A three-dimensional ultrasonic brain imaging method based on a full convolution network

Abstract

The application provides a three-dimensional ultrasonic skull brain imaging method based on a full convolution network, and aims to solve the technical problem that the existing ultrasonic imaging method cannot meet the skull ultrahigh resolution and 3D real-time imaging. The steps are as follows: firstly, the time domain ultrasonic signal is calculated according to the physical characteristics of biological tissues and skull brain ultrasonic signal propagation, and a numerical simulation database is constructed; and the ultrasonic signal data in the numerical simulation database is subjected to data preprocessing; secondly, a full convolution network is built, and the ultrasonic signal data after data preprocessing is used for network training to obtain a network model with an optimal network hyperparameter combination; finally, the skull brain ultrasonic signal to be predicted is input into the network model for ultrasonic skull brain imaging. The ultrasonic skull brain imaging technology is easy to implement, has high resolution and high speed.

Description

Technical Field

[0001] This invention relates to the field of cranial imaging technology, and in particular to a three-dimensional ultrasound cranial imaging method based on a fully convolutional network. Background Technology

[0002] Cranial imaging plays an irreplaceable role in clinical medicine, aiding in the early and comprehensive diagnosis and assessment of intracranial lesions, as well as the early determination of treatment plans, and has a significant impact on the world's population. Traditional imaging techniques, such as magnetic resonance imaging (MRI) and computed tomography (CT), are widely used in clinical imaging due to their high soft tissue resolution. However, MRI is not suitable for obese individuals or those with obesity. CT involves exposing the body to harmful ionizing radiation, especially in the presence of magnetic foreign bodies. Traditional B-mode ultrasound and ultrasound reflection tomography can be used for imaging the heart, abdomen, urinary system, and digestive system, but the mechanism of ultrasound interaction with the skull and soft tissues is complex. Ultrasound signals passing through the skull produce severe reflections, refractions, and scattering, resulting in a complex and distorted intracranial wave field. Strong, high-amplitude reflections from the skull can overwhelm the tiny pulses reflected from soft tissues, making high-resolution cranial imaging impossible. Furthermore, all these techniques require large and expensive equipment and must be operated by medical professionals.

[0003] One feasible method to address wavefield phase and wavefront distortions caused by the skull is to excite and receive transcranial ultrasound through the open fontanelle of the skull. However, this method is only suitable for infants, as the open fontanelle gradually closes with age. Another approach is to capture signals from intracranial tissues by piercing the skull with a photoacoustic probe to monitor and image various signals in the brain. However, this invasive method is difficult to perform in clinical applications, and even slight errors could lead to irreversible consequences. Another imaging technique is full waveform inversion, which monitors the wavefield around the brain, solves nonlinear local parameter optimization problems, and iteratively simulates model parameters to reduce the difference between the experimental unknown model wavefield and the simulated known model wavefield, thus enabling prediction of the physical model structure. However, the iterative solution process of model gradients and Hessian matrices and vector products in full waveform inversion consumes significant computational resources and time, lacking the real-time advantages of conventional ultrasound imaging, and therefore cannot be applied to clinical real-time medical imaging. Currently, no ultrasound imaging method can simultaneously achieve ultra-high resolution of the skull and real-time 3D imaging, which is a challenge facing cranial imaging. Summary of the Invention

[0004] To address the shortcomings of the aforementioned background technology, this invention proposes a three-dimensional ultrasound cranial imaging method based on a fully convolutional network, which solves the technical problem that existing ultrasound imaging methods cannot meet the requirements for ultra-high resolution and real-time 3D imaging of the skull.

[0005] The technical solution of this invention is implemented as follows:

[0006] A three-dimensional ultrasound cranial imaging method based on fully convolutional networks, the steps of which are as follows:

[0007] Step 1: Calculate the time-domain ultrasound signal based on the physical characteristics of ultrasound signal propagation in biological tissues and the brain, and construct a numerical simulation database;

[0008] Step 2: Perform data preprocessing on the ultrasonic signal data in the numerical simulation database;

[0009] Step 3: Build a fully convolutional network and train the network using the preprocessed ultrasound signal data to obtain a network model with the optimal combination of hyperparameters;

[0010] Step 4: Input the predicted cranial ultrasound signal into the network model for cranial ultrasound imaging.

