An adaptive weight gradient optimization method for bio-tissue ultrasound inversion

Abstract

The application relates to an adaptive weight gradient optimization method for biological tissue ultrasonic inversion, and the method comprises the following steps: S1, acquiring a rough tissue region division vector based on travel time tomography, which represents the position of a point in the region; S2, initializing the variational posterior distribution of each particle; S3, calculating a region-dependent weight; S4, obtaining simulated ultrasonic data corresponding to each sample sampled from the current variational posterior distribution; S5, calculating a joint loss function based on the current variational distribution, the sample, the simulated ultrasonic data and experimental observation data; S6, judging whether convergence is achieved based on the loss function, if yes, executing S7, otherwise, iteratively calculating based on the loss function to obtain a new variational posterior distribution, and returning to S3; S7, taking the current mean value as an inversion imaging result and taking the current standard deviation as the inversion uncertainty of the inversion imaging result. Compared with the prior art, the application has the advantages of solving the period jump problem in complex biological tissue inversion and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ultrasonic inversion imaging, in particular to an adaptive weight gradient optimization method for biological tissue ultrasonic inversion. BACKGROUND

[0002] As a non-invasive and real-time medical imaging technology, ultrasonic imaging has been widely used in the detection and diagnosis of various tissues in the human body. In biological tissue imaging, full waveform inversion (FWI) has become an important research direction in the field of ultrasonic imaging because it can fully utilize multiple scattering and diffraction information to provide higher resolution and more detailed imaging results than traditional methods. However, the complex and multi-level structure of biological tissues makes the traditional FWI method prone to cycle skipping problems during inversion, i.e., due to phase mismatch or initial model bias, the inversion result appears severe fluctuations or discontinuous jumps in some areas, which seriously interferes with the accuracy and physical reasonableness of parameter reconstruction.

[0003] In recent years, variational inference (VI) method as an efficient approximation form of Bayesian inference has gradually attracted attention in full waveform inversion. This method regards medium parameters as random variables and constructs a parameterizable approximation distribution to approximate the true posterior distribution, thereby realizing the quantification of parameter uncertainty. In the variational inference framework, the imaging result is usually taken as the posterior mean, and the standard deviation is used to characterize the inference confidence at different positions, thereby providing more uncertainty information for subsequent imaging. However, due to the multi-scale characteristics and spatial heterogeneity of biological tissues, the traditional VI method adopts a globally homogeneous prior modeling strategy, which fails to combine the differences in tissue structure to identify and optimize uncertainty in different regions, resulting in limited inversion accuracy and convergence speed in high complexity areas (such as the interface between soft and hard tissues). The traditional FWI algorithm flow chart is shown in Figure 1

[0004] Currently, some studies have attempted to alleviate the cycle skipping problem by introducing regularization, deep learning initialization, etc. For example, by adding smoothing or structural prior to constrain the model, or using neural networks to generate reasonable initial values to avoid error propagation. However, these methods have certain limitations: regularization will sacrifice the boundary clarity and image resolution to some extent; deep learning methods rely on training data sets, and have the problems of poor generalization ability and high computational resource requirements. In addition, existing methods generally fail to effectively utilize the standard deviation estimates obtained in variational inference, lack a dynamic regulation mechanism for uncertainty regions, and are difficult to meet the demand for high-precision and high-stability imaging under complex biological tissue structures. SUMMARY ​

[0005] The application aims to provide an adaptive weight gradient optimization method for biological tissue ultrasonic inversion.

[0006] The application aims to provide an adaptive weight gradient optimization method for biological tissue ultrasonic inversion.

