A breast electrical impedance imaging reconstruction method, device and storage medium

Abstract

The application discloses a breast electrical impedance imaging reconstruction method, equipment and storage medium. The method comprises the following steps: collecting boundary voltage response data of a breast region; based on the data, an initial conductivity image is obtained through forward transformation and inversion reconstruction; a corresponding breast X-ray image is acquired, preprocessed, structural information is extracted and spatial scale is matched, and a structural prior image is constructed; the initial conductivity image and the structural prior image are respectively input into an encoding network, initial conductivity features and structural prior features are extracted; the initial conductivity features and the structural prior features are fused to obtain fused features; and the fused features are input into a decoding network for layer-by-layer upsampling recovery, and a target conductivity distribution image is output. After the introduction of the molybdenum target image prior, the synergistic advantages of the Transformer and the UNet can be more fully utilized, background artifacts in the reconstruction process can be effectively reduced, and the lesion region positioning capability can be improved.

Description

Technical Field

[0001] This invention relates to a medical image reconstruction and bioelectrical impedance imaging technique, and more particularly to a breast electrical impedance imaging reconstruction method that integrates prior mammogram images with a Transformer-UNet network. Background Technology

[0002] Early diagnosis of breast cancer is crucial for improving patient survival rates. Currently, commonly used breast imaging techniques such as mammography, ultrasound, and magnetic resonance imaging (MRI) play an important role in clinical diagnosis. Although mammography is considered the gold standard for breast cancer screening and diagnosis, it is essentially a structural imaging technique and cannot provide functional information related to tissue physiological state and electrical properties. Ultrasound imaging is highly dependent on operator experience, while magnetic resonance imaging (MRI) is expensive and requires sophisticated infrastructure, which limits its application in large-scale early screening to some extent.

[0003] Electrical Impedance Tomography (EIT) is an emerging functional imaging method with advantages such as being non-invasive, low-cost, and portable, showing promising application potential in the early detection of breast diseases. EIT reconstructs the internal conductivity distribution of tissues using boundary voltage measurement data and is highly sensitive to changes in electrical properties caused by lesions. However, due to the strong nonlinearity and severe ill-conditioning of its inverse problem, traditional reconstruction algorithms, including Tikhonov, Newton One-Step Error Reconstructor (NOSER), and Total Variation (TV), still have significant limitations in terms of spatial resolution and quantitative reconstruction accuracy.

[0004] In recent years, deep learning has demonstrated strong effectiveness in addressing nonlinear and ill-conditioned problems in EIT modeling. By learning the nonlinear mapping relationship between boundary voltage measurements and conductivity distribution, deep learning methods can improve reconstruction quality. UNet-based network structures, due to their multi-scale feature extraction and spatial detail preservation capabilities, have been widely applied to EIT image reconstruction. Existing research proposes an improved R-UNet framework to enhance multi-scale representation capabilities, and UNet-based reconstruction and correction strategies can effectively suppress artifacts and improve boundary sharpness and contrast. Furthermore, combining the Kolmogorov-Arnold network with UNet for 3D EIT reconstruction achieves superior structural recovery capabilities and quantitative accuracy. These studies demonstrate that UNet and its variants possess strong robustness and scalability in complex EIT imaging scenarios.

[0005] However, convolutional network structures, due to the localization of the receptive field, have certain shortcomings in capturing long-range dependencies. In contrast, Transformer-based models are more advantageous in modeling global contextual relationships. Research by Chen H et al. shows that the Swin Transformer outperforms traditional convolutional neural networks in segmentation tasks, especially in small target regions and complex boundary detection. These studies indicate that different network structures have their own advantages in feature modeling, and a single network structure struggles to simultaneously achieve a complete representation of global contextual relationships and accurate recovery of local fine-grained details. This problem is particularly prominent in highly ill-conditioned EIT reconstruction tasks. Against this backdrop, existing deep learning-based EIT reconstruction methods can be broadly categorized into two types: end-to-end reconstruction methods, which directly learn the mapping from boundary voltage measurements to conductivity distribution images; and two-stage reconstruction methods, which utilize deep learning models to post-process the initial images obtained by traditional algorithms. However, most of these methods rely only on single-modal information, namely boundary voltage measurement data, failing to fully utilize complementary anatomical information or cross-modal priors.

[0006] To address this issue, researchers have begun exploring multimodal EIT reconstruction strategies, introducing complementary structural or anatomical information to mitigate the pathological nature of the inverse problem. Existing research discloses a kernel-based multimodal EIT reconstruction strategy that effectively fuses information from different modalities and enhances the stability and robustness of EIT imaging. Subsequently, Chen et al. proposed a structural modality-enhanced multimodal EIT imaging method, introducing external structural information constraints during pressure distribution reconstruction, thereby improving spatial localization accuracy and reconstruction consistency.

