A three-dimensional lung ventilation electrical impedance imaging method and system

Abstract

The application discloses a three-dimensional lung ventilation electrical impedance imaging method and system, and belongs to the technical field of electrical impedance imaging. The method comprises the following steps: obtaining denoised voltage measurement values by using a residual denoising network; sending random input quantities into a double-output neural network with an encoder and a decoder; attaching a full-connection branch to the output end of the encoder, outputting a latent variable, and generating a lung mask by using a lung mask generator, so as to serve as a generative model prior; the decoder generates a lung ventilation image, and the double-output network structure itself provides implicit regularization; based on the denoised voltage measurement values, the lung shape mask and the estimated boundary voltage, an image reconstruction loss function comprising a data consistency term and a lung shape loss term is adopted, the double-output neural network parameters are iteratively optimized by using the reconstruction loss function, and image reconstruction is performed on each instance in an unsupervised manner. The method improves the robust performance and generalization performance of lung ventilation imaging, and improves the image quality.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electrical impedance tomography, and particularly relates to a three-dimensional lung ventilation electrical impedance tomography method and system. BACKGROUND

[0002] Precise and real-time lung ventilation information is crucial for guiding mechanical ventilation, assessing lung function, and supporting clinical decision-making. Electrical impedance tomography (EIT) has become an important clinical bedside imaging technique due to its non-invasive, non-radiation, real-time, and portable advantages. In particular, EIT can reconstruct regional and dynamic lung ventilation images, providing real-time lung function information, making it a valuable tool for lung function assessment, personalized mechanical ventilation, acute respiratory distress syndrome monitoring, and chronic obstructive pulmonary disease detection. Most current clinical EIT applications are based on two-dimensional EIT, while three-dimensional (3D) EIT can represent more comprehensive spatial information and thus provide more accurate pathological change information, so 3D EIT has greater clinical potential. However, EIT image reconstruction needs to solve the inherent ill-posed inverse problem. In 3D EIT, the number of volumes to be reconstructed increases dramatically due to the increase in dimension, further exacerbating the ill-posedness. In addition, image reconstruction is highly sensitive to measurement noise, which can also severely degrade the quality of the reconstructed image.

[0003] Traditional regularization reconstruction methods have a complete mathematical framework, but the reconstruction results often have low spatial resolution and severe artifacts. Deep learning (DL) models can utilize prior knowledge in large datasets for complex representation, so more and more reconstruction methods use DL technology to improve the quality of reconstructed images. However, DL-based EIT faces the challenge of requiring large-scale paired "boundary voltage-conductivity distribution" datasets, which are difficult to obtain in clinical practice. Therefore, unsupervised reconstruction strategies are increasingly attracting attention. Current unsupervised EIT image reconstruction methods can be divided into three categories, one of which is the compressed sensing method based on a generative model (CSGM). The generative model is pre-trained to learn the conductivity distribution prior knowledge in a specific scenario, and then an optimization method based on compressed sensing is constructed for image reconstruction. However, when the target conductivity distribution is different from that in the pre-training stage, the performance of CSGM will decrease. Another class combines data-driven learning with physical models, incorporating linearized approximations or partial differential equation form forward models into the loss function. The reconstruction neural network is trained on large-scale real measurement data to utilize its nonlinear representation ability to construct a mapping from boundary measurements to conductivity distribution, but this method still requires a large amount of real boundary voltage data. SUMMARY

[0004] The present application provides a three-dimensional lung ventilation electrical impedance imaging method, which combines deep image prior and generative model prior for image reconstruction, can provide clearer lung boundary, reduce image artifacts, and improve three-dimensional lung ventilation imaging quality.

[0005] The three-dimensional lung ventilation electrical impedance imaging method provided by the present application comprises:

[0006] S1, using a residual denoising network to denoise the boundary voltage measurement value to obtain a denoised voltage measurement value;

[0007] S2, sending random input Gaussian noise into a double-output neural network based on an encoder and a decoder structure, the encoder output end is additionally connected with a full connection branch, outputs a latent variable, and generates a lung mask through a lung mask generator as a generative model prior; the decoder generates a lung ventilation image, and the double-output neural network structure itself provides implicit regularization to guide the reconstructed image, and the reconstructed image is estimated through an EIT forward model to obtain a boundary voltage;

[0008] S3, the VAE decoder generates a lung shape mask according to the latent variable; based on the denoised voltage measurement value, the lung shape mask and the estimated boundary voltage, an image reconstruction loss function is used, which comprises a data consistency term and a lung shape loss term , and the double-output neural network parameters are iteratively optimized through the reconstruction loss function to reconstruct the lung ventilation image of each instance in an unsupervised manner.

