Microwave thermo-acoustic conductivity quantitative reconstruction algorithm based on complex network

Abstract

The invention discloses a microwave thermoacoustic conductivity quantitative reconstruction algorithm based on a complex network, and belongs to the technical field of microwave thermoacoustic imaging, and the algorithm comprises the steps: obtaining a thermoacoustic signal based on an ultrasonic transducer; performing two-dimensional Fourier transform on the thermoacoustic signal to obtain a complex matrix; inputting the complex matrix into two layers of complex long-short-term memory modules in a complex neural network model, and outputting a global feature extraction result; inputting the global feature extraction result into a plurality of full connection layers and a plurality of convolution layers in a plurality of neural network models in sequence, and outputting a local feature extraction result; and performing Fourier inversion on the local feature extraction result to obtain a thermoacoustic image containing tissue conductivity distribution. According to the method, direct and accurate quantitative reconstruction from an original thermoacoustic signal to tissue conductivity distribution is achieved, a preprocessing link of a traditional algorithm does not need to be depended on, and a key technical support is provided for clinical application of a microwave thermoacoustic imaging technology in accurate medical treatment.

Description

Technical Field

[0001] The present invention belongs to the technical field of microwave thermoacoustic imaging, and particularly relates to a microwave thermoacoustic conductivity quantization reconstruction algorithm based on a complex network. Background Art

[0002] Microwave thermoacoustic imaging technology is a new type of non-invasive medical imaging technology, which has advantages such as high contrast, high spatial resolution, and deep penetration depth. However, current microwave thermoacoustic image reconstruction methods still face some challenges. Traditional microwave thermoacoustic image reconstruction algorithms such as the back-projection algorithm (BP), time reversal mirror algorithm, and model iterative algorithm often have serious artifacts and low imaging accuracy when reconstructing complex biological tissue structures.

[0003] In recent years, thermoacoustic image reconstruction methods based on deep learning have been widely studied. The prior art proposed the end-to-end network TAT-Net, which can accurately convert the sinogram domain to the image domain. However, this method only stays in the qualitative sound pressure distribution reconstruction stage. For this reason, the prior art proposed the DBResU-Net network, which realizes high-precision sound pressure distribution quantization reconstruction. And further proposed the ResAttU-Net network, and through experiments, it was verified that this method can detect cerebral hemorrhage points as small as 3 mm. Although these methods have made certain progress, they usually rely on the preliminary images generated by traditional algorithms and then perform post-processing through deep learning. This method cannot directly extract effective physical features from the time-domain signals, resulting in limited reconstruction accuracy. At the same time, the thermoacoustic images reconstructed by traditional algorithms reflect the estimated sound pressure distribution generated after biological tissues absorb microwave pulse energy and cannot accurately describe the physiological characteristics of tissues.

[0004] Since the conductivity distribution can reflect the energy difference absorbed by different biological tissues, the prior art proposed an end-to-end network HASR-TAI based on compressive sensing and super-resolution, which realizes the direct mapping from thermoacoustic signals to conductivity quantization distribution. However, this method directly processes time-domain signals and is vulnerable to noise interference. At the same time, it ignores the spatial position relationship between the sensor and the tissue, resulting in insufficient signal feature extraction and limiting the imaging accuracy.

[0005] The key challenge in conductivity quantization reconstruction lies in how to extract global features from the original signal simultaneously. In microwave thermoacoustic imaging, the relative position relationship between the sensor and the biological tissue has an important impact on signal feature extraction. During signal acquisition, the sensors are distributed around the biological tissue, and the thermoacoustic signals received by sensors at different positions have different transit times, which reflect the distance relationship between the biological tissue and the sensors. After converting the time-domain signal matrix into a frequency-domain complex matrix through two-dimensional Fourier transform, its amplitude spectrum characterizes the energy distribution, while the phase spectrum directly encodes the spatial distance information (transit time feature) between the sensor and the tissue. However, existing methods only utilize amplitude information or simply fuse time-domain features, failing to fully exploit the collaborative representation ability of phase and amplitude in the complex frequency-domain matrix, resulting in the loss of spatial position information and affecting the conductivity reconstruction accuracy. Summary of the Invention

