Muscle-bone ultrasonic multi-parameter imaging method based on physical information neural network

Abstract

The invention relates to a muscle-bone ultrasonic multi-parameter imaging method based on a physical information neural network. The method comprises the following steps: S1, obtaining an observation signal; s2, constructing a generator and a discriminator; s3, generating an analog ultrasonic signal generated at the position of the ultrasonic transducer; s4, inputting the simulated ultrasonic signal and the observation signal into a discriminator together, and training the discriminator based on a generative adversarial task objective function; and S5, updating the current sound velocity distribution and the current medium density distribution of the generator, generating the simulated ultrasonic signal again, if the difference between the simulated ultrasonic signal and the observation signal is smaller than a threshold value and does not reach the training period, returning to S3, and otherwise, taking the current sound velocity distribution and the current medium density distribution at the moment as a multi-parameter imaging result of the muscle bone tissue. Compared with the prior art, the method has the advantages that the physical consistency of imaging results is improved, and then multi-parameter imaging of sound velocity and density of muscle bone tissue is accurately achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of ultrasonic multi-parameter imaging, and in particular to a musculoskeletal ultrasonic multi-parameter imaging method based on physical information neural network. Background Art

[0002] Musculoskeletal tissue is an essential component of the human body and plays a vital role in human life. Pathologies of musculoskeletal tissue can directly lead to muscle loss, joint inflammation, and bone loss, seriously impacting the quality of daily life. Ultrasound imaging has gained increasing attention in the detection of musculoskeletal diseases due to its radiation-free, portable, and efficient nature. However, due to the strong scattering and high attenuation properties of bones in musculoskeletal tissue, ultrasound waves are distorted in phase after passing through bone tissue, and their energy is significantly attenuated, resulting in a low signal-to-noise ratio for the signal received by the ultrasonic transducer, making it impossible to accurately image the internal structure of musculoskeletal tissue.

[0003] Full-wave inversion technology has shown the advantage of high resolution in ultrasound imaging of musculoskeletal tissue. However, due to the large difference between the sound velocity of bone and the surrounding soft tissue in musculoskeletal tissue, this method is prone to falling into the trap of local extrema during the imaging process. This method also has problems such as dependence on the initial model and large amount of inversion calculation. Data-driven deep learning methods have achieved image quality enhancement in musculoskeletal ultrasound imaging, but this method has problems such as dependence on large-scale training data and weak interpretability. In addition, most existing methods focus on the high-resolution reconstruction of the sound velocity parameters of musculoskeletal tissue, ignoring the influence of other acoustic parameters of musculoskeletal tissue on ultrasound propagation. Traditional multi-parameter imaging methods based on alternating inversion strategies cannot accurately reflect the structure and state of musculoskeletal tissue due to problems such as parameter coupling and inversion ill-posedness. These problems limit the application of ultrasound imaging in the detection of clinical musculoskeletal diseases. Summary of the Invention

[0004] The present invention aims to alleviate the problem that the imaging results cannot accurately reflect the structure and state of musculoskeletal tissue due to parameter coupling and inversion ill-posedness. It provides a musculoskeletal ultrasonic multi-parameter imaging method based on physical information neural network. It introduces a two-dimensional elastic wave equation as a generator and then introduces physical information as a constraint, thereby improving the physical consistency of the imaging results and accurately realizing multi-parameter imaging of the sound velocity and density of musculoskeletal tissue.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A musculoskeletal ultrasound multi-parameter imaging method based on physical information neural network, the method comprising the following steps:

[0007] S1. The ultrasonic transducer array element transmits a pulse signal and receives and records the ultrasonic signal after passing through the medium to obtain an observation signal;

[0008] S2. Constructing a generator and a discriminator, wherein the generator includes a current sound speed and density model;

[0009] S3, the current sound velocity and density model obtains a simulated ultrasonic signal generated at the ultrasonic transducer position according to the current sound velocity distribution and the current medium density distribution;

[0010] S4, the simulated ultrasonic signal and the observed signal are input into the discriminator together, and the discriminator is trained based on the generative adversarial task objective function. After the iterative training is completed, the trained discriminator is obtained;

[0011] S5. Train the generator according to the generative adversarial task objective function, update the generator's current sound speed distribution and current medium density distribution, obtain a new sound speed and density model, use the new sound speed and density model as the current sound speed and density model, and generate a simulated ultrasound signal again. If the difference between the simulated ultrasound signal and the observed signal is less than the threshold and the training cycle has not been reached, return to S3. Otherwise, use the current sound speed distribution and current medium density distribution at this time as the multi-parameter imaging result of musculoskeletal tissue.

