CN116797852A - Ultrasound image recognition model and ultrasound image recognition method - Google Patents

Abstract

This application is applicable to the technical field of ultrasonic image recognition, and provides an ultrasonic image recognition model and an ultrasonic image recognition method. The ultrasonic image recognition model includes a residual module, a pyramid-like decomposition module, a probability pooling module, a channel attention module, and a first Feature image adjustment module, second feature image adjustment module and image recognition module; the residual module receives the ultrasound image, and the output terminals are respectively connected to the pyramid-like decomposition module, the first feature image adjustment module, the second feature image adjustment module, and the pyramid-like decomposition module The probability pooling module is connected, the probability pooling module is connected to the channel attention module, the channel attention module is connected to the first feature image adjustment module, the first feature image adjustment module is connected to the second feature image adjustment module, and the second feature image adjustment module is connected to the image Recognition module, the image recognition module outputs the recognition result of the ultrasound image. This application can improve the accuracy of ultrasound image recognition.

Description

Ultrasound image recognition model and ultrasound image recognition method

Technical field

The present application belongs to the technical field of ultrasonic image recognition, and in particular relates to an ultrasonic image recognition model and an ultrasonic image recognition method.

Background technique

Ultrasonic testing achieves analysis of the internal condition of an object by receiving and imaging the echoes of sound waves of a certain frequency on interfaces with differences in medium density. Because ultrasonic non-destructive testing has the advantages of dynamic detection, high sensitivity, high controllability and good economy, it has long had an irreplaceable position in many fields such as industry, medicine, and civil engineering. However, ultrasound images are formed by echo sampling, conversion, and assignment of fixed center frequencies. Not only are there a small number of channels, but there are also many noise disturbances, resulting in a low amount of effective information. In addition, because the identification of ultrasound images requires professionals in various fields, it not only relies on personal experience, but also the accuracy and efficiency cannot meet the needs of social development.

In recent years, artificial intelligence technology represented by neural networks has continued to make breakthroughs and is entering a period of rapid development, with its applications in various industries gradually deepening. Therefore, it is of great significance to introduce artificial intelligence methods in the recognition and classification tasks of ultrasound images, which will help professionals make decisions and alleviate a large number of phase problems. However, due to the large difference between the ultrasound image imaging process and conventional light imaging, it is difficult to directly use convolutional neural networks to extract sufficient features from ultrasound images, and the effect is not ideal. Therefore, the recognition accuracy of existing ultrasound image recognition methods needs to be improved.

Contents of the invention

This application provides an ultrasound image recognition model and an ultrasound image recognition method, which can solve the problem of low recognition accuracy of current ultrasound image recognition methods.

In the first aspect, this application provides an ultrasound image recognition model, including a residual module, a pyramid-like decomposition module, a probability pooling module, a channel attention module, a first feature image adjustment module, a second feature image adjustment module, and an image identification module;

The residual module is used to perform convolution processing on the input ultrasound image to obtain the initial multi-channel feature image corresponding to the input ultrasound image;

The pyramid-like decomposition module is used to perform pyramid decomposition on the feature image corresponding to each channel in the initial multi-channel feature image, and obtain multiple feature sub-images corresponding to each channel; the scales of the multiple feature sub-images are different from each other;

The probability pooling module is used to perform probability pooling on each feature sub-image in multiple feature sub-images to obtain the initial weight coefficient of each feature sub-image; probability pooling represents the pooling method corresponding to different feature sub-images. The probabilities are different from each other. The pooling methods include maximum pooling, average pooling, minimum pooling and random pooling. The probability obeys the standard normal distribution;

The channel attention module is used to calculate the channel weight of each channel based on the initial weight coefficient, and activate the channel weights of all channels to obtain the channel weight matrix of the initial multi-channel feature image;

The first feature image adjustment module is used to multiply the channel weight matrix and the initial multi-channel feature image to obtain the intermediate multi-channel feature image;

The second feature image adjustment module is used to add the intermediate multi-channel feature image and the initial multi-channel feature image to obtain the final multi-channel feature image;

The image recognition module is used to recognize the final multi-channel feature image and obtain the recognition result of the final multi-channel feature image.

