Ultrasonic image section parameter evaluation system

Abstract

The invention discloses an ultrasonic image section parameter evaluation system, which belongs to the technical field of ultrasonic detection and comprises a data set acquisition module, a classification identification module, a matching module, a parameter extraction module, a neural network training unit and an evaluation analysis unit. Statistics-based parameter extraction, structural similarity-based parameter extraction and natural scene analysis-based parameter extraction are performed through a parameter extraction module, so that section image quality parameters can be extracted from different angles, and different parameter advantages are fused and substituted into a neural network training unit for identification and verification. Secondly, an output result is input into an evaluation analysis unit for calculation and analysis, node parameters are randomly distributed, an output matrix of a hidden layer is calculated, an output weight is solved, the output result is used for representing an ultrasonic image section parameter evaluation value, and a traditional mode of adopting a single section parameter for evaluation and analysis is abandoned; and the evaluation accuracy is greatly improved.

Description

Technical Field

[0001] The invention belongs to the technical field of ultrasonic detection, and in particular relates to an ultrasonic image section parameter evaluation system. Background Art

[0002] In medical imaging, ultrasound imaging is the safest imaging method. It has the advantages of low cost, low radiation, and easy operation, and its application rate is much higher than other imaging methods. Echocardiography, which uses ultrasound to image the heart, has been widely used in clinical examinations and disease diagnosis.

[0003] The integrity of ultrasound standard section images and the evaluation of ultrasound image quality are important links in judging the condition. Traditional image quality evaluation mainly relies on manual means for identification and statistics. Manual image recognition, quality evaluation and statistics are quite time-consuming and prone to errors. Quality evaluation and statistics cannot be used for real-time diagnosis. Most of the existing section image quality evaluation algorithms are based on a single parameter. The image will be affected by changes in the cardiac cycle and other factors. The evaluation is not comprehensive, which greatly reduces its evaluation accuracy. Summary of the invention

[0004] In order to overcome the above-mentioned defects, the present invention provides an ultrasound image section parameter evaluation system, which solves the problem that most of the existing section image quality evaluation algorithms are based on a single parameter, the image will be affected by changes in the cardiac cycle and other factors, the evaluation is not comprehensive, and its evaluation accuracy is low.

[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: an ultrasound image section parameter evaluation system, comprising a data set acquisition module, a classification and recognition module, a matching module, a parameter extraction module, a neural network training unit and an evaluation and analysis unit, wherein the output end of the data set acquisition module is connected to the input end of the classification and recognition module, the output end of the classification and recognition module is connected to the input end of the matching module, the output end of the matching module is connected to the neural network training unit through the parameter extraction module, and the output end of the neural network training unit is connected to the input end of the evaluation and analysis unit;

[0006] The data set acquisition module classifies and annotates the images according to the ultrasound image section parameters to establish a standard image section classification data set;

[0007] The classification and recognition module is used to respectively recognize the data sets acquired in the data set acquisition module, and acquire the section classification feature component and the recognition feature component;

[0008] The matching module is used to match the section classification feature component and the identification feature component in the classification and identification module to obtain several groups of section classification results, and the matched single section classification feature component and single identification feature component are a group;

[0009] The parameter extraction module is used to extract each group of section parameters in the matching module, and the parameter extraction module includes parameter extraction based on statistics, parameter extraction based on structural similarity, and parameter extraction based on natural scene analysis.

[0010] As a further solution of the present invention: the specific implementation steps of the parameter extraction module are as follows:

[0011] Statistics-based parameter extraction:

[0012] For the M×N slice, the peak signal-to-noise ratio is calculated by the root mean square error method. The peak signal-to-noise ratio is defined as follows:

[0013] The root mean square error method is denoted by MES, and the peak signal-to-noise ratio is denoted by PSNR;

[0014]

[0015] Where R(m,n) is the gray value of the reference section x at the coordinate (m,n), I(m,n) is the gray value of the distorted image y at (m,n), and L is the peak signal;

[0016] Parameter extraction based on structural similarity:

[0017] The parameters of structural similarity are represented by SSIM, which compares the brightness, contrast and structural information of the sections. Let x and y be the reference section and the distorted section respectively, then:

[0018]

[0019]

[0020] Among them, μ x , μ y are the mean values ​​of the sections x and y, respectively, reflecting their brightness information, σ x and σ y are the variances of the sections x and y, respectively, reflecting their contrast information, σ xy is the correlation coefficient of x and y, the similarity of reaction structure information, C 1 , C 2 , C 3 is a positive constant close to zero to prevent abnormal results when the denominator is zero. The structural similarity of x and y is obtained by integrating brightness, contrast and structural information;

[0021] SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ

[0022] Among them, α, β, and γ are all greater than 0, which are used to adjust the relativity of brightness, contrast, and structural information;

[0023] Parameter extraction based on natural scene analysis:

[0024] VIF is used to analyze natural scenes and evaluate the quality of cross-sectional images by measuring two common information:

[0025] Information 1: The common information of the output cross-sectional image and the input cross-sectional image obtained through the undistorted channel, that is, the common information of C and x, is the reference cross-sectional image information;

[0026] Information 2: the common information of the output cross-sectional image and the input cross-sectional image obtained through the distortion channel, that is, the common information of C and y, which is the distorted cross-sectional image information;

[0027] Secondly, HVS is modeled. HVS is represented as an undistorted channel. The HVS section image information is filtered and degraded. The degraded information only contains noise items, that is, visual noise. HVS noise is modeled in the wavelet domain.

[0028] For the linear contrast enhanced cross-sectional image, the VIF value is greater than 1, and in other cases the VIF value is between 0 and 1.

[0029] As a further solution of the present invention: the neural network training unit uses the standard parameters of ultrasound image sections as the target training set, and as the input parameters of the neural network model. The target training set includes standard feature components and standard recognition feature components. A neural network model is established, and the neural network model is trained using the target training set to obtain an automatic classification and recognition neural network for ultrasound image sections.

[0030] As a further solution of the present invention: the evaluation and analysis unit is used to bring the characteristic parameters composed in the parameter extraction module into the neural network training unit as a set of data for identification and verification, and to learn the extreme learning machine including L hidden layer nodes and M output layer nodes;

[0031] The specific implementation steps of the evaluation and analysis unit are as follows:

[0032] Randomly assign node parameters: At the beginning of the calculation, the node parameters are randomly generated, that is, the node parameters are independent of the input data, and the random generation obeys any continuous probability distribution;

[0033] Calculate the output matrix of the hidden layer: The size of the hidden layer output matrix is ​​N rows and M columns, that is, the number of rows is the number of input training data, and the number of columns is the number of hidden layer nodes. The output matrix is ​​the result of mapping N input data to L nodes;

[0034] Solving the output weights: The size of the output weight matrix of the hidden layer is L rows and M columns, that is, the number of rows is the number of hidden layer nodes, and the number of columns is the number of output layer nodes. In the extreme learning machine algorithm, there is no error node in the output layer, so that the output weights are solved to minimize the error function. The output result is used to represent the evaluation value of the ultrasound image section parameters.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows: through the parameter extraction module, parameter extraction based on statistics, parameter extraction based on structural similarity and parameter extraction based on natural scene analysis are performed, so that section image quality parameters can be extracted from different angles, and the advantages of different parameters are integrated and brought into the neural network training unit for recognition and verification. Secondly, the output results are input into the evaluation and analysis unit for calculation and analysis, and node parameters are randomly assigned, the output matrix of the hidden layer is calculated, and the output weight is solved. The output results are used to represent the evaluation value of the ultrasound image section parameter, abandoning the traditional method of using a single section parameter for evaluation and analysis, thereby greatly improving the evaluation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The figure is a functional block diagram of the system of the present invention. DETAILED DESCRIPTION

[0037] The technical solution of the present application is further described in detail below in conjunction with specific implementation methods.

