Modulus inversion method and system based on traveling wave decomposition and machine learning

Abstract

This invention provides a modulus inversion method and system based on traveling wave decomposition and machine learning, comprising: Step S1: Imaging and acquiring the displacement field for wave displacement to generate training data based on traveling wave decomposition; Step S2: Constructing a complex covariance neural network; Step S3: Establishing a noisy training dataset generated by the traveling wave decomposition model; Step S4: Using the trained network for inversion and multi-frequency, multi-directional fusion. This invention solves the problem of modulus inversion for noisy, complex wave fields by constructing a complex covariance neural network and utilizing the noisy training dataset generated by the traveling wave decomposition model to mine the mapping operator from the wave field to elastic parameters.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and more specifically, to a modulus inversion method and system based on traveling wave decomposition and machine learning. Background Technology

[0002] Currently, in ultrasonic elastography, algorithms for inverting ultrasonic elastography problems mainly utilize finite element simulation data and neural networks. These algorithms include generating complex wavefields for specific scenarios using finite element methods and estimating modulus using multilayer neural networks. However, this method has limitations in training set generation, the modulus estimation results depend on the training scenario, and its generalization performance is generally poor.

[0003] Patent document CN109567872B discloses a method and system for vascular guided wave elastography based on machine learning. The method includes: numerically simulating the propagation of shear waves in a thin-layer system of blood vessels using finite element software and obtaining finite element analysis results; obtaining the velocity distribution of all nodes in the field based on the finite element analysis results, extracting the velocity of nodes on the midline of the thin layer, and obtaining a dispersion curve as the input signal of a neural network; obtaining the training set and test set of the neural network based on the dispersion curve, and training the neural network using a neural network method until the error on the training set is less than a preset value, so that the final neural network obtains a vascular guided wave elastography image.

[0004] Patent document CN108986909B discloses a method and apparatus for characterizing the elasticity and viscoelasticity of soft tissue based on ultrasonic elastography. The method includes: simulating ultrasonic elastography using the finite element method to obtain training data for an artificial neural network; preprocessing the training data to ensure that it contains valid information that meets preset conditions, and obtaining a preprocessed feature map; and training the preprocessed feature map to obtain a neural network that characterizes the elasticity and viscoelasticity of soft tissue and satisfies the preset neural network effect.

[0005] The aforementioned patent documents disclose algorithms for inverting ultrasonic elastography problems using finite element simulation data and neural networks, including generating complex wave fields for specific scenarios using finite element methods and estimating modulus using multilayer neural networks. However, the training set generation method of the above methods has limitations, the modulus estimation results depend on the training scenario, and the generalization performance is generally poor. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the purpose of this invention is to provide a modulus inversion method and system based on traveling wave decomposition and machine learning.

[0007] A modulus inversion method based on traveling wave decomposition and machine learning, provided by the present invention, includes:

[0008] Step S1: Imaging the displacement field for wave displacement and generating training data based on traveling wave decomposition;

[0009] Step S2: Construct a complex covariance neural network;

[0010] Step S3: Establish a noisy training dataset generated by the traveling wave decomposition model;

[0011] Step S4: Use the trained network to perform inversion and multi-frequency, multi-directional fusion.

[0012] Preferably, in step S1:

[0013] Step S1.1: Under the assumption of local homogeneity, the wave field in an incompressible isotropic medium is a synthesis of multiple traveling waves:

[0014]

[0015] Where r is the spatial position vector, U is the total wave field at r, M is the total number of traveling waves, and a m It is the complex amplitude of the m-th traveling wave. is the unit propagation direction of the m-th traveling wave, k = k′ + i·k″ is the local complex wave number, k′ is the real wave number, and k″ is an exponential term related to attenuation;

[0016] Set different numbers of traveling waves M and different complex amplitudes u m Different directions of dissemination A series of training data were simulated with different k;

[0017] Step S1.2: Add complex Gaussian noise to the wave field:

[0018]

[0019] in, b represents unit complex Gaussian noise, and b is the noise intensity.

[0020] Step S1.3: Normalize the frequency of k:

[0021]

[0022] The noisy wavefield and normalized wavenumber were used as the training set for the subsequent network. ω is the vibration frequency. For normalized complex wavenumber.

[0023] Preferably, in step S2:

[0024] Step S2.1: Perform covariance preprocessing on the network input, vec(U) H Where U is the complex wave field;

[0025] Step S2.2: Complex fully connected neural network, where the weights and biases in the network are all complex numbers, and the activation function is:

[0026]

[0027] in, z is a complex number input to the activation function, b is a real parameter to be learned, and θ z Let z be the phase angle;

[0028] Step S2.3: Divide and There are two estimation structure branches, and the two branches have the same structure. The output layer takes the real part and the imaginary part as the final output, respectively. The network training and inference branches are carried out independently.