Optionally, the input end of the residual module receives the ultrasound image, and the output end of the residual module is respectively connected to the input end of the pyramid-like decomposition module, the first input end of the first feature image adjustment module, and the third input end of the second feature image adjustment module. One input terminal, the output terminal of the pyramid-like decomposition module is connected to the input terminal of the probability pooling module, the output terminal of the probability pooling module is connected to the input terminal of the channel attention module, and the output terminal of the channel attention module is connected to the first feature image adjustment module The second input end of the first feature image adjustment module is connected to the second input end of the second feature image adjustment module. The output end of the second feature image adjustment module is connected to the input end of the image recognition module. The output of the image recognition module The terminal outputs the recognition results of the ultrasound image.

Optionally, perform pyramid decomposition on the feature image corresponding to each channel in the initial multi-channel feature image to obtain multiple feature sub-images corresponding to each channel, including:

For the feature image corresponding to each channel, perform the following steps:

Step i, use the feature image as the initial image, copy the initial image, and obtain the intermediate image;

Step ii: intercept the outermost pixels of the intermediate image with a preset step size to obtain the final image;

Step iii, if the scale of the final image is greater than or equal to the preset scale threshold, use the final image as the initial image in step i, and return to step i; otherwise, use the initial image and all final images obtained by executing step iii as each Multiple feature sub-images corresponding to the channel.

Optional, the calculation formula of scale threshold is:

E＝(min(M,N)-2)mod2*T

Among them _{, M}

Optionally, calculate the channel weight of each channel based on the initial weight coefficient, and perform one-dimensional convolution processing on the channel weights of all channels to obtain the channel weight matrix of the initial multi-channel feature image, including:

For each channel, add all the initial weight coefficients corresponding to the channel to obtain the channel weight;

Use one-dimensional convolution to convolve the channel weights of all channels to obtain the intermediate channel weight matrix; among them, if the total number of channels is less than or equal to 128, set the convolution kernel of the one-dimensional convolution to 5, otherwise set one The convolution kernel of dimensional convolution is set to 7;

Use the activation function to activate the intermediate channel weight matrix to obtain the channel weight matrix.

In the second aspect, this application provides an ultrasound image recognition method, including:

Obtain an ultrasound image training set; the ultrasound image training set includes multiple ultrasound images of known types;

Input multiple known types of ultrasound images into the ultrasound image recognition model one by one to obtain the recognition results of each ultrasound image; the ultrasound image recognition model is the above-mentioned ultrasound image recognition model;

According to the recognition results and the type of each ultrasound image, the loss of the ultrasound image recognition model is calculated, and the ultrasound image recognition model is backpropagated according to the loss until the ultrasound image recognition model is fitted, and the trained ultrasound image recognition model is obtained;

The ultrasound image to be recognized is input into the trained ultrasound image recognition model to obtain the recognition result of the ultrasound image to be recognized.

Optionally, input multiple ultrasound images of known types into the ultrasound image recognition model one by one to obtain the recognition results of each ultrasound image, including:

By calculation formula

Obtain the output probability P _k of each type; where, P _k represents the output probability of type k, k = 1, 2,..., Q, Q represents the total number of types of ultrasound images in the ultrasound image training set, and z _k represents The output value of the k-th node has a one-to-one correspondence between the number of nodes and the number of types, j represents the j-th node, n represents the number of output nodes, and softmax(·) represents the activation function;

The type with the highest output probability is used as the recognition result.

Optionally, calculate the loss of the ultrasound image recognition model based on the recognition result and the type of each ultrasound image, and perform backpropagation on the ultrasound image recognition model based on the loss until the ultrasound image recognition model is fitted to obtain the trained ultrasound image. Recognition models, including:

For each ultrasound image, calculate the formula

FL＝-a _k (1-p _k ) ^γ log(p _k )

Obtain the loss FL of the ultrasound image recognition model; where a _k represents the weight of type k, p _k represents the output probability of type k, γ represents the focus parameter, Y∈[0,5];

The ultrasound image recognition model is backpropagated according to the loss EL, until the loss value of the new ultrasound image recognition model obtained by backpropagation is less than the preset loss threshold, the fitting of the new ultrasound image recognition model is determined, and the trained ultrasound image recognition model is obtained. Model; among them, the momentum of back propagation is set to 0.9, the learning rate of back propagation is obtained using the thermal learning rate strategy, and the weight attenuation coefficient of L1 regularization of back propagation is set to 5e-4.