[0011] The implementation method for step one is as follows:

[0012] Considering the physical characteristics of ultrasound signal propagation in biological tissues and the brain, the propagation of ultrasound in the human brain is described by the 3D ultrasound motion equation of an isotropic medium, as shown in equation (1):

[0013]

[0014] Where p(r,t) is the pressure wave field at point source r at time t; ρ(r) is the density of point source r, and c(r) is the velocity of point source r; transforming equation (1) to the spatial frequency domain, we obtain the acoustic wave equation:

[0015]

[0016] Where, k r =2πf / c r The background wavenumber at point source r is represented by f, where f is the frequency and c is the frequency. r Let r be the background velocity at the point source, ψ be the pressure field in the spatial frequency domain, and O(r) be the mathematical expression for the scattering body, defined as:

[0017]

[0018] Among them, c u This represents the velocity of the background without scattering bodies. In ultrasound signal acquisition, the skull and soft tissue are used as the background model, and blood clots are used as the interference term in the forward modeling of ultrasound signals. The background ultrasound signal and the disturbed ultrasound signal are defined as follows:

[0019]

[0020]

[0021] Where δ is the Dirac function; Let i be an intermediate variable, i be the imaginary unit, and ψ be an intermediate variable. r Indicates the background wave field;

[0022] A numerical simulation database is established using the forward modeling algorithm based on equation (2-5).

[0023] The data preprocessing is a signal processing algorithm used to extract the preferred frequency signal.

[0024] The implementation method for step three is as follows:

[0025] Feature extraction is performed on the preferred frequency signal using a fully convolutional network. The calculation process is as follows:

[0026]

[0027] Where F(mn) represents the feature, R represents the activation function, and w u,v The weights of the feature extractor in the v-th row and u-th column are represented by S, where S is the input layer and B is the bias; m and n represent the m-th row and n-th column of the feature, respectively; and L represents the number of times the feature extractor is executed.

[0028] Perform dimensionality reduction on the features:

[0029]

[0030] Where n' represents the feature number, c represents the channel number, r' is the row number, h is the column number, kr∈[1,k] is the length of the dimensionality reduction window, and kw∈[1,k] is the width of the dimensionality reduction window; P(·) represents the dimensionality reduction result, X represents the feature, rs is the starting row number of the dimensionality reduction window, and ws is the starting column number of the dimensionality reduction window;

[0031] Define the objective function, and optimize the parameters and structure using an iterative method to minimize the error:

[0032]

[0033] in, A two-dimensional cross-sectional planar velocity diagram representing a three-dimensional model. The two-dimensional cross-sectional planar velocity map represents the predicted three-dimensional model, where n” is the number of two-dimensional cross-sectional planar velocity maps, and [·] denotes a matrix.

[0034] Compared with the prior art, the beneficial effects of the present invention are: the ultrasound cranial imaging technology of the present invention is easy to implement, has high resolution, and is fast. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the present invention.

[0037] Figure 2 This is a diagram of the fully convolutional network structure of the present invention.

[0038] Figure 3 This is a comparison of the training and validation losses for the fully convolutional network of this invention.

[0039] Figure 4 The imaging results are based on the known prior model of this invention.

[0040] Figure 5 This is a statistical analysis of the results of the real brain and the brain predicted by the fully convolutional network in the full dataset of this invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] like Figure 1 As shown, this embodiment of the invention provides a three-dimensional ultrasound cranial imaging method based on a fully convolutional network. The method establishes a database with a one-to-one correspondence between cranial ultrasound and three-dimensional cranial sound velocity models; preprocesses the data from both the cranial ultrasound and the three-dimensional cranial sound velocity models; constructs a fully convolutional neural network and uses the preprocessed cranial ultrasound signal and the three-dimensional cranial sound velocity model as the input and output of the fully convolutional neural network, respectively; trains the fully convolutional neural network to obtain a network model with the optimal combination of hyperparameters; this network model can then be used for cranial imaging. The specific steps are as follows:

[0043] Step 1: Calculate the time-domain ultrasound signal based on the physical characteristics of ultrasound signal propagation in biological tissues and the brain, and construct a numerical simulation database;