[0007] Existing multimodal EIT reconstruction methods typically rely on coarse structural priors or simple fusion strategies, lacking an effective mechanism to simultaneously model global contextual dependencies and local fine anatomical details. Although Transformer possesses strong global feature modeling capabilities and UNet exhibits good local detail recovery capabilities, simply fusing Transformer and UNet in breast electrical impedance imaging reconstruction tasks cannot consistently achieve ideal reconstruction results. This is because initial EIT images often suffer from low resolution, blurred lesion boundaries, significant background artifacts, and noise sensitivity. Directly using Transformer for global dependency modeling easily incorporates artifacts and invalid background information into the feature association process, making Transformer susceptible to noise and artifact interference when modeling global associations. While UNet possesses some local detail recovery capabilities, its representation of global structural consistency and long-distance dependencies remains insufficient in complex multi-lesion scenarios, and it struggles to accurately recover complex lesion boundaries without effective structural constraints. Therefore, direct fusion of the two methods easily leads to inaccurate target contour recovery, unclear local boundaries, and insufficient overall structural stability. Summary of the Invention

[0008] Objective of this invention: The objective of this invention is to provide a breast electrical impedance imaging (EIT) reconstruction method that addresses the problems of inaccurate target contour recovery, unclear local boundaries, and insufficient overall structural stability in existing multimodal EIT reconstruction methods. Another objective of this invention is to provide a computer device and storage medium that solves the problem of how to execute a breast electrical impedance imaging reconstruction method.

[0009] Technical solution: The present invention provides a method for breast electrical impedance imaging reconstruction, comprising the following steps: Collect boundary voltage response data of the breast region; Based on the boundary voltage response data, the initial conductivity image is obtained through forward transformation and inversion reconstruction. Obtain the mammogram corresponding to the boundary voltage response data, preprocess the mammogram, extract structural information and perform spatial scale matching to construct a structural prior image corresponding to the initial conductivity image; The initial conductivity image and the prior structure image are input into the encoding network to extract the initial conductivity features and the prior structure features, respectively. The initial conductivity features and structural prior features are fused based on the multi-head self-attention mechanism. Structural prior guiding weights are introduced during the fusion process to obtain the fused features. The fused features are input into the decoding network, and layer-by-layer upsampling is performed to recover the target conductivity distribution image. Preferably, the method for acquiring the boundary voltage response data is as follows: An electrode array is placed around the area of ​​the breast to be tested. A weak alternating current is sequentially applied to the breast tissue through a preset excitation mode, and the boundary voltage signal on each electrode is measured simultaneously to obtain boundary voltage response data.

[0010] Preferably, the forward transform is:

[0011] Where Ω represents the region of interest; function σ ( x This represents the location-dependent conductivity distribution within the region; μ ( x () represents the potential distribution; The inversion and reconstruction method is as follows: Under the linear approximation condition, the change in boundary voltage Δ V With conductivity perturbation Δ σ The following relationship must be satisfied:

[0012] Where, Δ V This represents the voltage difference between the measured boundary voltage and the simulated voltage under the background conductivity condition; J ∈R m×n The Jacobian matrix, denoted by Δ, reflects the sensitivity of voltage measurements to changes in conductivity. σ This represents the conductivity perturbation to be estimated.

[0013] Preferably, the method for preprocessing the mammogram includes at least one of grayscale normalization, noise suppression, region of interest cropping, size resampling, and structural enhancement; The structural information extraction includes: highlighting the boundaries of internal breast tissue and the morphological information of suspected lesions using edge enhancement or contour extraction methods; The spatial scale matching includes adjusting the mammogram to the same spatial resolution as the initial conductivity image through a size unification operation.

[0014] Preferably, the method for extracting the initial conductivity features and prior structural features is as follows: The Transformer coding network was used to extract the multi-scale conductivity distribution features of the initial conductivity image as the initial conductivity features; the Transformer coding network was used to extract the tissue structure and lesion contour features of the structural prior image as the structural prior features. During the encoding process, the input image can be divided into several local blocks and mapped to feature sequences, which are then input into multiple encoding layers for feature learning.

[0015] Preferably, the fusion of initial conductivity features and prior structural features based on a multi-head self-attention mechanism includes: The initial conductivity characteristic F e With structural prior features F p Weighted fusion is performed, and the result is input into a multi-head self-attention module for global modeling to obtain the fused feature F. f The formula is as follows:

[0016] in, λ ∈[0,1] represents the prior guidance weighting factor, used to control the degree of influence of structural prior information in the fusion process; MHSA represents multi-head self-attention operation; The multi-head self-attention mechanism includes the construction process of the query matrix, key matrix, and value matrix. It learns feature dependencies in different subspaces in parallel through multiple attention heads, and its output is expressed as follows:

[0017] in, Q , K and V These represent query, key-value projection, and h, respectively. n Indicates the first n The output of each attention head; The outputs of all attention heads are concatenated along the feature dimension, using a learnable weight matrix. W O Mapping to generate the final output; Each attention head is defined as follows:

[0018] in, d k The dimension scaling factor used for normalization. Q i For the query projection of the i-th attention head, K i Let T be the key-value projection of the i-th attention head, and T be the transposed rank.