[0007] An adaptive weight gradient optimization method for biological tissue ultrasonic inversion, the method comprising the following steps:

[0008] S1, obtaining a coarse tissue region division vector based on travel time tomography , denotes the position of a point in the region;

[0009] S2, initializing a variational posterior distribution , the initialized variational posterior distribution is modeled as a Gaussian distribution with a spatial dependence structure, wherein the variational distribution parameter , is the mean of the Gaussian distribution, is the standard deviation of the Gaussian distribution, and the preliminary sound velocity obtained by travel time tomography inversion is used as the mean of the initialized variational posterior distribution , the standard deviation of the initialized variational posterior distribution is obtained based on the inversion residual in the travel time imaging and the local ray path length , and the mean and the standard deviation are taken as the current mean and the current standard deviation respectively, and the variational posterior distribution is taken as the current variational posterior distribution, denotes the sound velocity parameter;

[0010] S3, calculating a region-dependent weight based on the coarse tissue region division vector ;

[0011] S4, sampling a sound velocity model sample from the current variational posterior distribution, performing forward simulation on the sound velocity model sample by using an acoustic wave equation, and obtaining simulated ultrasonic data corresponding to each sample sampled from the current variational posterior distribution ;

[0012] S5, calculating a joint loss function based on the current variational posterior distribution, the sample, the simulated ultrasonic data, and experimental observation data;

[0013] S6. Determine whether convergence has occurred based on the loss function. If convergence has occurred, proceed to S7. Otherwise, iteratively calculate the loss function to obtain the new mean, new standard deviation, and new variational posterior distribution. Use the new mean, new standard deviation, and new variational posterior distribution as the current mean, current standard deviation, and current variational posterior distribution, respectively, and return to S3.

[0014] S7. Use the current mean as the inversion imaging result and the current standard deviation as the inversion uncertainty of the inversion imaging result.

[0015] Furthermore, the coarse organization region partitioning vector for:

[0016]

[0017] Here, "boundary" represents the boundary region. Indicates the initial speed of sound. This indicates the sound velocity threshold of fat. The muscle sound velocity threshold. The threshold for sound velocity in hard tissue.

[0018] Furthermore, based on the inversion residuals in travel-time imaging With local ray path length Obtain the initialized variational posterior distribution Standard deviation The specific steps are as follows:

[0019] For each source-receive path Calculate the travel time residuals for the path, then calculate the normalized travel time residuals for each point based on the travel time residuals for the path, and finally obtain the initialized variational posterior distribution based on the normalized travel time residuals for each point. Standard deviation .

[0020] Furthermore, the travel time residual of the path is:

[0021]

[0022] in, The observation travel time corresponding to this path, The theoretical travel time is obtained from forward modeling calculations at the initial sound speed. For path The travel time residual.

[0023] Furthermore, the normalized travel time residual is:

[0024]

[0025] in, For path total length of the path, the traversal length of the path at position , is an indicator function.

[0026] Further, the standard deviation of the initialized variational posterior distribution is:

[0027]

[0028] Further, the region-dependent weight is:

[0029]

[0030] wherein, denotes the soft tissue weight, denotes the hard tissue weight, denotes a weight control parameter.

[0031] Further, the soft tissue weight is:

[0032]

[0033] wherein, is a hyperparameter that adjusts the slope, denotes the current standard deviation, is the standard deviation of the uncertainty within the tissue region, denotes the average uncertainty within the tissue region.

[0034] Further, the hard tissue weight is:

[0035]

[0036] wherein, denotes an exponent that controls the non-linear growth of the weight.

[0037] Further, the loss function is:

[0038]

[0039] wherein, denotes the log-likelihood function, is the experimental observation data, is the loss function, denotes the regularization strength, denotes the entropy term of the variational distribution, is the likelihood model with Gaussian noise, the prior distribution is the regularization term.

[0040] ​

[0041] n represents the total number of simulated ultrasound data sets, represents the i-th simulated ultrasound data set, is the experimental observation data;

[0042]

[0043] where the operator is the Hadamard product, represents the i-th sound velocity model sample;

[0044] According to the loss function iterative calculation, the specific steps of obtaining a new mean, a new standard deviation and a new variational posterior distribution are as follows:

[0045] According to the loss function, a new variational distribution parameter is calculated , a new mean and a new standard deviation are obtained, the new variational distribution parameter is substituted into the current variational posterior distribution, and a new variational posterior distribution is obtained.

[0046] Compared with the prior art, the present application has the following beneficial effects:

[0047] 1. Region-aware variational inference modeling improves imaging accuracy and reliability:

[0048] The present application obtains tissue region structure information based on travel time tomography, and realizes parameter initialization by constructing a spatial position related Gaussian distribution as a variational posterior and combining physiological characteristics of different tissue regions. Compared with the practice of using a global homogeneous prior distribution in the traditional variational inference method, the present application is more consistent with the real tissue structure, effectively improves the convergence of variational inference and the physical rationality of the inversion result.