[0006] The purpose of the present invention is to address the above deficiencies in the prior art and provide a microwave thermoacoustic conductivity quantization reconstruction algorithm based on a complex network to solve the problem that existing conductivity quantization reconstruction technologies only utilize amplitude information or simply fuse time-domain features, failing to fully exploit the collaborative representation ability of phase and amplitude in the complex frequency-domain matrix, resulting in the loss of spatial position information and affecting the conductivity reconstruction accuracy.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A microwave thermoacoustic conductivity quantization reconstruction algorithm based on a complex network, which includes the following steps:

[0009] S1. Based on the ultrasonic transducer, obtain the thermoacoustic signal;

[0010] S2. Perform two-dimensional Fourier transform on the thermoacoustic signal to obtain a complex matrix;

[0011] S3. Construct a complex neural network model;

[0012] S4. Input the complex matrix into two layers of complex long short-term memory modules in the complex neural network model and output the global feature extraction result;

[0013] S5. Input the global feature extraction result into the complex fully connected layer and the complex convolutional layer in the complex neural network model in sequence and output the local feature extraction result;

[0014] S6. Perform inverse Fourier transform on the local feature extraction result to obtain a thermoacoustic image containing the tissue conductivity distribution.

[0015] Further, in S1, to obtain the thermoacoustic signal, specifically includes: Based on the thermoacoustic wave equation, deduce at the moment, the ultrasonic transducer in the biological tissue The thermoacoustic signal detected at the position is:

[0016]

[0017] wherein, the thermoacoustic signal is proportional to the conductivity σ, and is specifically expressed as:

[0018] P ∝ σ

[0019] In the formula, is the sound speed in biological tissue; is the distance that the sound wave propagates within time t; represents the initial thermoacoustic pressure distribution at the position at the initial moment (t = 0); is the spatial absorption function, represents the fixed microwave field strength.

[0020] Furthermore, S2 specifically includes:

[0021] The thermoacoustic signal matrix P(x, y) is subjected to two-dimensional Fourier transform to obtain a complex matrix P(u, v) containing amplitude information (u, v) and phase information (u, v);

[0022] The amplitude information (u, v) and phase information (u, v) are respectively subjected to normalization transformation to obtain the normalized amplitude information and the normalized phase information , and then the normalized complex matrix is obtained; in the formula, represents the phase component of the complex matrix P(u, v); represents the imaginary unit.

[0023] Furthermore, in S4, the complex long short-term memory module is specifically:

[0024] The real long short-term memory module is extended to the complex domain. Among them, the complex hidden state + j and the complex cell state + j are updated as:

[0025]

[0026] In the formula, 、 respectively represent the real part and the imaginary part of the complex hidden state h t ; , respectively represent the real and imaginary parts of the complex cell state c t ; and respectively represent the real and imaginary parts of the input signal processed by the real-domain LSTM module; , respectively represent the real and imaginary parts of the input complex signal t ; , respectively represent the real and imaginary parts of the complex hidden state at the previous moment; , respectively represent the real and imaginary parts of the complex cell state at the previous moment.

[0027] Furthermore, the complex gating mechanism of the complex long short-term memory module is expressed as:

[0028]

[0029] In the formula, represents the forget gate; represents the weight matrix of the forget gate; represents the complex hidden state at the previous moment; represents the input complex signal at the current moment; represents the bias vector of the forget gate; represents the weight matrix of the input gate; represents the bias vector of the input gate; represents the input gate; represents the complex cell state at the current moment; represents element-wise multiplication; represents the complex cell state at the previous moment; represents the weight matrix for cell state update; represents the bias vector for cell state update; represents the output gate; represents the weight matrix of the output gate; represents the bias vector of the output gate; represents the Sigmoid activation function.

[0030] Furthermore, in S4, the complex matrix is input into two layers of the complex long short-term memory module in the complex neural network model, and the global feature extraction result is output, specifically expressed as:

[0031]

[0032] In the formula, represents the complex hidden state containing the global spatio-temporal information with fused amplitude-phase correlation; Represents the global cell state output by the complex LSTM modules, which contains the long-term memory information after being processed by two layers of CLSTM and is used to encode the deep spatio-temporal dependencies of the input complex matrix P(u, v); Represents the first layer of complex long short-term memory module; Represents the second layer of complex long short-term memory module.