The above solution of this application has the following beneficial effects:

The ultrasonic image recognition model provided by this application uses a pyramid-like decomposition module to perform pyramid decomposition on the feature image corresponding to each channel, which can make the subsequent extracted features more accurate, thereby improving the accuracy of ultrasonic image recognition. It uses a probability pooling module to decompose each channel. Probabilistic pooling of feature sub-images can prevent over-fitting when pooling feature images. Through a variety of pooling methods, important local features can be selectively extracted, and feature images can be identified based on these important local features. It can significantly improve the accuracy of ultrasound image recognition.

Other beneficial effects of the present application will be described in detail in the subsequent specific embodiments section.

Description of the drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the purpose of the present application. For some embodiments, for those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is a schematic structural diagram of an ultrasound image recognition model provided by an embodiment of the present application;

Figure 2 is a schematic structural diagram of a residual module provided by an embodiment of the present application;

Figure 3 is a schematic structural diagram of a feature image character provided by an embodiment of the present application;

Figure 4 is a flow chart of ultrasound image recognition provided by an embodiment of the present application.

Detailed ways

In the following description, for the purpose of explanation rather than limitation, specific details such as specific system structures and technologies are provided to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be practiced in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

It will be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements and/or components but does not exclude one or more other The presence or addition of features, integers, steps, operations, elements, components and/or collections thereof.

It will also be understood that the term "and/or" as used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

As used in this specification and the appended claims, the term "if" may be interpreted as "when" or "once" or "in response to determining" or "in response to detecting" depending on the context. ". Similarly, the phrase "if determined" or "if [the described condition or event] is detected" may be interpreted, depending on the context, to mean "once determined" or "in response to a determination" or "once the [described condition or event] is detected ]" or "in response to detection of [the described condition or event]".

In addition, in the description of this application and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

Reference in this specification to "one embodiment" or "some embodiments" or the like means that a particular feature, structure or characteristic described in connection with the embodiment is included in one or more embodiments of the application. Therefore, the phrases "in one embodiment", "in some embodiments", "in other embodiments", "in other embodiments", etc. appearing in different places in this specification are not necessarily References are made to the same embodiment, but rather to "one or more but not all embodiments" unless specifically stated otherwise. The terms “including,” “includes,” “having,” and variations thereof all mean “including but not limited to,” unless otherwise specifically emphasized.

In order to solve the problem of low recognition accuracy of current ultrasonic image recognition models, this application provides an ultrasonic image recognition model and an ultrasonic image recognition method. The ultrasonic image recognition model uses a pyramid-like decomposition module to pyramid the feature images corresponding to each channel. Decomposition can make the subsequent extracted features more accurate, thereby improving the accuracy of ultrasound image recognition. Using the probability pooling module to perform probability pooling on each feature sub-image can prevent overfitting during feature image pooling. Through a variety of The pooling method can selectively extract important local features, and identify feature images based on these important local features, which can significantly improve the accuracy of ultrasound image recognition.

The following is an exemplary description of the ultrasound image recognition model provided by this application.

As shown in Figure 1, the ultrasound image recognition model provided by this application includes a residual module (shown as 11 in Figure 1), a pyramid-like decomposition module (shown as 12 in Figure 1), and a probability pooling module (shown as 12 in Figure 1) (shown as 13 in Figure 1), channel attention module (shown as 14 in Figure 1), first feature image adjustment module (shown as 15 in Figure 1), second feature image adjustment module (shown as 16 in Figure 1 ) and image recognition module (shown as 17 in Figure 1).

For the convenience of explanation, in the embodiment of this application, only the functions of each module in the ultrasound image recognition model are explained, including the residual module, the pyramid-like decomposition module, the probability pooling module, the channel attention module, and the first feature image adjustment. The module, the second characteristic image adjustment module and the image recognition module serve as a set of functional units, and the ultrasonic image recognition model also includes multiple sets of the above functional units.