[0049] 2. Adaptive uncertainty weighting mechanism strengthens the inversion ability of low-confidence regions:

[0050] The present application innovatively introduces a region-aware uncertainty measurement method, uses the variational inference inversion standard deviation to measure the uncertainty estimation of each position point, and designs a weighting function for soft tissue, hard tissue and boundary regions according to the above. The weighting mechanism dynamically adjusts the regional sensitivity of the inversion gradient in the optimization process, gives higher weight to the regions with high uncertainty (low signal-to-noise ratio or fuzzy boundary), effectively improves the imaging accuracy of weak signals and edge structures, and at the same time suppresses the excessive iteration of high-quality regions, improves the overall inversion robustness and spatial resolution. In addition, when the standard deviation of uncertainty changes, the weight of the region with increased uncertainty increases exponentially, so that the weight of the region with high uncertainty increases greatly, and the region with high uncertainty is paid more attention in the optimization process.

[0051] 3. Joint loss optimization strategy, enhance model regularization ability and convergence stability:

[0052] The inversion objective function constructed by the present application comprehensively considers the log-likelihood error between the simulation data and the observation data, the weighted regularization term and the entropy term of the variational distribution, and realizes end-to-end joint optimization. The strategy introduces a regional adaptive regularization mechanism on the basis of preserving the consistency of physical data, effectively suppresses the gradient oscillation in the parameter updating process, enhances the convergence stability and anti-interference ability of the algorithm to noise, and avoids the discontinuous phenomena such as periodic jump easily appearing in the traditional inversion. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a traditional FWI algorithm flowchart;

[0054] Figure 2 is a flowchart of the adaptive weight gradient optimization inversion method proposed by the present application;

[0055] Figure 3 is a circular bone canal simulation model diagram. The small diamond-shaped blocks represent the positions of the transducers;

[0056] Figure 4 is a transmit waveform diagram;

[0057] Figure 5 is a traditional FWI inversion result;

[0058] Figure 6 is the mean result of the adaptive weight gradient optimization inversion method of the present application. DETAILED DESCRIPTION

[0059] The present application will be described in detail below in combination with the drawings and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present application, and gives a detailed implementation manner and specific operation process, but the protection scope of the present application is not limited to the following examples.

[0060] The application provides an ultrasound inversion method based on tissue region uncertainty weighting, aiming to improve the accuracy and stability of biological tissue sound velocity imaging. The method comprises the following steps: S1, constructing an initial sound velocity model based on travel time tomography and dividing a coarse tissue region to form a tissue mask; S2, constructing a Gaussian variational posterior distribution with a spatial dependence structure based on the initial sound velocity and travel time residual, which is used to describe the sound velocity and its uncertainty; S3, combining the tissue mask and the uncertainty estimation to design a region-aware adaptive weight function; S4, sampling the sound velocity sample from the variational distribution and inputting the forward simulation module to obtain simulated ultrasound data; S5, constructing a weighted loss function, combining the likelihood term and the prior term to measure the model fitting degree, and judging whether it is converged; S6, updating the variational distribution by using the adaptive weighted variational inference algorithm; S7, iteratively updating the variational distribution until convergence, and outputting the mean value and uncertainty estimation of the inversion result. Compared with the prior art, the application dynamically adjusts the region weight in the inversion process, significantly enhances the imaging ability of the high-uncertainty region, suppresses the local optimal trap problem caused by the cycle jump, and at the same time retains the stable estimation of the low-uncertainty region, effectively improves the continuity, noise resistance and calculation efficiency of the inversion result.

[0061] The object of the application can be achieved by the following technical solutions:

[0062] An adaptive weight gradient optimization method for biological tissue ultrasound inversion, the method comprising the following steps:

[0063] S1, obtaining a coarse tissue region mask vector based on travel time tomography , wherein represents the position of a point in the region, The number of elements in the region is . The specific calculation steps are as follows:

[0064] Using travel time tomography inversion, estimating the preliminary sound velocity . Set the sound velocity threshold range corresponding to the tissue: fat km / s, muscle km / s, and hard tissue such as bone km / s. According to the threshold, the tissue type is divided as follows:

[0065]

[0066] Wherein, boundary represents the boundary region.

