Multi-dimensional loss driven local lesion electrical impedance imaging algorithm

Abstract

The invention provides a multi-dimensional loss driven local lesion electrical impedance imaging algorithm, which comprises the following steps of S1, randomly obtaining a plurality of thoracic cavity models through normal distribution, applying excitation current to the thoracic cavity models by using electrodes to obtain an initial conductivity distribution diagram X, and inputting the initial conductivity distribution diagram X into a generator to generate a predicted conductivity distribution diagram Ypred; s2, obtaining a real conductivity distribution diagram Ytrue based on the thoracic cavity model; s3, based on the real conductivity distribution diagram Ytrue, through a discriminator, correlation calculation COR and relative error evaluation RE, obtaining a loss value of the predicted conductivity distribution diagram Ypred; and S4, based on the loss value, adjusting the weights of the generator and the discriminator in real time, training a local lesion electrical impedance imaging algorithm model, and realizing image restoration of the initial conductivity distribution diagram X.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical impedance tomography, and specifically relates to a multi-dimensional loss-driven local lesion electrical impedance tomography algorithm. Background Art

[0002] Electrical impedance tomography technology is an imaging technology based on measuring the internal electrical impedance distribution of an object. Its principle is to infer the internal tissue structure and lesion conditions of the object based on the voltage changes generated by different substances having different electrical conductivities for current. Traditional electrical impedance tomography technology usually injects current into the object to be imaged using multiple electrodes, then measures the voltage responses at different positions, and performs inversion calculations through these measurement data to finally reconstruct the electrical impedance distribution image of the object. However, traditional methods are often restricted by problems such as imaging resolution limitations, noise interference, and difficulties in image reconstruction under complex structures.

[0003] In recent years, with the rapid development of deep learning technology, deep learning-based electrical impedance tomography technology has gradually become a research hotspot. Deep learning technology learns complex features in a large amount of data by constructing deep neural networks and can achieve efficient non-linear mapping relationships. Compared with traditional methods, deep learning-based electrical impedance tomography technology has higher imaging resolution, stronger anti-noise ability, and faster imaging speed, which has greatly promoted the development and application of electrical impedance tomography technology.

[0004] In the medical field, electrical impedance tomography technology has broad application prospects. The impedance value of biological tissues changes with their type, density, size, and physiological state. Therefore, electrical impedance tomography technology can be used for the early diagnosis and monitoring of many diseases, such as lung diseases, brain diseases, and cardiovascular diseases, etc. Especially in the diagnosis of lung diseases, electrical impedance tomography technology can provide immediate, non-invasive, and low-cost imaging examinations and is expected to become an important tool for lung disease screening and monitoring.

[0005] When a tissue undergoes a lesion, its impedance value will change significantly, which provides a potential opportunity for early disease diagnosis. By using electrical impedance tomography technology for the chest cavity, it is possible to achieve early discrimination of the health status of the chest cavity. More importantly, the resistance values of some tissues in the potential lesion stage will also change, and this change is often difficult to effectively detect by traditional CT imaging technology. However, EIT technology can accurately display these early lesion characteristics and provide a new approach for early diagnosis.

[0006] Compared with traditional CT imaging technology, EIT has many advantages. First of all, the price of EIT is relatively low and the operation is simple and easy, which makes its application in the medical field more extensive and feasible. Secondly, the non-invasive nature of EIT technology makes it a more popular examination method among patients. It does not require the use of radioactive drugs or exposure to radiation risks, so it is also safer and more reliable.

[0007] Given the great potential of EIT technology in lesion diagnosis, it is expected to become an essential item for everyone's physical examination in the future. By diagnosing diseases in advance, EIT technology is expected to greatly improve the treatment effect and survival rate, thus having a profound impact on human health.

[0008] However, the current electrical impedance tomography technology still faces some challenges. For example, traditional methods are limited in aspects such as imaging resolution, noise interference, and imaging speed, resulting in poor image reconstruction under complex structures. Therefore, seeking new imaging methods and technologies is an urgent problem to be solved in this field. Summary of the Invention

[0009] The present invention is made to solve the above problems, and its purpose is to provide a multi-dimensional loss-driven local lesion electrical impedance tomography algorithm.

