organ shape solving module establishes a mapping relation between the initial EIT imaging and a patient organ shape graph; the conductivity reconstruction solving module establishes a mapping relation between the initial EIT image and a conductivity characteristic pattern; and the feature fusion module is used for reading the organ shape graph of the patient output by the organ shape solving module and the conductivity feature graph output by the conductivity reconstruction solving module, and outputting a tissue sensing conductivity distribution graph.

Description

Technical Field

[0001] This invention belongs to the technical field of medical imaging and deep learning, specifically relating to an organ shape solving system, a conductivity reconstruction solving system, an EIT imaging system, and a method. Background Technology

[0002] There are two main technical approaches to EIT imaging: The first method is to solve for the conductivity distribution based on the Jacobian matrix. For example, in reference 1: “Li Zhiwei, Yu Yao, Wu Yang, et al. Research progress of electrical impedance tomography in pulmonary function testing [J]. Mechanical Manufacturing and Automation, 2024, 53(1):1-9”, the basic idea is: a. Obtain images of the patient's biological organs (such as CT scans), and based on these images, establish an EIT simulation model region in COSMEL software; b. Perform simulation in COSMEL software to solve for the Jacobian matrix; c. The boundary voltage data of the patient is collected using an EIT device, and then the conductivity distribution is solved using an algorithm such as TK-Noser.

[0003] The second method uses neural networks to solve for the conductivity distribution. As described in reference 2: CN117274413B (Nanjing University of Aeronautics and Astronautics, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences), the basic idea behind the EIT imaging method based on neural networks is: a. Obtain images of the patient's biological organs (such as CT scans), and based on these images, establish an EIT simulation model region in COSMEL software; b. Perform simulations in COSMEL software to obtain multiple sets of boundary voltage data – boundary voltage datasets – to serve as training sets for subsequent neural networks; c. Training the neural network: Establishing a mapping relationship between boundary voltage data and conductivity distribution; d. Use EIT equipment to collect the patient's boundary voltage data; e. Using the trained neural network, the boundary voltage data obtained in step d is used as the input layer of the neural network to obtain the conductivity distribution, and the conductivity distribution results are used for EIT imaging.

[0004] In other words, both existing approaches require prior imaging of the patient to determine the conductivity distribution.

[0005] However, for some patients (such as children, pregnant women, and other special patients), it is not possible to take CT scans or other imaging. In this case, if there are no prior images of the patient, how to solve for the conductivity distribution becomes a problem. Summary of the Invention

[0006] The purpose of this invention is to solve the problems existing in the prior art and to provide an organ shape solving system and an organ shape solving method.

[0007] Another objective of this application is to provide a conductivity reconstruction solution system and reconstruction method.

[0008] Another objective of this application is to provide an EIT imaging system and method based on a hybrid of tissue information and conductivity sensing.

[0009] The technical solution of this application is as follows: An organ shape solving system, comprising: The storage module stores the patient's boundary voltage matrix ΔV and the Jacobian matrix J obtained by solving based on idealized prior information about the tissue contour. The initial EIT imaging module solves for the initial conductivity estimation matrix based on ΔV and J. σ 0, based on σ 0 yields the initial EIT image; The organ shape solving module establishes a mapping relationship between the initial EIT imaging and the patient's organ shape map; The organ shape solving module includes a sequentially connected network input layer, an overlapping 3D patch embedding layer, an encoder, a decoder, and a network output layer; The network input layer reads the three-dimensional image reconstructed based on the patient's boundary voltage matrix V and idealized prior tissue contour information; The overlapping 3D patch embedding layer divides the input 3D image into several small blocks; The encoder is composed of several 3D hybrid Transformer modules connected sequentially; the 3D hybrid Transformer module is composed of an efficient self-attention mechanism module, a three-dimensional hybrid feedforward network, and an overlapping 3D patch embedding layer connected sequentially; the encoder extracts local structural details and global semantic features across scales layer by layer to form a multi-scale volume feature representation from coarse to fine. The decoder employs a full multilayer perceptron, which integrates and restores features from different layers through step-by-step upsampling and channel fusion operations, thereby achieving high-resolution, semantically strong structure mask output. The network output layer outputs a high-resolution binary mask to accurately characterize the structural boundaries of biological tissue regions.

[0010] Furthermore, the 3D image input to the network input layer of the organ shape solving module has a length × width × height of 64 × 64 × 64, and the high-resolution binary mask output by the network output layer has a length × width × height of 64 × 64 × 64.

[0011] The method for determining organ shape includes the following steps: S100, the initial conductivity estimate matrix is ​​solved based on the patient's boundary voltage data ΔV and the Jacobian matrix J obtained from idealized prior tissue contour information. σ 0, based on σ 0 yields the initial EIT image; S200: Input the initial EIT image into the trained organ shape solving module to obtain the patient's organ tissue morphology map.

[0012] The conductivity reconstruction solution system includes: The storage module stores ΔV and J; ΔV is the patient's boundary voltage matrix, and J is the Jacobian matrix obtained based on idealized prior information about the tissue contour or based on the patient's tissue contour information. The initial EIT imaging module solves for the initial conductivity estimation matrix based on ΔV and J. σ 0, based on σ 0 yields the initial EIT image; The conductivity reconstruction solution module establishes a mapping relationship between the initial EIT imaging and the conductivity feature map; The conductivity reconstruction solution module adopts a Slim VSCAU-Net3D model, which has an overall U-shaped structure and includes sequentially connected components: The network input layer reads the three-dimensional image reconstructed based on the patient's boundary voltage matrix V and idealized prior information about the tissue contours. The length of the input image is [length value missing]. The encoder consists of four stages in sequence, each stage including: a 3D depthwise separable convolutional module and a Slim UNETR module; the 3D depthwise separable convolutional module is used for lightweight local feature extraction; the Slim UNETR module is used for efficient encoding of local and long-range information; a spatial dimension downsampling operation is performed after each stage. The decoder, symmetrical to the encoder, includes three sets of upsampling paths. Each set of upsampling paths includes, in sequence: a SlimUNETR module, a 3D transposed convolution module, and a VFSCA module. The SlimUNETR module is used to efficiently encode local and long-range information; the 3D transposed convolution module is used for spatial upsampling; and the VFSCA module is used to enhance the semantic information of the fused features. The network output layer consists of a set of 1×1×1 three-dimensional convolutions, which are used to generate the final conductivity reconstruction result map. Furthermore, the VFSCA module includes: Spatial attention path, based on the construction of local contextual attention using a 3D sliding window; and a local contextual attention graph is constructed based on the 3D sliding window mechanism to mine the correlation of voxel-level spatial neighborhoods; Channel attention path: It uses shared query and key vectors to construct non-linear attention relationships between different channels, thereby improving the ability to model inter-channel dependencies; The attention fusion layer fuses spatial attention and channel attention through element-wise multiplication or weighted addition to form a joint attention map; the attention fusion layer achieves joint spatial-channel enhancement, improving the global modeling capability of deep features; The convolutional dimensionality reduction layer uses 1×1×1 convolution operations to compress and reconstruct the fused high-dimensional features; The residual connection module is used to perform a weighted summation of the original features and the fused features to improve feature consistency and training stability. The VFSCA module captures the voxel spatial dependencies and inter-channel interactions in 3D medical images through parallel computation of spatial attention paths and channel attention paths, thereby enhancing feature representation capabilities and reconstruction robustness.