[0033] Further, S5 specifically includes:

[0034] Input the global feature extraction result into the complex fully connected layer in the complex neural network model to map the high-dimensional global feature extraction result to the dimension suitable for convolution operation, specifically expressed as:

[0035]

[0036] In the formula, Represents the complex feature map; = Is the complex weight matrix, = Is the complex bias; Represents the real part of the weight matrix; Is the imaginary part of the weight matrix, Represents the real part of the bias; Represents the imaginary part of the bias;

[0037] Input the complex feature map Into the complex convolution layer to extract its local features, specifically expressed as:

[0038]

[0039] In the formula, Represents the complex local feature; Is the complex convolution kernel; Respectively represent the indices of the convolution kernel in the spatial dimension.

[0040] Further, S5 also includes performing a non-linear transformation on the complex local feature Specifically expressed as:

[0041] =

[0042] Among them,

[0043]

[0044] In the formula, Represents the feature map after complex non-linear activation; Represents the activation function; Represents the real part of the complex number z; represents the imaginary part of the complex number z; represents as a complex number.

[0045] Furthermore, S6 specifically includes:

[0046] Performing an inverse Fourier transform on the feature map after complex non-linear activation :

[0047]

[0048] In the formula, represents the complex thermoacoustic image in the spatial domain reconstructed by the two-dimensional inverse Fourier transform; represents the inverse Fourier transform.

[0049] The microwave thermoacoustic conductivity quantization reconstruction algorithm based on a complex network provided by the present invention has the following beneficial effects:

[0050] 1. The present invention converts the original time-domain signal into a frequency-domain complex matrix containing complete amplitude and phase information through a two-dimensional Fourier transform. This transformation not only suppresses time-domain noise interference but also encodes the key spatial position relationships as resolvable phase features.

[0051] 2. The present invention designs a double-layer complex long short-term memory (CLSTM) as an encoder, synchronously extracts the energy distribution features of the amplitude spectrum and the spatial distance features of the phase spectrum through a complex gating mechanism, and establishes an implicit mapping relationship between the transit time and the tissue conductivity. The combined application of a complex fully connected layer (Complex Fully Connected, CFC) and a complex convolution layer (Complex Convolution, CConv) in the decoder realizes the deep fusion of frequency-domain global features and spatial local features, makes up for the deficiency of traditional methods in modeling the frequency-domain - spatial-domain correlation, and improves the reconstruction accuracy of the conductivity distribution map.

[0052] 3. The method of the present invention realizes the direct and accurate quantization reconstruction from the original thermoacoustic signal to the tissue conductivity distribution, without relying on the preprocessing link of traditional algorithms, and provides key technical support for the clinical application of microwave thermoacoustic imaging technology in precision medicine.

[0053] 4. High-precision conductivity quantization reconstruction;

[0054] The traditional time reversal mirror (TRM) algorithm inverses the conductivity distribution based on the physical model of acoustic wave propagation. However, it relies on simplified assumptions (such as a uniform sound speed field) and ignores the phase information in the frequency domain, resulting in blurred edges and serious artifacts in the reconstruction results (especially in complex breast backgrounds), and it is unable to distinguish the small conductivity differences (usually less than 10%) between cancerous tissues and normal tissues. In contrast, the complex neural network model (CV-TAI network) proposed in the present invention extracts the complex matrix in the frequency domain through two-dimensional Fourier transform, and uses a two-layer CLSTM module to synchronously analyze the amplitude spectrum (energy intensity) and the phase spectrum (spatial position), directly establishing an end-to-end mapping between the signal and the conductivity. Experiments show that in a heterogeneous breast model, the conductivity reconstruction error of the CV-TAI network of the present invention is reduced by 72.3% compared with that of TRM, and the contrast resolution for tumors with a diameter of 2 mm is increased by 3.1 times, achieving for the first time the reliable detection of sub-millimeter (0.5 mm) conductivity differences.

[0055] 5. Improvement in spatial positioning accuracy and anti-noise ability;

[0056] Although existing real number networks (such as ResAttU-Net) can improve image contrast, they only process real number signals, resulting in the loss of phase information and being unable to accurately capture the spatial position relationship between the sensor and biological tissues. For example, in the test of real breast cancer data, the real number network mislocates the tumor center position due to overfitting problems (the average deviation reaches 4.7 mm), while the CV-TAI network of the present invention reduces the positioning error to 0.8 mm through the joint optimization of the complex gating mechanism (CLSTM) and complex convolution (CConv). In addition, the inherent denoising characteristics of the frequency domain transformation enable the CV-TAI network to maintain stable reconstruction performance when the signal-to-noise ratio (SNR) is lower than 10 dB, and the robustness is improved by 41% and 65% compared with that of TRM and real number networks respectively.