[0012] Furthermore, the current sound speed and density model is:

[0013]

[0014] Among them, x and z form the xz two-dimensional plane, ρ represents the current medium density distribution, and u x and u z Respectively represent the fluctuation amplitude in the x-direction and z-direction with time t as the coordinate, σ xx ,σ xz ,σ zz are all stress components, corresponding to the normal stress in the x direction, the shear stress in the x and z directions, and the normal stress in the z direction, v is the current sound velocity distribution, λ and μ are the first and second Lame coefficients, respectively.

[0015] Furthermore, the simulated ultrasonic signal is composed of fluctuation amplitudes in the x-direction and the z-direction with time t as the coordinate.

[0016] Furthermore, the objective function of the generative adversarial task is:

[0017]

[0018] Among them, L(D φ ,G P ) represents the objective function of the generated adversarial task, G P Denotes the generator, D φ represents the discriminator, φ represents the parameters in the network, P truerepresents the probability distribution function of the observed signal, P gen Represents the probability distribution function of the simulated ultrasonic signal, u i and u j They are from P true and P gen Samples from the data distribution, P uni Indicates P true and P gen The linear interpolation mixture distribution of , η represents the weight coefficient of the gradient penalty term.

[0019] Furthermore, the discriminator includes a depthwise separable convolution module and a feature extraction module.

[0020] Furthermore, the depthwise separable convolution module includes 7×7 and 1×1 convolution layers, and the feature extraction module includes a 3×3 convolution layer and a 2×2 pooling layer.

[0021] Furthermore, the specific steps of transmitting a pulse signal by an ultrasonic transducer array element and receiving and recording the ultrasonic signal after passing through the medium are as follows:

[0022] The ultrasonic transducer array consists of N array elements. Array element No. 1 in the transducer array transmits a pulse signal, and all array elements receive and record the ultrasonic signal after passing through the medium. Then array element No. 2 transmits a pulse signal, and so on until the Nth array element transmits a pulse signal and records the ultrasonic signal.

[0023] Furthermore, the observation signals are all recorded ultrasonic signals.

[0024] Furthermore, the dimension of the observation signal is the same as that of the simulated ultrasonic signal, which is N s ×N r ×N t , N s and N r are the number of transmitting array elements and receiving array elements in the full matrix receiving mode of the ring ultrasonic transducer array, N s and N r Both are N, N t is the number of time samples.

[0025] Furthermore, the Adam optimizer is used in the process of updating the current sound speed distribution and the current medium density distribution of the generator.

[0026] The present invention combines the nonlinear mapping advantages of neural networks to alleviate the ill-posedness problem of traditional musculoskeletal ultrasound imaging; introduces physical information as a constraint in the adversarial training process, reduces the dependence on large-scale training data, and improves the physical consistency of the imaging results; uses automatic differentiation to independently update the parameter distribution, alleviating the multi-parameter coupling problem. As a result, the present invention can accurately achieve multi-parameter imaging of the sound velocity and density of musculoskeletal tissue. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flow chart of the musculoskeletal ultrasonic multi-parameter imaging method based on physical information neural network of the present invention;

[0028] Figure 2 are the actual sound velocity density distribution and the initial sound velocity density distribution of the tibia and fibula model of the present invention, wherein, Figure 2 (a) is the real sound speed distribution, Figure 2 (b) is the initial sound velocity distribution, Figure 2 (c) is the true density distribution, Figure 2 (d) is the initial density distribution;

[0029] Figure 3 This is a schematic diagram of the tibia and fibula imaging scene of the present invention;

[0030] Figure 4 The multi-parameter imaging results of the tibia and fibula model of the present invention and the traditional full-wave inversion are shown in FIG. Figure 4 (a) is the result of traditional full-wave inversion sound velocity imaging. Figure 4 (b) is the result of traditional full-wave inversion density imaging. Figure 4 (c) is the sound velocity imaging result of the present invention, Figure 4 (d) is the density imaging result of the present invention;