[0010] The present invention provides a multi-dimensional loss-driven local lesion electrical impedance tomography algorithm, which has the following characteristics and includes the following steps: S1, randomly obtain multiple thoracic models through normal distribution, apply excitation current to the thoracic models using electrodes, obtain the initial conductivity distribution map X, and input the initial conductivity distribution map X into the generator to generate the predicted conductivity distribution map Y pred ; S2, obtain the true conductivity distribution map Y based on the thoracic model true ; S3, based on the true conductivity distribution map Y true , through the discriminator, correlation calculation COR and relative error evaluation RE, obtain the loss value of the predicted conductivity distribution map Y pred ; S4, based on the loss value, adjust the weights of the generator and the discriminator in real time, train the local lesion electrical impedance tomography algorithm model, and realize the image restoration of the initial conductivity distribution map X.

[0011] In the multi-dimensional loss-driven local lesion electrical impedance tomography algorithm provided by the present invention, it may also have the following characteristics: Among them, step S1 specifically includes the following sub-steps: Set the conductivity of the lungs in the thoracic cavity model to 0.18 - 0.19 S / m, the conductivity of the heart to 0.2 - 0.25 S / m, the background to muscles and tendons with a conductivity of 0.36 - 0.37 S / m, and the conductivity of the lesion tissue to 2 - 3 times that of the lungs. The data all use human data at a frequency of 100 kHz. Uniformly attach 16 electrodes on the surface of the thoracic cavity model, apply an excitation current to each electrode in turn, use adjacent driving, input the excitation current at electrode No. 1, ground electrode No. 2, and measure the voltage between each pair of electrodes in turn, obtaining a total of 208 voltage data. Subtract the empty-field voltage value from the obtained boundary voltage value, multiply by the sensitivity matrix, set the conductivity outside the circular field to 0, and then perform normalization processing to obtain the 32*32 initial conductivity distribution map X.

[0012] In the multi-dimensional loss-driven local lesion electrical impedance tomography algorithm provided by the present invention, it may also have the following characteristics: Among them, step S2 specifically includes the following sub-steps: Randomly obtain the size, position, and rotated angle of the left and right lungs through a standard normal distribution, and set the upper and lower limits to ensure that the generated thoracic cavity model conforms to reality. Repeat the above operations to obtain a large number of corresponding initial conductivity distribution maps X and true conductivity distribution maps Y true , forming a data set.

[0013] In the multi-dimensional loss-driven local lesion electrical impedance tomography algorithm provided by the present invention, it may also have the following characteristics: Among them, in step S3, the function of calculating the correlation COR is specifically as follows: Among them, corrcoef is the built-in correlation calculation function in numpy.

[0014] In the multi-dimensional loss-driven local lesion electrical impedance tomography algorithm provided by the present invention, it may also have the following characteristics: Among them, in step S3, the function of evaluating the relative error RE is specifically as follows:

[0015]

[0016] Y pred is the predicted conductivity distribution map generated by the generative adversarial network, and Y p-norm is Y pred obtained after normalization, and Y t-norm is Y true obtained after normalization.

[0017] In the multi-dimensional loss-driven local lesion electrical impedance tomography algorithm provided by the present invention, it may also have the following characteristics: Among them, step S3 specifically includes the following sub-steps:

[0018] Input the initial conductivity distribution map X into the generator to obtain the predicted conductivity distribution map Y pred , and input it together with the true conductivity distribution map Y true into discriminator D1, discriminator D2, and calculate the RE and COR loss values between them. The obtained loss values are returned to the generator and gradient descent is performed. The loss function of the generator is

[0019]

[0020] where m is the number of samples in the current batch, α1 and α2 are the weight parameters of discriminator 1 and discriminator 2, β1 is the weight parameter of COR, β2 is the weight parameter of RE, X i is the i-th initial distribution map, Y pred(i) is the i-th generated distribution map, and Y pred(i) =G(x i ).

