WO2024051015A1

WO2024051015A1 - Image feature extraction and classification method based on muscle ultrasound

Info

Publication number: WO2024051015A1
Application number: PCT/CN2022/137851
Authority: WO
Inventors: 周永进; 邓妙琴
Original assignee: 深圳大学
Priority date: 2022-09-08
Filing date: 2022-12-09
Publication date: 2024-03-14
Also published as: CN116309250A

Abstract

An image feature extraction and classification method based on muscle ultrasound. The method comprises: taking a muscle ultrasound image as a training data set, and extracting a first image feature from the training data set (S100); performing feature selection and feature dimension reduction on the first image feature to obtain a second image feature (S200); training a classifier on the basis of the second image feature and performing model verification to obtain a classification model of the muscle ultrasound image (S300); and acquiring a muscle ultrasound image to be classified as a test data set, extracting a third image feature from the test data set, and inputting the third image feature into the classification model of the muscle ultrasound image for classification to obtain a classification result (S400). The method can achieve classification and evaluation of individual muscle ultrasound images, so that large-scale disease early analysis and risk assessment at an individual level are achieved.

Description

An image feature extraction and classification method based on muscle ultrasound

Technical field

The invention relates to the field of medical image recognition, and specifically relates to an image feature extraction and classification method based on muscle ultrasound.

Background technique

Alzheimer’s Disease (AD) is a neurodegenerative disease characterized by insidious onset and progressive impairment of behavioral and cognitive functions, which mainly occurs in people aged 65 and above. With the intensification of aging, the number of AD patients is increasing year by year. However, due to its insidious onset, high incidence, and complex pathophysiological changes, timely intervention in its early stages can help delay and control the development of the disease.

Neuroimaging testing is mainly based on structural and functional neuroimaging technologies such as CT, MRI, fMRI, and PET to non-invasively detect structural and functional changes in the brain in vitro. Voxel-based morphological measurements of CT and MRI can quantitatively calculate morphological changes such as thickness and volume of the cerebral cortex; fMRI and PET can provide anatomical and physiological information of the brain and detect areas of metabolic activity changes in the patient's brain and areas related to AD in the brain. of protein aggregates. Neuroimaging testing technology plays an important role in detecting early-stage AD patients to a certain extent, but the testing is expensive and costly and is not suitable for early large-scale screening.

Therefore, the existing technology still needs to be improved and developed.

Contents of the invention

The technical problem to be solved by the present invention is to provide an image feature extraction and classification method based on muscle ultrasound in view of the above-mentioned defects of the existing technology, aiming to solve the problem of expensive detection and high cost in the existing technology, which is not suitable for early stage Problems with mass screening.

The technical solutions adopted by the present invention to solve the technical problems are as follows:

In a first aspect, the present invention provides an image feature extraction and classification method based on muscle ultrasound, wherein the method includes:

Obtaining muscle ultrasound images, using the muscle ultrasound images as a training data set, and extracting first image features from the training data set;

Perform feature selection and feature dimensionality reduction on the first image features to obtain second image features;

Train a machine learning model and a deep learning model based on the second image features to obtain a classifier, and perform model verification on the classifier to obtain a classification model for muscle ultrasound images;

Obtain a muscle ultrasound image to be classified, use the muscle ultrasound image to be classified as a test data set, extract a third image feature from the test data set, and input the third image feature into the classification of the muscle ultrasound image The model performs classification and obtains classification results.

In one implementation, obtaining muscle ultrasound images and using the muscle ultrasound images as a training data set includes:

Set the detection mode of the ultrasound imaging system to the musculoskeletal detection mode;

Place the long axis of the ultrasonic probe parallel to the long axis of the muscle, and keep the ultrasonic probe at the first detection position by setting a mark;

Based on the detection mode, using real-time B-mode ultrasound imaging equipment to obtain the muscle ultrasound image of the first detection position while the subject is static;

The muscle ultrasound image of the first detection position is used as the training data set.

In one implementation, extracting the first image feature from the training data set includes:

Perform a normalized Raiden transformation on the training data set to obtain a Raiden transformation matrix;

Calculate the gradient of the Raiden transformation matrix and perform edge enhancement to obtain the Raiden transformation gradient matrix;

Perform binarization and clustering processing on the Leyden transform gradient matrix to obtain deep fascia feature points, muscle bundle feature points and superficial fascia feature points;

Perform precise division and inverse Leyden transformation on the deep fascia feature points, muscle bundle feature points and superficial fascia feature points to obtain muscle thickness, muscle fiber length and pennation angle features;

According to the muscle thickness, muscle fiber length and pennation angle characteristics, the muscle morphological characteristics are obtained;

Obtain the average frequency analysis feature of the training data set; wherein, the calculation formula of the average frequency analysis feature is

n, I, and f are the length, power, and frequency of the power density spectrum respectively;

The first image feature is obtained based on the muscle morphological feature and the average frequency analysis feature.