[0013] Furthermore, the VFSCA module adopts a design strategy of parallel extraction and fusion enhancement to form a structured attention enhancement module; First, the input features simultaneously enter the spatial attention path and the channel attention path, which are used to model the spatial dependencies between voxels and the feature relationships between channels, respectively. These two attention branches independently extract attention maps of their respective dimensions, and then integrate them through an attention fusion layer to obtain a joint attention map, thereby achieving joint modeling of spatial and channel information. The joint attention map is then channel-compressed and reconstructed through a 1×1×1 convolutional dimensionality reduction layer to reduce computational cost and improve the compactness of feature representation; Finally, the fused features and the original input features are weighted and superimposed in the residual connection module to achieve consistent matching between the enhanced features and the original semantics.

[0014] Furthermore, the VFSCA module operates as follows: Step a, Generate spatial attention-enhanced features through the spatial attention path Figure X s Input features Figure X Perform shared queries and key vector calculations to generate the query feature map Q. shared Key feature map K shared ; Input features Figure X Perform calculations to generate the Value vector V. spatial ; Spatial attention enhancement features Figure X s for: X s =SA(Q) shared K shared V spatial ), Where SA is the spatial attention module function; Step b, generate channel attention enhancement features Figure X c : Input features Figure X Perform shared queries and key vector calculations to generate the query feature map Q. shared Key feature map K shared ; Input features Figure X Calculate the value projection weight V of the generated channel path. channel ; Channel attention enhancement features Figure X c for: X c =CA(Q) shared K shared V channel ), Where CA is the channel attention module function; Step c, the attention fusion module outputs X by combining the spatial path and the channel path. s X c The fusion process yields the enhanced comprehensive feature X. fusion ; Step d, residual connection: X out =X fusion +X, Step e, output the fused and enhanced features Figure X out Moving on to the next module.

[0015] An EIT imaging system based on a hybrid of tissue information and conductivity sensing includes: a storage module, an initial EIT imaging module, an organ shape solving module, a conductivity reconstruction solving module, and a feature fusion module; The storage module is used to store the patient's boundary voltage matrix ΔV and the Jacobian matrix J obtained by solving based on idealized prior tissue contour information. The initial EIT imaging module solves for the initial conductivity estimation matrix based on ΔV and J. σ 0, based on σ 0 yields the initial EIT image; Among them, the organ shape solving module establishes the mapping relationship between the initial EIT imaging and the patient's organ shape map; Among them, the conductivity reconstruction solution module establishes the mapping relationship between the initial EIT imaging and the conductivity feature map; Among them, the feature fusion module reads the patient organ shape map output by the organ shape solving module and the conductivity feature map output by the conductivity reconstruction solving module, and outputs a tissue-sensing conductivity distribution map to realize deep collaborative modeling of multimodal features.

[0016] Furthermore, the feature fusion module includes, in sequence, an input layer, a channel stitching layer, and a 1×1×1 convolutional fusion layer. The input layer receives the patient organ structure map output from the organ shape solving module and the conductivity feature map from the conductivity reconstruction solving module 400. The channel stitching layer stitches the patient organ structure map and the conductivity feature map into a fused feature map. The 1×1×1 convolutional fusion layer reads the fused feature map and outputs a structure-aware conductivity distribution image.

[0017] Furthermore, the structure-sensing conductivity distribution image simultaneously characterizes ventilation or lesion information within the boundaries of the patient's lung structures.

[0018] Furthermore, the organ shape solving module includes a sequentially connected network input layer, an overlapping 3D patch embedding layer, an encoder, a decoder, and a network output layer; The network input layer reads the three-dimensional image reconstructed based on the patient's boundary voltage matrix V and idealized prior tissue contour information; The overlapping 3D patch embedding layer divides the input 3D image into several small blocks; The encoder is composed of several 3D hybrid Transformer modules connected sequentially; the 3D hybrid Transformer module is composed of an efficient self-attention mechanism module, a three-dimensional hybrid feedforward network, and an overlapping 3D patch embedding layer connected sequentially; the encoder extracts local structural details and global semantic features across scales layer by layer to form a multi-scale volume feature representation from coarse to fine. The decoder employs a full multilayer perceptron, which integrates and restores features from different layers through step-by-step upsampling and channel fusion operations, thereby achieving high-resolution, semantically strong structure mask output. The network output layer outputs a high-resolution binary mask to accurately characterize the structural boundaries of biological tissue regions.

[0019] Furthermore, the length × width × height of the 3D image input to the network input layer of the organ shape solving module is 64 × 64 × 64, and the high-resolution binary mask output by the network output layer is 64 × 64 × 64.