[0044] Considering the physical characteristics of ultrasound signal propagation in biological tissues and the brain, the propagation of ultrasound in the human brain is described by the 3D ultrasound motion equation of an isotropic medium, as shown in equation (1):

[0045]

[0046] Where p(r,t) is the pressure wave field at point source r at time t; ρ(r) is the density of point source r, and c(r) is the velocity of point source r; transforming equation (1) to the spatial frequency domain, we obtain the acoustic wave equation:

[0047]

[0048] Where, k r =2πf / c r The background wavenumber at point source r is represented by f, where f is the frequency and c is the frequency. r Let r be the background velocity at the point source, ψ be the pressure field in the spatial frequency domain, and O(r) be the mathematical expression for the scattering body, defined as:

[0049]

[0050] Among them, c u This represents the velocity of the background without scattering bodies. In ultrasound signal acquisition, the skull and soft tissue are used as the background model, and blood clots are used as the interference term in the forward modeling of ultrasound signals. The background ultrasound signal and the disturbed ultrasound signal are defined as follows:

[0051]

[0052]

[0053] Where δ is the Dirac function; Let i be an intermediate variable, i be the imaginary unit, and ψ be an intermediate variable. r Indicates the background wave field;

[0054] A numerical simulation database is established using the forward modeling algorithm according to equation (2-5), with the finite difference method being the preferred forward modeling algorithm.

[0055] Step 2: Perform data preprocessing on the ultrasonic signal data in the numerical simulation database; the data preprocessing is a signal processing algorithm used to extract the preferred frequency signal. The preferred signal processing algorithm is Fourier transform and normalization operation, and the normalized frequency domain amplitude corresponds to the preferred frequency signal.

[0056] Step 3: Build a fully convolutional network and train the network using the preprocessed ultrasound signal data to obtain a network model with the optimal combination of hyperparameters;

[0057] like Figure 2 As shown, the fully convolutional network structure in this invention includes an input layer, a convolutional layer, a pooling layer, and an output layer. The cranial ultrasound signal is input into the input layer of the fully convolutional network. The convolutional layer then extracts features from the cranial ultrasound signal, and the pooling layer downsamples and reduces the dimensionality of the feature map, simplifying the network complexity, improving the generalization ability of the network structure, reducing computational load, and saving computational resources.

[0058] The training method is as follows:

[0059] The preferred frequency signal is input into the input layer of the fully convolutional network. The convolutional layer is connected after the input layer and is used to extract features from the preferred frequency signal. The calculation process is as follows:

[0060]

[0061] Where F(mn) represents the feature, R represents the activation function, and w u,v The weights of the feature extractor in the v-th row and u-th column are represented by S, where S is the input layer and B is the bias; m and n represent the m-th row and n-th column of the feature, respectively; and L represents the number of times the feature extractor is executed.

[0062] Perform dimensionality reduction on the features:

[0063]

[0064] Where n' represents the feature number, c represents the channel number, r' is the row number, h is the column number, kr∈[1,k] is the length of the dimensionality reduction window, and kw∈[1,k] is the width of the dimensionality reduction window; P(·) represents the dimensionality reduction result, X represents the feature, rs is the starting row number of the dimensionality reduction window, and ws is the starting column number of the dimensionality reduction window;

[0065] Define the objective function, and optimize the parameters and structure using an iterative method to minimize the error:

[0066]

[0067] in, A two-dimensional cross-sectional planar velocity diagram representing a three-dimensional model. The two-dimensional cross-sectional planar velocity map represents the predicted three-dimensional model, where n” is the number of two-dimensional cross-sectional planar velocity maps, and [·] denotes a matrix.

[0068] Step 4: Input the predicted cranial ultrasound signal into the network model for cranial ultrasound imaging.

[0069] Figure 3 A comparison of the training and validation losses of the fully convolutional network in this invention is presented. The optimal network model is stored at iteration 2760. The root mean square error (RMSE) of the network model decreases during training and validation, showing a rapid decrease at the beginning of training and a stable decrease in the middle and end of training. Compared with the validation set, the RMSE of the training set shows a more stable trend and smaller numerical fluctuations. The RMSEs of the optimal model on the training and validation sets are 2.398 × 10⁻⁶. -4 and 2.429×10 -4 It meets the resolution requirements of 3D imaging in clinical medicine.