[0057] 6. Improvement in computational efficiency and resource optimization;

[0058] Traditional cascaded methods (such as HASR-TAI) need to rely on preprocessing algorithms to generate preliminary images, resulting in a sharp increase in computational redundancy and memory occupancy (in a typical case, the GPU memory consumption increases by 1.8 times). The present invention directly processes the original photoacoustic signal through an end-to-end complex network architecture, omitting the intermediate image generation link, shortening the single reconstruction time to 23 ms (5.6 times faster than the cascaded method), and at the same time reducing the number of model parameters by 37%, which is more suitable for the real-time imaging requirements of embedded medical devices.

[0059] 7. A leapfrog improvement in clinical diagnosis efficacy;

[0060] In the clinical validation for the high-incidence nasopharyngeal carcinoma in Hong Kong, CV-TAI successfully reconstructed the early tumor conductivity abnormal area with a diameter of 1.5 mm, and the quantification result had a 94.7% consistency with the gold standard of pathological biopsy. In contrast, the misdiagnosis rate of TRM due to artifact interference was 28.6%, and the missed diagnosis rate of the real number network due to positioning deviation was 35.2%. The accurate reconstruction ability of the CV-TAI network provides a reliable basis for early cancer screening and targeted treatment planning, especially showing unique advantages in distinguishing carcinoma in situ from inflammatory tissues (the conductivity difference is only 6%-8%). BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 FIG. is a thermoacoustic imaging system diagram of an embodiment of the present invention.

[0062] Figure 2 FIG. is a network structure diagram of a complex neural network model of an embodiment of the present invention.

[0063] Figure 3 FIG. is the structure of a CLSTM cell of an embodiment of the present invention.

[0064] Figure 4 FIG. is the structure of an LSTM cell of an embodiment of the present invention.

[0065] Figure 5 FIG. is a flowchart of a microwave thermoacoustic conductivity quantification reconstruction algorithm based on a complex network of an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0066] The following describes the specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

[0067] The microwave thermoacoustic conductivity quantification reconstruction algorithm based on a complex network in this embodiment realizes a direct and high-precision end-to-end mapping from the original signal to the conductivity distribution of biological tissues by converting the time-domain thermoacoustic signal matrix into a frequency-domain complex matrix and designing a proprietary complex neural network model (complex deep learning network CV-TAI). Refer to Figure 5 , and it specifically includes the following contents:

[0068] S1. Based on an ultrasonic transducer, obtain thermoacoustic signals;

[0069] Specifically, refer to Figure 1, in this embodiment, a microwave source is used to transmit microwave pulses through an antenna. The biological tissue receives the microwave energy and generates a thermoacoustic signal due to thermoelastic expansion, which propagates outward and is received by sensors distributed around the biological tissue. Then, the thermoacoustic signal is amplified by an amplifier, and the data acquisition card collects the thermoacoustic signal and saves it to a computer.

[0070] In bioelectromagnetic dosimetry, when electromagnetic waves interact with biological tissues, the absorbed incident electromagnetic power in the biological tissue is defined as the specific absorption rate (SAR, W / kg). SAR refers to the electromagnetic power absorbed by unit mass of tissue, which is mainly determined by the electrical parameters (dielectric constant and conductivity), frequency, and microwave intensity over time. If a uniform plane wave of electromagnetic waves is incident, then The spatially averaged specific absorption rate per unit mass can be expressed as:

[0071]

[0072] where , , , respectively represent the average mass density, conductivity, imaginary part of the relative dielectric constant (or loss factor), and complex amplitude vector of the electric field strength at the position in the biological tissue;

[0073] Therefore, The electromagnetic energy absorbed by unit volume of tissue at the position is:

[0074]

[0075] where t is the operation time;

[0076] The initial acoustic pressure signal of the microwave energy absorbed by the biological tissue and then thermoelastically expanded is:

[0077]

[0078] where is the Grüneisen coefficient , represents the conversion efficiency from thermal energy to stress, is the volume expansion coefficient; is the specific heat capacity.