The connection relationship between each module is as follows:

The input end of the residual module receives the ultrasound image, and the output end of the residual module is respectively connected to the input end of the pyramid-like decomposition module, the first input end of the first feature image adjustment module, and the first input end of the second feature image adjustment module, The output end of the pyramid-like decomposition module is connected to the input end of the probability pooling module, the output end of the probability pooling module is connected to the input end of the channel attention module, and the output end of the channel attention module is connected to the second input of the first feature image adjustment module terminal, the output terminal of the first characteristic image adjustment module is connected to the second input terminal of the second characteristic image adjustment module, the output terminal of the second characteristic image adjustment module is connected to the input terminal of the image recognition module, and the output terminal of the image recognition module outputs the ultrasound image. identification results.

Each module is illustratively described below.

The residual module is used to perform convolution processing on the input ultrasound image to obtain the initial multi-channel feature image corresponding to the input ultrasound image.

Illustratively, as shown in Figure 2, in an embodiment of the present application, the above-mentioned residual module includes a first convolution layer (shown as 21 in Figure 2), a first normalization layer (shown as 22 in Figure 2) shown), the first activation layer (shown as 23 in Figure 2), the second convolution layer (shown as 24 in Figure 2), the second normalization layer (shown as 25 in Figure 2), the second activation layer (shown as 26 in Figure 2), the third convolution layer (shown as 27 in Figure 2), and the third normalization layer (shown as 28 in Figure 2), the above layers are connected in sequence. Among them, the first convolution layer is a 1*1 convolution layer, the second convolution layer is a 3*3 convolution layer, the third convolution layer is a 1*1 convolution layer, and the step size of the second convolution layer is set is 1, this can reduce the loss of pixel related information caused by the stride length.

It should be noted that in order to reduce redundant parameters, in the embodiment of the present application, the bias terms of each convolution layer are removed.

The pyramid-like decomposition module is used to perform pyramid decomposition on the feature image corresponding to each channel in the initial multi-channel feature image, and obtain multiple feature sub-images corresponding to each channel.

Among them, the scales of multiple feature sub-images are different from each other.

Use the pyramid-like decomposition module to perform pyramid decomposition on the feature image corresponding to each channel in the initial multi-channel feature image. The specific process of obtaining multiple feature sub-images corresponding to each channel is as follows:

In step i, the feature image is used as the initial image, and the initial image is copied to obtain the intermediate image.

Step ii: intercept the outermost pixels of the intermediate image with a preset step size to obtain the final image.

Step iii, if the scale of the final image is greater than or equal to the preset scale threshold, use the final image as the initial image in step i, and return to step i; otherwise, use the initial image and all final images obtained by executing step iii as each Multiple feature sub-images corresponding to the channel. The calculation formula of scale threshold is

E＝(min(M,N)-2)mod2*T

Among them _{, M}

In addition, it can be calculated by the formula

Get the total number of feature sub-images Q. It should be understood that all feature sub-images are arranged in order from small to large scales, and a feature sub-image pyramid can be obtained, as shown in Figure 3.

When performing pyramid-like decomposition on feature images, the calculation formula can be used

P _C =C*P _Q

The number of pixels P _C after quasi-pyramid decomposition is obtained, where C represents the number of channels of the feature image, P _Q represents the number of pixels contained in the single-channel feature image, and k represents the k-th feature sub-image.

The probability pooling module is used to perform probability pooling on each feature sub-image in multiple feature sub-images to obtain the initial weight coefficient of each feature sub-image.

The above probability pooling means that the probabilities of the pooling methods corresponding to different feature sub-images are different from each other. The pooling methods include maximum pooling, average pooling, minimum pooling and random pooling, and the probabilities obey the standard normal distribution.

Since the results produced by the same pooling on feature images with central symmetry are also similar, if a unified pooling method is used, it is easy to over-amplify local features, which will not only cover other features, but also make it difficult to extract multi-dimensional information. . In view of this, this application uses different pooling methods for feature sub-images of different scales after pyramid decomposition.