[0009] The present application also provides a three-dimensional lung ventilation electrical impedance imaging system, which comprises:

[0010] A residual network module is used for denoising the boundary voltage measurement value in the data preprocessing stage;

[0011] A generative model module is used for generating a lung shape mask, and the module is pre-trained by using a lung binary three-dimensional image data set constructed by public chest CT data;

[0012] A double-output convolutional neural network module is used for reconstructing a lung ventilation image and predicting a latent variable input of a lung shape mask generator;

[0013] A loss calculation module is used for iteratively optimizing training to update the lung ventilation image.

[0014] The present application has the following beneficial effects compared with the prior art:

[0015] The mixed depth image prior and the generation model prior are used for bioelectric impedance imaging, the depth image prior guides the image to the natural structure by applying implicit regularization, the lung mask generator is constructed based on the latent variable, and the reconstructed image is explicitly constrained to have physiological rationality. Through data consistency loss and shape constraint, random initialization and iterative reconstruction are performed on each sample in an unsupervised manner. The development and design are carried out for clinical practice, and the imaging process does not depend on a large amount of labeled data, thereby improving the imaging quality and generalization performance. In addition, the residual denoising network as a signal preprocessing module effectively suppresses the measurement noise, thereby improving the robustness and stability of image reconstruction. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is a general block diagram of the present application;

[0017] Figure 2 It is a schematic diagram of 3D lung ventilation EIT excitation and measurement;

[0018] Figure 3 It is a schematic diagram of constructing a lung binary three-dimensional image data set;

[0019] Figure 4 It is a flow chart of lung ventilation image iterative reconstruction;

[0020] Figure 5 It is a three-dimensional lung ventilation image reconstruction result graph. DETAILED DESCRIPTION

[0021] The embodiments of the technical scheme of the present application will be described in detail below with reference to the drawings. Unless otherwise specified, the technical terms or scientific terms used in the present application are the general meanings understood by the skilled in the art.

[0022] Example 1, reference Figure 1 The imaging method combined in this embodiment includes a residual network module, a generation model module, a double-output convolutional neural network module and a loss calculation module.

[0023] The residual denoising network module is used for denoising the boundary voltage measurement value in the data preprocessing stage.

[0024] The generation model module is used for generating a lung mask, and the module is pre-trained using a lung binary three-dimensional image data set constructed using public chest CT data;

[0025] The double-output convolutional neural network module is used for generating a lung ventilation image and predicting the latent variable input of the lung shape mask generator;

[0026] The loss calculation module is used for iterative optimization of the double-output neural network parameters and reconstruction of the lung ventilation image.

[0027] The technical scheme adopted is as follows:

[0028] Based on the circuit reciprocity theorem, the original boundary voltage value of the EIT single measurement is divided into two corresponding complementary subsets and , as paired noise voltage data, for denoising network training, wherein the i, j electrode pair is excited, and the p, q electrode pair is measured, and the i, j electrode pair is excited, and the p, q electrode pair is measured, and the i, j electrode pair is excited, and the p, q electrode pair is measured.

[0029] Denoising is performed by a residual Noise2Noise (R-N2N) denoising network, and the boundary voltage measurement value is denoised in the data preprocessing stage, thereby improving the robustness of subsequent image reconstruction. The residual Noise2Noise denoising network uses parallel multi-scale dilated convolution and residual connection to capture different scale features of the one-dimensional input voltage signal ; wherein five different size convolution kernels (3, 5, 7, 9, 11) are used to process the input voltage signal in parallel, the first three are ordinary convolution, and the last two are expanded by dilated convolution without significantly increasing the parameters; after the channel dimension of all parallel outputs is spliced, four standard convolution layers are used for feature fusion and transformation; the original input and the noise features learned by the network are added through a residual connection, and the final network output is the noise component in the noise component. The training target is:

[0030]

[0031] The trained R-N2N denoiser is used to predict noise, and the noise component is removed to obtain the denoised voltage measurement value , and the denoised signal is used for subsequent EIT image reconstruction:

[0032]

[0033] wherein, represents a residual network, represents the parameters of the network, represents the expected mathematical symbol.