[0021] In the multi-dimensional loss-driven local lesion electrical impedance tomography algorithm provided by the present invention, it can also have the following feature: Among them, discriminator D1 generally judges the quality of the generated conductivity distribution map, splits the generated picture into 4 pieces and judges them in sequence, and gives the authenticity of each small piece of picture. The closer the value given is to 1, the higher the quality of the generated picture of this piece. Otherwise, the quality is lower. The loss function of discriminator D1 is

[0022]

[0023] The loss includes two parts, namely, judging whether the true distribution map and the initial distribution map correspond, and judging whether the generated distribution map and the true distribution map correspond, and performing gradient descent on discriminator D1 through cross-entropy loss.

[0024] In the multi-dimensional loss-driven local lesion electrical impedance tomography algorithm provided by the present invention, it can also have the following feature: Among them, a value between 0 and 1 is used to represent the discrimination result. The larger the value, the higher the quality of the generated one judged. The loss function of discriminator D2 is

[0025]

[0026] The loss function of discriminator D2 consists of three parts. ω1, ω2, and ω3 are the weights of the three parts of the loss. The first part is to judge the correspondence between the true conductivity distribution map Y true and the initial conductivity distribution map X. The second part is to judge the correspondence between the predicted conductivity distribution map Y pred and the true conductivity distribution map Y true . The third part is to use the non-corresponding true conductivity distribution map Y trueJudging with the initial conductivity distribution map X, Y true(i) is the next true conductivity distribution map corresponding to the true conductivity distribution map. During each round of training, the data order will be randomly shuffled. Therefore, Y here true(i) is a random true distribution map.

[0027] In the multi-dimensional loss-driven local lesion electrical impedance tomography algorithm provided by the present invention, it may also have the following characteristics: Among them, step S4 specifically includes the following sub-steps:

[0028] In the early stage of the training of the local lesion electrical impedance tomography algorithm model, set the weight of discriminator D1 to 2. In the later stage of the training of the local lesion electrical impedance tomography algorithm model, set the weight of discriminator D1 to 0.01. In the early stage of the training of the local lesion electrical impedance tomography algorithm model, set the weight α2 of discriminator D2 passed to the generator in the early stage to 0 or 0.005, so as to avoid the drastic impact of the large fluctuations of D2 on the generator. In the later stage of the training of the local lesion electrical impedance tomography algorithm model, raise the weight α2 of discriminator D2 to 1. In the early stage of the training of the local lesion electrical impedance tomography algorithm model, set ω1 of discriminator D2 to 1, ω2 to 1, and ω3 to 0. In the middle stage of the training of the local lesion electrical impedance tomography algorithm model, adjust ω1 of discriminator D2 to 0.5, ω2 to 0.25, and ω3 to 0.25.

[0029] Functions and effects of the invention

[0030] According to the multi-dimensional loss-driven local lesion electrical impedance tomography algorithm involved in the present invention, by inputting the initial conductivity distribution map into the trained generative adversarial network, the impedance tomography image can be obtained. Compared with the traditional method, the generative adversarial network of the present invention has stronger non-linear modeling ability and better edge feature processing ability, which can effectively improve the imaging resolution and anti-noise ability, so as to achieve more accurate image reconstruction under complex structures.

[0031] The present invention innovatively sets weight parameters for the discriminator and adjusts the weights in different periods to amplify the functions of different discriminators. Compared with the existing network, the training speed is greatly improved. At the same time, through the accuracy judgment of discriminator D2, the accuracy and precision of the generated pictures are ensured. Compared with the traditional method, the present invention has better anti-noise ability, robustness and image reconstruction results, and the model has strong adaptability, can be used in a variety of scenarios, has low requirements for the sensitivity matrix, and all the original information is included in 208 data, and has good generation ability for the actual scenario. Brief description of the drawings

[0032] Figure 1 is the flowchart of the electrical impedance tomography reconstruction method based on the generative adversarial network in the embodiment of the present invention;

[0033] Figure 2 is the overall architecture of the local lesion electrical impedance tomography algorithm model in the embodiments of the present invention; and

[0034] Figure 3 is a schematic diagram of the network model of the generator in the embodiments of the present invention. Detailed implementation manners

[0035] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the following embodiments will specifically describe the multi-dimensional loss-driven local lesion electrical impedance tomography algorithm of the present invention in conjunction with the accompanying drawings.