Specifically, the probabilities of different pooling methods are set to obey the standard normal distribution. For example, the probability of the pooling method (global maximum pooling) distributed in the interval (μ-σ,μ+σ) is 68%, and the probability of the pooling method (global average) distributed in the interval (-∞,μ-σ) The probability of pooling (pooling) is 16%, and the probability of pooling (global minimum pooling) distributed in the interval (μ+σ,+∞) is 12%. When it is located on μ-σ or μ+σ, the global Random pooling with 4% probability. In practical applications, global maximum pooling and global average pooling are preferred, because global maximum pooling can retain the representative information of feature images, and global average pooling can blur the information of feature images.

The channel attention module is used to calculate the channel weight of each channel based on the initial weight coefficient, and perform one-dimensional convolution processing on the channel weights of all channels to obtain the channel weight matrix of the initial multi-channel feature image.

In the embodiment of this application, the channel attention module calculates the channel weight of each channel based on the initial weight coefficient, and performs one-dimensional convolution processing on the channel weights of all channels to obtain the specific channel weight matrix of the initial multi-channel feature image. The process is as follows:

For each channel, add all the initial weight coefficients corresponding to the channel to obtain the channel weight.

For example, the channel weight Wi is obtained by calculating the formula _Wi = _wi,1 + _wi,2 +...+ _wi,j +...+wi _,Q ; where _Wi represents _the The channel weight of the i channel, w _i,j represents the j-th initial weight coefficient of the i-th channel.

One-dimensional convolution is used to convolve the channel weights of all channels to obtain the intermediate channel weight matrix. Among them, if the total number of channels is less than or equal to 128, the convolution kernel of the one-dimensional convolution is set to 5, otherwise the convolution kernel of the one-dimensional convolution is set to 7.

It is worth mentioning that the convolution kernel of one-dimensional convolution is set according to the total number of channels, which is beneficial to deal with the complex changes of ultrasound images. At the same time, one-dimensional convolution performs convolution processing on the channel weights of all channels, which can extract the dependence between adjacent channels and make the channel weight matrix more accurate.

Use the activation function to activate the intermediate channel weight matrix to obtain the channel weight matrix.

For example, the expression of the activation function is S(x)=1/(1+e ^-x ).

It should be noted that each row in the channel weight matrix corresponds to a channel weight.

The first feature image adjustment module is used to multiply the channel weight matrix and the initial multi-channel feature image to obtain the intermediate multi-channel feature image.

This part essentially weights the feature image of each channel. Each element in the channel weight matrix represents the weight value of the corresponding channel. By multiplying with the initial multi-channel feature image, the importance of each channel can be adjusted to highlight or suppress the feature information of different channels. This can extract more distinguishing and important features.

The second feature image adjustment module is used to add the intermediate multi-channel feature image and the initial multi-channel feature image to obtain the final multi-channel feature image.

This step can enhance the expressiveness and robustness of the feature image. Through the addition operation, the feature information in the intermediate multi-channel feature image can be fused with the original information in the initial multi-channel feature image, thereby obtaining a richer and more accurate feature representation.

The image recognition module is used to recognize the final multi-channel feature image and obtain the recognition result of the final multi-channel feature image.

In the embodiment of this application, the image recognition module is a fully connected layer. When performing feature image recognition, each neuron in the fully connected layer will output a probability, and the ultrasound image type corresponding to the neuron with the highest probability will be used as the recognition result. ; Among them, neurons and ultrasound image types have a one-to-one correspondence, that is, each neuron corresponds to an ultrasound image type.

It can be seen that the ultrasound image recognition model provided in this application uses a pyramid-like decomposition module to perform pyramid decomposition on the feature image corresponding to each channel, which can make the subsequently extracted features more accurate, thereby improving the accuracy of ultrasound image recognition, and uses the probability pooling module Probabilistic pooling of each feature sub-image can prevent over-fitting when pooling feature images. Through a variety of pooling methods, important local features can be selectively extracted, and feature images can be processed based on these important local features. Recognition can significantly improve the accuracy of ultrasound image recognition.

The following is an exemplary description of the ultrasound image recognition method provided by this application.

