An image-based intelligent auxiliary analysis and optimization method and device for pathological vocal cords

Abstract

The present invention provides an image-based intelligent assisted analysis and optimization method and device for pathological vocal cords. The method comprises: step 1, extracting input features; step 2, obtaining physical quantities as actual output data; step 3, partitioning the dataset into training, validation, and test sets; step 4, constructing and training a neural network based on the training and validation sets; and step 5, evaluating the neural network's predictive performance using the test set. The method can quickly and accurately predict the flow and pressure distribution corresponding to various pathological vocal cord conditions, and can be applied clinically for real-time, noninvasive, assisted diagnosis and treatment of pathological vocal cord conditions.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary field of computational fluid dynamics, artificial intelligence and biomedical engineering, and in particular relates to an image-based intelligent auxiliary analysis and optimization method and device for pathological vocal cords. Background Art

[0002] Speech is the most basic, effective, and important means of human communication. The vocal cords are the most crucial organ of articulation in speech production, yet they are also the most vulnerable to damage and pathology. At least 20% of the world's population suffers from varying degrees of voice disorders. Research on the vocalization process can improve protection and welfare for those suffering from voice disorders, and therefore has significant theoretical research significance and social application value.

[0003] The vocalization process of the human vocal cords involves a complex fluid-structure interaction (FSI) phenomenon. The interaction between airflow from the lungs and the soft tissue of the vocal cords triggers continuous vibration of the vocal cords, thereby regulating the amount of airflow through the glottis. Voice disorders are often associated with vocal cord pathology, such as nodules, cysts, scars, and paralysis. These pathologies alter the geometry and material properties of the vocal cord tissue, resulting in irregular mucosal waves. Currently, clinical examination of pathological vocal cords relies primarily on electronic laryngoscopes, an invasive procedure. Although local anesthesia is administered before the examination, patients with sensitive laryngopharyngeal reflexes may experience discomfort. Furthermore, currently commonly used electronic / stroboscopic laryngoscopes only provide two-dimensional assessment, lack a stereoscopic perspective, are relatively subjective, and are only suitable for periodic vocal cord vibrations. Laryngeal CT scans enable three-dimensional dynamic assessment of the vocal cords, detecting submucosal lesions that are invisible to laryngoscopes. Furthermore, this is a non-invasive procedure that does not cause discomfort to the patient. Furthermore, numerical FSI models for simulating vocal cord vibration have played a crucial role in understanding, diagnosing, and treating voice disorders, and numerous FSI models of varying complexity have been developed. However, the vocal cord vibration FSI models currently developed domestically and internationally either fail to guarantee the accuracy of the calculations due to model simplification or are prohibitively expensive to implement clinically. Clinical applications primarily refer to surgical management based on numerical simulations, such as predicting the outcomes of surgical interventions, optimizing surgical adjustment plans, and providing personalized treatment plans tailored to the specific patient's anatomy and needs.

[0004] In recent years, thanks to the rapid development of artificial intelligence and machine learning technologies, data-driven flow prediction models have gradually been developed. This type of model can significantly reduce calculation time while ensuring calculation accuracy, so it can be used to replace time-consuming first-principles-based flow models, greatly broadening the application scenarios. For example, the coronary CT flow reserve fraction (CT-FFR) algorithm based on CFD technology has significant potential in assessing myocardial ischemia, guiding treatment strategies, and providing prognostic information. However, since the CFD algorithm is based on the Navier-Stokes equation, it requires the use of a supercomputer for discrete solution, which is time-consuming and expensive, limiting its widespread clinical application. To this end, researchers have proposed a deep learning-based model that can non-invasively predict coronary artery FFR from CT scan images in real time. Compared with invasive FFR, the model shows higher diagnostic accuracy.