[0067] S2, initializing the variational posterior distribution , the variational posterior distribution is modeled as a Gaussian distribution with a spatial dependence structure, represents the sound velocity parameter, and the variational distribution parameter represents the mean value of the Gaussian distribution and , i.e. In variational inference, the mean and the standard deviation measure the uncertainty of the inverted sound speed The preliminary sound speed obtained from traveltime tomography inversion is used as the initial mean, i.e. The inversion residuals in traveltime imaging are mapped to the initial standard deviation of the sound speed at each spatial point. The detailed calculation steps are as follows:

[0068] First, for each source-receiver pair , the traveltime residual of the path is calculated as

[0069]

[0070] where is the observed traveltime of the path, and is the theoretical traveltime obtained from the forward calculation under the current sound speed model . Then, the path residual is projected to the spatial points it passes through, weighted by the path length, to define the normalized traveltime residual at each point as

[0071]

[0072] where is the total length of the path , is the traversal length of the path at position , and is the indicator function. From this residual estimate, the initial standard deviation of the variational distribution at spatial point is constructed as

[0073]

[0074] where is the average traversal length of all paths at position , and is the initial sound speed estimate obtained from traveltime inversion. The final constructed variational posterior distribution is initialized as

[0075]

[0076] ​S3, mask vector Combining uncertainty measure Design adaptive weights for different tissue regions, the specific steps are as follows:

[0077] In soft tissues such as fat and muscle, the sigmoid function is used to smooth the adjustment weight,

[0078]

[0079] Where, The average uncertainty in the tissue region, The standard deviation of the uncertainty in the tissue region. The hyperparameter for adjusting the slope, controlling the amplification degree of high uncertainty area, generally set to .

[0080] In hard tissues such as bone, the weight uses a power function to enhance high uncertainty,

[0081]

[0082] Where, The exponential is used to control the nonlinear growth of the weight, generally . Global maximum uncertainty, used for normalization.

[0083] Boundary region, pay attention to local sharp features, use two kinds of weighting mix,

[0084]

[0085] Where, The weight control parameter.

[0086] Based on the above, the region-dependent weight based on the tissue mask M(x) is constructed as follows ,

[0087]

[0088] S4, sample from the current variational posterior distribution Obtain Speed of sound model samples . For each sample , by numerically solving the acoustic wave equation, a forward simulation is performed to obtain the corresponding simulated waveform data . In the forward simulation setting, there are ultrasound transducers, each transducer is used as a source in turn, and the remaining transducers are used as receivers. The excitation source uses a Ricker wavelet with a center frequency of 0.5 MHz, and the simulation process uses a perfect matched layer as an absorbing boundary. Finally, we get Group simulation data , .

[0089] S5, the S4 obtained , sample , simulation model into the expression of the adaptive weighted loss function , determine whether to converge. Specifically, take a small amount , assuming the loss function value of the first step is , calculate whether to meet, if meet, terminate iteration, output results , using the mean imaging biological tissue, measure uncertainty; if not, continue to perform S6. Specifically, the expression of the loss function is

[0090]

[0091] In the above formula, denotes the regularization strength, the prior distribution is the regularization term, assuming Gaussian prior, there is

[0092]

[0093] Where, the operator is Hadamard product, indicating the corresponding position elements in the vector are multiplied. denotes the log-likelihood, assuming the error between the simulation data and the experimental observation data is Gaussian noise, indicating that the log-likelihood function is

[0094]

[0095] Where, is the variance of the observation noise, which is set to be a certain proportion of the amplitude of the observation data according to empirical knowledge, which is generally set to , with .

[0096] S6, using adaptive weight gradient inversion update, along the gradient descent update ,

[0097]

[0098]

[0099] S7, the updated is brought into , to get the updated variational distribution , return to S3, and repeat S3-S6 until the convergence condition is met.