As shown in Figure 4, the ultrasound image recognition method provided by this application includes the following steps:

Step 41: Obtain the ultrasound image training set.

The above-mentioned ultrasound image training set includes multiple ultrasound images of known types.

The following is an exemplary explanation of the process of obtaining the ultrasound image training set.

Step 41.1: Obtain 1500 ultrasound images from the ultrasound image open data set to form an ultrasound image training set.

Step 41.2: Experts in related fields identify the ultrasound image training set and determine the type of each ultrasound image.

For example, the ultrasound images in the ultrasound image training set are identified as ultrasound images corresponding to the liver, ultrasound images corresponding to the spleen, ultrasound images corresponding to the kidneys, ultrasound images corresponding to the intestines, ultrasound images corresponding to the gallbladder, and ultrasound images corresponding to the bladder, Ultrasound images corresponding to the calf muscles and ultrasonic images corresponding to the breast. Accordingly, the types of ultrasound images include liver, spleen, kidney, intestine, gallbladder, bladder, calf muscles, and breast.

Step 41.3: Perform data amplification on the classified ultrasound images.

Data amplification methods that can be used in the embodiments of this application include:

Mode 1, any angle rotation within 0°~30°;

Method 2: Randomly flip up and down or left and right;

Method 3: Maintain the aspect ratio of the image and enlarge it to 1.5 times the original size, and crop it to the original size;

Method 4: Crop out a small part of the original image and then enlarge it to the original size;

Method 5: Randomly select a pixel area of 20*20~50*50 in the image, and fill all the pixels in the area with the average pixel value of the area;

Method 6: Randomly use one of Sobel operator (Sobel), Roberts operator (Roberts operator), Prewitt (a differential operator for image edge detection) and Laplacian operator (Laplacian) for ultrasound The image is sharpened;

Method 7: Randomly add one of Gaussian noise, salt-and-pepper noise, or gamma noise to the ultrasound image.

Step 41.4: Scale the ultrasound image after data amplification.

Specifically, the size of the ultrasound images of each category can be scaled to 64*64, and then all ultrasound images of the same size can be standardized so that the mean of any pixel value of each ultrasound image is 0 and the standard deviation is 1. The calculation formula of this process is P _stand = (P _S -μ)/σ, where P _s represents the original pixel value, P _stand represents the new pixel value after normalization, μ represents the mean value of the pixel, and σ represents the standard deviation of the pixel.

Step 42: Input multiple ultrasound images of known types into the ultrasound image recognition model one by one to obtain the recognition result of each ultrasound image.

Wherein, the ultrasound image recognition model is the above-mentioned ultrasound image recognition model.

The following is an exemplary description of the process of step 42 (input multiple known types of ultrasound images into the ultrasound image recognition model one by one to obtain the recognition result of each ultrasound image).

Specifically, through calculation formula

Get the output probability P _k of each type.

Among them, P _k represents the output probability of type k, k = 1, 2,..., Q, Q represents the total number of types of ultrasound images in the ultrasound image training set, z _k represents the output value of the k-th node, the node's There is a one-to-one correspondence between the number and the number of types. j represents the j-th node, n represents the number of output nodes, and softmax(·) represents the activation function.

Exemplarily, in an embodiment of the present application, the output probability of the ultrasound image recognition model is

The type with the highest output probability is used as the recognition result. That is, the input ultrasound image is recognized as an ultrasound image corresponding to the calf muscle.

Step 43: Calculate the loss of the ultrasound image recognition model based on the recognition result and the type of each ultrasound image, and perform backpropagation on the ultrasound image recognition model based on the loss until the ultrasound image recognition model is fitted to obtain the trained ultrasound image recognition Model.

Specifically, for each ultrasound image, the calculation formula

FL＝-a _k (1-p _k ) ^γ log(p _k )

Obtain the loss FL of the ultrasound image recognition model; where a _k represents the weight of type k, p _k represents the output probability of type k, γ represents the focus parameter, γ∈[0,5];

The ultrasound image recognition model is backpropagated according to the loss FL, until the loss value of the new ultrasound image recognition model obtained by backpropagation is less than the preset loss threshold, the new ultrasound image recognition model is determined to fit, and the trained ultrasound image recognition model is obtained. Model; among them, the momentum of back propagation is set to 0.9, the learning rate of back propagation is obtained using the thermal learning rate strategy, and the weight attenuation coefficient of L1 regularization of back propagation is set to 5e-4.