Based on pixel grayscale distribution calculation, extract first-order statistical features from the training data set; wherein the first-order statistical features include integrated optical density, mean, standard deviation, variance, skewness, kurtosis and energy;

Based on gray level co-occurrence matrix calculation, Haralick features are extracted from the training data set; wherein the Haralick features include contrast, correlation, energy, entropy, homogeneity and symmetry;

Based on grayscale run length matrix calculation, Galloway features are extracted from the training data set; wherein, the Galloway features include short run advantage, long run advantage, grayscale non-uniformity, long run non-uniformity, and run percentage;

Based on the comparison results between the central pixel of the local area of the image and its neighborhood, local binary pattern features are extracted from the training data set; wherein the local binary pattern features include energy and entropy, and the energy is LBP _energy =∑ _i f _i ² , the entropy is LBP _entropy =-∑ _i ^{fi 2} _log ₂ (f _i ), f _i represents the corresponding frequency of the i-th block in the local area of the image;

The image texture analysis feature is obtained according to the first-order statistical feature, the Haralick feature, the Galloway feature and the local binary pattern feature;

The first image feature is obtained based on the image texture analysis feature.

In one implementation, performing feature selection and feature dimensionality reduction on the first image features to obtain second image features includes:

Using the t-test method to perform feature selection on the first image feature to obtain the screened first image feature;

Principal component analysis, linear discriminant analysis, t-SNE, multidimensional scaling method, isometric mapping method, local linear embedding, Laplacian feature mapping, and nearest neighbor component analysis methods are used to conduct the filtered first image features. Feature dimensionality reduction is performed to obtain the second image feature.

In one implementation, training a machine learning model and a deep learning model based on the second image features to obtain a classifier includes:

Train the machine learning model and the deep learning model to be trained based on the second image features to obtain training results;

The training results are evaluated through a first performance index to obtain the classifier; wherein the first performance index includes accuracy, precision, recall, F1 score, and area under the receiver operating curve.

In one implementation, performing model verification on the classifier to obtain a classification model for muscle ultrasound images also includes:

Obtain a verification data set, input the verification data set into the classifier, use 10-fold cross-validation for model verification, and obtain verification accuracy;

The classifier is adjusted according to the verification accuracy to obtain a classification model of the muscle ultrasound image.

In a second aspect, embodiments of the present invention also provide an image feature extraction and classification device based on muscle ultrasound. The device includes:

A first image feature acquisition module, configured to acquire muscle ultrasound images, use the muscle ultrasound images as a training data set, and extract first image features from the training data set;

A second image feature acquisition module is used to perform feature selection and feature dimensionality reduction on the first image features to obtain second image features;

A classification model acquisition module, configured to train a machine learning model and a deep learning model based on the second image features to obtain a classifier, and perform model verification on the classifier to obtain a classification model for muscle ultrasound images;

A classification module, used to obtain muscle ultrasound images to be classified, use the muscle ultrasound images to be classified as a test data set, extract third image features from the test data set, and input the third image features into the Classify the muscle ultrasound images using a classification model to obtain the classification results.

In a third aspect, embodiments of the present invention further provide an intelligent terminal, wherein the intelligent terminal includes a memory, a processor, and muscle ultrasound-based image feature extraction stored in the memory and operable on the processor. and a classification program. When the processor executes the muscle ultrasound-based image feature extraction and classification program, it implements the steps of the muscle ultrasound-based image feature extraction and classification method as described in any one of the above.

In a fourth aspect, embodiments of the present invention further provide a computer-readable storage medium, wherein a muscle ultrasound-based image feature extraction and classification program is stored on the computer-readable storage medium, and the muscle ultrasound-based image feature extraction And when the classification program is executed by the processor, the steps of the muscle ultrasound-based image feature extraction and classification method as described in any of the above are implemented.

Beneficial effects: Compared with the existing technology, the present invention provides an image feature extraction and classification method based on muscle ultrasound. The present invention first obtains muscle ultrasound images, uses the muscle ultrasound images as a training data set, and utilizes widely used And the muscle ultrasound images obtained by low-cost B-mode ultrasound imaging equipment can significantly reduce the image acquisition cost. Then extract the first image features from the training data set, and obtain the second image features by performing feature selection and feature dimensionality reduction on the first image features to avoid possible redundancy, high linear correlation, and model overfitting between features. Then, a machine learning model and a deep learning model are trained based on the second image features to obtain several classifiers. Then, through model verification of the classifiers, the classifier with the best performance is evaluated as a classification model for muscle ultrasound images. Finally, the third image features are extracted from the muscle ultrasound images to be classified, and the third image features are input into the classification model of the muscle ultrasound images for classification, thereby achieving early risk assessment at the individual level.

Description of the drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only These are some embodiments recorded in the present invention. 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 flowchart of an image feature extraction and classification method based on muscle ultrasound provided by an embodiment of the present invention.

Figure 2 is an original ultrasound image of the gastrocnemius muscle provided by an embodiment of the present invention.

Figure 3 is an example diagram of gastrocnemius ultrasonic image feature detection provided by an embodiment of the present invention.

Figure 4 is a flow chart of morphological parameter extraction provided by an embodiment of the present invention.

Figure 5 is a schematic block diagram of an image feature extraction and classification device based on muscle ultrasound provided by an embodiment of the present invention.

Figure 6 is a functional block diagram of the internal structure of an intelligent terminal provided by an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solution and effect of the present invention clearer and clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described here are only used to explain the present invention and are not intended to limit the present invention.