[0005] In clinical practice, non-invasive auxiliary diagnosis and treatment of pathological vocal cords can also be achieved by establishing corresponding data-driven models. However, there are still few studies in this area at home and abroad. The present invention constructs a three-dimensional convolutional neural network (3D-CNN) model based on the tomographic images extracted from the synthetic pathological vocal cord model library to quickly and accurately predict the flow and pressure distribution corresponding to various pathological vocal cords, and realize real-time non-invasive auxiliary diagnosis and treatment, which has certain theoretical research significance and practical application value. Summary of the Invention

[0006] Purpose of the invention: The purpose of the present invention is to overcome the shortcomings of low efficiency of traditional CFD methods in the non-invasive auxiliary diagnosis and treatment of pathological vocal cords. Based on the tomographic images extracted from the synthetic pathological vocal cord model library, a three-dimensional convolutional neural network (3D-CNN) model is proposed to establish a mapping relationship between the vocal cord tomographic images and the corresponding flow and pressure distributions. The actual flow and pressure distribution of the pathological vocal cords are obtained by a high-fidelity sharp interface immersed boundary solver. Using the trained model, the flow and pressure distribution corresponding to various pathological vocal cords can be predicted quickly and accurately, which is expected to be applied to the real-time non-invasive auxiliary diagnosis and treatment of pathological vocal cords in clinical practice.

[0007] The present invention specifically provides an image-based intelligent auxiliary analysis and optimization method for pathological vocal cords, comprising:

[0008] First, based on the modal equation of vocal cord vibration, a three-dimensional model library consisting of several synthetic pathological vocal cord shapes is generated. Various three-dimensional pathological vocal cords in the model library are sliced ​​two-dimensionally along the direction of mucosal wave propagation to obtain a series of grayscale tomographic images. Then, a high-fidelity sharp interface immersed boundary solver suitable for processing complex shapes is used to obtain the flow and pressure distribution of various pathological vocal cords. Then, a mapping relationship between the vocal cord tomographic images and the corresponding flow and pressure distributions is established through a three-dimensional convolutional neural network (3D-CNN) model. The neural network is trained and the hyperparameters are adjusted and optimized using the gradient descent algorithm to minimize the loss function. Finally, for any tomographic image of a pathological vocal cord, the trained model is used to quickly and accurately predict and reconstruct the corresponding flow and pressure distribution, and the prediction performance of the model is evaluated. The specific steps are as follows:

[0009] Step 1: Generate a three-dimensional geometric shape of the pathological vocal cords according to the vibration modal equation of the vocal cords to obtain a three-dimensional vocal cord model library;

[0010] Step 2, performing two-dimensional slicing on the vocal cord model in the vocal cord three-dimensional model library to obtain a three-dimensional input matrix;

[0011] In step 3, a high-fidelity immersed boundary solver is used to obtain real physical data. The solver is based on finite difference discretization, can capture sharp interfaces, and uses an efficient MPI (Message Passing Interface) parallel strategy. Its accuracy and reliability have been verified by a large number of numerical examples. Therefore, the physical quantities obtained by the solution can be used as real output data.

[0012] Step 4: Use a random segmentation algorithm to divide the data in the vocal cord three-dimensional model library to obtain a training set (128), a validation set (32), and a test set (20); the data in the three-dimensional model library consists of input and output matrices corresponding to 180 synthetic vocal cords of different shapes. Use a random data segmentation algorithm to use 20 groups of data as test sets to evaluate the prediction performance of the trained three-dimensional convolutional neural network. The remaining 160 groups of data are divided into training sets and validation sets at a ratio of 80% and 20% to participate in the training of the three-dimensional convolutional neural network.

[0013] Step 5: Based on the training set and the validation set, a three-dimensional convolutional neural network (CNN) is constructed and trained, and the weights of the optimal model obtained through training are saved.

[0014] Step 6: Use the test set to evaluate the prediction performance of the 3D convolutional neural network.

[0015] Step 1 includes: the mode shape g of the vocal cords can be regarded as the superposition of the initial shape ξ0 of the vocal cords and the shape ξ obtained by the vibration mode. Its expression is shown in formula (1), and the expressions of ξ0 and ξ are shown in formulas (2) and (3), respectively:

[0016] g(y,z,t)=ξ0(y,z)+ξ(y,z,t) (1)

[0017] ξ0(y,z)=(1-z / L)[ξ 02 +(ξ 01 -ξ 02 -4ξ B y / T)(1-y / T)] (2)

[0018] ξ(y,z,t)=ξ m sin(mπz / L)[sinωt-n(ω / c)(yy m )cosωt] (3)