[0031] Figure 5 The figure shows the comparison of the sound velocity and density imaging results at "z = 75 mm" between the present invention and the traditional full-wave inversion imaging results, as well as the relative error comparison with the actual sound velocity and density distribution. Figure 5 (a) is the comparison of sound velocity imaging results. Figure 5 (b) is the relative error calculation of sound velocity imaging, Figure 5 (c) is the comparison of density imaging results. Figure 5 (d) is the relative error calculation of density imaging. DETAILED DESCRIPTION

[0032] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0033] The present invention proposes a musculoskeletal ultrasonic multi-parameter imaging method based on a physical information neural network, which can alleviate the problems of traditional full-wave inversion methods that are prone to falling into local extreme values, multi-parameter coupling, and large-scale data dependence, and provide high-resolution imaging of musculoskeletal tissue sound velocity and density parameters.

[0034] In order to solve the above problems, the present invention provides a method for multi-parameter ultrasound imaging of musculoskeletal tissue based on physical information neural network, which can achieve high-resolution sound velocity and density imaging of musculoskeletal tissue. The flow chart is as follows: Figure 1 As shown, the following technical solutions are adopted:

[0035] Step 1: For a realistic musculoskeletal model, a non-focused full-matrix reception mode is used. The ultrasonic transducer array elements sequentially transmit pulse signals, and all elements receive the signals. The ultrasonic signals, after propagation through the medium, are received and stored as observation signals.

[0036] Step 2: Use a physics-based generator to calculate the signal received at the ultrasonic transducer array using the current sound velocity and density distribution. During the initial calculation, an initial sound velocity and density model is set as the current sound velocity and density distribution, and the elastic wave equation is introduced to form the physics-based generator. The current sound velocity and density model is used to calculate the simulated ultrasonic signal generated at the ultrasonic transducer location after the transmitted signal passes through the medium.

[0037] Step 3: Train a neural network-based discriminator to maximize the difference between the observed signal and the simulated ultrasound signal. The goal is to enable the discriminator to easily distinguish between the observed signal and the simulated ultrasound signal.

[0038] Step 4: Train the physics-based generator and use the optimizer to update the sound velocity and density distributions using the automatic differentiation gradient provided by the discriminator. The simulated ultrasonic signal at the ultrasonic transducer array is recalculated based on the current sound velocity and density distributions and the elastic wave equation.

[0039] Step 5: Determine whether the difference between the simulated ultrasound signal and the observed signal meets the requirements, or whether the training cycle is completed, and use the sound velocity and density distribution after the training cycle as the final multi-parameter imaging result.

[0040] Furthermore, in step 1, a non-focused full-matrix receiving mode is used, where the ultrasonic transducer array elements sequentially transmit pulse signals, and all elements receive the signals. The ultrasonic signals propagated through the medium are received and stored as observation signals. Specifically:

[0041] A circular ultrasonic transducer array consisting of N array elements is placed horizontally in an ultrasonic coupling environment, and a musculoskeletal tissue model is placed in the center of the circular ultrasonic transducer array.

[0042] Element 1 in the transducer array transmits a pulse signal toward the musculoskeletal tissue model, while all other elements receive and record the ultrasonic signal after it passes through the medium. Element 2 then transmits a pulse signal, and all other elements receive and record the ultrasonic signal until the Nth element completes its transmission.

[0043] The ultrasonic signal received in full matrix mode is named observation signal. The dimension of the observation signal after transmission and reception is N s ×N r ×N t , where N s and N r are the number of transmitting elements and receiving elements in the full matrix receiving mode of the ring ultrasonic transducer array, both of which are N. t is the number of time samples.

[0044] Furthermore, in step 2, the initial sound velocity and density model is set as the current sound velocity and density distribution during the first calculation, and the elastic wave equation is introduced to form a generator based on physical information. Specifically:

[0045] The physical information-based generator consists of two parts. The first part is the initial sound speed and density model. Here, the initial sound speed and density model is set to a constant velocity distribution and a constant density distribution according to the musculoskeletal tissue model for subsequent simulation signal generation and sound speed density update.