[0020] Furthermore, the organ shape solving module establishes a mapping relationship between the initial EIT imaging and the patient's organ shape map; The organ shape solving module includes a sequentially connected network input layer, an overlapping 3D patch embedding layer, an encoder, a decoder, and a network output layer; The network input layer reads the three-dimensional image reconstructed based on the patient's boundary voltage matrix V and idealized prior tissue contour information; The overlapping 3D patch embedding layer divides the input 3D image into several small blocks; The encoder is composed of several 3D hybrid Transformer modules connected sequentially; the 3D hybrid Transformer module is composed of an efficient self-attention mechanism module, a three-dimensional hybrid feedforward network, and an overlapping 3D patch embedding layer connected sequentially; the encoder extracts local structural details and global semantic features across scales layer by layer to form a multi-scale volume feature representation from coarse to fine. The decoder employs a full multilayer perceptron, which integrates and restores features from different layers through step-by-step upsampling and channel fusion operations, thereby achieving high-resolution, semantically strong structure mask output. The network output layer outputs a high-resolution binary mask to accurately characterize the structural boundaries of biological tissue regions.

[0021] Furthermore, the conductivity reconstruction solution module adopts a Slim VSCAU-Net3D model, which has an overall U-shaped structure and includes sequentially connected components: The network input layer reads the three-dimensional image reconstructed based on the patient's boundary voltage matrix V and idealized prior information about the tissue contours. The length of the input image is [length value missing]. The encoder consists of four stages in sequence, each stage including: a 3D depthwise separable convolutional module and a Slim UNETR module; the 3D depthwise separable convolutional module is used for lightweight local feature extraction; the Slim UNETR module is used for efficient encoding of local and long-range information; a spatial dimension downsampling operation is performed after each stage. The decoder, symmetrical to the encoder, includes three sets of upsampling paths. Each set of upsampling paths includes, in sequence: a SlimUNETR module, a 3D transposed convolution module, and a VFSCA module. The SlimUNETR module is used to efficiently encode local and long-range information; the 3D transposed convolution module is used for spatial upsampling; and the VFSCA module is used to enhance the semantic information of the fused features. The network output layer consists of a set of 1×1×1 three-dimensional convolutions, which are used to generate the final conductivity reconstruction result map, i.e., the conductivity feature map.

[0022] An EIT imaging method based on a hybrid approach of tissue information and conductivity sensing includes the following steps: S100, the initial conductivity estimate matrix is ​​solved based on the patient's boundary voltage data ΔV and the Jacobian matrix J obtained from idealized prior tissue contour information. σ 0, based on σ 0 yields the initial EIT image; S200: Input the initial EIT image into the trained organ shape solving module to obtain the patient's organ tissue morphology map; S300: Input the initial EIT image into the trained conductivity reconstruction solution module to obtain the patient's conductivity feature map; S400: Input the organ tissue morphology map obtained in S200 and the conductivity feature map obtained in S300 into the feature fusion module to obtain the tissue-sensing conductivity distribution map.

[0023] The advantages of the technical solution of this invention are mainly reflected in: (1) This application proposes an EIT imaging system based on a hybrid of tissue information and conductivity sensing, which is used to solve the problem of obtaining conductivity images from prior detection images without patients. High-resolution, multiphase conductivity distribution three-dimensional imaging is achieved by using different loss functions and a dual-branch neural network.

[0024] This method combines the advantages of physical modeling and deep learning. First, it pre-reconstructs the boundary voltage using the sensitivity matrix method to generate a preliminary conductivity image. Then, the organ solving module and the conductivity reconstruction solving module extract the tissue structure features and conductivity features of the target region, respectively, and optimize them through the feature fusion module. By fusing multimodal information from the two branches, the network can more comprehensively understand the correlation between tissue structure and electrical properties, thereby effectively improving the clarity of image boundaries and the robustness of overall reconstruction, and achieving accurate three-dimensional conductivity imaging in complex physiological regions.

[0025] It should be noted that the network structure algorithm proposed in this application is not limited to the organ solving module and the conductivity reconstruction solving module.

[0026] (2) The organ shape solving system of this application can function as an independent solving system. Its core lies in the organ shape solving module, and the neural network architecture it uses possesses the following innovations: 2.1 Structure-aware Transformer skeleton design: This network adopts a multi-level 3D Mix Transformer encoder structure, which integrates the Efficient Self-Attention (ESA) mechanism and the Mix-FFN module to capture local boundary and global anatomical context features while maintaining a low number of parameters.

[0027] 2.2 Overlapping 3D patch embedding mechanism: The input σ0 is divided into small patches through an overlapping sliding window and projected into a high-dimensional feature space, effectively preserving the continuity of local structure and providing prior guarantees for complex shape recognition.

[0028] 2.3 Lightweight Decoder Design (All-MLP): The All-MLP structure is used to replace the traditional deconvolution decoder, which greatly reduces the model complexity and computational load. The structure segmentation output is achieved while significantly reducing the number of parameters and inference latency.

[0029] 2.4 Multi-scale semantic fusion mechanism: The network introduces cross-layer connections (SkipConnection) between different Transformer layers, and strengthens the contextual expressiveness through channel dimension splicing and MLP fusion mechanism to achieve accurate reconstruction of multi-scale lung morphology.

[0030] 2.5. Universality of structure-guided EIT inversion: The output organ tissue (e.g., lung structure, brain structure) mask can be independently called as organ prior, adapting to the structural constraint input of different reconstruction models, and has good module universality and transferability.

[0031] This neural network can serve as an independent organ morphology recognition system, independent of specific EIT reconstruction methods, and possesses broad engineering applicability and innovative value.

[0032] In its use, firstly, the initial conductivity estimate matrix is ​​solved based on the patient's boundary voltage data ΔV and the Jacobian matrix J obtained from idealized prior tissue contour information. σ 0, based on σ 0. Obtain the initial EIT image; then, input the initial EIT image into the trained organ shape solving module to obtain the patient's organ tissue morphology map.

[0033] (3) The conductivity reconstruction solution system of this application can also be used as an independent solution system. Its core point is a novel three-dimensional electrical impedance image reconstruction neural network architecture (conductivity reconstruction solution module) proposed in this application, which balances accuracy and efficiency in structural design. The network uses depthwise separable convolution for lightweight modeling, effectively compressing the number of parameters and computational load; at the same time, it introduces a voxel-focused spatial-channel attention mechanism (VFSCA) to enable the model to fully capture the spatial distribution features and channel response differences in the lung structure, and enhance the ability to identify and express key regions. Through the above design, this method achieves the advantages of both high image resolution and fast inference output in three-dimensional EIT image reconstruction tasks, breaking the bottleneck of traditional methods that are difficult to balance speed and quality, and has good clinical application potential.