[0029] Preferably, in step S3:

[0030] Different frequency components train different D to In the TWENN neural network, during data inversion, data of the same frequency but different directions share the same TWENN block, while different frequencies use TWENNs trained at the corresponding frequencies.

[0031] Preferably, in step S4:

[0032] The fusion formula is as follows, where G′ and G″ represent the energy storage modulus and loss modulus, respectively.

[0033]

[0034] Where i is the imaginary unit, n is the frequency number, and f n Let m be the nth frequency, N be the total number of frequencies, m be the wave propagation direction number, and d be the frequency of the nth frequency. m Let m be the direction of wave propagation, and M be the total number of directions.

[0035] A modulus inversion system based on traveling wave decomposition and machine learning, according to the present invention, includes:

[0036] Module M1: Imaging and acquiring displacement fields for wave-like displacement, generating training data based on traveling wave decomposition;

[0037] Module M2: Construct a complex covariance neural network;

[0038] Module M3: Creates a noisy training dataset generated by the traveling wave decomposition model;

[0039] Module M4: Utilizes the trained network for inversion and multi-frequency, multi-directional fusion.

[0040] Preferably, in module M1:

[0041] Module M1.1: Under the assumption of local homogeneity, the wave field in an incompressible isotropic medium is a synthesis of multiple traveling waves:

[0042]

[0043] Where r is the spatial position vector, U is the total wave field at r, M is the total number of traveling waves, and a m It is the complex amplitude of the m-th traveling wave. is the unit propagation direction of the m-th traveling wave, k = k′ + i·k″ is the local complex wave number, k′ is the real wave number, and k″ is an exponential term related to attenuation;

[0044] Set different numbers of traveling waves M and different complex amplitudes u m Different directions of dissemination A series of training data were simulated with different k;

[0045] Module M1.2: Adds complex Gaussian noise to the wave field:

[0046]

[0047] in, b represents unit complex Gaussian noise, and b is the noise intensity.

[0048] Module M1.3: Normalizes the frequency of k:

[0049]

[0050] The noisy wavefield and normalized wavenumber were used as the training set for the subsequent network. ω is the vibration frequency. For normalized complex wavenumber.

[0051] Preferably, in module M2:

[0052] Module M2.1: Performs covariance preprocessing on the network input, vec(U) H Where U is the complex wave field;

[0053] Module M2.2: Complex fully connected neural network, where the weights and biases are all complex numbers, and the activation function is:

[0054]

[0055] in, z is a complex number input to the activation function, b is a real parameter to be learned, and θ z Let z be the phase angle;

[0056] Module M2.3: divided into and There are two estimation structure branches, and the two branches have the same structure. The output layer takes the real part and the imaginary part as the final output, respectively. The network training and inference branches are carried out independently.

[0057] Preferably, in module M3:

[0058] Different frequency components train different D to In the TWENN neural network, during data inversion, data of the same frequency but different directions share the same TWENN block, while different frequencies use TWENNs trained at the corresponding frequencies.

[0059] Preferably, in module M4:

[0060] The fusion formula is as follows, where G′ and G″ represent the energy storage modulus and loss modulus, respectively.

[0061]

[0062] Where i is the imaginary unit, n is the frequency number, and f n Let m be the nth frequency, N be the total number of frequencies, m be the wave propagation direction number, and d be the frequency of the nth frequency. m Let m be the direction of wave propagation, and M be the total number of directions.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] This invention solves the problem of modulus inversion in noisy complex wave fields by constructing a complex covariance neural network and mining the mapping operator from the wave field to elastic parameters using a noisy training dataset generated by a traveling wave decomposition model. Attached Figure Description

[0065] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0066] Figure 1 This is a schematic diagram of the TWENN network structure;

[0067] Figure 2 This is a schematic diagram of the multi-frequency, multi-directional inversion process. Detailed Implementation

[0068] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0069] Example 1:

[0070] For imaging of wave displacement, such as displacement fields acquired by imaging methods like magnetic resonance, ultrasound, and optics, a complex covariance neural network is constructed, a noisy training dataset generated by a traveling wave decomposition model is established, and a mapping operator from wave field to elastic parameters is constructed to realize the inversion calculation of the measured target modulus and support multi-frequency and multi-directional synthesis.