As a common aging phenotype, sarcopenia is defined as the loss of muscle structure and function and is closely related to Alzheimer’s disease (AD), mild cognitive impairment, and cognitive decline. Furthermore, in many older adults, impaired motor function precedes and predicts cognitive decline, mild cognitive impairment, and AD.

Taking AD as an example, there are currently three main methods for clinical diagnosis: the first is neuropsychological testing based on multiple psychiatric and behavioral symptom rating scales, and the second is based on samples such as cerebrospinal fluid, blood and urine. Biochemical testing, and the third type is neuroimaging testing based on CT, MRI, PET and other technologies.

The method of neuropsychological testing refers to the use of multiple psychiatric and behavioral symptom rating scales to detect cognitive function decline and assess the degree of decline and dementia. The detection of cognitive function decline through testing can provide a more objective basis for screening and diagnosis. It is also helpful in the differential diagnosis of dementia. Commonly used test scales can be divided into the following categories according to clinical use and purpose of use: cognitive impairment screening scale, cognitive function assessment scale, daily living ability assessment scale, psychiatric behavioral symptom assessment scale, and overall function assessment. scales, dementia grading scales, and identification and exclusion diagnostic scales, such as the Mini-mental State Examination (MMSE), the most widely used cognitive screening scale at present, and the Montreal Cognitive Assessment Scale ( montreal cognitive assessment scale (MocA), clock drawing test (CDT), etc. Although there are currently a variety of scales specifically used for cognitive function and dementia detection, which are helpful in identifying early AD patients, most of them are for AD. The sensitivity and specificity of early diagnosis are not ideal, and the scale test is time-consuming and partially susceptible to interference from education, culture and other factors.

Biochemical detection of relevant biological markers mainly uses cerebrospinal fluid, peripheral blood and urine as sample sources, and is based on abnormal deposition of amyloid β-protein (Aβ), hyperphosphorylation of Tau protein, and immune-inflammatory response. , mitochondrial dysfunction, oxidative stress, etc. [5] are recognized molecular pathogenic mechanisms of AD at home and abroad, and potential biomarkers are detected. At present, the detection of AD-related biological markers is mainly based on cerebrospinal fluid examination. Although its core biomarkers T-tau, P-tau and _Aβ42 have shown high sensitivity and specificity in the identification of AD patients, cerebrospinal fluid collection It is an invasive sampling method that may cause discomfort or side effects to some subjects; and cerebrospinal fluid biomarkers have high variability and are easily affected by transportation, experimental methods, and reference values, and clinical limit values are difficult to determine; In addition, it is difficult to popularize it and is not suitable for large-scale early screening. In addition to cerebrospinal fluid testing, there are also biomarker tests such as peripheral blood and urine. Although the collection of these two samples is extremely simple, the blood components are complex and the content of blood proteins and compounds that interfere with the measurement is high, and the potential AD biomarkers exist in the blood at low concentrations and are affected by peripheral circulation; urine, as a potential source of brain disease biomarkers, can accumulate a variety of disease-related metabolic changes and is non-invasive, convenient and accessible. It has advantages such as repetition, but the urine components metabolize quickly and have low stability, and the related detection technology requires high technical personnel and equipment, which limits its clinical promotion and application.

Neuroimaging testing is mainly based on structural and functional neuroimaging technologies such as CT, MRI, fMRI, and PET to non-invasively detect structural and functional changes in the brain in vitro. Voxel-based morphological measurements of CT and MRI can quantitatively calculate morphological changes such as thickness and volume of the cerebral cortex; fMRI and PET can provide anatomical and physiological information of the brain and detect areas of metabolic activity changes in the patient's brain and areas related to AD in the brain. of protein aggregates. Neuroimaging testing technology plays an important role in the early diagnosis and differential diagnosis of AD patients to a certain extent. However, the testing is expensive and costly and is not suitable for early large-scale screening.

Therefore, in order to solve the above problems, this embodiment provides an image feature extraction and classification method based on muscle ultrasound. First, a muscle ultrasound image is obtained, and the muscle ultrasound image is used as a training data set, using the widely used and low-cost B-type Muscle ultrasound images acquired by ultrasound imaging equipment are used to assess individual risks, enabling non-invasive, convenient, and low-cost early large-scale screening and risk assessment of diseases. Then extract the first image features from the training data set, and obtain the second image features by performing feature selection and feature dimensionality reduction on the first image features to avoid possible redundancy, high linear correlation, and model overfitting between features. Then, a machine learning model and a deep learning model are trained based on the second image features to obtain several classifiers. Then, through model verification of the classifiers, the classifier with the best performance is evaluated as a classification model for muscle ultrasound images. Finally, the third image feature is extracted from the muscle ultrasound image to be classified, and the third image feature is input into the classification model of the muscle ultrasound image for classification, so as to obtain the classification result through a non-invasive and convenient muscle ultrasound image classification method to achieve Early risk assessment at the individual level.