[0036] The multi-dimensional loss-driven local lesion electrical impedance tomography algorithm in this embodiment includes the following steps:

[0037] S1. Randomly obtain multiple thoracic models through normal distribution, apply excitation current to the thoracic models using electrodes to obtain an initial conductivity distribution map X, and input the initial conductivity distribution map X into the generator to generate a predicted conductivity distribution map Y pred .

[0038] Figure 1 is a flowchart of the electrical impedance tomography reconstruction method based on a generative adversarial network in the embodiments of the present invention.

[0039] Step S1 specifically includes the following sub-steps:

[0040] As Figure 1 shown, in this embodiment, the process of generating the predicted conductivity distribution map Y through the thoracic model pred is specifically as follows:

[0041] (1) Set the conductivity at each location inside the thoracic model; (2) Generate random lungs, hearts, and tumor sizes using the standard normal distribution; (3) Generate a dataset of the thoracic model; (4) Measure the voltage values at the thoracic boundary; (5) Form a dataset based on the boundary voltage values and the true conductivity distribution map; (6) Train the generative adversarial network model and save the parameters; (7) Debug the generative adversarial network according to the boundary voltage values; (8) Obtain the predicted conductivity distribution map Y of the measured area pred .

[0042] Set the conductivity of the lungs in the thoracic model randomly around 0.189 S / m, the conductivity of the heart randomly around 0.215 S / m, the background is set as muscle and tendon, and the conductivity is set randomly around 0.361 S / m. The conductivity of the tumor lesion location is set randomly around 0.5 S / m. The data is the human body conductivity at 100 kHz.

[0043] In this embodiment, 16 electrodes were attached to the surface of the chest model, with a measurement frequency of 100 kHz. An excitation current was applied to each electrode in sequence, using adjacent drive. The excitation current was applied to electrode 1, while electrode 2 was grounded. The voltage between each electrode was measured sequentially, yielding a total of 208 voltage data points. These data points were multiplied by the sensitivity matrix to obtain g0, which is 812 x 1, where 812 represents the coordinates of the circular field within the 32 x 32 square field. Missing positions were padded with 0s, and the data was normalized to the range 0-1, resulting in a 32 x 32 initial distribution map. The sensitivity matrix linearly maps the 208 data points to the 812 data points. The 208 original data points already contain all the information within the field, and this linear mapping does not destroy the information contained within. Generative adversarial networks have a strong learning capability and can learn the linear mapping from the original data to the 812 data points after several iterations. Therefore, this model does not rely on the accuracy of the sensitivity matrix. A high-precision sensitivity matrix ensures that the initial data better reflects the actual distribution map, making learning and training relatively easy. The low-precision sensitivity matrix makes the initial distribution relatively abstract, making learning more difficult. However, as the number of training iterations increases, this does not affect the final results. This method is applied to each model in turn, resulting in a corresponding dataset.

[0044] The resulting initial conductivity distribution map X is input into a generative adversarial network and optimized using a loss function. In this example, cross-entropy loss is used. Cross-entropy loss simplifies gradient calculation during backpropagation. It also provides a strong penalty for large deviations, shortening the model training cycle.

[0045] Figure 3 2 is a schematic diagram of a generative adversarial network model of a generator in an embodiment of the present invention.

[0046] like Figure 3 As shown in the figure, the generator's generative adversarial network consists of two parts: encoding and decoding. The input image is a 32*32 grayscale image. The input image is encoded into a 2*2 image with 1024 channels through batch convolution, batchnorm, ReLU, and pooling. The decoder then decodes it, performing batch deconvolution, batchnorm, ReLU, and convolution, converting it back to a 32*32 grayscale image. This multi-layered encoding and decoding allows the generator to better learn the features in the distribution map and adapt to a wider range of scenarios.

[0047] Both input and output are 32*32 distribution maps to achieve the effect of image reconstruction.

[0048] The generator is updated using the loss function, using backpropagation in this case. The principle is to adjust the parameters in the direction of the negative gradient of the target based on the gradient descent strategy.

[0049] The loss function of the generator consists of four parts, namely, the discrimination loss of discriminator D1, the discrimination loss of discriminator D2, and the RE and COR losses of the generated images.