[0019] where g(y,z,t) represents the mode function of the vocal cords, y represents the inferior-superior direction, and z represents the anterior-posterior direction of the vocal cord vibration, respectively. Up and down, and anterior and posterior represent the endpoints of the three-dimensional geometry of the vocal cords on the y-axis and the z-axis, respectively. Up represents y_max, down represents y_min, anterior represents z_max, and posterior represents z_min. t represents time. 01 ,ξ 02 ,ξ B Corresponding to the glottal inlet half-width, glottal outlet half-width and surface convexity coefficient, T and L represent the thickness and length of the glottis, respectively. m represents the modal displacement amplitude, y m represents the inflection point of the vertical half-wavelength, ω represents the angular frequency, c represents the propagation velocity of the mucosal wave, and m and n represent the number of modes vibrating along the z and y directions, respectively;

[0020] By changing the parameters in formulas (1), (2), and (3), including ξ 01 ,ξ 02 ,ξ B 、y m , ω, etc., to synthesize vocal cords of various shapes and form a three-dimensional vocal cord model library.

[0021] Step 2 includes: performing two-dimensional slicing of various three-dimensional pathological vocal cords in the vocal cord three-dimensional model library along the direction of mucosal wave propagation (the axial plane of the vocal cords), extracting a series of grayscale tomographic images, converting each grayscale tomographic image slice into a binary matrix, and stacking them along the direction of mucosal wave propagation (the axial plane of the vocal cords) to obtain a three-dimensional input matrix X that can fully reflect the geometric information of the vocal cords; to make the neural network easier to train, the three-dimensional input matrix X is normalized.

[0022] Step 3 includes establishing the governing equations for glottal flow as follows:

[0023]

[0024] where u i ,ρ,p,ν respectively represent the velocity, density, pressure and kinematic viscosity coefficient of the fluid, represents the partial derivative; the control equation is discretized and solved using the sharp interface immersed boundary method suitable for the complex vocal cord shape. After grid independence verification, an appropriate grid scale (the minimum grid scale is 0.01 cm) is selected for calculation to obtain real physical data, which are stored in the output matrix y. The real physical data include the flow rate, average pressure distribution, and glottal surface pressure field of each vocal cord in the vocal cord three-dimensional model library.

[0025] Step 5 includes: establishing a mapping relationship between the input three-dimensional matrix X and the output matrix y by constructing a three-dimensional convolutional neural network CNN to complete the offline training process. The mathematical expression of the mapping relationship is:

[0026] y=f(X;θ) (5)

[0027] Where θ represents all learnable convolution kernels and weights, and f represents a three-dimensional convolutional neural network;

[0028] Construct a three-dimensional convolutional neural network for predicting flow and average pressure distribution and a three-dimensional convolutional neural network for reconstructing the pressure field of the left and right vocal cords;

[0029] The 3D convolutional neural networks used to predict flow and average pressure distribution and to reconstruct the left and right vocal cord pressure fields both use four layers of 3D convolutional layers (Conv3D). Each Conv3D layer is followed by a 3D maximum pooling layer (MaxPool3D) to further reduce the dimensionality of the feature information. The stride of each MaxPool3D layer is 2. The first three 3D convolutional layers (Conv3D) use non-zero padding (valid padding), and the last 3D convolutional layer (Conv3D) uses zero padding (same padding). After the 3D convolutional layer (Conv3D), a 3D global average pooling layer (GlobalAveragePooling3D) is used instead of the traditional fully connected layer to reduce overfitting that may occur during training.

[0030] The 3D convolutional neural network used to reconstruct the pressure field of the left and right vocal cords has a fully connected layer FC set between the GlobalAveragePooling3D layer and the output layer;

[0031] In the 3D convolutional neural network used to predict flow and average pressure distribution and the 3D convolutional neural network used to reconstruct the left and right vocal cord pressure fields, ReLU (Rectified Linear Unit) is used as the activation function between each convolution layer to introduce nonlinearity:

[0032] σ(x)=ReLU(x)=max(x,0) (6)

[0033] Where x represents the input feature, σ represents the activation function, and max represents the maximum value;

[0034] Using forecast data The mean squared sum of the errors between the real data y and the real data y is used as the loss function J.

[0035]

[0036] Where N is the number of data points, λ is the regularization coefficient to avoid overfitting of the model, and w represents the convolution kernel and weight;

[0037] The Adam stochastic gradient descent optimization algorithm is used to adjust the weights to minimize the loss function, and the weights of the optimal model obtained through training are saved.

[0038] Step 6 includes: using the trained three-dimensional convolutional neural network to online predict the flow and pressure distribution of the vocal cords in the test set, and comparing the predicted results with the real physical data to evaluate the prediction performance.