Example methods

This embodiment provides an image feature extraction and classification method based on muscle ultrasound. During specific implementation, as shown in Figure 1, the method includes the following steps:

Step S100: Obtain a muscle ultrasound image, use the muscle ultrasound image as a training data set, and extract the first image feature from the training data set;

In one implementation, step S100 specifically includes the following steps:

Step S101: Set the detection mode of the ultrasound imaging system to the musculoskeletal detection mode;

Step S102: Place the long axis of the ultrasonic probe parallel to the long axis of the muscle, and keep the ultrasonic probe at the first detection position by setting a mark;

Step S103. Based on the detection mode, use real-time B-mode ultrasound imaging equipment to obtain the muscle ultrasound image of the first detection position while the subject is static;

Step S104: Use the muscle ultrasound image of the first detection position as the training data set.

B-mode ultrasound imaging equipment, as the most widely used and simplest ultrasound equipment in clinical practice, has the potential to detect the fine structure of muscles and allows the visualization and quantification of muscle structure. Using muscle ultrasound images acquired by widely used and low-cost B-mode ultrasound imaging equipment to assess individual risks can achieve non-invasive, convenient, and low-cost early large-scale screening and risk assessment of diseases.

Specifically, while the subject is static, real-time B-mode ultrasound imaging equipment is used to obtain muscle ultrasound images. The ultrasound imaging system selects the muscle-bone detection mode. The long axis of the ultrasound probe should be parallel to the long axis of the muscle and placed in the muscle belly. or other specific positions, that is, the first detection position in this embodiment; apply an appropriate amount of ultrasound gel coupling agent to ensure the acoustic coupling between the probe and the skin; the ultrasound probe can be adjusted to optimize the contrast of the muscle bundles in the ultrasound image , and locations can be marked to ensure the probe is placed in the same location every time. An example of an ultrasound image of the gastrocnemius muscle is shown in Figure 2.

It should be noted that in terms of muscle ultrasound image acquisition, in addition to collecting static muscle ultrasound images of the subject, it can also collect muscle ultrasound images during the dynamic structural changes caused by the subject's muscle stretching; in addition to the B-type acquisition equipment In addition to ultrasound equipment, shear wave elastography equipment or other ultrasound imaging methods can also be used for acquisition.

Step S105: Perform normalized Redon transformation on the training data set to obtain a Redon transformation matrix;

Step S106: Calculate the gradient of the Redon transformation matrix and perform edge enhancement to obtain the Redon transformation gradient matrix;

Step S107: Perform binarization and clustering processing on the Leyden transform gradient matrix to obtain deep fascia feature points, muscle bundle feature points and superficial fascia feature points;

Step S108: Perform precise division and inverse Leiden transformation on the deep fascia feature points, muscle bundle feature points and superficial fascia feature points to obtain muscle thickness, muscle fiber length and pennation angle features;

Step S109: Obtain the muscle morphological characteristics according to the muscle thickness, muscle fiber length and pennation angle characteristics;

Specifically, the morphological characteristics of muscles include muscle thickness, muscle fiber length, pennation angle, etc. This embodiment uses a feature detection method based on the Redden transform gradient matrix to estimate the morphological parameters of the muscle. As shown in Figure 3, the muscle layer of the gastrocnemius contains superficial fascia, muscle fascia area, and deep fascia. According to the arrangement direction of muscle fibers and the direction of deep fascia, muscle fascia line L1 and deep fascia line L2 can be drawn. Perform the normalized Leiden transformation on the image, find the gradient of the transformation matrix, highlight the edge feature points of the muscle structure elements, and then accurately divide the deep and superficial fascial edge feature points and muscle bundle edge feature points, and finally classify the feature points. The inverse Leiden transform realizes the precise positioning of muscle structural elements, thereby calculating the above-mentioned morphological parameters. The specific flow chart of muscle morphological feature extraction is shown in Figure 4.

Step S110: Obtain the average frequency analysis feature of the training data set; wherein the calculation formula of the average frequency analysis feature is

Step S111: Obtain the first image feature based on the muscle morphological feature and the average frequency analysis feature.

Specifically, the mean frequency analysis feature (MFAF), as an effective parameter related to muscle mass and expected to describe differences in skeletal muscle structure, will not be significantly affected by the configuration of different ultrasound equipment. The calculation formula is shown below:

Among them, n, I, and f are the length, power and frequency of the power density spectrum respectively.

Step S112: Extract first-order statistical features from the training data set based on pixel grayscale distribution calculation; wherein the first-order statistical features include integrated optical density, mean, standard deviation, variance, skewness, and kurtosis. and energy;

Step S113: Extract Haralick features from the training data set based on gray level co-occurrence matrix calculation; wherein the Haralick features include contrast, correlation, energy, entropy, homogeneity and symmetry;

Step S114: Extract Galloway features from the training data set based on grayscale run length matrix calculation; wherein, the Galloway features include short run advantage, long run advantage, grayscale non-uniformity, long run non-uniformity, and run percentage. ;

Step S115: Extract local binary pattern features from the training data set based on the comparison results between the central pixel of the local area of the image and its neighborhood; wherein the local binary pattern features include energy and entropy, and the energy is LBP _energy =∑ _i _fi ² , the entropy is LBP _entropy =-∑ _i _fi ² log ₂ ( _fi ), _fi represents the corresponding frequency of the i-th block in the local area of the image;

Step S116: Obtain the image texture analysis feature according to the first-order statistical feature, the Haralick feature, the Galloway feature and the local binary pattern feature;

Step S117: Obtain the first image feature based on the image texture analysis feature.