Step 44: Input the ultrasound image to be recognized into the trained ultrasound image recognition model to obtain the recognition result of the ultrasound image to be recognized.

The above-mentioned ultrasound image to be recognized refers to an ultrasound image of an uncertain type.

In the embodiment of the present application, after the ultrasound image recognition model is trained, in order to further improve the accuracy of the ultrasound image recognition model, the average accuracy, precision, sensitivity and F-1 score (which is a statistic) corresponding to the recognition results can be calculated. (an index used in science to measure the accuracy of a binary classification model) to measure the effect of the ultrasound image recognition model, and select the ultrasound image with the best effect (highest average accuracy, highest precision, highest sensitivity, and highest F-1 score) The recognition model serves as the final ultrasound image recognition model.

The average accuracy is calculated as follows:

The calculation formula for accuracy is as follows:

The sensitivity is calculated as follows:

The F-1 score is calculated as follows:

Among them, represents the number of samples correctly predicted to be positive, represents the number of samples incorrectly predicted to be positive, represents the number of samples incorrectly predicted to be negative, and represents the number of samples correctly predicted to be negative.

The comparison of the recognition effects of the ultrasound image recognition model and ultrasound image recognition method provided by this application with other existing technologies is as shown in the following table:

It can be seen from the results in the above table that the ultrasonic image recognition model and ultrasonic image recognition method provided by this application are more advanced than other existing technologies, and the ultrasonic image recognition effect is more accurate.

The above is the preferred embodiment of the present application. It should be pointed out that for those of ordinary skill in the art, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications can also be made. should be regarded as the scope of protection of this application.

Claims (8)