[0039] The present invention also provides an image-based intelligent auxiliary analysis and optimization device for pathological vocal cords, comprising:

[0040] A feature extraction module is used to generate a three-dimensional geometric shape of the pathological vocal cords according to the vibration modal equation of the vocal cords, and obtain a three-dimensional vocal cord model library;

[0041] The physical quantity calculation module is used to perform two-dimensional slicing on the vocal cord model in the three-dimensional vocal cord model library to obtain a three-dimensional input matrix and use a high-fidelity immersed boundary solver to obtain real physical data;

[0042] The data set partitioning module is used to partition the data in the vocal cord 3D model library using a random segmentation algorithm to obtain a training set, a validation set, and a test set;

[0043] The neural network construction and training module is used to build and train a three-dimensional convolutional neural network based on the training set and the validation set, and save the weights of the optimal model obtained through training;

[0044] The performance evaluation module is used to evaluate the prediction performance of the 3D convolutional neural network using a test set.

[0045] The present invention also provides a storage medium storing a computer program or instruction. When the computer program or instruction is executed, the image-based intelligent auxiliary analysis and optimization method for pathological vocal cords is implemented.

[0046] Beneficial effects: The method of the present invention establishes a three-dimensional convolutional neural network (3D-CNN) model, which can directly infer the corresponding flow and pressure distribution from the synthesized pathological vocal cord tomographic images. The method has the following advantages: (1) For the first time, the three-dimensional convolutional neural network (3D-CNN) is applied to the prediction of the physical quantities of pathological vocal cords, overcoming the shortcomings of the traditional CFD method in the non-invasive auxiliary diagnosis of pathological vocal cords, and achieving satisfactory results; (2) The input features use a set of tomographic images that can fully reflect the geometric information of the vocal cords, avoiding the problems of omissions or partial generalizations caused by relying on prior results or experience to extract features, and can be directly applied to clinical intelligent auxiliary diagnosis based on laryngeal CT scan images, with strong portability; (3) Unlike most current vocal cord prediction models based on fully connected neural networks, this model can not only predict the average pressure distribution on the vocal cord surface, but also reconstruct the pressure field on the left and right vocal cord surfaces in real time, thereby providing more comprehensive flow field information to achieve better auxiliary diagnosis effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more apparent.

[0048] Figure 1 It is the overall flow chart of the present invention.

[0049] Figure 2Schematic diagram of the input to the pathological glottal flow model.

[0050] Figure 3 This is a diagram showing the computational domain settings.

[0051] Figure 4 It is a structural diagram of the pathological glottal flow model.

[0052] Figure 5 It is a scatter plot of the actual and predicted values ​​of flow rate and mean pressure distribution.

[0053] Figure 6 This is a schematic diagram of the predicted results of the average pressure distribution on the vocal cord surface.

[0054] Figure 7 It is a schematic diagram of the pressure prediction results corresponding to the grid points on the left and right surfaces of the vocal cords.

[0055] Figure 8 This is a schematic diagram comparing the pressure fields predicted and reconstructed on the left and right surfaces of the vocal cords with the actual pressure fields.

[0056] Figure 9 This is a diagram showing the statistical results of the relative MSE errors of the left and right vocal cords. DETAILED DESCRIPTION

[0057] like Figure 1 As shown, the present invention specifically provides an image-based intelligent auxiliary analysis and optimization method for pathological vocal cords, comprising:

[0058] Step 1: The mode shape g of the vocal cords is considered as the superposition of the initial shape ξ0 of the vocal cords and the shape ξ obtained by the vibration mode. The expression is shown in Equation (1), where y and z represent the up-down direction and the front-back direction of the vocal cord vibration, respectively, and t represents time. The expressions of ξ0 and ξ are shown in Equations (2) and (3), respectively, where ξ 01 ,ξ 02 ,ξ B They correspond to the glottal inlet half-width, glottal outlet half-width and surface convexity coefficient respectively, T and L represent the thickness and length of the glottis respectively, ξ m represents the modal displacement amplitude, y m represents the inflection point of the vertical half-wavelength, ω represents the angular frequency, c represents the propagation velocity of the mucosal wave, and m and n represent the number of modes vibrating along the z and y directions, respectively.