[0071] According to the present invention, a modulus inversion method based on traveling wave decomposition and machine learning is provided, such as... Figures 1-2 As shown, it includes:

[0072] Step S1: Imaging the displacement field for wave displacement and generating training data based on traveling wave decomposition;

[0073] Specifically, in step S1:

[0074] Step S1.1: Under the assumption of local homogeneity, the wave field in an incompressible isotropic medium is a synthesis of multiple traveling waves:

[0075]

[0076] Where r is the spatial position vector, U is the total wave field at r, M is the total number of traveling waves, and a m It is the complex amplitude of the m-th traveling wave. is the unit propagation direction of the m-th traveling wave, k = k′ + i·k″ is the local complex wave number, k′ is the real wave number, and k″ is an exponential term related to attenuation;

[0077] Set different numbers of traveling waves M and different complex amplitudes u m Different directions of dissemination A series of training data were simulated with different k;

[0078] Step S1.2: Add complex Gaussian noise to the wave field:

[0079]

[0080] in, b represents unit complex Gaussian noise, and b is the noise intensity.

[0081] Step S1.3: Normalize the frequency of k:

[0082]

[0083] The noisy wavefield and normalized wavenumber were used as the training set for the subsequent network. ω is the vibration frequency. For normalized complex wavenumber.

[0084] Step S2: Construct a complex covariance neural network;

[0085] Specifically, in step S2:

[0086] Step S2.1: Perform covariance preprocessing on the network input, vec(U) H Where U is the complex wave field;

[0087] Step S2.2: Complex fully connected neural network, where the weights and biases in the network are all complex numbers, and the activation function is:

[0088]

[0089] in, z is a complex number input to the activation function, b is a real parameter to be learned, and θ z Let z be the phase angle;

[0090] Step S2.3: Divide and There are two estimation structure branches, and the two branches have the same structure. The output layer takes the real part and the imaginary part as the final output, respectively. The network training and inference branches are carried out independently.

[0091] Step S3: Establish a noisy training dataset generated by the traveling wave decomposition model;

[0092] Specifically, in step S3:

[0093] Different frequency components train different D to In the TWENN neural network, during data inversion, data of the same frequency but different directions share the same TWENN block, while different frequencies use TWENNs trained at the corresponding frequencies.

[0094] Step S4: Use the trained network to perform inversion and multi-frequency, multi-directional fusion.

[0095] Specifically, in step S4:

[0096] The fusion formula is as follows, where G′ and G″ represent the energy storage modulus and loss modulus, respectively.

[0097]

[0098] Where i is the imaginary unit, n is the frequency number, and f n Let m be the nth frequency, N be the total number of frequencies, m be the wave propagation direction number, and d be the frequency of the nth frequency. m Let m be the direction of wave propagation, and M be the total number of directions.

[0099] Example 2:

[0100] Example 2 is a preferred embodiment of Example 1, and is used to illustrate the present invention in more detail.

[0101] The present invention also provides a modulus inversion system based on traveling wave decomposition and machine learning. The modulus inversion system based on traveling wave decomposition and machine learning can be implemented by executing the process steps of the modulus inversion method based on traveling wave decomposition and machine learning. That is, those skilled in the art can understand the modulus inversion method based on traveling wave decomposition and machine learning as a preferred embodiment of the modulus inversion system based on traveling wave decomposition and machine learning.

[0102] A modulus inversion system based on traveling wave decomposition and machine learning, according to the present invention, includes:

[0103] Module M1: Imaging and acquiring displacement fields for wave-like displacement, generating training data based on traveling wave decomposition;

[0104] Specifically, in module M1:

[0105] Module M1.1: Under the assumption of local homogeneity, the wave field in an incompressible isotropic medium is a synthesis of multiple traveling waves:

[0106]

[0107] Where r is the spatial position vector, U is the total wave field at r, M is the total number of traveling waves, and a m It is the complex amplitude of the m-th traveling wave. is the unit propagation direction of the m-th traveling wave, k = k′ + i·k″ is the local complex wave number, k′ is the real wave number, and k″ is an exponential term related to attenuation;

[0108] Set different numbers of traveling waves M and different complex amplitudes u m Different directions of dissemination A series of training data were simulated with different k;

[0109] Module M1.2: Adds complex Gaussian noise to the wave field:

[0110]

[0111] in, b represents unit complex Gaussian noise, and b is the noise intensity.

[0112] Module M1.3: Normalizes the frequency of k:

[0113]

[0114] The noisy wavefield and normalized wavenumber were used as the training set for the subsequent network. ω is the vibration frequency. For normalized complex wavenumber.