[0019] Preferably, the step of inputting the fused features into the decoding network and performing layer-by-layer upsampling recovery includes: The decoding network is a UNet decoding network, which recovers the spatial details of the target image layer by layer by multi-level upsampling and combining the skip connection features output from the corresponding coding layer.

[0020] in, Indicates the firstl The decoder features of the layer, which are derived from the features of the next deeper layer. After upsampling using the upsampling function Up(·), it is then compared with the corresponding encoder features. Obtained by adding the jump connections; After layer-by-layer decoding, the target conductivity distribution image is output.

[0021] This invention introduces mammography images as structural prior information. By preprocessing the mammography images, extracting structural features, and performing spatial scale matching, a structural prior image corresponding to the initial conductivity image is constructed. During the feature fusion stage, this structural prior is used to guide the global modeling process of the Transformer, while simultaneously working with UNet to restore local boundary details. Because mammography images provide clearer breast tissue contours and lesion boundary information, they can effectively suppress erroneous feature associations caused by artifacts, enhance the network's ability to focus on real lesion areas, thereby improving lesion boundary clarity, increasing structural consistency, and enhancing overall reconstruction accuracy and stability.

[0022] Preferably, the above method further includes the following model training step: During training, a composite loss function is used to optimize the network parameters; The composite loss function is expressed as follows:

[0023] in, N This represents the total number of pixels in the image. Indicates the first i The predicted conductivity value for each pixel. σ i This represents the corresponding actual conductivity value, and SSIM represents the structural similarity index. α and β These are the weighting coefficients of the loss function. σ This represents the overall conductivity value. This represents the overall predicted conductivity value.

[0024] A second aspect of the present invention discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0025] A third aspect of the present invention provides a computer storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described method.

[0026] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) This invention proposes a novel structure prior-guided deep learning EIT reconstruction method. This method introduces high-resolution structure priors from mammogram images into a Transformer-UNet hybrid network, combining the global modeling capabilities of Transformer with the local feature extraction capabilities of UNet. The structure priors extracted from mammogram images guide the feature learning process, thereby enhancing the network's ability to represent the internal anatomical structures and lesion areas of breast tissue. Experimental results show that after introducing the mammogram image prior, the synergistic advantages of Transformer and UNet are more fully utilized. The reconstructed images show significant improvements in lesion localization, boundary restoration, and background artifact suppression. It can effectively reduce background artifacts during the reconstruction process, improve the lesion area localization capability, and enhance the structural fidelity and spatial resolution of the reconstructed images.

[0027] (2) This invention introduces high-resolution structural information from breast X-ray images into the breast electrical impedance imaging reconstruction process, and constructs a reconstruction mechanism driven by "electrical measurement information + structural prior information", which can effectively alleviate the reconstruction instability problem caused by the severe pathological nature of the traditional EIT inverse problem.

[0028] (3) The present invention outperforms the classic EIT reconstruction algorithm and existing deep learning methods in terms of structural similarity index and peak signal-to-noise ratio, and can more accurately restore the internal conductivity distribution and morphological characteristics of the lesion area in the breast, thus providing a non-invasive, low-cost and high-precision technical solution for early screening and auxiliary diagnosis of breast tumors. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the hardware system of the present invention; Figure 2 Here is a diagram of the Transformer-UNet network architecture; Figure 3 Comparison of imaging results for different reconstruction methods with varying numbers of targets; Figure 4 Comparison of imaging results of various deep learning EIT reconstruction methods at different training iteration stages; Figure 5 The graph shows the evolution curves of evaluation metrics for different EIT deep learning reconstruction methods during the training process; the left graph shows the change of SSIM with the number of iterations, and the right graph shows the change curve of PSNR. Figure 6 A comparison of conductivity reconstruction results for various electrical impedance imaging reconstruction methods under different noise levels (SNR = 10 dB, 20 dB, and 30 dB); Figure 7The figure shows the distribution of SSIM and PSNR of different EIT reconstruction methods on the test set under the condition of 10 random seed repeated experiments; the left figure is a comparison of box plots on SSIM of each method, and the right figure is a comparison of box plots on PSNR. Figure 8 Comparison of electrical impedance imaging reconstruction results under ablation conditions of different network structures and key modules. Detailed Implementation