[0034] The lung shape mask generator is implemented by a VAE (based on a variational autoencoder), and the VAE includes an encoder and a decoder, which needs to be pre-trained. When pre-training, the lung image processed by the public data set is input into the encoder, and the encoder outputs the latent variable , The decoder outputs the lung mask; minimize and loss between the training VAE, the trained VAE decoder part is used as a lung mask generator for generating lung mask. In the lung ventilation image reconstruction process, only the lung mask generator is used, and the input of the lung mask generator is the latent variable output by the dual-output neural network , and the output is the lung mask.

[0035] The VAE structure can be built according to actual needs. For example: the encoder of the VAE extracts the features of the three-dimensional lung image through the step convolution layer of the three-dimensional convolutional neural network, and then projects the features into the 128-dimensional latent space through the fully connected layer. The decoder mirrors the encoder through the transpose convolution.

[0036] In the training process of the VAE, based on the lung binary three-dimensional image dataset, the VAE is trained through binary cross entropy and KL divergence:

[0037]

[0038] In the formula, represents the total loss for training the VAE, and represents the binary three-dimensional lung image input into the VAE and the VAE output, represents the binary cross entropy loss, represents the KL divergence loss, is the encoder, is a standard Gaussian distribution, is the latent variable, is a hyperparameter.

[0039] In this embodiment, the disclosed chest CT data is used to construct a lung binary three-dimensional image dataset to pre-train the VAE. The image features are extracted from the disclosed CT data and the latent representation of the lung shape is learned, wherein each CT image uses a standardized process to obtain a binary lung mask: cropping, segmentation, removing small non-connected domains, forming entities, surface smoothing, binarization and direction alignment, and then resampling to 32x32x32 images. The constructed dataset is used to pre-train the VAE, and the decoder of the VAE is used as a generative model prior (GMP) in the image reconstruction process.

[0040] In the lung ventilation image reconstruction stage, a dual-output convolutional neural network based on an encoder-decoder structure is constructed, and the dual-output convolutional neural network takes the fixed random variable as the input of the network, propagates through the encoder-decoder with 3D convolutional layers, pooling and skip connection, generates a 3D lung conductivity image matched with the measurement voltage, wherein a fully connected branch is added to the output end of the encoder, and the output is the latent variable , a deep image prior (DIP) with a dual-output convolutional neural network is imposed to make the reconstructed image natural and smooth, and then a generative model prior (GMP) is imposed to constrain the lung shape. The output of the dual-output convolutional neural network is represented as:

[0041]

[0042] wherein, is the reconstructed lung ventilation image, denotes a randomly initialized neural network, is a random variable inputting the network, representing a random input Gaussian noise;

[0043] The latent variable is sampled in the learned VAE latent space, and a 3D lung binary image is generated by a lung shape mask generator, which is used as a GMP to constrain the reconstructed lung ventilation image. Both the lung mask and the predicted conductivity image are normalized to the same range to ensure stable optimization and effective backpropagation.

[0044] In the process of reconstructing the lung ventilation image, the parameters of the VAE decoder and the R-N2N denoiser network are fixed, and the R-N2N denoiser network is used to denoise the to obtain the denoised voltage measurement The VAE decoder generates a lung mask according to the latent variable The parameters of the dual-output neural network (DOCNN) are iteratively optimized to minimize the loss function, and the measured value The lung ventilation image is iteratively reconstructed in an unsupervised manner without the need for real conductivity images for training. The loss function includes a data consistency term and a lung shape loss term, wherein the data consistency term estimates the boundary voltage of the reconstructed conductivity image through the EIT forward model , and calculates the Euclidean norm between the voltage measurement and the denoised voltage measurement The loss function is defined as:

[0045]

[0046]

[0047] wherein, denotes the lung shape loss, denotes the reconstructed lung ventilation image, denotes the lung shape mask generator, denotes the inner product at the voxel level, is a small constant that ensures numerical stability. represents the EIT forward model, represents the final loss for image reconstruction, and λ1 is used to balance the data consistency term and the regularization term.

[0048] Referring to Figure 1 , the embodiment provides a three-dimensional lung ventilation electrical impedance imaging method combining a deep image prior and a generative model prior, mainly including R-N2N denoising, generative model prior construction, and image reconstruction based on a mixed deep image prior and generative model prior.

[0049] The method is described in combination with the accompanying drawings:

[0050] Step one: as shown in Figure 2 , a 32-electrode EIT system collects boundary voltages in a calibration stage through an adjacent excitation and an adjacent measurement mode, and each frame of data contains 928 measured voltage values. Based on the circuit reciprocity theorem, the original boundary voltage values measured by the 32-electrode EIT system in a single measurement are divided into two corresponding complementary subsets and , as shown in Figure 1 (a), where the i, j electrode pair is excited, and the p, q electrode pair is measured, and the i, j electrode pair is excited, and the p, q electrode pair is measured. The paired noise data and are used for residual denoising network training.