[0038] As shown in the figure, the present invention provides a technical solution: an ultrasound image section parameter evaluation system, comprising a data set acquisition module, a classification and recognition module, a matching module, a parameter extraction module, a neural network training unit and an evaluation and analysis unit, wherein the output end of the data set acquisition module is connected to the input end of the classification and recognition module, the output end of the classification and recognition module is connected to the input end of the matching module, the output end of the matching module is connected to the neural network training unit through the parameter extraction module, and the output end of the neural network training unit is connected to the input end of the evaluation and analysis unit;

[0039] The data set acquisition module classifies and annotates the images according to the ultrasound image section parameters and establishes a standard image section classification data set;

[0040] The classification and recognition module is used to respectively recognize the data sets obtained in the data set acquisition module, and obtain the section classification feature components and the recognition feature components;

[0041] The matching module is used to match the section classification feature component and the identification feature component in the classification and identification module to obtain several groups of section classification results, and the matched single section classification feature component and single identification feature component are grouped together;

[0042] The parameter extraction module is used to extract each set of slice parameters in the matching module. The parameter extraction module includes parameter extraction based on statistics, parameter extraction based on structural similarity, and parameter extraction based on natural scene analysis.

[0043] The specific implementation steps of the parameter extraction module are as follows:

[0044] Statistics-based parameter extraction:

[0045] For the M×N slice, the peak signal-to-noise ratio is calculated by the root mean square error method. The peak signal-to-noise ratio is defined as follows:

[0046] The root mean square error method is denoted by MES, and the peak signal-to-noise ratio is denoted by PSNR;

[0047]

[0048] Where R(m,n) is the gray value of the reference section x at the coordinate (m,n), I(m,n) is the gray value of the distorted image y at (m,n), and L is the peak signal;

[0049] Parameter extraction based on structural similarity:

[0050] The parameters of structural similarity are represented by SSIM, which compares the brightness, contrast and structural information of the sections. Let x and y be the reference section and the distorted section respectively, then:

[0051]

[0052] Among them, μ x , μ y are the mean values ​​of the sections x and y, respectively, reflecting their brightness information, σ x and σ y are the variances of the sections x and y, respectively, reflecting their contrast information, σ xy is the correlation coefficient of x and y, the similarity of reaction structure information, C 1 , C 2 , C 3 is a positive constant close to zero to prevent abnormal results when the denominator is zero. The structural similarity of x and y is obtained by integrating brightness, contrast and structural information;

[0053] SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ

[0054] Among them, α, β, and γ are all greater than 0, which are used to adjust the relativity of brightness, contrast, and structural information;

[0055] Parameter extraction based on natural scene analysis:

[0056] VIF is used to analyze natural scenes and evaluate the quality of cross-sectional images by measuring two common information:

[0057] Information 1: The common information of the output cross-sectional image and the input cross-sectional image obtained through the undistorted channel, that is, the common information of C and x, is the reference cross-sectional image information;

[0058] Information 2: the common information of the output cross-sectional image and the input cross-sectional image obtained through the distortion channel, that is, the common information of C and y, which is the distorted cross-sectional image information;

[0059] Secondly, HVS is modeled. HVS is represented as an undistorted channel. The HVS section image information is filtered and degraded. The degraded information only contains noise items, that is, visual noise. HVS noise is modeled in the wavelet domain.

[0060] For the linear contrast enhanced cross-sectional image, the VIF value is greater than 1, and in other cases the VIF value is between 0 and 1.

[0061] The neural network training unit uses the standard parameters of ultrasound image sections as the target training set and as the input parameters of the neural network model. The target training set includes standard feature components and standard recognition feature components. A neural network model is established and trained through the target training set to obtain an automatic classification and recognition neural network for ultrasound image sections.