Image texture analysis features mainly include first-order statistical features and high-order texture features. On the one hand, first-order statistical features can effectively and quantitatively describe the ultrasound echo intensity of skeletal muscles; in addition, there are differences in the ultrasound echo intensity information of skeletal muscles of different ages or groups, and these features can also provide some muscle status-related information. Structural information, thereby providing effective information for muscle damage assessment. On the other hand, high-order texture features, such as Haralick features, Galloway features and Local Binary Pattern (LBP) features, can perform better in fine tasks such as muscle gender recognition than first-order statistical features. better.

Specifically, first-order statistical features are features calculated directly based on the pixel grayscale distribution of the original image, including features such as integrated optical density, mean, standard deviation, variance, skewness, kurtosis, and energy. The Haralick feature is calculated from the gray level co-occurrence matrix. The gray level co-occurrence matrix is a matrix function of the pixel distance and direction. Calculating the correlation between the gray level values of two points at a given spatial distance d and direction θ is a method through A common method to describe image texture by studying the spatial correlation characteristics of grayscale. Typical Haralick features include contrast, correlation, energy, entropy, homogeneity and symmetry, etc. Typically, each Haralick feature contains four directions: 0°, 45°, 90° and 135°. The Galloway feature is calculated based on the grayscale run length matrix. The grayscale run length matrix represents the regularity of texture changes of an image. Its size is determined by the gray level and image size of the image, including short run length advantages and long run length advantages. , grayscale non-uniformity, long run non-uniformity, run percentage and other texture statistical features, and each Galloway feature also contains four directions of 0°, 45°, 90° and 135°. LBP features are obtained by comparing the central pixel of a local area of the image with its neighborhood. It mainly describes the local texture features of the image. It has significant advantages such as rotation invariance and grayscale invariance, including energy and entropy. , the specific calculation formula is shown below:

LBP _energy =∑ _i f _i ² ,

LBP _entropy =-∑ _i f _i ² log ₂ (f _i ),

where _fi represents the corresponding frequency of the i-th block in the local area of the image.

It should be noted that when extracting muscle morphological features from muscle ultrasound images, such as pennation angle, muscle thickness, muscle fiber angle, muscle bundle length, muscle physiological cross-sectional area, etc., in addition to using features based on the Redden transform gradient matrix In addition to detection methods, methods based on Hough transform, deep learning, etc. can also be used to locate various structural elements of muscle tissue, thereby achieving automatic measurement of morphological parameters.

Step S200: Perform feature selection and feature dimensionality reduction on the first image features to obtain second image features.

Specifically, with the development of the fields of machine learning and pattern recognition, the number of features has increased dramatically. Especially in medical image analysis, traditional algorithms often encounter the curse of dimensionality, and there may be redundancy and high linear correlation between features, which can easily lead to The model is overfitting. Therefore, it is necessary to select and reduce the dimensionality of all features, screen out a set of related feature subsets that are most sensitive to the disease, and reduce the dimensionality of the data, which can speed up the learning process, effectively improve the efficiency of data analysis, and improve Classifier model performance.

In one implementation, step S200 specifically includes the following steps:

Step S201: Use the t-test method to perform feature selection on the first image features to obtain the filtered first image features;

Step S202: Use principal component analysis, linear discriminant analysis, t-SNE, multidimensional scaling method, isometric mapping method, local linear embedding, Laplacian eigenmap, and nearest neighbor component analysis methods to analyze the filtered first The image features undergo feature dimensionality reduction to obtain the second image features.

Specifically, for the feature selection part, this embodiment uses the t-test method in the Filter method to measure their discriminability by analyzing each feature, and then selects the most discriminative feature subset through sorting. Use principal component analysis, linear discriminant analysis, t-SNE, multidimensional scaling method, isometric mapping method, local linear embedding, Laplacian feature mapping, nearest neighbor component analysis and other methods to select features of the first image The final features are dimensionally reduced to obtain the second image features.

Step S300: Train a machine learning model and a deep learning model based on the second image features to obtain a classifier, and perform model verification on the classifier to obtain a classification model for muscle ultrasound images;

Specifically, machine learning algorithms can be divided into supervised learning, semi-supervised learning, unsupervised learning and reinforcement learning based on learning methods. Currently commonly used machine learning algorithms include support vector machines, decision trees, random forests, logistic regression, Gaussian Bayesian classifiers, Bernoulli Bayesian classifiers, artificial neural networks, etc. With the development of big data and deep learning, algorithms based on deep learning can overcome the limitations of traditional shallow machine learning in solving complex tasks and improve discrimination capabilities. By training a machine learning model and a deep learning model based on the second image features, a classifier can be obtained, and then the classifier can be model verified, and finally a classification model of muscle ultrasound images can be obtained.

In one implementation, step S300 specifically includes the following steps:

Step S301: Train the machine learning model and deep learning model to be trained based on the second image features to obtain training results;

Step S302: Evaluate the training results through a first performance index to obtain the classifier; wherein the first performance index includes accuracy, precision, recall, F1 score, and area under the receiver operating curve.