[0070] Figure 4 The imaging results of the known prior model in this invention are presented, and the 3D cranial model is displayed through 2D slices. Figure 4 (a) and (b) show a real 2D slice of the brain and a 2D slice of the brain predicted by a fully convolutional network, respectively. Figure 4 It is evident that fully convolutional networks can effectively distinguish blood clots from brain tissue. Figure 4 (e)-(f) are Figure 4 (c)-(d) Imaging results of cross-sections at the vertical and horizontal lines of the blood clot. It can be seen that the boundary between the blood clot and the tissue is clear, and the blood clot can be predicted. The overlap between the real brain and the brain predicted by the fully convolutional network is 97.88%.

[0071] Figure 5 Statistical results are presented on the comparison between real brain images on the full dataset and brain images predicted by the fully convolutional network. The average overlap between real brain images on the full dataset and brain images predicted by the fully convolutional network is greater than 92%, demonstrating the robustness and stability of the network structure.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A three-dimensional ultrasound cranial imaging method based on fully convolutional networks, characterized in that, The steps are as follows: Step 1: Calculate the time-domain ultrasound signal based on the physical characteristics of ultrasound signal propagation in biological tissues and the brain, and construct a numerical simulation database; Considering the physical characteristics of ultrasound signal propagation in biological tissues and the brain, the propagation of ultrasound in the human brain is described by the 3D ultrasound motion equation of an isotropic medium, as shown in equation (1): Where p(r,t) is the pressure wave field at point source r at time t; ρ(r) is the density of point source r; and c(r) is the velocity of point source r. Transforming equation (1) to the spatial frequency domain, we obtain the acoustic wave equation: Where, k r =2πf / c r The background wavenumber at point source r is represented by f, where f is the frequency and c is the frequency. r Let r be the background velocity at the point source, ψ be the pressure field in the spatial frequency domain, and O(r) be the mathematical expression for the scattering body, defined as: Among them, c u This represents the velocity of the background without scattering bodies. In ultrasound signal acquisition, the skull and soft tissue are used as the background model, and blood clots are used as the interference term in the forward modeling of ultrasound signals. The background ultrasound signal and the disturbed ultrasound signal are defined as follows: Where δ is the Dirac function; Let i be an intermediate variable, i be the imaginary unit, and ψ be an intermediate variable. r Indicates the background wave field; A numerical simulation database is established using forward modeling algorithms based on equations (2) to (5); Step 2: Perform data preprocessing on the ultrasonic signal data in the numerical simulation database; Step 3: Build a fully convolutional network and train the network using the preprocessed ultrasound signal data to obtain a network model with the optimal combination of hyperparameters; Step 4: Input the predicted cranial ultrasound signal into the network model for cranial ultrasound imaging.

2. The three-dimensional ultrasound cranial imaging method based on fully convolutional networks according to claim 1, characterized in that, The data preprocessing is a signal processing algorithm used to extract the preferred frequency signal.

3. The three-dimensional ultrasound cranial imaging method based on fully convolutional networks according to claim 2, characterized in that, The implementation method for step three is as follows: Feature extraction is performed on the preferred frequency signal using a fully convolutional network. The calculation process is as follows: Where F(mn) represents the feature, R represents the activation function, and w u,v The weights of the feature extractor in the v-th row and u-th column are represented by S, where S is the input layer and B is the bias; m and n represent the m-th row and n-th column of the feature, respectively; and L represents the number of times the feature extractor is executed. Perform dimensionality reduction on the features: Where n' represents the feature number, c represents the channel number, r' is the row number, h is the column number, kr∈[1,k] is the length of the dimensionality reduction window, and kw∈[1,k] is the width of the dimensionality reduction window; P(·) represents the dimensionality reduction result, X represents the feature, rs is the starting row number of the dimensionality reduction window, and ws is the starting column number of the dimensionality reduction window; Define the objective function, and optimize the parameters and structure using an iterative method to minimize the error: in, A two-dimensional cross-sectional planar velocity diagram representing a three-dimensional model. The two-dimensional cross-sectional planar velocity map represents the predicted three-dimensional model, where n” is the number of two-dimensional cross-sectional planar velocity maps, and [·] denotes a matrix.

Images