[0034] (4) The feature fusion module 500 of this application serves as the core bridge connecting the organ shape solving module 300 and the conductivity reconstruction solving module 400, aiming to achieve deep integration of multimodal features, thereby improving the boundary clarity and region perception capability of EIT imaging. The feature fusion module 500 receives high-order semantic features output from two sub-networks: one is the tissue morphology map from the organ shape solving module 300, and the other is the electrical feature map from the conductivity reconstruction solving module 400. Both are three-dimensional tensors of the same size. By jointly modeling organ tissue information and conductivity supervision, the consistency of image spatial structure is improved, while the fine-grained expression capability of conductivity distribution is enhanced, thereby achieving higher-precision three-dimensional EIT imaging results. To achieve cross-modal semantic alignment, the tissue morphology map and electrical feature map are first stitched together in the channel dimension. Then, a 1×1×1 convolution operation is used to perform channel compression and nonlinear transformation on the fused features. This hybrid fusion mechanism effectively overcomes the bottleneck of decoupling structural and electrical information in traditional EIT methods, improving the network's imaging robustness and reconstruction accuracy in complex tissue regions. Ultimately, the output structure-aware conductivity distribution map reflects both the boundary information of the tissue's anatomical structure and preserves the local variation characteristics of multiphase conductive regions. Attached Figure Description

[0035] The present application will be further described in detail below with reference to the embodiments in the accompanying drawings, but this does not constitute any limitation on the present application.

[0036] Figure 1 A schematic diagram illustrating the idealized priori setting of tissue contour information.

[0037] Figure 2 The imaging results are illustrated using the conductivity distribution matrix reconstructed using prior information about the tissue contour.

[0038] Figure 3 The image of the research and development target is shown.

[0039] Figure 4 The diagram illustrates the overall architecture of the EIT imaging system based on a hybrid of tissue information and conductivity sensing proposed in this application.

[0040] Figure 5 The network structure diagram of the organ shape solving module is shown.

[0041] Figure 6 The diagram illustrates the architecture of the organ shape solving module.

[0042] Figure 7 The network structure diagram of the conductivity reconstruction solution module is shown.

[0043] Figure 8 The diagram illustrates the structure of the VFSCA module.

[0044] Figure 9 The diagram illustrates the architecture of the conductivity reconstruction solution module.

[0045] Figure 10 The method for generating the dataset in this application is illustrated.

[0046] Figure 11 The diagram illustrates the application results of the method described in this application.

[0047] Figure 12 The physical meaning of the symbols used in this application is illustrated. Detailed Implementation

[0048] The objectives, advantages, and features of this invention will be explained through the following non-limiting description of preferred embodiments. These embodiments are merely typical examples of applying the technical solutions of this invention, and all technical solutions formed by equivalent substitutions or equivalent transformations fall within the scope of protection claimed by this invention.

[0049] <Part 1: Research and Development Strategy> Because images of the patient's biological organs are unavailable, only idealized, pre-defined tissue contour information (such as...) can be used. Figure 1 In the diagram, (a) represents the electric field line distribution in the 3D model. Current is injected from the red electrode, and the blue electrode is the grounding terminal. (b) represents the refined finite element model of the thoracic cavity (source: https: / / eidors3d.sourceforge.net / index.shtml) as a biological organ model.

[0050] Figure 2The imaging results are illustrated using the conductivity distribution matrix reconstructed using prior information about the tissue contour. Figure 3 The image of the research and development target is shown.

[0051] Compare Figure 2 , Figure 3 It can be seen that: (1) Figure 2 It can only roughly discern the outline of the lungs, but cannot accurately determine their boundaries. This is because the neural network's input is a boundary voltage vector, and its output is a reconstructed three-dimensional conductivity matrix. The training samples use idealized or prior-defined tissue contour information. Since this process lacks the ability to perceive the actual organ structure, the network is prone to problems such as structural ambiguity, unclear boundaries, and localization errors when processing multiphase conductive tissues or complex anatomical regions.

[0052] (2) Insufficient coupling between biological tissues (i.e. organs) and electrical conductivity information. Figure 3 The condition of the lungs can be seen (red indicates areas of lung damage; blue indicates areas of lung ventilation), and Figure 2 The above information is missing.

[0053] <II. Overall Architecture> Figure 4 The overall architecture diagram of this application is shown. An EIT imaging system based on the hybrid of tissue information and conductivity sensing includes: a storage module, an initial EIT imaging module 200, an organ shape solving module 300, a conductivity reconstruction solving module 400, and a feature fusion module 500.

[0054] The storage module 100 is used to store the patient's boundary voltage matrix ΔV and the Jacobian matrix J obtained by solving based on idealized prior set tissue contour information (this contour information is not the patient's contour information, but can be regarded as a kind of general contour information); The initial EIT imaging module 200 solves for the initial conductivity estimation matrix based on ΔV and J. σ 0, based on σ 0 yields the initial EIT image; Among them, the organ shape solving module 300 establishes the mapping relationship between the initial EIT imaging and the patient's organ shape map; Among them, the conductivity reconstruction solution module 400 establishes the mapping relationship between the initial EIT imaging and the conductivity feature map; Among them, the feature fusion module 500 reads the patient organ shape map output by the organ shape solving module 300 and the conductivity feature map output by the conductivity reconstruction solving module 400, and outputs a tissue-sensing conductivity distribution map to realize deep collaborative modeling of multimodal features.