[0046] The second part introduces physical information. For ultrasonic multi-parameter imaging of musculoskeletal tissue, the two-dimensional elastic wave equation is used as the physical description of ultrasound propagation in the musculoskeletal tissue model. Its expression can be expressed based on the momentum conservation equation and the stress-strain relationship:

[0047]

[0048]

[0049] Among them, x and z form the xz two-dimensional plane, and ρ represents the density distribution matrix of the medium. x and u z Respectively represent the fluctuation amplitude in the x-direction and z-direction with time t as the coordinate. xx ,σ xz ,σ zz are stress components, corresponding to the normal stress in the x direction, the shear stress in the x and z directions, and the normal stress in the z direction. The sound velocity distribution can be expressed according to the first Lame coefficient λ and the second Lame coefficient μ as follows:

[0050] Furthermore, in step 2, the simulated ultrasonic signal generated at the ultrasonic transducer position after the transmitted signal passes through the medium is calculated based on the current sound speed and density model. Specifically:

[0051] According to the current sound velocity and density distribution, the above two-dimensional elastic wave equation is used to calculate the ultrasonic signal received at each element of the annular ultrasonic transducer array, and the calculation result is used as a simulated ultrasonic signal for subsequent calculations. The dimension of the simulated ultrasonic signal is N s ×N r ×N t , which is consistent with the observed signal.

[0052] Furthermore, in step 3, a neural network-based discriminator is trained to quantify the difference between the observed signal and the simulated ultrasound signal to the greatest extent possible. Specifically:

[0053] The discriminator's network architecture is based on a deep stack of depthwise separable convolutional modules and feature extraction modules. The depthwise separable convolutional modules are constructed using 7×7 and 1×1 convolutional layers, followed by Gaussian error linear units (LEUs), with residual connections between input and output. The feature extraction module is constructed using 3×3 convolutional layers and 2×2 pooling layers, followed by leaky linear rectifier units (LRUs).

[0054] The simulated ultrasonic signal and the observed signal are used as the input of the discriminator. After passing through the depthwise separable convolution and feature extraction modules, the results are flattened and reshaped using a fully connected layer to finally obtain the quantitative difference between the two, which is maximized during the training process.

[0055] Furthermore, in step 4, the generator based on physical information is trained, and the speed of sound and density distribution are updated separately using the optimizer through the automatic differentiation gradient provided by the discriminator. Specifically:

[0056] The discriminator and the physics-based generator are trained using an asymmetric adversarial training strategy. Specifically, the discriminator is trained multiple times to distinguish between simulated and observed ultrasound signals. The weight parameters of each layer in the network are adjusted based on the automatic differentiation gradient, so that the discriminator output maximizes the difference between the observed and simulated ultrasound signals. The generator is then trained again, using the Adam optimizer to update the values ​​at the imaging locations in the sound velocity and density distributions based on the automatic differentiation gradient provided by the discriminator, minimizing the difference in the discriminator output.

[0057] Furthermore, in step 4, the simulated ultrasonic signal at the ultrasonic transducer array is calculated again based on the current sound velocity and density distribution and the elastic wave equation. Specifically:

[0058] The generator based on physical information can update the sound velocity and density distribution values ​​of the current medium, and use the updated sound velocity and density distribution combined with the two-dimensional elastic wave equation to recalculate the simulated ultrasonic signal at each array element of the annular ultrasonic transducer array, and use it and the observed data as the input of the discriminator when training the generator.

[0059] The training process of the physical information-based generator and discriminator can be described as the following generative adversarial task:

[0060]

[0061] Among them, L(D φ ,G P ) represents the joint optimization task of sound velocity and density in multi-parameter imaging. G P Denotes the generator that introduces the physical information of the two-dimensional constant density wave equation, D φ P represents the mapping function of the discriminator neural network, and φ represents the parameters in the network. true represents the probability distribution function of the observed signal under the real sound speed and density distribution, P gen Represents the probability distribution function of the simulated ultrasonic signal calculated using the two-dimensional elastic wave equation under the current sound velocity and density distribution. i and u j They are from P true and P gen A sample from the data distribution. uni Indicates P true and P gen The linear interpolation mixture distribution of , η represents the weight coefficient of the gradient penalty term.