[0062] The evaluation and analysis unit is used to bring the characteristic parameters composed in the parameter extraction module into the neural network training unit as a set of data for identification and verification, and to learn the extreme learning machine including L hidden layer nodes and M output layer nodes;

[0063] The specific implementation steps of the evaluation and analysis unit are as follows:

[0064] Randomly assign node parameters: At the beginning of the calculation, the node parameters are randomly generated, that is, the node parameters are independent of the input data, and the random generation obeys any continuous probability distribution;

[0065] Calculate the output matrix of the hidden layer: The size of the hidden layer output matrix is ​​N rows and M columns, that is, the number of rows is the number of input training data, and the number of columns is the number of hidden layer nodes. The output matrix is ​​the result of mapping N input data to L nodes;

[0066] Solving the output weights: The size of the output weight matrix of the hidden layer is L rows and M columns, that is, the number of rows is the number of hidden layer nodes, and the number of columns is the number of output layer nodes. In the extreme learning machine algorithm, there is no error node in the output layer, so that the output weights are solved to minimize the error function. The output result is used to represent the evaluation value of the ultrasound image section parameters.

[0067] Embodiment 1:

[0068] An ultrasonic image section parameter evaluation system comprises a data set acquisition module, a classification and recognition module, a matching module, a parameter extraction module, a neural network training unit and an evaluation and analysis unit, wherein the output end of the data set acquisition module is connected to the input end of the classification and recognition module, the output end of the classification and recognition module is connected to the input end of the matching module, the output end of the matching module is connected to the neural network training unit through the parameter extraction module, and the output end of the neural network training unit is connected to the input end of the evaluation and analysis unit;

[0069] The data set acquisition module classifies and annotates the images according to the ultrasound image section parameters, and establishes a standard image section classification data set. The classification and recognition module is used to separately identify the data sets obtained in the data set acquisition module, and obtain section classification feature components and recognition feature components. The matching module is used to match the section classification feature components and recognition feature components in the classification and recognition module to obtain several groups of section classification results, and the matched single section classification feature components and single recognition feature components are grouped together. The neural network training unit uses the ultrasound image section standard parameters as the target training set and as the input parameters of the neural network model. The target training set includes standard feature components and standard recognition feature components. A neural network model is established, and the neural network model is trained through the target training set to obtain an automatic classification and recognition neural network for ultrasound image sections.

[0070] Embodiment 2:

[0071] The parameter extraction module is used to extract each group of slice parameters in the matching module. The parameter extraction module includes parameter extraction based on statistics, parameter extraction based on structural similarity, and parameter extraction based on natural scene analysis. The specific implementation steps of the parameter extraction module are as follows:

[0072] Statistics-based parameter extraction:

[0073] For the M×N slice, the peak signal-to-noise ratio is calculated by the root mean square error method. The peak signal-to-noise ratio is defined as follows:

[0074] The root mean square error method is denoted by MES, and the peak signal-to-noise ratio is denoted by PSNR;

[0075]

[0076] Where R(m,n) is the gray value of the reference section x at the coordinate (m,n), I(m,n) is the gray value of the distorted image y at (m,n), and L is the peak signal;

[0077] Parameter extraction based on structural similarity:

[0078] The parameters of structural similarity are represented by SSIM, which compares the brightness, contrast and structural information of the sections. Let x and y be the reference section and the distorted section respectively, then:

[0079]

[0080] Among them, μ x , μ y are the mean values ​​of the sections x and y, respectively, reflecting their brightness information, σ x and σ y are the variances of the sections x and y, respectively, reflecting their contrast information, σ xy is the correlation coefficient of x and y, the similarity of reaction structure information, C 1 , C 2 , C 3 is a positive constant close to zero to prevent abnormal results when the denominator is zero. The structural similarity of x and y is obtained by integrating brightness, contrast and structural information;

[0081] SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ

[0082] Among them, α, β, and γ are all greater than 0, which are used to adjust the relativity of brightness, contrast, and structural information;

[0083] Parameter extraction based on natural scene analysis:

[0084] VIF is used to analyze natural scenes and evaluate the quality of cross-sectional images by measuring two common information:

[0085] Information 1: The common information of the output cross-sectional image and the input cross-sectional image obtained through the undistorted channel, that is, the common information of C and x, is the reference cross-sectional image information;

[0086] Information 2: the common information of the output cross-sectional image and the input cross-sectional image obtained through the distortion channel, that is, the common information of C and y, which is the distorted cross-sectional image information;

[0087] Secondly, HVS is modeled. HVS is represented as an undistorted channel. The HVS section image information is filtered and degraded. The degraded information only contains noise items, that is, visual noise. HVS noise is modeled in the wavelet domain.