[0030] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0031] This invention provides a breast electrical impedance imaging reconstruction method that integrates prior information from mammogram images with a Transformer-UNet network. Based on boundary voltage data acquired by a breast electrical impedance imaging system and combined with prior structural information provided by mammogram images, it reconstructs an image of the conductivity distribution within the breast, thereby improving the spatial resolution, structural fidelity, and reconstruction accuracy of the lesion area. The specific steps are as follows: (1) Building a hardware platform The breast electrical impedance imaging system used in this invention consists of an STM32 core processor, a current source module, a signal amplification circuit, a multiplexing module, and an ADC acquisition module. Its overall structure... Figure 1 As shown, the current source module is built based on a DDS circuit and is powered by a ±7.5V dual power supply to provide a stable and controllable AC excitation current to the device under test (DUT). The voltage signal generated after the current passes through the DUT is first amplified by an AD8421ARZ-R7 to suppress common-mode noise, and then amplified and differential-to-single-ended by an AD8275ARMZ-RL and an INA592IDR to meet the amplitude requirements of subsequent analog-to-digital conversion. The system uses four ADG1406 multiplexers as multiplexing modules for switching control of current injection and voltage acquisition channels. The digital signal after ADC conversion is processed by an STM32 microcontroller and transmitted to the host computer via serial port.

[0032] The breast phantom used in subsequent experiments was made of PLA material, with a height of 45.5 mm and a radius of 60 mm, and filled with porcine fat to simulate the electrical properties of breast tissue. The electrodes were made of copper foil and uniformly distributed on the model surface, with an angular distance of 22.5° between adjacent electrodes to ensure geometric symmetry during current injection and voltage acquisition.

[0033] (2) Boundary voltage response data acquisition and initial conductivity image acquisition First, an excitation current is applied to the breast region to be tested using a breast electrical impedance imaging system, and the corresponding boundary voltage response data is acquired. The breast electrical impedance imaging system includes a current excitation module, a voltage acquisition module, an electrode array module, and a data processing module. During data acquisition, the electrode array is positioned around the breast region to be tested. Weak alternating currents are sequentially applied to the breast tissue using a preset excitation mode, and the boundary voltage signals on each measuring electrode are measured simultaneously to obtain the boundary voltage response data.

[0034] Subsequently, to obtain an initial image suitable for input to a deep learning network, a forward transform is first performed based on the boundary voltage response data, followed by the use of a linearized inversion algorithm or a conventional electrical impedance imaging reconstruction algorithm to obtain the initial conductivity image. The specific process is as follows: The formula for the positive transformation is:

[0035] In the formula, Ω represents the region of interest; the function σ ( x This represents the location-dependent conductivity distribution within the region; μ ( x ) represents the electric potential distribution. This represents the gradient.

[0036] The linearized inversion algorithm is as follows: Under the linear approximation condition, the change in boundary voltage Δ V With conductivity perturbation Δ σ The following relationship must be satisfied:

[0037] In the formula, Δ V This represents the voltage difference between the measured boundary voltage and the simulated voltage under the background conductivity condition; J ∈R m×n The Jacobian matrix represents the sensitivity of voltage measurements to changes in conductivity; Δ σ This represents the conductivity perturbation to be estimated.

[0038] (3) Preprocessing of mammogram images, extraction of structural features and construction of structural priors, specifically: Obtain the mammogram I corresponding to the boundary voltage response data. xray To ensure that mammogram images can be used as effective structural priors in the reconstruction network, preprocessing operations are performed on them. These preprocessing operations include, but are not limited to, grayscale normalization, noise suppression, region of interest cropping, size resampling, and structural enhancement.

[0039] Specifically, the mammogram image can first be normalized to map the pixel value distribution to a uniform range, thereby reducing the grayscale differences caused by different acquisition conditions. Then, the region of interest in the breast is extracted, and the background and irrelevant areas are removed. Next, the image is adjusted to the same spatial resolution as the initial conductivity image through a size unification operation, i.e., spatial scale matching operation. If necessary, edge enhancement or contour extraction can also be used to highlight the boundaries of internal breast tissue and the morphological information of suspected lesions, i.e., structural information extraction operation.

[0040] After preprocessing, a priori structural image corresponding to the initial conductivity image is obtained:

[0041] in, I x Represents a raw mammogram. I eit This represents the initial conductivity image reconstructed based on boundary voltage measurement data. P(·) indicates the preprocessing operation, F(·) indicates the structural information extraction operation, and A(·) indicates the spatial scale matching operation. I p This represents the constructed prior image of the structure.

[0042] The structural prior image mainly contains anatomical information of breast tissue and spatial contour information of lesion areas, which are used to guide the network to focus on areas that better match the actual distribution of breast structure during subsequent reconstruction.