[0055] Figure 5 The network structure diagram of the organ shape solving module is shown. Figure 6 The diagram illustrates the architecture of the organ shape solving module. The organ shape solving module 300 is a neural network structure; it includes: The network input layer reads a 3D image reconstructed based on the patient's boundary voltage matrix V and idealized prior tissue contour information. The input 3D image has a size of 64 × 64 × 64, representing the initial volume conductivity distribution, which serves as the basis for structural segmentation. Overlapping 3D Patch Embedding layer divides the input 3D image into several small patches (i.e., overlapping 3D Patch Embedding operation) and maps them to a high-dimensional feature space, effectively preserving local continuity information; The encoder consists of several 3D Mix Transformer blocks connected sequentially. Each 3D Mix Transformer block is composed of an Efficient Self-Attention mechanism, a 3D Mix-FFN network, and an Overlapped 3D Patch Embedding layer connected sequentially. The encoder extracts local structural details and global semantic features across scales layer by layer to form a multi-scale volumetric feature representation from coarse to fine. The decoder employs an All-MLP (All-Multi-Layer Perceptron) to integrate and recover features from different levels through progressive upsampling and channel fusion operations, thereby achieving high-resolution, semantically strong structural mask output. The network output layer outputs a high-resolution binary mask of 64 × 64 × 64, which accurately delineates the structural boundaries of biological tissue regions (such as the lungs). This mask provides anatomical guidance for 3D EIT imaging, effectively improving the accuracy and robustness of the overall reconstruction.

[0056] Instructions for constructing the dataset (dividing the dataset into training, validation, and test sets in an 8:1:1 ratio): S1, the Jacobian matrix J is obtained by solving based on the idealized prior set of tissue contour information; S2, Measure the boundary voltage data ΔV1~ΔV for N testers. N The initial conductivity estimate matrix is ​​obtained by using J. σ 01~ σ 0N ,based on σ01~ σ 0N N initial EIT images were obtained; S3 uses CT or MRI to obtain the organ shapes of N test subjects. Multiple initial EIT images and corresponding organ shape diagrams constitute the dataset.

[0057] Figure 7 The network structure diagram of the conductivity reconstruction solution module is shown. The conductivity reconstruction solution module 400 employs a resource-constrained Slim VSCAU-Net3D (Voxel-Focused Spatial-Channel Attention U-Net3D) model, which has an overall U-shaped structure, including: The network input layer reads the three-dimensional image reconstructed based on the patient's boundary voltage matrix V and idealized prior information about the tissue contour. The input is a four-dimensional tensor with dimensions of 64 (H) × 64 (W) × 64 (D) × 1 (C), where H, W, and D represent the height, width, and depth of the image, respectively, and C is the number of channels. The encoder consists of four stages, each of which includes: a 3D depthwise separable convolution (DepthwiseConv3D) module, a lightweight residual structure block (Slim UNETR Block, which simplifies the number of Transformer layers and convolution parameters, thereby significantly reducing parameter overhead while maintaining structural modeling capabilities; see reference: Pang Y, Liang J, Huang T, Chen H, Li Y, Li D, Huang L, Wang Q. Slim UNETR: Scale Hybrid Transformers to Efficient 3D Medical ImageSegmentation Under Limited Computational Resources. IEEE Trans Med Imaging.2024 Mar;43(3):994-1005. doi:10.1109 / TMI.2023.3326188. Epub 2024 Mar 5. PMID:37862274.); the 3D depthwise separable convolution module is used for lightweight local feature extraction; the lightweight residual structure block (Slim UNETR Block) The UNETR Block is used to efficiently encode local and long-range information. It performs a spatial dimension downsampling operation once after each stage, so that the feature size is gradually downsampled from the input H×W×D to (H / 32)×(W / 32)×(D / 32), while the number of channels increases layer by layer.

[0058] The decoder, symmetrical to the encoder, includes three sets of upsampling paths. Each set of upsampling paths sequentially includes: a SlimUNETR module, a 3D transposed convolutional module (ConvTranspose3D), and a Voxel-Focused Spatial-Channel Attention Unit (VFSCA) module. The SlimUNETR module is used for efficient encoding of local and long-range information; the 3D transposed convolutional module is used for spatial upsampling (i.e.,...). Figure 7 The TS Conv 3D deconvolution in the model is used for feature map upsampling; the VFSCA module is used to enhance the semantic information of the fused features; the output of each decoder is spliced ​​and fused with the output of the corresponding stage of the encoder through skip connections; The network output layer, which includes a set of 1×1×1 three-dimensional convolutions, is used to generate the final conductivity reconstruction result or semantic segmentation prediction map, and its output shape is 64 (H)×64 (W)×64 (D)×1. Figure 8 The diagram illustrates the structure of the VFSCA module. The VFSCA module includes: Spatial attention path: Construct local contextual attention based on a 3D sliding window; Construct a local contextual attention graph based on a 3D sliding window mechanism (e.g., 3×3×3 convolution) to mine the correlation of voxel-level spatial neighborhood; Channel attention paths employ shared query and key vectors to construct non-linear attention relationships between different channels, thereby enhancing the ability to model inter-channel dependencies. The attention fusion layer fuses spatial attention and channel attention through element-wise multiplication or weighted addition to form a joint attention map; the attention fusion layer achieves joint spatial-channel enhancement, improving the global modeling capability of deep features; The convolutional dimensionality reduction layer uses 1×1×1 convolution operations to compress and reconstruct the fused high-dimensional features; The residual connection module is used to perform a weighted summation of the original features and the fused features to improve feature consistency and training stability. The VFSCA module captures voxel spatial dependencies and inter-channel interactions in 3D medical images through parallel computation of the Spatial Attention Path and the Channel Attention Path, thereby enhancing feature representation capabilities and reconstruction robustness. The information flow design of the VFSCA module is as follows: 1) The input voxel feature map is: X∈R C×D×H×W , in, C Where D represents the number of channels, and H, W represents the depth, height, and width, respectively. 2) Spatial Attention Path Shared query and key vectors are used to improve computational efficiency and enhance information coupling between the two branches. Simultaneously, a value vector V is calculated and generated. spatial : Q shared =W Q ·X, K shared =W K ·X, V spatial =W Vs ·X, Among them, W Q W ｋ ∈R C×C Both are learnable weight matrices, Q shared K shared For the generated query and key feature map; W Vs It is the value projection weight of the spatial path, V spatial This indicates the generation of a Value vector; Obtaining spatial attention-enhanced feature maps: X s =SA(Q) shared K shared V spatial ), Where SA is the spatial attention module function, X s Enhanced feature maps for spatial attention; To enhance the spatial attention module's ability to model long-range dependencies, a pooling path can be introduced. Specifically, the input voxel feature map is first subjected to global average pooling to extract global context features, and then passed through two linear transformation modules W. Kp and W Vp These are mapped to the keys and values ​​of attention, respectively. The K generated by this pooling path... pool V pool After being merged with the main path Key / Value, they jointly participate in attention calculation, achieving a fusion of local and global context perception and significantly improving the ability of spatial attention to express complex structures. Pooled paths are introduced for global context modeling, enhancing contextual representation. K pool = W Kp ·σ(X), V pool= W Vp ·σ(X), Where σ(X) is the global average pooling operation, W Kp W Vp These are learnable parameters for the pooling path; 4) Channel Attention Path This path is used to model the global dependencies between different channels, calculated as follows: V channel =W Vc ·X, X c =CA(Q) shared K shared V channel ), Where CA is the channel attention module function, W Vc It is the value projection weight of the channel path, X c Enhanced feature maps for channel attention.