[0050] Among them, the discrimination loss of discriminator D1 is the cross-entropy loss between the discrimination result of discriminator D1 for the generated distribution map and 1. The same applies to discriminator D2.

[0051] The loss of RE is the cross-entropy loss between the RE value of the generated distribution map and 0. The COR loss is the cross-entropy loss between the COR value of the generated distribution map and 1. RE and COR are evaluation indices. The smaller the RE, the closer it is to 0, and the larger the COR, the closer it is to 1, indicating that the quality of the generated distribution map is better.

[0052] The generator hopes to deceive the discriminator so that the discriminator's discrimination results are all true, that is, 1. While the discriminator hopes to accurately discriminate, identifying each generated distribution map as false, that is, 0, and identifying all real distribution maps as 1.

[0053] In addition to discriminating the generated distribution maps and real distribution maps, discriminator D2 also needs to discriminate the real distribution maps that do not correspond to the initial distribution map. This is to improve the discrimination accuracy of the discriminator and prevent the generator from generating distribution maps that are true but incorrect.

[0054] S2. Obtain the real conductivity distribution map Y based on the thoracic cavity model ture 。

[0055] Step S2 specifically includes the following sub-steps:

[0056] Randomly obtain the sizes, positions, and rotated angles of the left and right lungs through the standard normal distribution, and set the upper and lower limits to ensure that the generated thoracic cavity model conforms to the actual situation. Repeat the above operations to obtain a large number of corresponding initial conductivity distribution maps X and real conductivity distribution maps Y true ,forming a data set.

[0057] Figure 2 This is the overall architecture of the local lesion electrical impedance tomography algorithm model in the embodiments of the present invention.

[0058] S3. As Figure 2 shown, based on the real conductivity distribution map Y true ,obtain the predicted conductivity distribution map Y pred through the discriminator, correlation calculation COR, and relative error evaluation RE, and obtain the loss value of

[0059] Step S3 specifically includes the following sub-steps:

[0060] The function of correlation calculation COR is specifically as follows:

[0061]

[0062] corrcoef is a built-in function for calculating correlation in numpy.

[0063] The function for evaluating the relative error RE is as follows:

[0064]

[0065] Y pred is the predicted conductivity distribution map generated by the generative adversarial network. Y p-norm is Y pred obtained by normalizing Y t-norm is Y true obtained by normalizing.

[0066] Input the initial conductivity distribution map X into the generator to obtain the predicted conductivity distribution map Y pred , and input it together with the real conductivity distribution map Y true into discriminator D1, discriminator D2, and calculate the RE and COR loss values between them. The obtained loss value is returned to the generator and gradient descent is performed. The loss function of the generator is:

[0067]

[0068] where m is the number of samples in the current batch, α1α2 are the weight parameters of discriminator 1 and discriminator 2, β1 is the weight parameter of COR, β2 is the weight parameter of RE, X i is the i-th initial distribution map, T pred(i) is the i-th generated distribution map, T pred(i) = G(x i ).

[0069] Discriminator D1 generally judges the quality of the generated conductivity distribution map. The generated picture is split into 4 pieces and judged in turn, and the authenticity of each small piece of picture is given. The closer the value given is to 1, the higher the quality of the generated picture of this piece. Otherwise, the quality is lower. The loss function of discriminator D1 is:

[0070]

[0071] The loss includes two parts, namely, judging whether the real distribution map and the initial distribution map correspond, and judging whether the generated distribution map and the real distribution map correspond, and performing gradient descent on discriminator D1 through cross-entropy loss.

[0072] The structure of discriminator D2 is relatively complex. It judges the accuracy of the generated picture, but is relatively weak in overall control. The discriminant result is represented by a value between 0 and 1. The larger the value, the higher the quality of the generated picture judged. The loss function of discriminator D2 is:

[0073]

[0074] The loss function of discriminator D2 consists of three parts, and ω1, ω2, and ω3 are the weights of the three parts of the loss.

[0075] The first part is to judge the correspondence between the true conductivity distribution map Y true and the initial conductivity distribution map X.

[0076] The second part is to judge the correspondence between the predicted conductivity distribution map Y pred and the true conductivity distribution map Y true of.