[0100] The actual experiment is carried out as follows:

[0101] S1, initialization of the tissue region mask. Using the travel time tomography method, the initial observation data is quickly inverted to obtain a rough sound speed distribution . According to the typical sound speed range of tissue types (such as fat 1.4 km / s, muscle 1.5 km / s, and bone tissue 2.0 km / s), the spatial region is divided into different tissue types, and a tissue region mask vector is constructed accordingly, which is used for subsequent regional weighting and differential modeling.

[0102] S2, initialization of the spatial-dependent variational posterior distribution. The variational posterior distribution is modeled as a spatial-dependent Gaussian distribution , where represents the mean field, which is initially taken as the travel time inversion obtained ; is the standard deviation field, which is mapped to the spatial position by weighted averaging the travel time residuals , combined with the path length normalization, to construct the initial uncertainty , ensuring and reflecting the uncertainty size of the inversion sound speed.

[0103] S3, construction of the region-aware weight function. According to the mask and uncertainty estimation , three types of adaptive weighting functions for tissue regions are designed. For soft tissue regions, a sigmoid function is used for smooth weighting; for high-impedance tissues such as bone, a power function is used to emphasize high-uncertainty regions; for boundary regions, a weighted combination of sigmoid and power functions is used. Thus, a unified spatial weighting field is formed, which is used to control the gradient optimization direction and regularization strength.

[0104] S4, sampling of sound speed samples from the current variational distribution and forward modeling. The current variational posterior is sampled to generate sound speed model samples , and numerical solution of the acoustic wave equation is used for forward modeling to obtain the simulated ultrasonic data corresponding to each sample. In the forward modeling configuration, a Ricker wavelet with a center frequency of 0.5 MHz is used as the excitation signal, all transducers are used as transmitters in turn, and the rest are used as receivers, with the boundary condition set as a perfect matched layer.

[0105] S5, will ,sample Simulation data and experimental observation data Substituting these into the joint loss function, we find a log-likelihood term, a region-weighted prior regularization term, and a variational entropy term.

[0106]

[0107] in This is the regularization strength coefficient. It calculates the change in loss value between the current round and the previous round. Is it less than the convergence threshold? (like (Determine whether to terminate).

[0108] S6. Adaptive weighted gradient update of parameters: If convergence is not achieved, adjust the variational distribution parameters according to the loss function. Take the derivative and update using the following formula:

[0109]

[0110] in,

[0111]

[0112] S7. Iterate and optimize until convergence. (Updated version) Substitute Return to step S3 to recalculate the weight function, and repeat steps S3–S6 until the convergence condition is met. The final output is the variational distribution mean. As a result of inversion imaging, the standard deviation This indicates uncertainty in the inversion process.

[0113] To demonstrate the invention more intuitively, this embodiment uses a circular bone tube simulation model for demonstration (see...). Figure 3 In this model, the inner radius of the bone canal is 0.3 cm, the outer radius is 0.48 cm, and the wall thickness is 0.18 cm. The sound velocity inside the bone canal is set to 3 km / s, while the sound velocity in the background medium is set to 1.5 km / s, simulating the sound velocity of water. This difference in sound velocity leads to changes in wave propagation. In the initial stage of the inversion, if the difference in sound velocity between the model and the actual bone canal is large, it will cause a periodic jump phenomenon, leading to the inversion results getting trapped in local minima, such as... Figure 5 As shown. Figure 6 The mean result of the adaptive weight gradient optimization inversion method of the present invention is compared with... Figure 5 In comparison, the results are significantly improved and are closer to the real model.

[0114] The embodiment adopts a two-dimensional grid model, and the grid size is set to 201 201, the spatial step is 0.3 mm, and the total area of the calculation region is 6 cm 6 cm. In the experiment, 32 sources and 32 receivers are used, and all the sources are uniformly distributed in a circular region with a radius of 2.55 cm, and the center of the circle is located at the center of the grid. The layout of the receivers is the same as that of the sources. The excitation source waveform is selected as a Ricker wavelet (see Figure 4 ), the center frequency of which is 0.5 MHz, and the duration is 0.1 ms.

[0115] In this embodiment, the Langevin Stein Variational Gradient Descent (LSVGD) algorithm is taken as an example to perform sound velocity inversion on a bone pipe simulation model. The present application is applicable to various variational inference methods including Stein Variational Gradient Descent (SVGD), Stochastic Variational Inference (SVI), Automatic Differentiation Variational Inference (ADVI), and the specific algorithm can be flexibly selected according to the specific application scene.