[0059] g(y,z,t)=ξ0(y,z)+ξ(y,z,t) (1)

[0060] ξ0(y,z)=(1-z / L)[ξ 02 +(ξ 01 -ξ 02 -4ξ B y / T)(1-y / T)] (2)

[0061] ξ(y,z,t)=ξ m sin(mπz / L)[sinωt-n(ω / c)(yy m )cosωt] (3)

[0062] By changing the parameters in the above expression, vocal cords of various shapes can be synthesized. In the present invention, a total of 180 vocal cord shapes are synthesized to form a three-dimensional vocal cord model library.

[0063] Step 2: For various vocal cord shapes in the model library, the three-dimensional vocal cord model is evenly cut into 32 sections along the mucosal wave propagation direction (y), and each section forms a binary image, such as Figure 2 shown.

[0064] Step 3: Convert each cross-sectional binary image generated in Step 2 into an 81×81 matrix and perform appropriate preprocessing, including normalization. Stack these 32 identical matrices to form a 81×81×32 3D matrix, which, as the input matrix, fully represents the 3D vocal cord geometry.

[0065] Step 4: The governing equation for glottal flow is the incompressible Navier-Stokes equation, which is expressed as follows:

[0066]

[0067] where u i ,ρ,p,ν represent the velocity, density, pressure and kinematic viscosity coefficient of the fluid. The glottal inlet pressure is 1kPa. In order to obtain the real flow and pressure distribution of each vocal cord in the model library, it is necessary to discretely solve equation (4). The present invention adopts the sharp interface immersed boundary method suitable for processing complex boundaries to discretely solve the control equation. The calculation domain is set as follows Figure 3 As shown in the figure, the computational domain dimensions along the x, y, and z directions are 1.5 cm, 21 cm, and 1.5 cm, respectively. After verifying mesh independence, an appropriate mesh size is selected for calculation. Real physical data, such as flow rate, average pressure distribution, and glottal surface pressure field, are obtained for each vocal cord in the 3D vocal cord model library and stored in the output matrix.

[0068] Step 5: Use a random data splitting algorithm to divide the dataset consisting of 180 input and output matrices into training, validation, and test sets. Twenty of these data sets serve as test sets to evaluate the predictive performance of the trained model. The remaining 160 data sets are divided into training and validation sets at an 80% / 20% ratio for training and training the neural network model.

[0069] Step 6: Figure 4As shown in the figure, by constructing a three-dimensional convolutional neural network CNN to establish a mapping relationship between input and output, the offline training process is completed. The three-dimensional convolutional neural network CNN structure used to predict flow and average pressure distribution is shown in Table 1, and the three-dimensional convolutional neural network CNN structure used to reconstruct the left and right vocal cord pressure fields is shown in Table 2. As can be seen from the table, both three-dimensional convolutional neural network CNN structures use 4 layers of three-dimensional convolutional layers Conv3D. Each Conv3D layer is followed by a three-dimensional maximum pooling layer MaxPool3D to further reduce the dimension of the feature information and prevent overfitting. The step size of each MaxPool3D is 2. The first three three-dimensional convolutional layers Conv3D use non-zero padding (valid padding), and the last three-dimensional convolutional layer Conv3D uses zero padding (same padding). After the three-dimensional convolutional layer Conv3D, the three-dimensional global average pooling layer GlobalAveragePooling3D is used to reduce training parameters and reduce overfitting. Because the output dimension of the structure in Table 2 is much larger than that in Table 1, an additional fully connected layer (FC) is added between the GlobalAveragePooling3D layer and the output layer to achieve better training results. Reluctant Linear Unit (ReLU) activation functions are used between each layer to introduce nonlinearity and improve the model's expressiveness. The mean squared error (MSE) of the corresponding points between the predicted and real data is used as the loss function. The Adam stochastic gradient descent optimization algorithm is used to adjust the weights to minimize the loss function. The weights of the optimal model obtained through training are saved in a file.