Step S303: Obtain a verification data set, input the verification data set into the classifier, and use 10-fold cross-validation to perform model verification to obtain verification accuracy;

Step S304: Adjust the classifier according to the verification accuracy to obtain a classification model of the muscle ultrasound image.

Specifically, in this embodiment, the current mainstream and commonly used machine learning and deep learning models are used for training, and the performance is evaluated through performance indicators such as accuracy, precision, recall, F1 score, and area under the subject operating curve. optimal classifier.

Specifically, cross-validation is a statistical validation technique for evaluating and comparing model performance. It uses a subset of the data set, trains it, and then evaluates the performance of the model using a complementary subset of the data set that was not used for training. It can guarantee that the model correctly captures patterns from the data, regardless of interference from the data. . In this embodiment, the muscle ultrasound image to be classified is used as a test data set, and the third image feature is input into the classification model of the muscle ultrasound image for classification, and a classification result is obtained. The most classic method in cross-validation is k-fold cross-validation, which divides the data set into k subsets, makes a test set for each subset, and uses the rest as a training set. Then the cross-validation is repeated k times, and the average value is calculated as the score. This embodiment adopts 10-fold cross-validation, which is currently most commonly used in the field of machine learning when establishing models and verifying model parameters, to obtain verification accuracy; and then adjusts the classifier according to the verification accuracy to obtain a classification model for muscle ultrasound images.

Step S400: Obtain the muscle ultrasound image to be classified, use the muscle ultrasound image to be classified as a test data set, extract third image features from the test data set, and input the third image feature into the muscle ultrasound image. The image classification model performs classification and obtains the classification result.

Specifically, the third image feature of the muscle ultrasound image to be classified is extracted, and the third image feature is input into the trained muscle ultrasound image classification model to perform classification, thereby obtaining the prediction result or category, thereby achieving early individual-level risk assessment.

Exemplary device

This embodiment also provides an image feature extraction and classification device based on muscle ultrasound. The device includes:

The first image feature acquisition module 10 is used to acquire muscle ultrasound images, use the muscle ultrasound images as a training data set, and extract the first image features from the training data set;

The second image feature acquisition module 20 is used to perform feature selection and feature dimensionality reduction on the first image features to obtain second image features;

The classification model acquisition module 30 is used to train a machine learning model and a deep learning model based on the second image features to obtain a classifier, and perform model verification on the classifier to obtain a classification model for muscle ultrasound images;

The classification module 40 is used to obtain muscle ultrasound images to be classified, use the muscle ultrasound images to be classified as a test data set, extract third image features from the test data set, and input the third image features into the test data set. Classify the muscle ultrasound images using the classification model described above to obtain the classification results.

In one implementation, the first image feature acquisition module 10 includes:

a detection mode setting unit, used to set the detection mode of the ultrasonic imaging system to the musculoskeletal detection mode;

An ultrasonic probe placement unit is used to place the long axis of the ultrasonic probe parallel to the long axis direction of the muscle, and maintain the ultrasonic probe at the first detection position by setting a mark;

A muscle ultrasound image acquisition unit, configured to acquire the muscle ultrasound image of the first detection position using a real-time B-mode ultrasound imaging device when the subject is static based on the detection mode;

A training data set acquisition unit is configured to use the muscle ultrasound image of the first detection position as the training data set.

A Raiden transformation matrix acquisition unit, used to perform a normalized Raiden transformation on the training data set to obtain a Raiden transformation matrix;

A Redden transform gradient matrix acquisition unit, used to obtain the gradient of the Redden transform matrix and perform edge enhancement to obtain the Redden transform gradient matrix;

A clustering unit is used to perform binarization and clustering processing on the Leyden transform gradient matrix to obtain deep fascia feature points, muscle bundle feature points and superficial fascia feature points;

An inverse Raiden transform unit is used to accurately divide and inverse Raiden transform the deep fascia feature points, muscle bundle feature points and superficial fascia feature points to obtain muscle thickness, muscle fiber length and pennation angle features;

A muscle morphological feature acquisition unit is used to obtain the muscle morphological features based on the muscle thickness, muscle fiber length and pennation angle features;

An average frequency analysis feature acquisition unit is used to obtain the average frequency analysis feature of the training data set; wherein the calculation formula of the average frequency analysis feature is:

A first-order statistical feature extraction unit is used to extract first-order statistical features from the training data set based on pixel grayscale distribution calculation; wherein the first-order statistical features include integrated optical density, average value, standard deviation, Variance, skewness, kurtosis, and energy;

A Haralick feature extraction unit is used to extract Haralick features from the training data set based on gray level co-occurrence matrix calculation; wherein the Haralick features include contrast, correlation, energy, entropy, homogeneity and symmetry;

The Galloway feature extraction unit is used to extract Galloway features from the training data set based on grayscale run length matrix calculation; wherein the Galloway features include short run advantages, long run advantages, grayscale non-uniformity, and long run non-uniformity. sex, run percentage;

A local binary pattern feature extraction unit, configured to extract local binary pattern features from the training data set based on a comparison result between the central pixel of the local area of the image and its neighborhood; wherein the local binary pattern features include energy and entropy , the energy is LBP _energy =∑ _i _fi ² , the entropy is LBP _entropy =-∑ _i _fi ² log ₂ ( _fi ), _fi represents the corresponding frequency of the i-th block in the local area of the image;

An image texture analysis feature acquisition unit, configured to obtain the image texture analysis feature based on the first-order statistical features, the Haralick features, the Galloway features and the local binary pattern features;

A first image feature acquisition unit is configured to obtain the first image feature based on the muscle morphology feature, the average frequency analysis feature and the image texture analysis feature.