[0077] The third part is to judge with the non-corresponding true conductivity distribution map Y true and the initial conductivity distribution map X.

[0078] Y true(i) is the next true conductivity distribution map corresponding to the true conductivity distribution map. During each round of training, the data order will be randomly shuffled. Therefore, Y here true(i) is a random true distribution map.

[0079] Through the loss of the third part, the discrimination accuracy of the discriminator can be greatly improved, avoiding the generated distribution map by the generator being "real" but not corresponding.

[0080] S4. Based on the loss value, adjust the weights of the generator and the discriminator in real time, train the local lesion electrical impedance imaging algorithm model, and realize the image restoration of the initial conductivity distribution map X.

[0081] Step S4 specifically includes the following sub-steps:

[0082] The quality of the reconstructed image can be evaluated through RE and COR. When RE is close to 0 and COR is close to 1, it indicates that the restoration degree of the image reconstruction is very high.

[0083] RE and COR can be used as indicators to evaluate the generated pictures, judge the quality of the generated distribution maps, and can also be used for the training of the model. In the early and middle stages of model training, calculate the COR and RE values of the generated pictures respectively. Transmit the loss between them and the expected values to the generator to accelerate the convergence speed of the generator. RE and COR can evaluate the quality of the generated pictures to a certain extent, but when the generated pictures are gradually approaching the true distribution map, the losses of RE and COR cannot well reflect the loss of the generator. Therefore, in the later stage of model training, the loss weights of RE and COR should be down-regulated.

[0084] In the early stage of training the local lesion electrical impedance imaging algorithm model:

[0085] Since the accuracy of the distribution map generated by the generator is relatively low, the loss function at this time is mainly composed of the discrimination of D1. Therefore, set α1 of the generator to 2 and α2 to 0.

[0086] The discriminator D1 has a relatively simple structure, converges rapidly, and can generally give a good judgment on the quality of the generated pictures, which can greatly help the generator converge in the early stage. Therefore, set the weight of the discriminator D1 to 2.

[0087] The discriminator D2 has a strong discrimination ability for details, but the generative adversarial network has many parameters and converges very slowly in the early stage. Therefore, set the weight ɑ2 passed to the generator by the discriminator D2 in the early stage to 0 or 0.005 to avoid the strong impact on the generator caused by the large fluctuations of D2.

[0088] At the same time, the discriminator D2 has weak discrimination ability in the early stage, and its own loss weight should also be adjusted accordingly. This is to avoid large fluctuations in the loss value of the discriminator D2. Set ω1 of the discriminator D2 to 1, ω2 to 1, and ω3 to 0 in the early stage of training the local lesion electrical impedance tomography algorithm model.

[0089] In the middle stage of training the local lesion electrical impedance tomography algorithm model:

[0090] When the local lesion electrical impedance tomography algorithm model trains to a certain accuracy and can restore the contour of the distribution map and some unclear tumors, the composition of the loss function can be changed to mainly RE and COR. Set α1 to 0.5, α2 to 0.02, β1 to 0.8, and β2 to 0.8. In this way, continue to improve the accuracy of the model.

[0091] At this time, adjust ω1 of the discriminator D2 to 0.5, ω2 to 0.25, and ω3 to 0.25. Since the cross-entropy loss is used for the loss, when it has the discrimination ability, the loss value is likely to rise sharply, resulting in the divergence of the discriminator D2. Therefore, the overall weight is lowered here.

[0092] In the later stage of training the local lesion electrical impedance tomography algorithm model:

[0093] When the generator can already restore the distribution map relatively clearly, change the composition of the loss function to mainly the discriminator D2.

[0094] Set α1 to 0, α2 to 1, β1 to 0.05, and β2 to 0.05. In this way, continue to improve the restoration ability of the generator.

[0095] The distribution map generated by the generator is already very close to the true conductivity distribution map Y generally. true, in the later stage of the training of the local lesion electrical impedance tomography algorithm model, the weight α2 of the discriminator D2 is increased to 1, and D2 is mainly responsible for discriminating and training the generator.

[0096] Since the generated distribution map is already relatively close to the real distribution map, and the structure of the discriminator D1 is relatively simple, it is difficult to give a detailed judgment. In the later stage of model training, it can no longer play a major role. Therefore, the weight of the discriminator D1 is set to 0.01 at this time.