[0070] As shown in Table 1, this is the 3D convolutional neural network (CNN) structure used to predict flow and average pressure distribution. Table 2 is the 3D convolutional neural network (CNN) structure used to reconstruct the pressure field of the left and right vocal cords:

[0071] Table 1

[0072]

[0073]

[0074] Table 2

[0075]

[0076]

[0077] Step 7: Use the trained neural network to predict the flow and pressure distribution of 20 groups of vocal cords in the test set online, establish an error estimation function, and evaluate the prediction performance of the model. The scatter distribution of the actual value and the predicted value of the flow and average pressure distribution is as follows Figure 5 The predicted results of the average pressure distribution corresponding to two randomly selected vocal cords in the test set are shown as follows. Figure 6The pressure prediction results of the left and right surface grid points corresponding to two randomly selected vocal cords in the test set are shown in Figure 7 The comparison results of the pressure field reconstructed by predicting the left and right surfaces of a randomly selected vocal cord in the test set and the real pressure field are shown in Figure 8 The relative MSE (mean square error) error statistics of the left and right vocal cords on the test set are shown as follows. Figure 9 As shown in Figure 2, the maximum relative MSE is approximately 6% and the average relative MSE is approximately 1%. These results indicate that the image-based intelligent auxiliary diagnosis method for pathological vocal cords proposed in this invention can very accurately predict the flow and pressure corresponding to pathological vocal cords, and is therefore expected to be promoted for clinical application.

[0078] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium is capable of storing a computer program that, when executed by the data processing unit, can execute the invention content of the image-based intelligent assisted analysis and optimization method for pathological vocal cords provided by the present invention and some or all of the steps in each embodiment. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0079] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, in essence or in other words, the part that contributes to the prior art, can be embodied in the form of a computer program, i.e., a software product. The computer program software product can be stored in a storage medium and includes several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, a MUU, or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0080] The present invention provides an image-based intelligent assisted analysis and optimization method and apparatus for pathological vocal cords. There are numerous methods and approaches for implementing this technical solution. The foregoing description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Any components not specified in this embodiment may be implemented using existing technologies.

Claims

1. A method and device for intelligent auxiliary analysis and optimization of pathological vocal cords based on images, characterized in that: The specific steps include: Step 1: Generate a three-dimensional geometric shape of the pathological vocal cords according to the vibration modal equation of the vocal cords to obtain a three-dimensional vocal cord model library; Step 2, performing two-dimensional slicing on the vocal cord model in the vocal cord three-dimensional model library to obtain a three-dimensional input matrix; Step 3: Use a high-fidelity immersed boundary solver to obtain real physical data; Step 4: Use a random segmentation algorithm to divide the data in the vocal cord 3D model library to obtain a training set, a validation set, and a test set; Step 5: Based on the training set and the validation set, a three-dimensional convolutional neural network is constructed and trained, and the weights of the optimal model obtained through training are saved; Step 6: Evaluate the prediction performance of the 3D convolutional neural network using the test set; Step 1 includes: the mode shape g of the vocal cords is regarded as the superposition of the initial shape ξ0 of the vocal cords and the shape ξ obtained by the vibration mode. The expression is shown in formula (1). The expressions of ξ0 and ξ are shown in formulas (2) and (3), respectively: g(y,z,t)=ξ0(y,z)+ξ(y,z,t) (1) ξ0(y,z)=(1-z / L)[ξ 02 +(ξ 01 −ξ 02 −4ξ B y / T)(1-y / T)] (2) ξ(y,z,t)=ξ m sin(mπz / L)[sinωt-n(ω / c)(yy m )cosωt] (3) where g(y,z,t) represents the mode function of the vocal cords, y and z represent the up-down and front-back directions of the vocal cord vibration, respectively. Up and down, front and back represent the endpoints of the three-dimensional geometry of the vocal cords on the y-axis and the z-axis, respectively. Up represents y_max, down represents y_min, front represents z_max, back represents z_min, and t represents time. 01 ,ξ 02 ,ξ B Corresponding to the glottal inlet half-width, glottal outlet half-width and surface convexity coefficient, T and L represent the thickness and length of the glottis, respectively. m represents the modal displacement amplitude, y m represents the inflection point of the vertical half-wavelength, ω represents the angular frequency, c represents the propagation velocity of the mucosal wave, and m and n represent the number of modes vibrating along the z and y directions, respectively; By changing the parameters in formulas (1), (2), and (3), including ξ 01 ,ξ 02 ,ξ B 、y m ,ω, synthesize vocal cords of various shapes to form a three-dimensional vocal cord model library; Step 2 includes: performing two-dimensional slicing of various three-dimensional pathological vocal cords in the vocal cord three-dimensional model library along the direction of mucosal wave propagation, extracting a series of grayscale tomographic images, converting each grayscale tomographic image slice into a binary matrix and stacking them along the direction of mucosal wave propagation to obtain a three-dimensional input matrix X that can fully reflect the geometric information of the vocal cords; and performing a normalization operation on the three-dimensional input matrix X; Step 3 includes establishing the governing equations for glottal flow as follows: where u i ,ρ,p,ν respectively represent the velocity, density, pressure and kinematic viscosity coefficient of the fluid, represents the partial derivative; the control equation is discretized and solved using the sharp interface immersed boundary method suitable for the complex vocal cord shape. After grid independence verification, an appropriate grid scale (the minimum grid scale is 0.01 cm) is selected for calculation to obtain real physical data, which are stored in the output matrix y. The real physical data include the flow rate, average pressure distribution, and glottal surface pressure field of each vocal cord in the vocal cord three-dimensional model library.