[0088] For the linear contrast enhanced cross-sectional image, the VIF value is greater than 1, and in other cases the VIF value is between 0 and 1.

[0089] Embodiment 3:

[0090] The evaluation and analysis unit is used to bring the characteristic parameters composed in the parameter extraction module into the neural network training unit as a set of data for identification and verification, and to learn the extreme learning machine including L hidden layer nodes and M output layer nodes;

[0091] The specific implementation steps of the evaluation and analysis unit are as follows:

[0092] Randomly assign node parameters: At the beginning of the calculation, the node parameters are randomly generated, that is, the node parameters are independent of the input data, and the random generation obeys any continuous probability distribution;

[0093] Calculate the output matrix of the hidden layer: The size of the hidden layer output matrix is ​​N rows and M columns, that is, the number of rows is the number of input training data, and the number of columns is the number of hidden layer nodes. The output matrix is ​​the result of mapping N input data to L nodes;

[0094] Solving the output weights: The size of the output weight matrix of the hidden layer is L rows and M columns, that is, the number of rows is the number of hidden layer nodes, and the number of columns is the number of output layer nodes. In the extreme learning machine algorithm, there is no error node in the output layer, so that the output weights are solved to minimize the error function. The output result is used to represent the evaluation value of the ultrasound image section parameters.

[0095] The working principle of the present invention is:

[0096] The data set acquisition module classifies and annotates the images according to the ultrasound image section parameters to establish a standard image section classification data set. The classification and recognition module recognizes the data sets obtained in the data set acquisition module respectively to obtain section classification feature components and recognition feature components. The matching module matches the section classification feature components and recognition feature components in the classification and recognition module to obtain several groups of section classification results, and the matched single section classification feature components and single recognition feature components are grouped together. The parameter extraction module is used to perform parameter extraction based on statistics, parameter extraction based on structural similarity, and parameter extraction based on natural scene analysis.

[0097] The neural network training unit uses the standard parameters of ultrasound image sections as the target training set and as the input parameters of the neural network model. The target training set includes standard feature components and standard recognition feature components. A neural network model is established and trained through the target training set to obtain an automatic classification and recognition neural network for ultrasound image sections. The final evaluation and analysis unit brings the feature parameters composed in the parameter extraction module as a set of data into the neural network training unit for recognition and verification, and learns the extreme learning machine containing L hidden layer nodes and M output layer nodes. The output result is used to represent the evaluation value of the ultrasound image section parameters.

[0098] From the above we know:

[0099] The parameter extraction module is used to perform parameter extraction based on statistics, parameter extraction based on structural similarity, and parameter extraction based on natural scene analysis. It is therefore possible to extract section image quality parameters from different angles, and integrate the advantages of different parameters into the neural network training unit for recognition and verification. Secondly, the output results are input into the evaluation and analysis unit for computational analysis, and node parameters are randomly assigned, the output matrix of the hidden layer is calculated, and the output weights are solved. The output results are used to represent the evaluation values ​​of ultrasound image section parameters, abandoning the traditional method of using a single section parameter for evaluation and analysis, greatly improving the evaluation accuracy.

[0100] The preferred implementation modes of the present application are described in detail above, but the present application is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present application.

Claims

1. An ultrasound image section parameter evaluation system, comprising a data set acquisition module, a classification and recognition module, a matching module, a parameter extraction module, a neural network training unit and an evaluation and analysis unit, characterized in that: The output end of the data set acquisition module is connected to the input end of the classification and recognition module, the output end of the classification and recognition module is connected to the input end of the matching module, the output end of the matching module is connected to the neural network training unit through the parameter extraction module, and the output end of the neural network training unit is connected to the input end of the evaluation and analysis unit; The data set acquisition module classifies and annotates the images according to the ultrasound image section parameters to establish a standard image section classification data set; The classification and recognition module is used to respectively recognize the data sets acquired in the data set acquisition module, and acquire the section classification feature component and the recognition feature component; The matching module is used to match the section classification feature component and the identification feature component in the classification and identification module to obtain several groups of section classification results, and the matched single section classification feature component and single identification feature component are a group; The parameter extraction module is used to extract each group of section parameters in the matching module, and the parameter extraction module includes parameter extraction based on statistics, parameter extraction based on structural similarity, and parameter extraction based on natural scene analysis.