[0043] (4) Dual-input feature encoding Initial conductivity image I eit With structural prior image I p Input the features into the feature encoding network (i.e., the Transformer network) respectively, and extract the corresponding initial conductivity features and structural prior features.

[0044] For the input initial conductivity image, the encoder of the Transformer network extracts its multi-scale conductivity distribution features to characterize the electrical information obtained from the boundary voltage inversion and outputs the initial conductivity features. For the input structural prior image, the encoder of the Transformer network extracts its tissue structure and lesion contour features to characterize the anatomical prior information in the mammogram and outputs structural prior features.

[0045] During the encoding process, the input image can be divided into several local blocks and mapped to feature sequences, which are then input into multiple encoding layers for feature learning.

[0046] After encoding by the Transformer network, the initial conductivity features F are obtained respectively. eand structural prior features F p :

[0047] in, This refers to the encoder that performs feature encoding on the initial conductivity image. This represents the encoding of features from a structural prior image.

[0048] (5) Global modeling and feature fusion based on Transformer.

[0049] To fully utilize the structural prior information provided by mammogram images, a feature fusion mechanism is introduced based on the encoded features.

[0050] Specifically, the initial conductivity characteristic F e With structural prior features F p Weighted fusion is performed, and the result is input into a multi-head self-attention module for global modeling to obtain the fused feature F. f :

[0051] In the formula, λ ∈[0,1] represents the prior guidance weighting factor, which is used to control the degree of influence of structural prior information in the fusion process; MHSA represents multi-head self-attention operation.

[0052] The multi-head self-attention mechanism includes a query matrix. Q Key matrix K Sum matrix V The construction process involves learning feature dependencies in different subspaces in parallel using multiple attention heads. Its output can be represented as:

[0053] In the formula, Q , K and V These represent query, key, and value projections, respectively; h n Indicates the first n The output of the attention head, concat, means concatenating the above three tensors together in the same dimension to generate a larger tensor.

[0054] The outputs of all attention heads are concatenated along the feature dimension, and then passed through a learnable weight matrix. W O Map to generate the final output.

[0055] Each attention head is defined as follows:

[0056] in, d k The dimension scaling factor used for normalization. Q i For the query projection of the i-th attention head, K i Let T be the key-value projection of the i-th attention head, and T be the transposed rank.

[0057] Through the above mechanism, the network can establish a dependency relationship between electrical characteristics and structural prior characteristics on a global scale, thereby enhancing the comprehensive representation ability of the overall morphology of breast tissue, lesion location and complex boundaries.

[0058] (6) Multi-scale recovery and reconstruction output based on UNet decoder F fusion feature F f The input is processed by a UNet decoding network, which performs layer-by-layer upsampling to recover the target conductivity distribution image. To balance global semantic information and local spatial details, a skip connection mechanism is incorporated into the decoding stage to pass shallow detail features from the corresponding coding layers to the decoder, thereby enhancing the recovery of edges and local structures. l Layer decoding features can be represented as:

[0059] In the formula, Indicates the first l The decoder features of the layer, which are derived from the features of the next deeper layer. After upsampling using the upsampling function Up(·), it is then compared with the corresponding encoder features. It is obtained by adding the jump connections.

[0060] After layer-by-layer decoding, the final output is the target conductivity distribution image. σ Compared to the initial conductivity image, this output image demonstrates superior performance in lesion contour clarity, background artifact suppression, and structural consistency. The overall network structure is as follows: Figure 2 As shown.

[0061] (7) Network training and loss function design To improve the pixel-level accuracy and structural consistency of the reconstruction results, a composite loss function is used to optimize the network parameters during training. This composite loss function consists of a pixel-level error term and a structural similarity constraint term, and can be expressed as:

[0062] in, N This represents the total number of pixels in the image. Indicates the first i The predicted conductivity value for each pixel.σ i This represents the corresponding actual conductivity value, and SSIM represents the structural similarity index. α and β These are the weighting coefficients of the loss function. σ This represents the overall conductivity value. This represents the overall predicted conductivity value.

[0063] (8) Output and application of reconstruction results.

[0064] After training, the boundary voltage response data of the test sample and the corresponding mammogram image are input into the trained reconstruction network (i.e., the Transformer-UNet hybrid network structure mentioned above), which outputs an image of the conductivity distribution within the breast. Based on the differences in conductivity among different tissues and lesions, suspected lesion areas can be located and displayed. The reconstruction results can be used for early screening, auxiliary diagnosis, and lesion area analysis of breast tumors.