2. The method according to claim 1, characterized in that Step 5 includes: establishing a mapping relationship between the input three-dimensional matrix X and the output matrix y by constructing a three-dimensional convolutional neural network CNN to complete the offline training process. The mathematical expression of the mapping relationship is: y=f(x;θ) (5) Where θ represents all learnable convolution kernels and weights, and f represents a three-dimensional convolutional neural network; Construct a three-dimensional convolutional neural network for predicting flow and average pressure distribution and a three-dimensional convolutional neural network for reconstructing the pressure field of the left and right vocal cords; The 3D convolutional neural network used to predict flow and average pressure distribution and the 3D convolutional neural network used to reconstruct the left and right vocal cord pressure fields both use four layers of 3D convolutional layers (Conv3D). Each Conv3D layer is followed by a 3D maximum pooling layer (MaxPool3D) to further reduce the dimensionality of the feature information. The stride of each 3D maximum pooling layer (MaxPool3D) is 2. The first three 3D convolutional layers (Conv3D) use non-zero padding (valid padding), and the last 3D convolutional layer (Conv3D) uses zero padding (same padding). After the 3D convolutional layer (Conv3D), a 3D global average pooling layer (GlobalAveragePooling3D) is used instead of the traditional fully connected layer to reduce overfitting that may occur during training. The 3D convolutional neural network used to reconstruct the pressure field of the left and right vocal cords has a fully connected layer (FC) between the GlobalAveragePooling3D layer and the output layer. In the 3D convolutional neural network used to predict flow and average pressure distribution and the 3D convolutional neural network used to reconstruct the left and right vocal cord pressure fields, ReLU (Rectified Linear Unit) is used as the activation function between each convolution layer to introduce nonlinearity: σ(x)=ReLU(x)=max(x,0) (6) Where x represents the input feature, σ represents the activation function, and max represents the maximum value; Using forecast data The mean squared sum of the errors between the real data y and the real data y is used as the loss function J. Where N is the number of data points, λ is the regularization coefficient to avoid overfitting of the model, and w represents the convolution kernel and weight; The Adam stochastic gradient descent optimization algorithm is used to adjust the weights to minimize the loss function, and the weights of the optimal model obtained through training are saved.

3. The method according to claim 2, characterized in that Step 6 includes: using the trained three-dimensional convolutional neural network to online predict the flow and pressure distribution of the vocal cords in the test set, and comparing the predicted results with the real physical data to evaluate the prediction performance.

4. An image-based intelligent auxiliary analysis and optimization device for pathological vocal cords implemented by the method according to any one of claims 1 to 3, characterized in that: include: A feature extraction module is used to generate a three-dimensional geometric shape of the pathological vocal cords according to the vibration modal equation of the vocal cords, and obtain a three-dimensional vocal cord model library; The physical quantity calculation module is used to perform two-dimensional slicing on the vocal cord model in the three-dimensional vocal cord model library to obtain a three-dimensional input matrix and use a high-fidelity immersed boundary solver to obtain real physical data; The data set partitioning module is used to partition the data in the vocal cord 3D model library using a random segmentation algorithm to obtain a training set, a validation set, and a test set; The neural network construction and training module is used to build and train a three-dimensional convolutional neural network based on the training set and the validation set, and save the weights of the optimal model obtained through training; The performance evaluation module is used to evaluate the prediction performance of the 3D convolutional neural network using a test set.

5. A storage medium, characterized in that: A computer program or instruction is stored, and when the computer program or instruction is executed, the method according to any one of claims 1 to 3 is implemented.

Images