[0097] The discriminator D1 mainly ensures the generality of the generation range. The discriminator D2 mainly ensures the accuracy of the position of the generated distribution map. By dynamically adjusting the weights imposed on the generator by the two discriminators, not only can the training time of the model be greatly shortened, but also the accuracy and precision of the generated pictures can be guaranteed at the same time.

[0098] In this embodiment, the trained generative adversarial network can save the parameters, directly use them when calling for detection, and export the results in a timely manner. It can bring great convenience to future medical work.

[0099] The discriminator D1 is composed of a batch convolution layer, LeakRulu, batchnorm, and a pooling layer. The input initial distribution map and the distribution map to be judged are combined into a 2*32*32 picture, and a 2*2 matrix is output through layers of networks. They respectively represent the authenticity of the four regions of the judged picture. The number of hidden layers is small, and the convergence speed is fast, which is used for rough judgment in the early stage. However, the judgment upper limit is not high, the accuracy is low, and it has little effect in the later stage of training.

[0100] The discriminator D2 is similar in structure to the discriminator D1. The input 2*32*32 distribution map becomes a 128*4*4 matrix after passing through the batch convolution layer, Relu, batchnorm, and pooling layer. Then there are three fully connected layers, and finally a sigmod function is connected. The final output size is a 1*2 matrix. The first number in the matrix is the probability of judging that the input distribution map is real, and the second number is the probability of judging that the input is generated by the generator.

[0101] The fully connected layer makes the network structure more abundant and improves the discrimination ability of the discriminator. To avoid overfitting, a dropout mechanism is added to the fully connected layer, and the dropout rate is set to 0.2.

[0102] After the iterative training is completed, the parameters are saved. During image reconstruction, the measured actual boundary voltage value is input into the network, and the trained parameters are called. The output distribution map size is the same as the input distribution map size, so as to achieve image reconstruction.

[0103] Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional loss-driven local lesion electrical impedance tomography algorithm, characterized in that, It includes the following steps: S1. Randomly obtain multiple chest models through normal distribution. Apply an excitation current to the chest models using electrodes to obtain an initial conductivity distribution map X. Input the initial conductivity distribution map X into a generator to generate a predicted conductivity distribution map Y pred ; S2. Obtain the true conductivity distribution map Y based on the thoracic cavity model true ; S3, based on the true conductivity distribution map Y true , through the discriminator, correlation calculation COR, and relative error evaluation RE, obtain the loss value of the predicted conductivity distribution map Y pred ; S4. Based on the loss value, adjust the weights of the generator and the discriminator in real time, train the local lesion electrical impedance tomography algorithm model, and realize the image restoration of the initial conductivity distribution map X.

2. The multi-dimensional loss-driven local lesion electrical impedance tomography algorithm according to claim 1, wherein: Among them, The step S1 specifically includes the following sub-steps: Set the conductivity of the lungs in the thoracic cavity model to 0.18 - 0.19 S / m, the conductivity of the heart to 0.2 - 0.25 S / m, the background to muscle and tendon with a conductivity of 0.36 - 0.37 S / m, and the conductivity of the lesion tissue to 2 - 3 times that of the lungs. The data all adopt human data at a frequency of 100 kHz. Attach 16 electrodes to the surface of the thoracic cavity model. Apply the excitation current to each electrode in turn, adopt adjacent driving, input the excitation current at electrode No. 1, ground electrode No. 2, and measure the voltage between each pair of electrodes in turn. A total of 208 voltage data are obtained. Subtract the open-field voltage value from the obtained boundary voltage value, multiply by the sensitivity matrix, set the conductivity outside the circular field to 0, and then perform normalization processing to obtain the 32*32 initial conductivity distribution map X.

3. The multi-dimensional loss-driven local lesion electrical impedance tomography algorithm according to claim 1, wherein: Among them, The step S2 specifically includes the following sub-steps: Randomly obtain the sizes, positions, and rotated angles of the left and right lungs through the standard normal distribution, and set the upper and lower limits to ensure that the generated thoracic cavity model conforms to reality. Repeat the above operations to obtain a large number of corresponding initial conductivity distribution maps X and true conductivity distribution maps Y. true , and form a data set.