In one implementation, the second image feature acquisition module 20 includes:

A feature screening unit, used to perform feature selection on the first image feature using the t-test method to obtain the screened first image feature;

Feature dimensionality reduction unit, used to use principal component analysis, linear discriminant analysis, t-SNE, multidimensional scaling method, isometric mapping method, local linear embedding, Laplacian eigenmap, nearest neighbor component analysis method to filter the Perform feature dimensionality reduction on the passed first image features to obtain the second image features.

In one implementation, the classification model acquisition module 30 includes:

A model training unit, configured to train the machine learning model and the deep learning model to be trained based on the second image features, and obtain training results;

A classifier acquisition unit, used to evaluate the training results through a first performance indicator to obtain the classifier; wherein the first performance indicator includes accuracy, precision, recall, F1 score, and receiver operating curve lower area.

A verification accuracy acquisition unit is used to obtain a verification data set, input the verification data set into the classifier, and use 10-fold cross-validation to perform model verification to obtain verification accuracy;

A classification model acquisition unit is used to adjust the classifier according to the verification accuracy to obtain a classification model of the muscle ultrasound image.

Based on the above embodiments, the present invention also provides an intelligent terminal, the functional block diagram of which can be shown in Figure 5 . The intelligent terminal includes a processor, memory, network interface, display screen, and temperature sensor connected through a system bus. Among them, the processor of the smart terminal is used to provide computing and control capabilities. The memory of the smart terminal includes non-volatile storage media and internal memory. The non-volatile storage medium stores operating systems and computer programs. This internal memory provides an environment for the execution of operating systems and computer programs in non-volatile storage media. The network interface of the smart terminal is used to communicate with external terminals through network connections. When the computer program is executed by the processor, it implements an image feature extraction and classification method based on muscle ultrasound. The display screen of the smart terminal can be a liquid crystal display or an electronic ink display. The temperature sensor of the smart terminal is pre-set inside the smart terminal for detecting the operating temperature of the internal device.

Those skilled in the art can understand that the principle block diagram shown in Figure 5 is only a block diagram of a partial structure related to the solution of the present invention, and does not constitute a limitation on the smart terminal to which the solution of the present invention is applied. Specific smart terminals may include more or fewer components than shown, or combine certain components, or have a different arrangement of components.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage. In the media, when executed, the computer program may include the processes of the above method embodiments. Any reference to memory, storage, database or other media used in the various embodiments provided by the present invention may include non-volatile and/or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Synchlink DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

In summary, the present invention discloses an image feature extraction and classification method based on muscle ultrasound. The method includes acquiring muscle ultrasound images and using the muscle ultrasound images as a training data set; extracting the first image feature from the training data set; Perform feature selection and feature dimensionality reduction on the first image feature to obtain the second image feature; train a classifier based on the second image feature and perform model verification to obtain a classification model for muscle ultrasound images; obtain the muscle ultrasound image to be classified as a test data set, The third image feature is extracted from the test data set, and the third image feature is input into the classification model of the muscle ultrasound image for classification, and the classification result is obtained. The present invention can collect low-cost muscle ultrasound images, collect a wider range of muscle structure and functional indicators, and use radiomics combined with artificial intelligence methods such as machine learning and deep learning to build a classification model, thereby realizing the classification and classification of individual muscle ultrasound images. Evaluate.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be used Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent substitutions are made to some of the technical features; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

An image feature extraction and classification method based on muscle ultrasound, characterized in that the method includes:

Obtaining muscle ultrasound images, using the muscle ultrasound images as a training data set, and extracting first image features from the training data set;

Perform feature selection and feature dimensionality reduction on the first image features to obtain second image features;

Train a machine learning model and a deep learning model based on the second image features to obtain a classifier, and perform model verification on the classifier to obtain a classification model for muscle ultrasound images;

Obtain a muscle ultrasound image to be classified, use the muscle ultrasound image to be classified as a test data set, extract a third image feature from the test data set, and input the third image feature into the classification of the muscle ultrasound image The model performs classification and obtains classification results.
The image feature extraction and classification method based on muscle ultrasound according to claim 1, characterized in that the acquisition of muscle ultrasound images and using the muscle ultrasound images as a training data set include:

Set the detection mode of the ultrasound imaging system to the musculoskeletal detection mode;

Place the long axis of the ultrasonic probe parallel to the long axis of the muscle, and keep the ultrasonic probe at the first detection position by setting a mark;

Based on the detection mode, using real-time B-mode ultrasound imaging equipment to obtain the muscle ultrasound image of the first detection position while the subject is static;