[0115] Module M2: Construct a complex covariance neural network;

[0116] Specifically, in module M2:

[0117] Module M2.1: Performs covariance preprocessing on the network input, vec(U) H Where U is the complex wave field;

[0118] Module M2.2: Complex fully connected neural network, where the weights and biases are all complex numbers, and the activation function is:

[0119]

[0120] in, z is a complex number input to the activation function, b is a real parameter to be learned, and θ z Let z be the phase angle;

[0121] Module M2.3: divided into and There are two estimation structure branches, and the two branches have the same structure. The output layer takes the real part and the imaginary part as the final output, respectively. The network training and inference branches are carried out independently.

[0122] Module M3: Creates a noisy training dataset generated by the traveling wave decomposition model;

[0123] Specifically, in module M3:

[0124] Different frequency components train different D to In the TWENN neural network, during data inversion, data of the same frequency but different directions share the same TWENN block, while different frequencies use TWENNs trained at the corresponding frequencies.

[0125] Module M4: Utilizes the trained network for inversion and multi-frequency, multi-directional fusion.

[0126] Specifically, in module M4:

[0127] The fusion formula is as follows, where G′ and G″ represent the energy storage modulus and loss modulus, respectively.

[0128]

[0129] Where i is the imaginary unit, n is the frequency number, and f n Let m be the nth frequency, N be the total number of frequencies, m be the wave propagation direction number, and d be the frequency of the nth frequency. m Let m be the direction of wave propagation, and M be the total number of directions.

[0130] Example 3:

[0131] Example 3 is a preferred example of Example 1, and is used to illustrate the present invention in more detail.

[0132] A complex covariance neural network inversion algorithm based on a traveling wave decomposition model mainly includes the generation of training data based on traveling wave decomposition, the construction and training of the complex covariance network, and the use of the trained network for inversion and multi-frequency, multi-directional fusion. Specifically:

[0133] Step S1: Generation of training data based on traveling wave decomposition.

[0134] Step S2: Construct a complex covariance neural network. Figure 1 The structure diagram of the TWENN neural network is shown.

[0135] Step S3: Multi-frequency and multi-directional fusion estimation. Figure 2 Demonstrates a multi-directional, multi-frequency fusion estimation process.

[0136] Step S1 includes the following steps:

[0137] Step S1.1: Under the assumption of local homogeneity, the wave field in an incompressible isotropic medium can be considered as the synthesis of multiple traveling waves:

[0138]

[0139] Where r is the spatial position vector, U is the total wave field at r, M is the total number of traveling waves, and a m It is the complex amplitude of the m-th traveling wave. is the unit propagation direction of the m-th traveling wave, k = k′ + i·k″ is the local complex wave number, k′ is the real wave number, and k″ is a term related to attenuation.

[0140] By setting different numbers of traveling waves M and different complex amplitudes a m Different directions of dissemination A series of training data were simulated with different k values.

[0141] Step S1.2: Add complex Gaussian noise to the wave field. is unit complex Gaussian noise. b is the noise intensity.

[0142] Step S1.3: Normalize the frequency of k. The noisy wavefield and normalized wavenumber were used as the training set for the subsequent network. ω is the vibration frequency. For normalized complex wavenumber.

[0143] Step S2 includes the following steps:

[0144] Step S2.1: Perform covariance preprocessing on the network input, vec(U) H , where U is the complex wave field.

[0145] Step S2.2: Complex Fully Connected Neural Network. The weights and biases in the network are all complex numbers; the newly designed activation function is... in The network has 6 to 10 layers. z is the complex number input to the activation function, b is the real parameter to be learned, and θ... z Let z be the phase angle.

[0146] Step S2.3: Divide and The network has two estimation structure branches, both with identical structures, but the output layer uses the real and imaginary parts respectively as the final output. The network training and inference branches are performed independently.

[0147] Step S3 includes the following steps:

[0148] Step S3.1: Train different TWENNs for different frequency components. In actual data inversion, data of the same frequency but different directions share the same TWENN block, while different frequencies use TWENNs trained at the corresponding frequencies. The TWENN is the one proposed in this method that transforms data from D to... Neural networks.

[0149] Step S3.2: The fusion formula is as follows, where G′ and G″ represent the energy storage modulus and loss modulus, respectively.

[0150]

[0151] Where i is the imaginary unit, n is the frequency number, and f n Let m be the nth frequency, N be the total number of frequencies, m be the wave propagation direction number, and d be the frequency of the nth frequency. m Let m be the direction of wave propagation, and M be the total number of directions.