[0116] It should be noted that in LSVGD, the variational distribution is represented by a set of particles , the entropy term of the variational distribution does not exist, and LSVGD updates a sample particle instead of the distribution . The specific implementation of the present application will be described in detail below, and the specific flowchart is shown in Figure 2 . The particle is the sound velocity distribution.

[0117] S1, construct a coarse tissue region mask M(x). Based on the travel time tomography method, perform preliminary sound velocity inversion on the target region to obtain an initial sound velocity field . Set different region typical sound velocity threshold values: the background medium sound velocity range km / s, and the high sound velocity bone pipe sound velocity is at least greater than km / s. According to this, the region is divided into three categories and a tissue mask vector is formed:

[0118]

[0119] where boundary denotes the boundary region.

[0120] S2, initializing the particle distribution . The initial velocity model is modeled as a Gaussian distribution with spatial dependence structure. The preliminary velocity estimates obtained from traveltime tomography inversion are used as the initial mean value, i.e. The inversion residuals in traveltime imaging are mapped to the initial standard deviation of the distribution. The specific calculation steps are as follows:

[0121] First, for each source-receiver pair path , the traveltime residual of the path is calculated,

[0122]

[0123] where is the observed traveltime of the path, is the theoretical traveltime obtained by forward calculation under the current velocity model . Subsequently, the path residual is projected to the spatial points it passes through in a path length weighted manner, and the normalized traveltime residual of each point is defined as as follows,

[0124]

[0125] where is the total length of the path , is the traversal length of the path at position , and is the indicator function. According to this residual estimate, the initial standard deviation of the distribution at the spatial point is constructed as ,

[0126]

[0127] where is the average traversal length of all paths at position , and is the initial velocity estimate obtained by traveltime inversion. The finally constructed distribution is initialized as

[0128]

[0129] S3, design region-aware adaptive weights . combined with masks and standard deviations , set the differentiated weighting strategy. For background media ( ), use the Sigmoid function to smooth the weight; for high sound speed regions ( ), use the power function to enhance high uncertainty; for boundary regions ( ), use the weighted mixed harmonic control.

[0130] S4, acoustic velocity sample sampling and forward calculation. Input particles to the two-dimensional acoustic wave equation forward simulation module, use the center frequency of 0.5 MHz Ricker wavelet as the excitation source, calculate and obtain groups of simulation data .

[0131] S5, construct the loss function and judge convergence. In LSVGD, the entropy term of the variational distribution does not exist, assuming that the error between the simulation data and the experimental observation data is Gaussian noise, the prior distribution is Gaussian distribution, then

[0132]

[0133] where, is the Hadamard product, which means the multiplication of the elements at the corresponding positions in the vector. is the variance of the observation noise, set to . Assuming that the loss function value of the step is , calculate whether it meets , if it meets, terminate the iteration, output the result , and calculate the mean and standard deviation of . Use the mean to image biological tissues, measure uncertainty; if it does not meet, continue to execute step S6.

[0134] S6, adaptive weight gradient update

[0135]

[0136] where,

[0137]

[0138] and for Gaussian white noise obeying standard normal distribution, for iteration step length.

[0139] S7, calculate the mean and standard deviation of the updated n particles , and return to S3.

[0140] The application proposes an ultrasound probability full waveform inversion method based on regional uncertainty perception, which combines travel time tomography coarse organization structure perception, variational inference modeling and regional adaptive weight mechanism, aiming to improve the stability and precision of complex biological tissue imaging. First, using the low resolution wave velocity model obtained by travel time tomography inversion, a coarse organization mask is constructed, combined with various types of tissue characteristic wave velocity (such as fat, muscle, bone, etc.), the preliminary division of the region is realized, and the structure prior is provided for the subsequent inversion. Secondly, the sound velocity field is modeled by variational distribution , and the spatial dependent standard deviation is constructed according to the inversion residual, so as to estimate the regional uncertainty. Further, according to the physiological properties and signal characteristics of different tissue regions, the weight function of sigmoid type, power function type and its mixed form is designed, and the adaptive weight mechanism of regional perception is constructed, so that the high uncertainty region obtains greater gradient response in optimization, and the stable region is inhibited, and the local focusing ability of inversion is strengthened. Finally, the weight mechanism is introduced into the optimization framework to form the uncertainty perception weighted loss function, which effectively guides the inference process to concentrate on the key region, relieves the period jump, and enhances the noise resistance of the model. The method is suitable for SVI, ADVI, SVGD, LSVGD and other inference algorithms, and has good universality and practicality.