4. The multi-dimensional loss-driven local lesion electrical impedance tomography algorithm according to claim 1, wherein: Among them, In the step S3, the function of the correlation calculation COR is specifically as follows: corrcoef is the built-in correlation calculation function in numpy.

5. The multi-dimensional loss-driven local lesion electrical impedance tomography algorithm according to claim 4, wherein: Among them, In the step S3, the function of the relative error evaluation RE is specifically as follows: Y pred is the predicted conductivity distribution map generated by the generative adversarial network, Y p-norm is Y pred obtained through normalization, Y t-norm is Y true obtained through normalization.

6. The multi-dimensional loss-driven local lesion electrical impedance tomography algorithm according to claim 5, wherein: Among them, The step S3 specifically includes the following sub-steps: Input the initial conductivity distribution map X into the generator to obtain the predicted conductivity distribution map Y pred , and input it together with the true conductivity distribution map Y true into discriminators D1 and D2, calculate the RE and COR loss values between them, return the obtained loss values to the generator, and perform gradient descent. The loss function of the generator is where m is the number of samples in the current batch, α1 and α2 are the weight parameters of discriminator 1 and discriminator 2, β1 is the weight parameter of COR, β2 is the weight parameter of RE, and X i is the i-th initial distribution map, and Y pred(i) is the i-th generated distribution map, and Y pred(i) = G(x i ).

7. The multi-dimensional loss-driven local lesion electrical impedance tomography algorithm according to claim 6, wherein: Among them, The discriminator D1 generally judges the quality of the generated conductivity distribution map. Split the generated picture into 4 pieces and judge them in turn, and give the authenticity of each small piece of picture. The closer the value given is to 1, the higher the quality of the generated picture of this piece. Otherwise, the quality is lower. The loss function of the discriminator D1 is The loss includes two parts, namely, judging whether the real distribution map and the initial distribution map correspond, and judging whether the generated distribution map and the real distribution map correspond, and performing gradient descent on the discriminator D1 through the cross-entropy loss.

8. The multi-dimensional loss-driven local lesion electrical impedance tomography algorithm according to claim 6, wherein: Among them, Use a value between 0 and 1 to represent the discrimination result. The larger the value, the higher the quality of the generated one judged. The loss function of the discriminator D2 is The loss function of the discriminator D2 consists of three parts. ω1, ω2, and ω3 are the weights of the three parts of the loss. The first part is to judge the correspondence between the true conductivity distribution map Y true and the initial conductivity distribution map X The second part is to judge the correspondence between the predicted conductivity distribution map Y pred and the true conductivity distribution map Y true of. The third part is to use the real conductivity distribution map Y that does not correspond true to judge with the initial conductivity distribution map X Y true(i) For the real conductivity distribution map that corresponds to the subsequent real conductivity distribution map, the data order will be randomly shuffled during each round of training. Therefore, Y here true(i) is a random real distribution map.

9. The multi-dimensional loss-driven local lesion electrical impedance imaging algorithm according to claim 1, characterized in that: Among them, The step S4 specifically includes the following sub-steps: In the early stage of the training of the local lesion electrical impedance imaging algorithm model, set the weight of the discriminator D1 to 2, and in the later stage of the training of the local lesion electrical impedance imaging algorithm model, set the weight of the discriminator D1 to 0.

01. In the early stage of the training of the local lesion electrical impedance imaging algorithm model, set the weight ɑ2 passed from the discriminator D2 to the generator in the early stage to 0 or 0.005, so as to avoid the drastic influence of the large fluctuations of D2 on the generator. In the later stage of the training of the local lesion electrical impedance imaging algorithm model, increase the weight ɑ2 of the discriminator D2 to 1. In the early stage of the training of the local lesion electrical impedance imaging algorithm model, set ω1 of the discriminator D2 to 1, ω2 to 1, and ω3 to 0. In the middle stage of the training of the local lesion electrical impedance imaging algorithm model, adjust ω1 of the discriminator D2 to 0.5, ω2 to 0.25, and ω3 to 0.25.

Images