[0062] Furthermore, in step five, it is determined whether the difference between the simulated ultrasound signal and the observed signal meets the requirements, or whether the training cycle is completed, and the sound velocity and density distribution after the training cycle is completed is used as the final multi-parameter imaging result. Specifically:

[0063] The difference between the simulated ultrasound signal and the observed signal is determined based on the output of the discriminator. When the discriminator cannot distinguish between the two, or the training cycle has ended, the most recently updated sound velocity and density distribution are used as the multi-parameter imaging results of musculoskeletal tissue.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] This method combines the advantages of nonlinear mapping in neural networks to alleviate the ill-posedness problem of traditional musculoskeletal ultrasound imaging. It also introduces physical information as a constraint during adversarial training, reducing the reliance on large-scale training data and improving the physical consistency of the imaging results. It also uses automatic differentiation to independently update parameter distributions, alleviating the problem of multi-parameter coupling. This method can accurately achieve multi-parameter imaging of the sound velocity and density of musculoskeletal tissue.

[0066] Next, we will describe the application of the invention method for multi-parameter imaging of the tibia and fibula model. The structure of the tibia and fibula model is derived from the results of high-resolution peripheral quantitative computer scanning, and the speed of sound and density parameters are allocated based on the actual conditions of human musculoskeletal tissue. Specifically, Figure 2 As shown, the actual size of the tibia and fibula model is 10.5 cm × 10.5 cm, the average speed of sound of the muscle tissue is set to 1600 m / s, and the average density is set to 1080 kg / m 3 The sound velocity range of bone tissue is set to 1600m / s~2800m / s, and the density range is set to 1080kg / m 3 ~2000kg / m 3 .

[0067] In step 1, for a realistic musculoskeletal tissue model, a non-focused full-matrix reception mode is used, with the ultrasonic transducer array elements sequentially transmitting pulse signals, and all elements receiving the signals. The ultrasonic signals propagated through the medium are received and stored as observation signals. Specifically:

[0068] Tibial and fibula ultrasound multi-parameter imaging environment Figure 3 As shown, the tibia and fibula are placed vertically in the center of the annular ultrasonic transducer. The annular ultrasonic transducer consists of 128 array elements and can transmit and receive ultrasonic signals within a two-dimensional cross-section of the tibia and fibula with a diameter of 10 cm.

[0069] Starting from array element 1, a pulsed ultrasound signal with a center frequency of 0.5 MHz was transmitted to the tibia and fibula model. All array elements simultaneously received 100 μs echo and transmission signals at an acquisition frequency of 20 MHz. Then, array element 2 transmitted a pulse signal. A total of 128 transmission and reception operations were performed to obtain the observation signal after the ultrasound propagated through the tibia and fibula model. The signal dimension was 128 × 128 × 2000.

[0070] In step 2, a generator based on physical information is used to calculate the signal received at the ultrasonic transducer array using the current sound velocity and density distribution. During the first calculation, the initial sound velocity and density model is set as the current sound velocity and density distribution, and the elastic wave equation is introduced to form a generator based on physical information. The simulated ultrasonic signal generated at the ultrasonic transducer position after the transmitted signal passes through the medium is calculated based on the current sound velocity and density model. Specifically:

[0071] The current sound velocity and density distribution are discretized into a 210×210 grid, and the actual physical size of each grid is 0.5mm×0.5mm. The initial sound velocity model is set to a fixed sound velocity value of 1600m / s, and the initial density model is set to a fixed density value of 1080kg / m 3 Based on the elastic wave equation, a perfectly matched layer condition was applied to the boundaries of the entire imaging area. While ensuring numerical stability, a simulated ultrasound signal with a total length of 100 μs was calculated based on the initial sound velocity and density model. Its dimensionality was consistent with the observed signal.

[0072] In step three, a neural network-based discriminator is trained to quantify the difference between the observed signal and the simulated ultrasound signal to the greatest extent possible. The goal is to enable the discriminator to easily distinguish between the observed signal and the simulated ultrasound signal. Specifically:

[0073] The observed signal and the simulated ultrasonic signal are divided into 8 batches and input into the discriminator network to improve the computational efficiency and reduce the memory consumption. The learning rate of the discriminator network is set to 10 -3 , used to update the weight parameters of the discriminator network. The gradient penalty weight is set to 10 to enhance stability during training. The difference between the observed signal and the simulated signal is maximized during the discriminator training process.