[0141] The technical scheme of the application mainly includes the following aspects:

[0142] 1. Coarse organization region mask construction mechanism: first, the target region is roughly inversed by travel time tomography to obtain a preliminary sound velocity field . Using the typical sound velocity range of biological tissues such as fat, muscle and bone, the region is divided into soft tissue, hard tissue and boundary region by setting threshold conditions, forming a mask vector perceived by the organization structure. The mask serves as the basis for subsequent adaptive weighting and gradient guidance.

[0143] 2. Variational posterior distribution initialization method: construct a spatially dependent variational posterior distribution , wherein the mean is initialized by the travel time inversion sound velocity, and the standard deviation is initialized by the travel time residual Mapping generation, measure inversion uncertainty. This initialization scheme can significantly improve the convergence and physical consistency of optimization.

[0144] 3. Region-aware uncertainty weighting function design: according to the mask Classification, design three types of weighting strategies: for soft tissue regions, use sigmoid function to regulate medium-high uncertainty; for hard tissue regions such as bones, use power function to emphasize high uncertainty area; for boundary transition area, fuse the two to construct hybrid weighting. Thus construct region-dependent weight distribution , used to regulate gradient contribution in optimization process.

[0145] 4. Combine particle sampling with forward simulation process: sample multiple sound speed models from the variational posterior distribution , solve the wave equation for each model to generate simulated data , compare with observed data , used for subsequent loss function calculation.

[0146] 5. Uncertainty-aware weighted variational objective function construction: use the above weighting function Introduced into the variational inference objective function, combined with the log-likelihood term and the regularization term, so that the inversion pays more attention to the high uncertainty area, suppresses overfitting and oscillation, and effectively alleviates the period jump problem.

[0147] 6. Adapt to various variational inference algorithm optimization update mechanism: the proposed method supports parameterized distribution (such as SVI, ADVI) and particle method (such as SVGD, LSVGD). In particular, under the LSVGD framework, by introducing uncertainty weight to each particle for gradient correction and random disturbance, guide the particle swarm to converge to the high confidence area, enhance the robustness and generalization performance of the inversion.

[0148] The present application starts from the physical prior of sound speed modeling, introduces the rough organization area division strategy based on travel time tomography, and combines the variational inference framework and uncertainty-aware mechanism, realizes an adaptive gradient optimization method for biological tissue structure difference. This method dynamically focuses on high uncertainty areas during inversion, effectively alleviates the period jump phenomenon in traditional ultrasound inversion, and improves the coherence and robustness of imaging. Compared with existing methods, the present application significantly reduces the computational cost while improving the inversion accuracy, and is suitable for sound speed reconstruction tasks under complex tissue structure, especially in medical ultrasound image reconstruction, tumor boundary detection and early lesion identification, etc. Clinical scenes have broad application prospects and practical value.

[0149] The preferred embodiments of the present application have been described above in detail. It should be understood that modifications and variations to the present application can be affected by those skilled in the art without departing from the scope of the application. Accordingly, it is intended that all of the subject matter of the above description and the claims be interpreted to encompass all such modifications and changes.