[0059] 5) Attention Fusion Module By fusing the outputs of spatial paths and channel paths, the comprehensive feature representation is enhanced. X fusion =conv3D(concat(X) s X c )), By fusing spatial and channel features through 3D convolution, channel compression and interactive enhancement can be achieved. `concat` (channel concatenation) is a feature fusion operation that uses 3D convolution to fuse spatial and channel features, achieving channel compression and interactive enhancement. `Conv3D` represents a three-dimensional convolutional layer.

[0060] 6) Residual Connection To improve training stability and feature consistency, the VFSCA module introduces a residual structure: X out =X fusion +X, Achieving feature reuse and gradient smoothing effectively alleviates the degradation problem of deep networks. 7) Output fusion enhancement features Figure X out Moving on to the next module.

[0061] The VFSCA module employs a parallel extraction and fusion enhancement design strategy, forming a structured attention enhancement module. First, the input feature map simultaneously enters both the spatial attention path and the channel attention path, used to model spatial dependencies between voxels and feature relationships between channels, respectively. These two attention branches independently extract attention maps of their respective dimensions, which are then integrated through an attention fusion layer to obtain a joint attention map, achieving joint modeling of spatial and channel information. This joint attention map is then subjected to channel compression and reconstruction through a 1×1×1 convolutional dimensionality reduction layer to reduce computational cost and improve the compactness of feature representation. Finally, the fused features and the original input features are weighted and superimposed in the residual connection module to achieve consistency matching between the enhanced features and the original semantics, thereby improving the network's stability and expressive power in conductivity reconstruction. Therefore, the VFSCA module as a whole follows a three-stage design logic of "parallel extraction—fusion enhancement—residual connection," rather than a single sequential structure.

[0062] Explanation of the dataset construction for the conductivity reconstruction solution module (the dataset is divided into training, validation, and test sets in an 8:1:1 ratio): S1. Use CT or MRI to obtain the organ shapes of N test subjects and import the organ shapes of N test subjects into the simulation model. S2, the Jacobian matrix J is obtained by solving based on the idealized prior information of the tissue contour; S3, in the simulation model, several boundary voltage data ΔV ~ conductivity characteristic map can be obtained from the simulation; S4, the initial conductivity estimation matrix is ​​solved by ΔV and J to generate the initial EIT image; S5, the initial EIT image corresponding to the same boundary voltage data V and the corresponding conductivity feature map constitute the dataset of the conductivity reconstruction solution module.

[0063] The feature fusion module 500 includes, in sequence, an input layer, a channel splicing layer, and a 1×1×1 convolutional fusion layer. The input layer receives a patient organ structure map output from the organ shape solving module 300 and a conductivity feature map from the conductivity reconstruction solving module 400; The channel splicing layer splices the patient's organ structure map (64 × 64 × 64) and conductivity characteristic map (64 × 64 × 64) into a fused characteristic map; The 1×1×1 convolutional fusion layer reads the fusion feature map and outputs a 64×64×64 structure-aware conductivity distribution image. This image can simultaneously characterize the patient's lung structure boundaries and their internal ventilation or lesions, improving the boundary clarity and spatial consistency of the final image.

[0064] It should be noted that the initial conductivity estimation matrix includes the node coordinates and the conductivity value corresponding to each coordinate point.

[0065] For the initial conductivity estimation module 200, it uses the existing TK-Noser method to solve the problem (reference: CN115177234A). Taking lung detection as an example, the specific working process is as follows: S201, Based on the simulation model of idealized prior set tissue contour information, determine the Jacobian matrix J; S202, Read the patient's boundary voltage data ΔV from the storage module and solve using the following formula. σ 0; σ 0=(J T ·J+k t ·I+k n W) -1 ·J T ·ΔV; J is a Jacobian matrix, J T It is the transpose of J, k t k n These are the regularization parameters, where W represents a diagonal matrix of the same order and diagonal elements as J; and I represents an identity matrix with the same number of columns as J.

[0066] <Part Three: Experimental Comparison> Static examination of three patients with pneumothorax: All three patients were diagnosed with pneumothorax by CT scan prior to EIT measurement. After wiping the skin of the chest surface with normal saline, an EIT bandage was applied to the 4th-5th intercostal space. The bandage electrodes were connected to the EIT measurement system via lead wires. All patients were in a supine position, and the raw EIT voltage signal was recorded during 5 minutes of spontaneous breathing. The raw voltage signal was processed using a Butterworth low-pass filter with a cutoff frequency of 0.6Hz, and the differential voltage signals at the end of inspiration and end of expiration were input into the sensitivity matrix method for pre-reconstruction.

[0067] Figure 11 Visualizations of EIT reconstruction from three patients with pneumothorax are presented (from left to right: CT slice, TK-Noser method, UNET3D method, and the method described in this paper). It can be seen that the neural network method and the proposed method can reconstruct the pneumothorax region within the patient's pleural cavity, and the predicted pneumothorax location is consistent with the CT image. The pneumothorax region predicted by the neural network method is larger than the actual area, which can easily lead to misdiagnosis. The method proposed in this application predicts the pneumothorax region more closely to the actual area, and the reconstruction results are superior to the TK-Noser method and the UNET3D method.

[0068] It should be noted that the UNET3D method is the method proposed in the paper "Zhiwei Li, Yang Wu, Tiecheng Xu, JiajuanRen, Kai Liu, Qiuju Cheng, Hao Wang, Bo sun, Jiafeng Yao*, A Fast 3D LungImage Reconstruction Method Based on CT Pixel Matrices Learning with Electrical Impedance Tomography, Measurement, 251, 117176, 2025".