[0065] Comparison of imaging results of different reconstruction methods with different numbers of targets: The method employed in this invention is comprehensively compared with traditional regularized reconstruction methods Tikhonov, NOSER, and TV, as well as typical deep learning reconstruction methods UNet-EIT, Deep-EIT, and cGAN-EIT. Among these, the Tikhonov method is a classic example. L 2. Regularized reconstruction methods [1], NOSER method is an electrical impedance imaging algorithm based on Newton one-step error reconstructor [2], TV method enhances boundary preservation capability by introducing total variational regularization term [3]; UNet-EIT is an EIT reconstruction model based on U-shaped convolutional encoder-decoder structure [4], Deep-EIT is a deep learning reconstruction network designed for electrical impedance imaging tasks [5], and cGAN-EIT is an EIT image reconstruction method based on conditional generative adversarial network [6]. References are as follows: [1] B. Sun, S. Yue, Z. Hao, et al., "An improved Tikhonovregularization method for lung cancer monitoring using electrical impedancetomography," Sensors, vol. 19, no. 8, pp. 3049–3057, 2019. [2] A. Abbasi and BV Vahdat, “A non-iterative linear inversesolution for the block approach in EIT,” J. Comput. Sci., vol. 1, no. 4, pp.190–196, 2010. [3] DM Nguyen, P. Qian, T. Barry, et al., "Self-weighted NOSER-prior electrical impedance tomography using internal electrodes in cardiacradiofrequency ablation," Physiol. Meas., vol. 40, no. 6, Art. no. 065006, 2019. [4] Cao H, Wang Y, Chen J, et al. Swin-unet: Unet-like puretransformer for medical image segmentation[C] / / European conference oncomputer vision. Cham: Springer Nature Switzerland, 2022: 205-218. [5] Wang Z, Zhang T, Zhang X, et al. Unsupervised Sensitivity PriorGeneration With Diffusion Model for EIT Image Reconstruction[J]. IEEETransactions on Industrial Informatics, 2026. [6] Deng J, Pang G, Zhang Z, et al. cGAN based facial expression recognition for human-robot interaction[J]. IEEE Access, 2019, 7: 9848-9859. Under the same experimental conditions, the differences between the various methods in terms of reconstruction accuracy, error performance, and computational complexity were analyzed. The results are as follows:Figure 3 As shown. By Figure 3 As can be seen, in Case 1 and Case 2, PriorEIT-TransUNet in this invention can accurately recover the target location and conductivity distribution, with clear boundaries in high conductivity regions, smooth backgrounds, and fewer artifacts. In contrast, the Tikhonov and NOSER methods exhibit significant boundary blurring and speckle artifacts. Although the TV method improves edge structure recovery, it still suffers from insufficient detail recovery in multi-target scenarios. Among deep learning methods, UNet-EIT, Deep-EIT, and cGAN-EIT are prone to target location shifts or background fluctuations under complex distribution conditions, while PriorEIT-TransUNet demonstrates more stable structure reconstruction performance. Quantitative results show that this method achieves the highest or near-highest SSIM and PSNR values ​​in both scenarios. Furthermore, in Case 2, PriorEIT-TransUNet maintains stable and accurate reconstruction performance.

[0066] Comparing the imaging results of various deep learning EIT reconstruction methods at different training iteration stages, such as... Figure 4 As shown, Figure 4 From left to right, they are: PriorEIT-TransUNet, UNet-EIT, Deep-EIT, and cGAN-EIT. Figure 5 Evolution curves of evaluation metrics for different EIT deep learning reconstruction methods during the training process. Figure 5 The left-middle figure shows the change of SSIM with the number of iterations, and the right-middle figure shows the change curve of PSNR.

[0067] Considering that the performance of PriorEIT-TransUNet in this invention may vary depending on the training phase, Figure 4 Further analysis was conducted on the reconstruction quality at different training epochs. The results show that in the initial training phase (Epoch = 200), all methods exhibited some artifacts and structural distortion. UNet-EIT and cGAN-EIT produced more pronounced oscillation artifacts, while PriorEIT-TransUNet demonstrated relatively stable structural reconstruction capabilities. When the training epoch increased to Epoch = 400, the reconstruction quality significantly improved, exhibiting clearer target boundaries, smoother backgrounds, and superior SSIM and PSNR compared to the comparative methods.

[0068] To further examine its performance changes during the later training phase, Figure 5The changes in evaluation metrics over 600 training epochs are presented. PriorEIT-TransUNet achieves optimal performance at approximately Epoch = 397, with a peak SSIM of 0.8971 and a PSNR of 25.8093 dB.

[0069] The results above demonstrate that the Transformer-Unet hybrid structure with fusion structure priors can effectively balance global contextual information and local detailed features during the mid-training phase, thereby achieving optimal reconstruction results.

[0070] The conductivity reconstruction results of various electrical impedance imaging reconstruction methods under different noise levels (SNR = 10 dB, 20 dB, and 30 dB) are compared, and the results are as follows: Figure 6 As shown.