1. An ultrasound image recognition model, characterized in that the ultrasound image recognition model includes a residual module, a pyramid-like decomposition module, a probability pooling module, a channel attention module, a first feature image adjustment module, a second feature image Adjustment module and image recognition module; The residual module is used to perform convolution processing on the input ultrasound image to obtain an initial multi-channel feature image corresponding to the input ultrasound image; The pyramid-like decomposition module is used to perform pyramid decomposition on the feature image corresponding to each channel in the initial multi-channel feature image to obtain multiple feature sub-images corresponding to each channel; the scales of the multiple feature sub-images different from each other; The probability pooling module is used to perform probability pooling on each feature sub-image in the plurality of feature sub-images to obtain the initial weight coefficient of each feature sub-image; the probability pooling represents different feature sub-images. The probabilities of the pooling methods corresponding to the images are different from each other. The pooling methods include maximum pooling, average pooling, minimum pooling and random pooling. The probabilities obey the standard normal distribution; The channel attention module is used to calculate the channel weight of each channel according to the initial weight coefficient, and perform one-dimensional convolution processing on the channel weights of all channels to obtain the channel weight matrix of the initial multi-channel feature image; The first feature image adjustment module is used to multiply the channel weight matrix and the initial multi-channel feature image to obtain an intermediate multi-channel feature image; The second feature image adjustment module is used to add the intermediate multi-channel feature image and the initial multi-channel feature image to obtain a final multi-channel feature image; The image recognition module is used to recognize the final multi-channel feature image and obtain the recognition result of the final multi-channel feature image. 2. The ultrasonic image recognition model according to claim 1, characterized in that the input end of the residual module receives the ultrasonic image, and the output end of the residual module is respectively connected to the input end of the pyramid-like decomposition module. The first input end of the first feature image adjustment module and the first input end of the second feature image adjustment module, the output end of the pyramid-like decomposition module is connected to the input end of the probability pooling module, the The output end of the probability pooling module is connected to the input end of the channel attention module, and the output end of the channel attention module is connected to the second input end of the first feature image adjustment module. The first feature image adjustment module The output end of the second feature image adjustment module is connected to the second input end of the second feature image adjustment module. The output end of the second feature image adjustment module is connected to the input end of the image recognition module. The output end of the image recognition module outputs an ultrasound image. identification results. 3. The ultrasonic image recognition model according to claim 1, characterized in that, the feature image corresponding to each channel in the initial multi-channel feature image is pyramid decomposed to obtain a plurality of feature elements corresponding to each channel. Images, including: For the feature image corresponding to each channel, perform the following steps: Step i, use the characteristic image as an initial image, copy the initial image to obtain an intermediate image; Step ii: intercept the outermost pixels of the intermediate image with a preset step size to obtain the final image; Step iii, if the scale of the final image is greater than or equal to the preset scale threshold, then use the final image as the initial image in step i, and return to step i; otherwise, use the initial image and the step i iii. All final images obtained are as multiple feature sub-images corresponding to each channel. 4. The ultrasonic image recognition model according to claim 3, characterized in that the calculation formula of the scale threshold is: E＝(min(M,N)-2)mod 2*T Among them _{, M} 5. The ultrasonic image recognition model according to claim 4, characterized in that the channel weight of each channel is calculated according to the initial weight coefficient, and the channel weights of all channels are subjected to one-dimensional convolution processing to obtain the The channel weight matrix of the initial multi-channel feature image includes: For each channel, add all the initial weight coefficients corresponding to the channel to obtain the channel weight; Use one-dimensional convolution to perform convolution processing on the channel weights of all channels to obtain the intermediate channel weight matrix; where, if the total number of channels is less than or equal to 128, the convolution kernel of the one-dimensional convolution is set to 5 , otherwise set the convolution kernel of the one-dimensional convolution to 7; The intermediate channel weight matrix is activated using an activation function to obtain the channel weight matrix. 6. An ultrasound image recognition method, characterized by comprising: Obtaining an ultrasound image training set; the ultrasound image training set includes multiple ultrasound images of known types; The multiple known types of ultrasound images are input into the ultrasound image recognition model one by one to obtain the recognition result of each ultrasound image; the ultrasound image recognition model is the ultrasound image according to any one of claims 1 to 5 identification model; According to the recognition result and the type of each ultrasound image, the loss of the ultrasound image recognition model is calculated, and the ultrasound image recognition model is backpropagated according to the loss until the ultrasound image recognition model is fitted, Obtain the trained ultrasound image recognition model; The ultrasound image to be recognized is input into the trained ultrasound image recognition model to obtain the recognition result of the ultrasound image to be recognized. 7. The ultrasonic image recognition method according to claim 6, characterized in that the plurality of known types of ultrasonic images are input into the ultrasonic image recognition model one by one to obtain the recognition result of each ultrasonic image, including : By calculation formula Obtain the output probability P _k of each type; where, P _k represents the output probability of type K, k = 1, 2,..., Q, Q represents the total number of types of ultrasound images in the ultrasound image training set, z _k represents the output value of the k-th node, j represents the j-th node, n represents the number of output nodes, and softmax(·) represents the activation function; The type with the highest output probability is used as the recognition result. 8. The ultrasonic image recognition method according to claim 7, wherein the loss of the ultrasonic image recognition model is calculated based on the recognition result and the type of each ultrasound image, and the loss is calculated based on the loss. The ultrasound image recognition model is backpropagated until the ultrasound image recognition model is fitted, and a trained ultrasound image recognition model is obtained, including: For each ultrasound image, calculate the formula FL＝-a _k (1-p _k ) ^γ log(p _k ) Obtain the loss FL of the ultrasound image recognition model; where a _k represents the weight of type k, p _k represents the output probability of type k, γ represents the focus parameter, γ∈[0,5]; The ultrasound image recognition model is back-propagated according to the loss FL until the loss value of the new ultrasound image recognition model obtained by back-propagation is less than the preset loss threshold, and the fitting of the new ultrasound image recognition model is determined to obtain The ultrasonic image recognition model after training; wherein, the momentum of back propagation is set to 0.9, the learning rate of back propagation is obtained using the thermal learning rate strategy, and the weight attenuation coefficient of L1 regularization of back propagation is set to 5e-4 .