2. The ultrasonic image section parameter evaluation system according to claim 1, characterized in that: The specific implementation steps of the parameter extraction module are as follows: Statistics-based parameter extraction: For the M×N slice, the peak signal-to-noise ratio is calculated by the root mean square error method. The peak signal-to-noise ratio is defined as follows: The root mean square error method is denoted by MES, and the peak signal-to-noise ratio is denoted by PSNR; Where R(m,n) is the gray value of the reference section x at the coordinate (m,n), I(m,n) is the gray value of the distorted image y at (m,n), and L is the peak signal; Parameter extraction based on structural similarity: The parameters of structural similarity are represented by SSIM, which compares the brightness, contrast and structural information of the sections. Let x and y be the reference section and the distorted section respectively, then: Among them, μ x , μ y are the mean values ​​of the sections x and y, respectively, reflecting their brightness information, σ x and σ y are the variances of the sections x and y, respectively, reflecting their contrast information, σ xy is the correlation coefficient of x and y, reflecting the similarity of structural information. C1, C2, and C3 are positive constants close to zero to prevent abnormal results when the denominator is zero. The structural similarity of x and y is obtained by integrating brightness, contrast, and structural information. SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ Among them, α, β, and γ are all greater than 0, which are used to adjust the relativity of brightness, contrast, and structural information; Parameter extraction based on natural scene analysis: VIF is used to analyze natural scenes and evaluate the quality of cross-sectional images by measuring two common information: Information 1: The common information of the output cross-sectional image and the input cross-sectional image obtained through the undistorted channel, that is, the common information of C and x, is the reference cross-sectional image information; Information 2: the common information of the output cross-sectional image and the input cross-sectional image obtained through the distortion channel, that is, the common information of C and y, which is the distorted cross-sectional image information; Secondly, HVS is modeled. HVS is represented as an undistorted channel. The HVS section image information is filtered and degraded. The degraded information only contains noise items, that is, visual noise. HVS noise is modeled in the wavelet domain. For the linear contrast enhanced cross-sectional image, the VIF value is greater than 1, and in other cases the VIF value is between 0 and 1.

3. The ultrasonic image section parameter evaluation system according to claim 2, characterized in that: The neural network training unit uses the standard parameters of ultrasound image sections as the target training set and as the input parameters of the neural network model. The target training set includes standard feature components and standard recognition feature components. The neural network model is established and trained through the target training set to obtain an automatic classification and recognition neural network for ultrasound image sections.

4. The ultrasonic image section parameter evaluation system according to claim 3, characterized in that: The evaluation and analysis unit is used to bring the characteristic parameters composed in the parameter extraction module as a set of data into the neural network training unit for identification verification, and to learn the extreme learning machine including L hidden layer nodes and M output layer nodes; The specific implementation steps of the evaluation and analysis unit are as follows: Randomly assign node parameters: At the beginning of the calculation, the node parameters are randomly generated, that is, the node parameters are independent of the input data, and the random generation obeys any continuous probability distribution; Calculate the output matrix of the hidden layer: The size of the hidden layer output matrix is ​​N rows and M columns, that is, the number of rows is the number of input training data, and the number of columns is the number of hidden layer nodes. The output matrix is ​​the result of mapping N input data to L nodes; Solving the output weights: The size of the output weight matrix of the hidden layer is L rows and M columns, that is, the number of rows is the number of hidden layer nodes, and the number of columns is the number of output layer nodes. In the extreme learning machine algorithm, there is no error node in the output layer, so that the output weights are solved to minimize the error function. The output result is used to represent the evaluation value of the ultrasound image section parameters.

Images