[0074] In step 4, the generator based on physical information is trained, and the optimizer is used to update the sound velocity and density distribution respectively through the automatic differentiation gradient provided by the discriminator. The simulated ultrasonic signal at the ultrasonic transducer array is recalculated based on the current sound velocity and density distribution and the elastic wave equation. Specifically:

[0075] Using an asymmetric adversarial training strategy, the discriminator is trained six times in a single training cycle, followed by the generator. The generator uses two gradient descent-based Adam optimizers with adaptive learning rate adjustment to update the sound velocity and density distributions. The momentum parameters β1 and β2 are set to 0.5 and 0.9, respectively. The learning rates for sound velocity updates are set to 30 and 10, respectively. The generator updates the sound velocity and density distributions to minimize the difference between the observed and simulated signals.

[0076] After the physics-based generator is updated, the simulated ultrasonic signal is recalculated using the two-dimensional elastic wave equation based on the updated sound velocity and density distribution.

[0077] In step five, it is determined whether the difference between the simulated ultrasound signal and the observed signal meets the requirements, or whether the training cycle is completed, and the sound velocity and density distribution after the training cycle is completed is used as the final multi-parameter imaging result.

[0078] The discriminator's difference quantification results determine whether the simulated ultrasound signal is consistent with the observed signal. If they are consistent, the current sound velocity and density distributions are used as the final imaging results. The training cycle is set to 200. When the training cycle ends, the difference between the simulated ultrasound signal and the observed signal is considered to have converged to a minimum value. The sound velocity and density distributions at this point can also be used as the final imaging results of the tibia and fibula.

[0079] Figure 4 This is the result of multi-parameter ultrasonic imaging of the tibia and fibula model using the traditional full-wave inversion method and the method of the present invention. The imaging results show that the traditional full-wave inversion method can roughly image the internal structure of bone tissue, but due to problems such as instability and parameter coupling during the imaging process, the final imaging effect is affected. In contrast, the imaging results of the present invention can accurately image the cortical bone and cancellous bone of the tibia and fibula, and its internal microstructural information such as trabeculae is accurately reconstructed. At the same time, the values ​​of the sound velocity and density distribution are closest to the real results, the structural similarity index exceeds 0.94, and the relative error is kept within 8%, which can achieve accurate multi-parameter imaging of musculoskeletal tissue.

[0080] Figure 5 Comparison of the multi-parameter imaging results of the tibia and fibula at z = 75 mm with the true sound velocity and density distribution. This comparison demonstrates that the sound velocity and density imaging results of the present invention are closer to the actual values. The maximum relative error between the sound velocity distribution and the true value is less than 15%, and the maximum relative error between the density distribution and the true value is less than 20%.

[0081] The present invention proposes a musculoskeletal ultrasonic multi-parameter imaging method based on a physical information neural network. The method comprises the following steps: Step 1: Acquire ultrasonic observation signals using a full-matrix receiving mode; Step 2: Generate simulated ultrasonic signals based on a physical information generator; Step 3: Train a discriminator to maximize the distinction between the observation signal and the simulated signal; Step 4: Train the generator to update the sound velocity and density distribution and generate a simulated signal accordingly; Step 5: Determine whether the signal difference meets the requirements or whether the training cycle has ended, and use the final updated sound velocity and density distribution as the multi-parameter imaging result of musculoskeletal tissue. Compared with the existing technology, the present invention can alleviate the ill-posedness and parameter coupling problems of full-wave inversion, while reducing the dependence of data-driven deep learning on large-scale training data, thereby achieving high-resolution imaging of musculoskeletal tissue sound velocity and density parameters.