[0152] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0153] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A modulus inversion method based on traveling wave decomposition and machine learning, characterized in that, include: Step S1: Imaging the displacement field for wave displacement and generating training data based on traveling wave decomposition; Step S2: Construct the complex covariance neural network TWENN; Step S3: Establish a noisy training dataset generated by the traveling wave decomposition model; Step S4: Use the trained network to perform inversion and multi-frequency, multi-directional fusion; In step S1: Step S1.1: Under the assumption of local homogeneity, the wave field in an incompressible isotropic medium is a synthesis of multiple traveling waves: in, It is a spatial position vector. Is The total wave field at that location, It is the total number of traveling waves. It is the first The complex amplitude of a traveling wave, It is the first The direction of propagation of a single traveling wave. It is the local complex wavenumber. It is the real wave number. It is an exponential term related to decay; The imaginary unit; Set different numbers of traveling waves Different complex amplitudes Different directions of dissemination and different A series of training data were simulated; Step S1.2: Add complex Gaussian noise to the wave field: in, It is a unit complex Gaussian noise. Noise intensity; Step S1.3: For Normalize the frequency: The noisy wavefield and normalized wavenumber were used as the training set for the subsequent network. , The vibration frequency, For normalized complex wavenumber; In step S2: Step S2.1: Perform covariance preprocessing on the network input. ,in For complex wave fields; Step S2.2: Complex fully connected neural network, where the weights and biases in the network are all complex numbers, and the activation function is: in, , The complex number is the input to the activation function. Let be the real-valued parameters to be learned. for The phase angle; Step S2.3: Divide and There are two estimation structure branches, and the two branches have the same structure. The output layer takes the real part and the imaginary part as the final output, respectively. The network training and inference branches are carried out independently.

2. The modulus inversion method based on traveling wave decomposition and machine learning according to claim 1, characterized in that, In step S3: Different frequency components train different In the TWENN neural network, during data inversion, data of the same frequency but different directions share the same TWENN block, while different frequencies use TWENNs trained at the corresponding frequencies.

3. The modulus inversion method based on traveling wave decomposition and machine learning according to claim 1, characterized in that, In step S4: The fusion formula is as follows: and These represent the energy storage modulus and the loss modulus, respectively. in, The imaginary unit, For frequency numbering, For the first One frequency, The total frequency count, Number the direction of wave propagation. For the first One direction of wave propagation, The total number of directions, For normalized complex wavenumber.

4. A modulus inversion system based on traveling wave decomposition and machine learning, characterized in that, include: Module M1: Imaging and acquiring displacement fields for wave-like displacement, generating training data based on traveling wave decomposition; Module M2: Constructs the complex covariance neural network TWENN; Module M3: Creates a noisy training dataset generated by the traveling wave decomposition model; Module M4: Utilizes the trained network for inversion and multi-frequency, multi-directional fusion; In module M1: Module M1.1: Under the assumption of local homogeneity, the wave field in an incompressible isotropic medium is a synthesis of multiple traveling waves: in, It is a spatial position vector. Is The total wave field at that location, It is the total number of traveling waves. It is the first The complex amplitude of a traveling wave, It is the first The direction of propagation of a single traveling wave. It is the local complex wavenumber. It is the real wave number. It is an exponential term related to decay; The imaginary unit; Set different numbers of traveling waves Different complex amplitudes Different directions of dissemination and different A series of training data were simulated; Module M1.2: Adds complex Gaussian noise to the wave field: in, It is a unit complex Gaussian noise. Noise intensity; Module M1.3: For Normalize the frequency: The noisy wavefield and normalized wavenumber were used as the training set for the subsequent network. , The vibration frequency, For normalized complex wavenumber; In module M2: Module M2.1: Performs covariance preprocessing on the network input. ,in For complex wave fields; Module M2.2: Complex fully connected neural network, where the weights and biases are all complex numbers, and the activation function is: in, , The complex number is the input to the activation function. Let be the real-valued parameters to be learned. for The phase angle; Module M2.3: divided into and There are two estimation structure branches, and the two branches have the same structure. The output layer takes the real part and the imaginary part as the final output, respectively. The network training and inference branches are carried out independently.

5. The modulus inversion system based on traveling wave decomposition and machine learning according to claim 4, characterized in that, In module M3: Different frequency components train different In the TWENN neural network, during data inversion, data of the same frequency but different directions share the same TWENN block, while different frequencies use TWENNs trained at the corresponding frequencies.

6. The modulus inversion system based on traveling wave decomposition and machine learning according to claim 4, characterized in that, In module M4: The fusion formula is as follows: and These represent the energy storage modulus and the loss modulus, respectively. in, The imaginary unit, For frequency numbering, For the first One frequency, The total frequency count, Number the direction of wave propagation. For the first One direction of wave propagation, The total number of directions, For normalized complex wavenumber.

Images