[0069] The prediction results include a multi-class label distribution at the voxel level, with the following categories: Category 0: background; Category 1: normal lung area; Category 2: lung injury area.

[0070] It should be noted that the conductivity reconstruction solution module can be directly applied to situations where patient organ contour information can be obtained.

[0071] At this point: The network input layer reads the three-dimensional image reconstructed based on the patient's boundary voltage matrix ΔV and the patient's tissue contour information. The input is a four-dimensional tensor with dimensions of 64 (H) × 64 (W) × 64 (D) × 1 (C), where H, W, and D represent the height, width, and depth of the image, respectively, and C is the number of channels.

[0072] At this point, the conductivity reconstruction solution module can independently complete the imaging effect.

[0073] Explanation of the dataset construction for the conductivity reconstruction solution module (the dataset is divided into training, validation, and test sets in an 8:1:1 ratio): S1. Use CT or MRI to obtain the shape of the patient's organs and import the shape of the patient's organs into the simulation model; S2, based on the shape of the patient's organ, obtains its corresponding Jacobian matrix J; S3, in the simulation model, several boundary voltage data ΔV ~ conductivity characteristic map can be obtained from the simulation; S4, the initial conductivity estimation matrix is ​​solved by ΔV and J to generate the initial EIT image; S5, the initial EIT image corresponding to the same boundary voltage data V and the corresponding conductivity feature map constitute the dataset of the conductivity reconstruction solution module.

[0074] It should be noted that the conductivity reconstruction solution system is applied to situations where patient organ contour information can be obtained, including: The storage module stores ΔV and J; ΔV is the patient's boundary voltage matrix, and J is the Jacobian matrix obtained based on the patient's tissue contour information. The initial EIT imaging module solves for the initial conductivity estimation matrix based on ΔV and J. σ 0, based on σ 0 yields the initial EIT image; The conductivity reconstruction solution module establishes a mapping relationship between the initial EIT image and the conductivity feature map; its specific structure still adopts the method shown in the attached figure. Figure 7 The conductivity reconstruction solution module mentioned above.

[0075] The above-described embodiments are preferred embodiments of the present invention and are only used to facilitate the illustration of the present invention. They are not intended to limit the present invention in any way. Any person skilled in the art who makes local modifications or alterations to the technical content disclosed in the present invention without departing from the scope of the technical features of the present invention shall still fall within the scope of the technical features of the present invention.

Claims

1. An organ shape solving system, characterized in that, include: The storage module stores the patient's boundary voltage matrix ΔV and the Jacobian matrix J obtained by solving based on idealized prior information about the tissue contour. The initial EIT imaging module solves for the initial conductivity estimation matrix based on ΔV and J. σ 0, based on σ 0 yields the initial EIT image; The organ shape solving module establishes a mapping relationship between the initial EIT imaging and the patient's organ shape map; The organ shape solving module includes a sequentially connected network input layer, an overlapping 3D patch embedding layer, an encoder, a decoder, and a network output layer; The network input layer reads the three-dimensional image reconstructed based on the patient's boundary voltage matrix V and idealized prior tissue contour information; The overlapping 3D patch embedding layer divides the input 3D image into several small blocks; The encoder is composed of several 3D hybrid Transformer modules connected sequentially; the 3D hybrid Transformer module is composed of an efficient self-attention mechanism module, a three-dimensional hybrid feedforward network, and an overlapping 3D patch embedding layer connected sequentially; the encoder extracts local structural details and global semantic features across scales layer by layer to form a multi-scale volume feature representation from coarse to fine. The decoder employs a full multilayer perceptron, which integrates and restores features from different layers through step-by-step upsampling and channel fusion operations, thereby achieving high-resolution, semantically strong structure mask output. The network output layer outputs a high-resolution binary mask to accurately characterize the structural boundaries of biological tissue regions.

2. The organ shape solving system as described in claim 1, characterized in that, The 3D image input to the network input layer of the organ shape solving module has dimensions of 64 × 64 × 64, and the high-resolution binary mask output by the network output layer has dimensions of 64 × 64 × 64.

3. A method for solving organ shape, characterized in that, Includes the following steps: S100, Solving for the initial EIT imaging: Based on the patient's boundary voltage data ΔV and the Jacobian matrix J obtained from idealized prior tissue contour information, solve for the initial conductivity estimation matrix. σ 0, based on σ 0 yields the initial EIT image; S200, invoke the organ shape solving system as described in claim 1 to solve for the patient's organ tissue morphology map: input the initial EIT image into the organ shape solving module of the trained organ shape solving system as described in claim 1 to obtain the patient's organ tissue morphology map.

4. A conductivity reconstruction solution system, characterized in that, include: The storage module stores ΔV and J. ΔV is the patient's boundary voltage matrix, and J is the Jacobian matrix obtained based on idealized prior information about tissue contours or based on patient tissue contour information. The initial EIT imaging module solves for the initial conductivity estimation matrix based on ΔV and J. σ 0, based on σ 0 yields the initial EIT image; The conductivity reconstruction solution module establishes a mapping relationship between the initial EIT imaging and the conductivity feature map; The conductivity reconstruction solution module adopts a Slim VSCAU-Net3D model, which has an overall U-shaped structure and includes sequentially connected components: The network input layer reads the three-dimensional image reconstructed based on the patient's boundary voltage matrix V and idealized prior information about the tissue contours. The length of the input image is [length value missing]. The encoder consists of four stages in sequence, each stage including: a 3D depthwise separable convolutional module and a SlimUNETR module; the 3D depthwise separable convolutional module is used for lightweight local feature extraction; the SlimUNETR module is used for efficient encoding of local and long-range information; a spatial dimension downsampling operation is performed after each stage. The decoder, symmetrical to the encoder, includes three sets of upsampling paths. Each set of upsampling paths includes, in sequence: a SlimUNETR module, a 3D transposed convolution module, and a VFSCA module. The SlimUNETR module is used to efficiently encode local and long-range information; the 3D transposed convolution module is used for spatial upsampling; and the VFSCA module is used to enhance the semantic information of the fused features. The network output layer consists of a set of 1×1×1 three-dimensional convolutions, which are used to generate the final conductivity reconstruction result map.