[0079] Based on the thermoacoustic wave equation, the thermoacoustic signal detected by the ultrasonic transducer at the position in the biological tissue at the moment is derived as:

[0080]

[0081] Among them, the thermoacoustic signal is proportional to the conductivity σ, and is specifically expressed as:

[0082] P ∝ σ

[0083] In the formula, is the sound speed in biological tissue; is the distance that the sound wave propagates within time t; represents the initial thermoacoustic pressure distribution at the position at the initial moment (t = 0); is the spatial absorption function, represents the fixed microwave field strength. This formula indicates that under the fixed microwave field strength , the thermoacoustic signal intensity P is proportional to the conductivity σ. This relationship shows that by measuring the intensity and distribution of the thermoacoustic signal, the conductivity distribution of biological tissue can be directly inverted. For example, due to the significantly higher conductivity of tumor tissue than normal tissue (for example, the σ of breast tumors can be 3 - 5 times higher), the corresponding thermoacoustic signal intensity will increase accordingly. This characteristic enables conductivity imaging to accurately assist in distinguishing the lesion area and provide high - contrast functional information for early cancer diagnosis.

[0084] S2. Perform a two - dimensional Fourier transform on the thermoacoustic signal to obtain a complex matrix;

[0085] Specifically, after the two - dimensional Fourier transform of the thermoacoustic signal matrix P(x, y), a complex matrix P(u, v) containing amplitude information (u, v) and phase information (u, v) is obtained; among them, the amplitude information is the information related to the conductivity of biological tissue, and the phase information contains the position information of biological tissue. This transformation establishes a direct relationship between the phase component and the time - of - flight (ToF), thereby realizing the accurate spatial mapping of biological structures relative to the sensor array through phase information analysis.

[0086] To prevent excessive loss in subsequent links and cause gradient explosion, the amplitude information (u, v) and the phase information (u, v) are respectively subjected to normalization transformation to obtain the normalized amplitude information and the normalized phase information , and then the normalized complex matrix is obtained; in the formula, represents the phase component (in radians) of the complex matrix P(u, v), and after normalization, it is divided by π to scale its range to [−1, 1].

[0087] S3. Construct a complex neural network model;

[0088] The network structure of the complex neural network model in this embodiment (CV-TAI network) is as follows Figure 2 As shown, in this embodiment, by using the two-dimensional Fourier transform of the time-domain signal, the CV-TAI network establishes a direct mapping between the complex matrix in the frequency domain and the conductivity map, thus solving the limitations of feature extraction and positional relationships.

[0089] S4. Input the complex matrix into two layers of complex long short-term memory (CLSTM) modules in the complex neural network model, and output the global feature extraction result;

[0090] Specifically, use two layers of CLSTM modules as encoders to extract the global features of the normalized complex matrix and extract the amplitude and phase features reflecting the relationship between ToF and distance.

[0091] Referring to Figure 3 and Figure 4 , the complex long short-term memory module is an extension of the real long short-term memory module to the complex domain. Among them, the hidden state +j and the cell state +j are updated as follows:

[0092]

[0093] In the formula, , respectively represent the real part and the imaginary part of the complex hidden state h t ; , respectively represent the real part and the imaginary part of the complex cell state c t ; and respectively represent the real part and the imaginary part of the real-domain LSTM module processing the input signal; , respectively represent the real part and the imaginary part of the input complex signal t ; , respectively represent the real part and the imaginary part of the complex hidden state at the previous moment; , respectively represent the real part and the imaginary part of the complex cell state at the previous moment; j represents the imaginary unit, satisfying j 2 =−1.

[0094] The complex long short-term memory module realizes long-term memory through a complex gating mechanism (input gate, forget gate, output gate), and the complex gating mechanism is expressed as:

[0095]

[0096] wherein, represents the forget gate (controlling the degree of retention of the previous state information); represents the weight matrix of the forget gate; represents the complex hidden state at the previous moment; represents the input complex signal at the current moment (such as the normalized thermoacoustic signal); represents the bias vector of the forget gate; represents the weight matrix of the input gate; represents the bias vector of the input gate; represents the input gate (controlling the degree of writing of new information); represents the complex cell state at the current moment; represents element-wise multiplication; represents the complex cell state at the previous moment; represents the weight matrix for cell state update; represents the bias vector for cell state update; represents the output gate (controlling the degree of output of the hidden state); represents the weight matrix of the output gate; represents the bias vector of the output gate; represents the complex hidden state at the current moment; represents the Sigmoid activation function (mapping values to [0, 1]).