[0051] Step two: as shown in Figure 1 (a), , denoising is performed through an R-N2N denoising network. Wherein different scale features in are extracted by using a parallel multi-scale dilated convolution and a residual connection structure, five different size convolution kernels (3, 5, 7, 9, 11) are used to process the input in parallel, the first three are ordinary convolutions, and the last two are dilated convolutions that expand the receptive field without significantly increasing the parameters, after the parallel outputs are spliced in the channel dimension, the features are fused through four standard convolutions, the original input and the noise features learned by the network are added through the residual connection, and the final network output is the noise component in . The training target is:

[0052]

[0053] wherein, represents a residual network, represents the parameters of the network

[0054] Step three: as shown in Figure 3The binary three-dimensional lung image dataset was constructed using the disclosed chest CT data. Each CT image was binarized using a standardized pipeline: cropping, segmentation, removal of small non-connected regions, solidification, surface smoothing, binarization, and orientation alignment, and then resampled to a 32x32x32 three-dimensional image. The VAE was pre-trained based on this dataset.

[0055] Step four: As shown in Figure 1 (b), the VAE model was pre-trained. The encoder of the VAE extracted the features of the three-dimensional lung image through a three-dimensional convolutional neural network, and then projected the features into a 128-dimensional latent space through a fully connected layer. The decoder mirrored the encoder through transposed convolution. The VAE was trained through binary cross-entropy and KL divergence:

[0056]

[0057] wherein, represents the total loss for training the VAE, and represents the binary three-dimensional lung image input into the VAE and the VAE output, represents the binary cross-entropy loss, represents the KL divergence loss, is a standard Gaussian distribution, is a hyperparameter. The decoder of the VAE model is used as a lung shape mask generator.

[0058] Step five: As shown in Figure 1 (c), a fixed random 3D tensor z is input into the dual-output convolutional neural network, which is propagated through an encoder-decoder with 3D convolutional layers, pooling, and skip connections to generate a 3D lung conductivity image matching the measured voltage. A fully connected branch is added to the output end of the encoder to output the latent variable The DIP is used to impose an implicit prior, making the reconstructed image natural and smooth, and then combined with the generative model prior to impose an explicit lung shape constraint. The output of the dual-output convolutional neural network is represented as:

[0059]

[0060] wherein, is the reconstructed lung ventilation image, represents a randomly initialized neural network, represents a random variable input into the network.

[0061] Step six: As shown in Figure 1 (c), the latent variable A 3D lung mask image is generated by a lung shape mask generator as a GMP to constrain the reconstructed ventilation image. Both the lung mask and the predicted conductivity image are normalized to the same range, and the lung shape loss is defined as:

[0062]

[0063] where, represents the inner product at the voxel level, is a small constant to ensure numerical stability.

[0064] Step seven: The loss function is designed to contain a data consistency term and a lung shape loss term. The data consistency term calculates the L2 norm loss between the reconstructed conductivity image and the denoised voltage measurement estimated by the forward model The loss function is defined as:

[0065]

[0066] where λ1 is used to balance the data consistency term and the regularization term.

[0067] Step eight: As shown in Figure 4 , the R-N2N denoiser network and the lung mask generator parameters are fixed, and the lung ventilation image is updated by iteratively optimizing the parameters of the dual-output neural network to minimize the loss function from the denoised voltage Unsupervised iterative reconstruction of lung ventilation images.

[0068] Figure 5 A three-dimensional lung ventilation reconstruction image of a test case is shown. The first row is the ground truth (GT) image, and the reconstructed image is shown in axial slices with a step size of 4 from 2 to 30. It can be seen that the proposed method provides clearer lung boundaries and reduces image artifacts compared to other methods, and is closer to the ground truth.

[0069] The above has been disclosed as a preferred embodiment of the present application, but is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed structure and technical content without departing from the scope of the present application, and equivalent embodiments with equivalent changes are still within the scope of the present application.