[0082] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A musculoskeletal ultrasound multi-parameter imaging method based on physical information neural network, characterized in that: The method comprises the following steps: S1. The ultrasonic transducer array element transmits a pulse signal and receives and records the ultrasonic signal after passing through the medium to obtain an observation signal; S2. Constructing a generator and a discriminator, wherein the generator includes a current sound speed and density model; S3, the current sound velocity and density model obtains a simulated ultrasonic signal generated at the ultrasonic transducer position according to the current sound velocity distribution and the current medium density distribution; S4, the simulated ultrasonic signal and the observed signal are input into the discriminator together, and the discriminator is trained based on the generative adversarial task objective function. After the iterative training is completed, the trained discriminator is obtained; S5. Train the generator according to the generative adversarial task objective function, update the generator's current sound speed distribution and current medium density distribution, obtain a new sound speed and density model, use the new sound speed and density model as the current sound speed and density model, and generate a simulated ultrasound signal again. If the difference between the simulated ultrasound signal and the observed signal is less than the threshold and the training cycle has not been reached, return to S3. Otherwise, use the current sound speed distribution and current medium density distribution at this time as the multi-parameter imaging result of musculoskeletal tissue.

2. The method for musculoskeletal ultrasound multi-parameter imaging based on physical information neural network according to claim 1, characterized in that: The current sound speed and density model is: Among them, x and z form the xz two-dimensional plane, ρ represents the current medium density distribution, and u x and u z Respectively represent the fluctuation amplitude in the x-direction and z-direction with time t as the coordinate, σ xx ,σ xz ,σ zz are all stress components, corresponding to the normal stress in the x direction, the shear stress in the x and z directions, and the normal stress in the z direction, v is the current sound velocity distribution, λ and μ are the first and second Lame coefficients, respectively.

3. The method for musculoskeletal ultrasound multi-parameter imaging based on physical information neural network according to claim 2, characterized in that: The simulated ultrasonic signal is composed of fluctuation amplitudes in the x-direction and the z-direction with time t as the coordinate.

4. The method for musculoskeletal ultrasound multi-parameter imaging based on physical information neural network according to claim 3, characterized in that: The objective function of the generative adversarial task is: Among them, L(D φ ,G P ) represents the objective function of the generated adversarial task, G P Denotes the generator, D φ represents the discriminator, φ represents the parameters in the network, P true represents the probability distribution function of the observed signal, P gen Represents the probability distribution function of the simulated ultrasonic signal, u i and u j They are from P true and P gen Samples from the data distribution, P uni Indicates P true and P gen The linear interpolation mixture distribution of , η represents the weight coefficient of the gradient penalty term.

5. The method for musculoskeletal ultrasound multi-parameter imaging based on physical information neural network according to claim 1, characterized in that: The discriminator includes a depthwise separable convolution module and a feature extraction module.

6. The method for musculoskeletal ultrasound multi-parameter imaging based on physical information neural network according to claim 5, characterized in that: The depthwise separable convolution module includes 7×7 and 1×1 convolution layers, and the feature extraction module includes a 3×3 convolution layer and a 2×2 pooling layer.

7. The method for musculoskeletal ultrasound multi-parameter imaging based on physical information neural network according to claim 1, characterized in that: The specific steps of the ultrasonic transducer array element transmitting a pulse signal and receiving and recording the ultrasonic signal after passing through the medium are as follows: The ultrasonic transducer array consists of N array elements. Array element No. 1 in the transducer array transmits a pulse signal, and all array elements receive and record the ultrasonic signal after passing through the medium. Then array element No. 2 transmits a pulse signal, and so on until the Nth array element transmits a pulse signal and records the ultrasonic signal.

8. The method for musculoskeletal ultrasound multi-parameter imaging based on physical information neural network according to claim 7, characterized in that: The observation signals are all recorded ultrasonic signals.

9. The method for musculoskeletal ultrasound multi-parameter imaging based on physical information neural network according to claim 8, characterized in that: The dimension of the observed signal is the same as that of the simulated ultrasonic signal, both are N s ×N r ×N t , N s and N r are the number of transmitting array elements and receiving array elements in the full matrix receiving mode of the ring ultrasonic transducer array, N s and N r Both are N, N t is the number of time samples.

10. The method for musculoskeletal ultrasound multi-parameter imaging based on physical information neural network according to claim 1, characterized in that: The Adam optimizer is used in the process of updating the current sound speed distribution and the current medium density distribution of the generator.

Images