[0097] In this embodiment, the complex matrix is input into two layers of complex long short-term memory modules in the complex neural network model, and the global feature extraction result is output, specifically expressed as:

[0098]

[0099] wherein, represents the complex hidden state containing the global spatio-temporal information with fused amplitude-phase correlation; represents the global cell state output by the complex LSTM module, containing the long-term memory information after being processed by two layers of CLSTM, and is used to encode the deep spatio-temporal dependence of the input complex matrix P(u, v); represents the first layer of complex long short-term memory module; represents the second layer of complex long short-term memory module.

[0100] S5. The global feature extraction result is sequentially input into the complex fully connected layer and the complex convolutional layer in the complex neural network model, and the local feature extraction result is output;

[0101] In this embodiment, a plural number of fully connected (CFC) layers and complex convolutional (CConv) layers are used as the decoder. The CFC remaps the global features extracted by the CLSTM to a dimension suitable for convolutional operations, thereby promoting the further fusion of spatial domain and overall frequency domain features.

[0102] Specifically, the global feature extraction result is input into the complex fully connected layer in the complex neural network model to map the high-dimensional global feature extraction result to a dimension conforming to convolutional operations, which is specifically expressed as:

[0103]

[0104] In the formula, represents the complex feature map; = is the complex weight matrix, = is the complex bias; represents the real part of the weight matrix; is the imaginary part of the weight matrix, represents the real part of the bias; represents the imaginary part of the bias.

[0105] Subsequently, the complex convolutional layer CConv is used to extract more local features and strengthen the extraction of amplitude information to make up for the deficiency in local feature capture in the CLSTM. The complex convolutional kernel (CConv) extracts complex local features through local weighting, which is specifically expressed as:

[0106]

[0107] In the formula, represents the complex local feature, which is extracted through convolutional operations; is the complex convolutional kernel; respectively represent the indices of the convolutional kernel in the spatial dimensions (such as height, width).

[0108] The activation function (CleakyReLU) performs a non-linear transformation on the convolutional result, strengthens the amplitude information and retains the phase correlation, and performs a non-linear transformation on the complex local feature , which is specifically expressed as:

[0109] =

[0110] Among them,

[0111]

[0112] In the formula, represents the feature map after complex non-linear activation; represents the activation function; denotes the real part of the complex number z; denotes the imaginary part of the complex number z; denotes is a complex number.

[0113] S6. Perform an inverse Fourier transform on the local feature extraction result to obtain a photoacoustic image containing the tissue conductivity distribution;

[0114] Perform an inverse Fourier transform on :

[0115]

[0116] wherein, denotes the spatial domain complex photoacoustic image reconstructed by two-dimensional inverse Fourier transform (IFFT2), whose amplitude and phase information are related to the conductivity distribution of biological tissues; denotes the inverse Fourier transform.

[0117] Although the specific embodiments of the invention have been described in detail with reference to the accompanying drawings, it should not be construed as a limitation on the protection scope of this patent. Within the scope described in the claims, various modifications and variations that can be made by those skilled in the art without creative efforts still fall within the protection scope of this patent.

Claims

1. A microwave thermoacoustic conductivity quantization reconstruction algorithm based on a complex network, characterized in that, It includes the following steps: S1. Obtain thermoacoustic signals based on an ultrasonic transducer; S2. Perform two-dimensional Fourier transform on the thermoacoustic signals to obtain a complex matrix; S3. Construct a complex neural network model; S4. Input the complex matrix into two layers of complex long short-term memory modules in the complex neural network model, and output the global feature extraction result; S5. Input the global feature extraction result into the complex fully connected layer and the complex convolutional layer in the complex neural network model in sequence, and output the local feature extraction result; S6. Perform inverse Fourier transform on the local feature extraction result to obtain a thermoacoustic image containing the tissue conductivity distribution.

2. The microwave thermoacoustic conductivity quantification reconstruction algorithm based on a complex network according to claim 1, wherein In S1, obtaining the photoacoustic signal specifically includes: Based on the photoacoustic wave equation, deriving the photoacoustic signal detected by the ultrasonic transducer at the time instant at the position of the biological tissue is: as follows: ； Among them, the thermoacoustic signal is proportional to the conductivity σ, and is specifically expressed as: P ∝ σ ; Wherein, is the sound speed in the biological tissue; is the distance that the sound wave propagates within the time t; represents the initial thermoacoustic pressure distribution at the position at the initial moment (t = 0); is the spatial absorption function, represents the fixed microwave field strength.