5. The conductivity reconstruction solution system according to claim 4, characterized in that, The VFSCA module includes: Spatial attention path, based on the construction of local contextual attention using a 3D sliding window; and a local contextual attention graph is constructed based on the 3D sliding window mechanism to mine the correlation of voxel-level spatial neighborhoods; Channel attention path: It uses shared query and key vectors to construct non-linear attention relationships between different channels, thereby improving the ability to model inter-channel dependencies; The attention fusion layer fuses spatial attention and channel attention through element-wise multiplication or weighted addition to form a joint attention map; the attention fusion layer achieves joint spatial-channel enhancement, improving the global modeling capability of deep features; The convolutional dimensionality reduction layer uses 1×1×1 convolution operations to compress and reconstruct the fused high-dimensional features; The residual connection module is used to perform a weighted summation of the original features and the fused features to improve feature consistency and training stability. The VFSCA module captures the voxel spatial dependencies and inter-channel interactions in 3D medical images through parallel computation of spatial attention paths and channel attention paths, thereby enhancing feature representation capabilities and reconstruction robustness.

6. The conductivity reconstruction solution system according to claim 5, characterized in that, The VFSCA module works as follows: Step a, Generate a spatial attention-enhanced feature map X through a spatial attention path. s Input feature map X to perform shared query and key vector calculation to generate query feature map Q shared Key feature map K shared ; Input feature map X to calculate and generate Value vector V spatial ; Spatial attention enhancement feature map X s for: X s =SA(Q shared ,K shared ,V spatial ) Where SA is the spatial attention module function; Step b, generate channel attention-enhanced feature map X c : Input feature map X to perform shared query and key vector calculation to generate query feature map Q shared Key feature map K shared ; Input feature map X to calculate the value projection weight V of the channel path. channel ; Channel attention enhancement feature map X c for: X c =CA(Q shared ,K shared ,V channel ), Where CA is the channel attention module function; Step c, the attention fusion module outputs X by combining the spatial path and the channel path. s X c The fusion process yields the enhanced comprehensive feature X. fusion ; Step d, residual connection: X out =X fusion +X , Step e, output the fused and enhanced feature map X out Moving on to the next module.

7. An EIT imaging system based on a hybrid approach of tissue information and conductivity sensing, characterized in that, include: Storage module, initial EIT imaging module, organ shape solving module, conductivity reconstruction solving module, feature fusion module; The storage module is used to store the patient's boundary voltage matrix ΔV and the Jacobian matrix J obtained by solving based on idealized prior tissue contour information. The initial EIT imaging module solves for the initial conductivity estimation matrix based on ΔV and J. σ 0, based on σ 0 yields the initial EIT image; Among them, the organ shape solving module establishes the mapping relationship between the initial EIT imaging and the patient's organ shape map; Among them, the conductivity reconstruction solution module establishes the mapping relationship between the initial EIT imaging and the conductivity feature map; Among them, the feature fusion module reads the patient organ shape map output by the organ shape solving module and the conductivity feature map output by the conductivity reconstruction solving module, and outputs a tissue-sensing conductivity distribution map to realize deep collaborative modeling of multimodal features.

8. The EIT imaging system based on a hybrid of tissue information and conductivity sensing as described in claim 7, characterized in that, The organ shape solving module includes a sequentially connected network input layer, an overlapping 3D patch embedding layer, an encoder, a decoder, and a network output layer; The network input layer reads the three-dimensional image reconstructed based on the patient's boundary voltage matrix V and idealized prior tissue contour information; The overlapping 3D patch embedding layer divides the input 3D image into several small blocks; The encoder is composed of several 3D hybrid Transformer modules connected sequentially; the 3D hybrid Transformer module is composed of an efficient self-attention mechanism module, a three-dimensional hybrid feedforward network, and an overlapping 3D patch embedding layer connected sequentially; the encoder extracts local structural details and global semantic features across scales layer by layer to form a multi-scale volume feature representation from coarse to fine. The decoder employs a full multilayer perceptron, which integrates and restores features from different layers through step-by-step upsampling and channel fusion operations, thereby achieving high-resolution, semantically strong structure mask output. The network output layer outputs a high-resolution binary mask to accurately characterize the structural boundaries of biological tissue regions.

9. The EIT imaging system based on a hybrid of tissue information and conductivity sensing as described in claim 7, characterized in that, The conductivity reconstruction solution module adopts a Slim VSCAU-Net3D model, which has an overall U-shaped structure and includes sequentially connected components: The network input layer reads the three-dimensional image reconstructed based on the patient's boundary voltage matrix V and idealized prior tissue contour information, and its input image size is length; The encoder consists of four stages in sequence, each stage including: a 3D depthwise separable convolutional module and a SlimUNETR module; the 3D depthwise separable convolutional module is used for lightweight local feature extraction; the SlimUNETR module is used for efficient encoding of local and long-range information; a spatial dimension downsampling operation is performed after each stage. The decoder, symmetrical to the encoder, includes three sets of upsampling paths. Each set of upsampling paths includes, in sequence: a SlimUNETR module, a 3D transposed convolution module, and a VFSCA module. The SlimUNETR module is used to efficiently encode local and long-range information; the 3D transposed convolution module is used for spatial upsampling; and the VFSCA module is used to enhance the semantic information of the fused features. The network output layer consists of a set of 1×1×1 three-dimensional convolutions, which are used to generate the final conductivity reconstruction result map, i.e., the conductivity feature map.

10. An EIT imaging method based on a hybrid approach of tissue information and conductivity sensing, characterized in that, Includes the following steps: S100, the initial conductivity estimate matrix is ​​solved based on the patient's boundary voltage data ΔV and the Jacobian matrix J obtained from idealized prior tissue contour information. σ 0, based on σ 0 yields the initial EIT image; S200: Input the initial EIT image into the trained organ shape solving module to obtain the patient's organ tissue morphology map; S300: Input the initial EIT image into the trained conductivity reconstruction solution module to obtain the patient's conductivity feature map; S400: Input the organ tissue morphology map obtained in S200 and the conductivity feature map obtained in S300 into the feature fusion module to obtain the tissue-sensing conductivity distribution map.

Images