Claims

1. An adaptive weight gradient optimization method for bio-tissue ultrasound inversion, characterized in that, The method comprises the following steps: S1, obtaining a coarse organization region division vector based on travel time tomography , denotes the position of a point in the region; S2. Initialize the variational posterior distribution The initial variational posterior distribution Modeled as a Gaussian distribution with spatially dependent structure, where the variational distribution parameters , The mean of a Gaussian distribution is given. The standard deviation of the Gaussian distribution is used to obtain the preliminary sound velocity through travel-time tomography inversion. The variational posterior distribution used for initialization mean Based on the inversion residuals in travel-time imaging With local ray path length Obtain the initialized variational posterior distribution Standard deviation and the mean and standard deviation The variational posterior distribution is represented by the current mean and the current standard deviation, respectively. As the current variational posterior distribution This represents the speed of sound parameter; S3, dividing vector based on coarse texture region Current mean and current standard deviation calculation region dependent weights ; S4, sample a sample of the sound velocity model from the current variational posterior distribution, perform forward simulation on the sample of the sound velocity model using the acoustic wave equation to obtain simulated ultrasonic data corresponding to each sample sampled from the current variational posterior distribution ; S5, calculating a joint loss function based on the current variational posterior distribution, the sample, the simulated ultrasonic data and the experimental observation data; S6, judging whether to converge based on the loss function, if yes, executing S7, otherwise, iteratively calculating based on the loss function to obtain a new mean value, a new standard deviation and a new variational posterior distribution, taking the new mean value, the new standard deviation and the new variational posterior distribution as the current mean value, the current standard deviation and the current variational posterior distribution respectively, and returning to S3; S7, taking the current mean value as the inversion imaging result and the current standard deviation as the inversion uncertainty of the inversion imaging result.

2. The adaptive weight gradient optimization method for bio-tissue ultrasonic inversion according to claim 1, characterized in that, Coarse organization area division vector is: ； wherein boundary represents a boundary region, represents a preliminary speed of sound, represents a fat speed of sound threshold, is a muscle speed of sound threshold, is a hard tissue speed of sound threshold.

3. The adaptive weight gradient optimization method for bio-tissue ultrasonic inversion according to claim 1, characterized in that, Inversion residuals in traveltime imaging With local ray path lengths Obtaining an initialized variational posterior distribution Standard deviation of The specific steps are as follows: For each source-receiver pair path , a traveltime residual for the path is computed, a normalized traveltime residual for each point is computed based on the traveltime residual for the path, an initialized variational posterior distribution is obtained based on the normalized traveltime residual for each point .

4. The adaptive weight gradient optimization method for bio-tissue ultrasonic inversion according to claim 3, characterized in that, The travel time residual of the path is: ； wherein, is the observed travel time for the path, is the theoretical travel time obtained by forward calculation under the initial sound speed, is the travel time residual of the path .

5. The adaptive weight gradient optimization method for bio-tissue ultrasonic inversion according to claim 4, characterized in that, The normalized travel time residual is: ； wherein is the total length of the path , is the traversal length of the path at position , is the indicator function.

6. The adaptive weight gradient optimization method for bio-tissue ultrasonic inversion according to claim 5, characterized in that, initialized variational posterior distribution the standard deviation of is: 。 7. The adaptive weight gradient optimization method for bio-tissue ultrasonic inversion according to claim 1, characterized in that, Region-dependent weights are: ； wherein, represents a soft tissue weight, represents a hard tissue weight, represents a weight control parameter.

8. The adaptive weight gradient optimization method for bio-tissue ultrasonic inversion according to claim 7, characterized in that, The soft tissue weight is: ； wherein, is a hyperparameter to adjust the slope, denotes the current standard deviation, is the standard deviation of the uncertainty within the tissue region, denotes the average uncertainty within the tissue region.

9. The adaptive weight gradient optimization method for bio-tissue ultrasound inversion according to claim 8, wherein, The hard tissue weight is: ； wherein, denotes an exponential that controls the non-linear growth of the weights.

10. The adaptive weight gradient optimization method for bio-tissue ultrasound inversion according to claim 8, wherein, The loss function is: ； wherein, denotes the log-likelihood function, is the experimental observation data, is the loss function, denotes the regularization strength, denotes the entropy term of the variational distribution, is the likelihood model with Gaussian noise, the prior distribution is the regularization term; ； n denotes the total number of simulated ultrasound data sets, denotes the i-th simulated ultrasound data set, is the experimental observation data; ； wherein, is a Hadamard product, denotes the i-th sound velocity model sample; The specific steps of iteratively calculating based on the loss function to obtain a new mean value, a new standard deviation and a new variational posterior distribution are: calculating new variational distribution parameters according to the loss function , obtaining a new mean and a new standard deviation, substituting the new variational distribution parameters into the current variational posterior distribution, and obtaining a new variational posterior distribution.

Images