[0071] Depend on Figure 6 As can be seen, the reconstruction quality of all methods improves with increasing signal-to-noise ratio (SNR). However, the robustness of each method to noise varies significantly. Traditional methods typically exhibit blurred boundaries and noticeable background artifacts under high noise conditions, and their ability to reconstruct multi-target structures remains limited even under high SNR conditions. Deep learning methods generally outperform traditional algorithms. Specifically, UNet-EIT and Deep-EIT can locate the main target under moderate noise conditions, but they still suffer from structural distortion or over-smoothing under low or high SNR conditions. Although cGAN-EIT can achieve relatively high PSNR in some scenarios, it is often accompanied by significant local artifacts and boundary inconsistencies. In contrast, PriorEIT-TransUNet exhibits more stable reconstruction performance under different noise levels. Under high noise conditions, this method effectively suppresses artifacts; under low noise conditions, it still maintains strong structural consistency and clear boundary features. Quantitative results show that the proposed method achieves the highest or near-highest SSIM and PSNR values ​​under various signal-to-noise ratio conditions, verifying the effectiveness of the Transformer-UNet architecture with fused structural priors in improving EIT reconstruction accuracy and noise robustness.

[0072] Under the condition of 10 random seed replication experiments, the SSIM and PSNR distributions of different EIT reconstruction methods on the test set were compared, and the results are as follows: Figure 7 As shown, Figure 7 The left image shows box plots on SSIM for each method, while the right image shows a comparison of PSNR box plots.

[0073] Figure 7The results show that the differences in SSIM and PSNR among different reconstruction methods are generally small, with performance improvements mainly reflected in more stable quantitative accuracy and better preservation of fine-grained structural details. To reduce the impact of randomness in a single experiment on the conclusions and to verify the statistical consistency of the proposed method's performance, repeated experiments were conducted using different random seeds. From the SSIM distribution, PriorEIT-TransUNet achieved the highest median and the most concentrated distribution, indicating that it can maintain high structural similarity relatively stably under different random initialization conditions. In contrast, UNet-EIT and Deep-EIT have a wider SSIM distribution range and are more sensitive to random seeds. Although cGAN-EIT can achieve high SSIM in some experiments, its overall stability remains limited. Regarding the PSNR, PriorEIT-TransUNet also exhibits the highest median and the smallest variance, while Deep-EIT and cGAN-EIT show significant fluctuations. The combined results of the two metrics show that PriorEIT-TransUNet exhibits better stability and repeatability under random perturbation conditions, further verifying the robustness of the proposed network architecture in EIT reconstruction tasks.

[0074] To evaluate the contribution of each component to the reconstruction performance, ablation experiments were conducted to compare the electrical impedance tomography reconstruction results under different network structures and key module ablation conditions. The methods are as follows: Except for the module being evaluated, all other training settings remained consistent, including data partitioning, input resolution, optimizer and learning rate strategy, and stopping criteria. Ablation targets included structural prior inputs and composite loss function terms. All settings were evaluated on the same test set, using SSIM and PSNR as performance metrics. First, the impact of structural prior inputs on imaging performance was analyzed. Representative reconstructed images were selected and compared with corresponding results without structural prior input; the results are shown below. Figure 8 As shown.

[0075] Depend on Figure 8As can be seen, under the complete model configuration, PriorEIT-TransUNet can accurately reconstruct the spatial location and shape of multiple lesions, with clearer boundaries and fewer artifacts, and its SSIM and PSNR are significantly higher than the comparative methods. When the structural prior is removed, the reconstruction quality of all methods decreases, mainly manifested in increased background noise and blurred lesion contours. However, PriorEIT-TransUNet still maintains relatively high stability, indicating that it still has a certain degree of robustness when prior information is lacking. In contrast, when the composite loss function is replaced with the basic pixel-wise L1 loss, although the reconstruction results can still roughly locate the lesions, there are problems of decreased contrast and deterioration of structural fidelity. This phenomenon indicates that the proposed composite loss function plays an important role in enhancing structural consistency and suppressing reconstruction artifacts.

[0076] In summary, this invention is the first to introduce prior structural information from mammogram images into the EIT deep learning reconstruction framework, and combines the Transformer's ability to model global features to establish an effective mapping relationship between electrical measurement data, structural constraint information and conductivity distribution, thereby significantly enhancing the model's ability to represent global consistency and long-distance dependencies.

Claims

1. A method for breast electrical impedance imaging reconstruction, characterized in that, Includes the following steps: Collect boundary voltage response data of the breast region; Based on the boundary voltage response data, the initial conductivity image is obtained through forward transformation and inversion reconstruction. Obtain the mammogram corresponding to the boundary voltage response data, preprocess the mammogram, extract structural information and perform spatial scale matching to construct a structural prior image corresponding to the initial conductivity image; The initial conductivity image and the prior structure image are input into the encoding network to extract the initial conductivity features and the prior structure features, respectively. The initial conductivity features and structural prior features are fused based on the multi-head self-attention mechanism. Structural prior guiding weights are introduced during the fusion process to obtain the fused features. The fused features are input into the decoding network, and layer-by-layer upsampling is performed to recover the target conductivity distribution image.