The muscle ultrasound image of the first detection position is used as the training data set.
The image feature extraction and classification method based on muscle ultrasound according to claim 1, wherein the extracting the first image feature from the training data set includes:

Perform a normalized Raiden transformation on the training data set to obtain a Raiden transformation matrix;

Calculate the gradient of the Raiden transformation matrix and perform edge enhancement to obtain the Raiden transformation gradient matrix;

Perform binarization and clustering processing on the Leyden transform gradient matrix to obtain deep fascia feature points, muscle bundle feature points and superficial fascia feature points;

Perform precise division and inverse Leyden transformation on the deep fascia feature points, muscle bundle feature points and superficial fascia feature points to obtain muscle thickness, muscle fiber length and pennation angle features;

According to the muscle thickness, muscle fiber length and pennation angle characteristics, the muscle morphological characteristics are obtained;

Obtain the average frequency analysis feature of the training data set; wherein, the calculation formula of the average frequency analysis feature is
n, I, and f are the length, power, and frequency of the power density spectrum respectively;

The first image feature is obtained based on the muscle morphological feature and the average frequency analysis feature.
The image feature extraction and classification method based on muscle ultrasound according to claim 1, wherein the extracting the first image feature from the training data set includes:

Based on pixel grayscale distribution calculation, extract first-order statistical features from the training data set; wherein the first-order statistical features include integrated optical density, mean, standard deviation, variance, skewness, kurtosis and energy;

Based on gray level co-occurrence matrix calculation, Haralick features are extracted from the training data set; wherein the Haralick features include contrast, correlation, energy, entropy, homogeneity and symmetry;

Based on grayscale run length matrix calculation, Galloway features are extracted from the training data set; wherein, the Galloway features include short run advantage, long run advantage, grayscale non-uniformity, long run non-uniformity, and run percentage;

Based on the comparison results between the central pixel of the local area of the image and its neighborhood, local binary pattern features are extracted from the training data set; wherein the local binary pattern features include energy and entropy, and the energy is LBP energy =∑ i f i 2 , the entropy is LBP entropy =-∑ i fi 2 log 2 (f i ), f i represents the corresponding frequency of the i-th block in the local area of the image;

The image texture analysis feature is obtained according to the first-order statistical feature, the Haralick feature, the Galloway feature and the local binary pattern feature;

The first image feature is obtained based on the image texture analysis feature.
The image feature extraction and classification method based on muscle ultrasound according to claim 1, wherein the second image feature is obtained by performing feature selection and feature dimensionality reduction on the first image feature, including:

Using the t-test method to perform feature selection on the first image feature to obtain the screened first image feature;

Principal component analysis, linear discriminant analysis, t-SNE, multidimensional scaling method, isometric mapping method, local linear embedding, Laplacian feature mapping, and nearest neighbor component analysis methods are used to conduct the filtered first image features. Feature dimensionality reduction is performed to obtain the second image feature.
The image feature extraction and classification method based on muscle ultrasound according to claim 1, characterized in that the training of a machine learning model and a deep learning model based on the second image features to obtain a classifier includes:

Train the machine learning model and the deep learning model to be trained based on the second image features to obtain training results;

The training results are evaluated through a first performance index to obtain the classifier; wherein the first performance index includes accuracy, precision, recall, F1 score, and area under the receiver operating curve.
The image feature extraction and classification method based on muscle ultrasound according to claim 1, characterized in that, performing model verification on the classifier to obtain a classification model of muscle ultrasound images also includes:

Obtain a verification data set, input the verification data set into the classifier, use 10-fold cross-validation for model verification, and obtain verification accuracy;

The classifier is adjusted according to the verification accuracy to obtain a classification model of the muscle ultrasound image.
An image feature extraction and classification device based on muscle ultrasound, characterized in that the device includes:

A first image feature acquisition module, configured to acquire muscle ultrasound images, use the muscle ultrasound images as a training data set, and extract first image features from the training data set;

A second image feature acquisition module is used to perform feature selection and feature dimensionality reduction on the first image features to obtain second image features;

A classification model acquisition module, configured to train a machine learning model and a deep learning model based on the second image features to obtain a classifier, and perform model verification on the classifier to obtain a classification model for muscle ultrasound images;

A classification module, used to obtain muscle ultrasound images to be classified, use the muscle ultrasound images to be classified as a test data set, extract third image features from the test data set, and input the third image features into the Classify the muscle ultrasound images using a classification model to obtain the classification results.
An intelligent terminal, characterized in that the intelligent terminal includes a memory, a processor, and an image feature extraction and classification program based on muscle ultrasound that is stored in the memory and can be run on the processor, and the processor When the muscle ultrasound-based image feature extraction and classification program is executed, the steps of the muscle ultrasound-based image feature extraction and classification method as described in any one of claims 1 to 7 are implemented.
A computer-readable storage medium, characterized in that a muscle ultrasound-based image feature extraction and classification program is stored on the computer-readable storage medium, and when the muscle ultrasound-based image feature extraction and classification program is executed by a processor , the steps of implementing the image feature extraction and classification method based on muscle ultrasound as described in any one of claims 1-7.