Claims

1. A three-dimensional lung ventilation electrical impedance imaging method, characterized in that: Includes the following steps: S1. Use a residual denoising network to denoise the boundary voltage measurement value to obtain the denoised voltage measurement value; S2. Random Gaussian noise is fed into a dual-output neural network based on an encoder and decoder structure. The encoder output is supplemented with a fully connected branch to output latent variables, which are then used by a lung mask generator to generate a lung mask as a prior for the generative model. The decoder generates a lung ventilation image. The dual-output neural network structure itself provides implicit regularization to guide the reconstruction image. The reconstructed image is used to estimate the boundary voltage using the EIT forward model. S3. The lung mask generator is implemented by the VAE decoder, which generates the lung mask based on latent variables. Based on the denoised voltage measurement, the lung mask, and the estimated boundary voltage, an image reconstruction loss function including a data consistency term and a lung shape loss term is used. By reconstructing the loss function The parameters of the dual-output neural network are optimized using minimum iterations to reconstruct lung ventilation images for each instance in an unsupervised manner. The VAE consists of a three-dimensional convolutional neural network that extracts features from three-dimensional lung images through convolution, and then projects these features into a multi-dimensional latent space through fully connected layers. The decoder is a transposed convolutional mirror encoder. The VAE is trained based on a binary three-dimensional lung image dataset using binary cross-entropy and KL divergence. In the formula, This represents the total loss used to train the VAE. Represents a binary three-dimensional lung image. This represents the image output by the VAE during training. This represents the binary cross-entropy loss. This represents the KL divergence loss. For encoder, It is a standard Gaussian distribution. As a latent variable, It's a hyperparameter; Lung shape loss is defined as: in, This indicates loss of lung shape. This represents a reconstructed lung ventilation image. This indicates a lung shape mask generator. Represents the inner product at the voxel level. It is a small constant that ensures numerical stability; In data consistency items and The voltage measurement value is estimated, and the deviation between the voltage measurement value and the denoised voltage measurement value is calculated using the Euclidean norm. The lung shape loss is... The image reconstruction loss function is defined as: in, This represents the loss function used for image reconstruction. This indicates the measured value of the noise-reducing voltage. Let λ1 represent the EIT forward model, where λ1 is used to balance the data consistency term and the regularization term. The parameters of the untrained dual-output neural network are iteratively optimized to minimize the image reconstruction loss, and each instance is reconstructed under unsupervised conditions.

2. The three-dimensional lung ventilation impedance imaging method according to claim 1, characterized in that: The residual denoising network is pre-trained, and the training process includes the following steps: Based on the circuit reciprocity theorem, the original boundary voltage values ​​measured by EIT are divided into two corresponding complementary subsets. and The residual denoising network is trained using paired noisy data. Denoising is performed using a residual denoising network; prediction is also performed. The noise component in the training objective is: in, This represents a residual denoising network. Represents the parameters of the network. Mathematical symbols representing expectations.

3. The three-dimensional lung ventilation electrical impedance imaging method according to claim 2, characterized in that: The residual denoising network structure is as follows: features at different scales in a one-dimensional input voltage signal are captured through parallel multi-scale dilated convolutions and residual connections; five different sizes of convolution kernels are used to process the input voltage signal in parallel, the first three of which are ordinary convolutions, and the latter two expand the receptive field without significantly increasing the parameters through dilated convolutions; all parallel outputs are concatenated in the channel dimension and then subjected to four layers of standard convolutions for feature fusion and transformation; finally, the original input and the noise features learned by the network are added together through residual connections.

4. A three-dimensional pulmonary ventilation impedance imaging method according to claim 2 or 3, characterized in that, The noise reduction voltage measurement value ,in, This indicates the measured value of the noise-reduced voltage. This represents a residual denoising network.

5. The three-dimensional pulmonary ventilation impedance imaging method according to claim 1, characterized in that: The process of constructing the binary 3D lung image dataset is as follows: each CT image is processed using a standardized procedure to obtain a binary lung mask: cropping, segmentation, removal of small non-connected regions, solid formation, surface smoothing, binarization, and orientation alignment, and then resampled into an image of a certain size.

6. The three-dimensional pulmonary ventilation impedance imaging method according to claim 1, characterized in that: The process of constructing a dual-output convolutional neural network is as follows: (The process involves) fixing random variables... As input, the signal is propagated through an encoder-decoder with 3D convolutional layers, pooling, and skip connections to generate a 3D lung conductivity image that matches the measured voltage. A fully connected branch is appended to the encoder output to output the latent variable. This makes the reconstructed result lean towards a natural and smooth image pattern. Furthermore, by combining the prior of the generative model with explicit lung shape constraints, the output of the dual-output convolutional neural network is represented as: in, For reconstructed lung ventilation images, This represents a randomly initialized neural network. As a latent variable, This represents random input Gaussian noise.

Images