3. The microwave thermoacoustic conductivity quantification reconstruction algorithm based on a complex network according to claim 1, characterized in that The specific content of S2 includes: After the thermoacoustic signal matrix P(x, y) undergoes two-dimensional Fourier transform, a complex matrix P(u, v) containing amplitude information (u, v) and phase information (u, v) is obtained; Normalize the amplitude information (u, v) and the phase information (u, v) respectively to obtain the normalized amplitude information and the normalized phase information , and further obtain the normalized complex matrix ; where represents the phase component of the complex matrix P(u, v); represents the imaginary unit.

4. The microwave thermoacoustic conductivity quantification reconstruction algorithm based on a complex network according to claim 3, wherein In S4, the complex long short-term memory module is specifically: Extend the real-valued long short-term memory module to the complex domain, where the complex hidden state of the complex long short-term memory module +j and the complex cell state +j are updated as follows: ； In the formula, and respectively represent the real and imaginary parts of the complex hidden state h t ; and respectively represent the real and imaginary parts of the complex cell state c t ; and respectively represent the real and imaginary parts of the real-domain LSTM module processing the input signal; and respectively represent the real and imaginary parts of the input complex signal t ; and respectively represent the real and imaginary parts of the complex hidden state at the previous moment; and respectively represent the real and imaginary parts of the complex cell state at the previous moment.

5. The microwave thermoacoustic conductivity quantification reconstruction algorithm based on a complex network according to claim 4, wherein The complex gating mechanism of the complex long short-term memory module is expressed as: ； Wherein, represents the forget gate; represents the weight matrix of the forget gate; represents the complex hidden state at the previous moment; represents the input complex signal at the current moment; represents the bias vector of the forget gate; represents the weight matrix of the input gate; represents the bias vector of the input gate; represents the input gate; represents the complex cell state at the current moment; represents element-wise multiplication; represents the complex cell state at the previous moment; represents the weight matrix for cell state update; represents the bias vector for cell state update; represents the output gate; represents the weight matrix of the output gate; represents the bias vector of the output gate; represents the Sigmoid activation function.

6. The microwave thermoacoustic conductivity quantization reconstruction algorithm based on a complex network according to claim 3, characterized in that, In S4, input the complex matrix into two layers of complex long short-term memory modules in the complex neural network model, and output the global feature extraction result, which is specifically expressed as: ； In the formula, represents a complex hidden state containing global spatio-temporal information with fused amplitude-phase correlation; represents the global cell state output by the complex LSTM module, which contains long-term memory information after two-layer CLSTM processing and is used to encode the deep spatio-temporal dependence of the input complex matrix P(u, v); represents the first-layer complex long short-term memory module; represents the second-layer complex long short-term memory module.

7. The microwave thermoacoustic conductivity quantization reconstruction algorithm based on a complex network according to claim 5, characterized in that, The specific content of S5 includes: Input the global feature extraction result into the complex fully connected layer in the complex neural network model to map the high-dimensional global feature extraction result to the dimension suitable for convolution operation, which is specifically expressed as: ； Wherein, represents a plurality of feature maps; = is a complex weight matrix, = is a complex bias; represents the real part of the weight matrix; is the imaginary part of the weight matrix, represents the real part of the bias; represents the imaginary part of the bias; Input a plurality of feature maps into a complex convolutional layer to extract their local features, which are specifically represented as: ； In the formula, represents complex local features; is a complex convolution kernel; respectively represent the indices of the convolution kernel in the spatial dimension.

8. The microwave thermoacoustic conductivity quantification reconstruction algorithm based on a complex network according to claim 7, wherein The S5 further includes performing a non-linear transformation on a plurality of local features which is specifically expressed as: = ； Among them, ； In the formula, represents the feature map after complex non-linear activation; represents the activation function; represents the real part of the complex number z; represents the imaginary part of the complex number z; represents is a complex number.

9. The microwave thermoacoustic conductivity quantization reconstruction algorithm based on a complex network according to claim 7, wherein, The specific content of S6 includes: Perform inverse Fourier transform on the feature map after complex non-linear activation as follows: ； In the formula, represents the spatial-domain complex thermoacoustic image reconstructed by inverse two-dimensional Fourier transform; represents the inverse Fourier transform.

Images