2. The breast electrical impedance imaging reconstruction method according to claim 1, characterized in that, The method for acquiring the boundary voltage response data is as follows: An electrode array is placed around the area of ​​the breast to be tested. A weak alternating current is applied to the breast tissue in sequence through a preset excitation mode, and the boundary voltage signal on each electrode is measured simultaneously to obtain boundary voltage response data.

3. The breast electrical impedance imaging reconstruction method according to claim 1, characterized in that, The positive transform is: Where Ω represents the region of interest; function σ ( x This represents the location-dependent conductivity distribution within the region; μ ( x () represents the potential distribution; The inversion and reconstruction method is as follows: Under the linear approximation condition, the change in boundary voltage Δ V With conductivity perturbation Δ σ The following relationship must be satisfied: Where, Δ V This represents the voltage difference between the measured boundary voltage and the simulated voltage under the background conductivity condition; J ∈R m×n The Jacobian matrix, denoted by Δ, reflects the sensitivity of voltage measurements to changes in conductivity. σ This represents the conductivity perturbation to be estimated.

4. The breast electrical impedance imaging reconstruction method according to claim 1, characterized in that, The method for preprocessing mammogram images includes at least one of grayscale normalization, noise suppression, region of interest cropping, size resampling, and structural enhancement. The structural information extraction includes: highlighting the boundaries of internal breast tissue and the morphological information of suspected lesions using edge enhancement or contour extraction methods; The spatial scale matching includes adjusting the mammogram to the same spatial resolution as the initial conductivity image through a size unification operation.

5. The breast electrical impedance imaging reconstruction method according to claim 1, characterized in that, The method for extracting the initial conductivity features and prior structural features is as follows: The Transformer coding network was used to extract the multi-scale conductivity distribution features of the initial conductivity image as the initial conductivity features; the Transformer coding network was used to extract the tissue structure and lesion contour features of the structural prior image as the structural prior features. During the encoding process, the input image can be divided into several local blocks and mapped to feature sequences, which are then input into multiple encoding layers for feature learning.

6. The breast electrical impedance imaging reconstruction method according to claim 1, characterized in that, The fusion of initial conductivity features and prior structural features based on the multi-head self-attention mechanism includes: The initial conductivity characteristic F e With structural prior features F p Weighted fusion is performed, and the result is input into a multi-head self-attention module for global modeling to obtain the fused feature F. f The formula is as follows: in, λ ∈[0,1] represents the prior guidance weighting factor, used to control the degree of influence of structural prior information in the fusion process; MHSA represents multi-head self-attention operation; The multi-head self-attention mechanism includes the construction process of the query matrix, key matrix, and value matrix. It learns feature dependencies in different subspaces in parallel through multiple attention heads, and its output is expressed as follows: in, Q , K and V These represent query, key-value projection, and h, respectively. n Indicates the first n The output of each attention head; The outputs of all attention heads are concatenated along the feature dimension, using a learnable weight matrix. W O Mapping to generate the final output; Each attention head is defined as follows: in, d k The dimension scaling factor used for normalization. Q i For the query projection of the i-th attention head, K i Let T be the key-value projection of the i-th attention head, and T be the transposed rank.

7. The breast electrical impedance imaging reconstruction method according to claim 1, characterized in that, The step of inputting the fused features into the decoding network and performing layer-by-layer upsampling recovery includes: The decoding network is a UNet decoding network, which recovers the spatial details of the target image layer by layer by multi-level upsampling and combining the skip connection features output from the corresponding coding layer. in, Indicates the first l The decoder features of the layer, which are derived from the features of the next deeper layer. After upsampling using the upsampling function Up(·), it is then compared with the corresponding encoder features. Obtained by adding the jump connections; After layer-by-layer decoding, the target conductivity distribution image is output.

8. The breast electrical impedance imaging reconstruction method according to claim 1, characterized in that, It also includes the following model training steps: During training, a composite loss function is used to optimize the network parameters; The composite loss function is expressed as follows: in, N This represents the total number of pixels in the image. Indicates the first i The predicted conductivity value for each pixel. σ i This represents the corresponding actual conductivity value, and SSIM represents the structural similarity index. α and β These are the weighting coefficients of the loss function. σ This represents the overall conductivity value. This represents the overall predicted conductivity value.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-8.

10. A computer storage medium, characterized in that, The computer storage medium stores instructions that, when executed on the computer, cause the computer to perform the method as described in any one of claims 1-8.

Images