Feature extraction and recognition method for minimal lesion in medical image

Abstract

The invention discloses a feature extraction and recognition method for minimal lesions in a medical image, and relates to the technical field of medical image processing, and the method comprises the specific steps: carrying out the preprocessing of an input medical image, the preprocessing comprises the denoising and contrast enhancement, and obtaining a preprocessed image; according to the method, comprehensive characterization of minimal lesion features is realized through a micro-texture and macro-morphology combined feature extraction strategy, micro-texture features can deeply reflect fine structure information such as cells and blood vessels in the lesion, and the micro-texture features and the micro-texture features can be combined with the micro-texture and macro-morphology combined feature extraction strategy, so that the micro-texture and macro-morphology combined feature extraction strategy is realized. The method is advantaged in that lesion pathological property determination is facilitated, the position, the range and the general shape of the lesion are determined on the whole through the macroscopic morphological characteristics, lesion positioning and overall evaluation are facilitated, one-sidedness of a single characteristic is avoided through organic combination of the macroscopic morphological characteristics and the macroscopic morphological characteristics, and the identification result is more accurate and reliable.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method for extracting and identifying features of microlesions in medical images. Background Art

[0002] In the field of modern medicine, medical imaging technologies such as CT, MRI, and PET have become important means of disease diagnosis. These technologies can obtain detailed image information of the internal structure and tissue of the human body, providing doctors with intuitive visual basis for detecting lesions. With the continuous development of medical technology, the demand for early detection and accurate diagnosis of tiny lesions is increasing. Tiny lesions are often the early manifestation of the disease. Timely and accurate identification of them is extremely critical for the effective treatment of the disease and the improvement of the patient's prognosis.

[0003] However, in terms of feature extraction and identification of tiny lesions in medical images, traditional methods focus on a single type of feature. For technologies that rely only on microscopic texture features, it is difficult to grasp the macroscopic morphological information of the shape, size, and position of the lesions as a whole, resulting in a lack of accuracy in lesion localization and overall assessment. Methods that focus on extracting macroscopic morphological features ignore the differences in the microscopic structure of the lesions, and cannot effectively distinguish tiny lesions with similar macroscopic morphology but different pathological properties. It is difficult to fully reflect the overall characteristics of tiny lesions, which can easily lead to missed diagnosis and misdiagnosis.

[0004] In summary, the existing medical image micro-lesion feature extraction and recognition technology is difficult to meet the needs of accurate and comprehensive diagnosis of micro-lesions in clinical practice. Therefore, there is an urgent need for an innovative method that can comprehensively consider micro-texture and macro-morphological features and integrate multimodal medical image information to improve the accuracy and reliability of micro-lesion feature extraction and recognition. Summary of the invention

[0005] The purpose of the present invention is to make up for the shortcomings of the prior art and provide a method for feature extraction and identification of tiny lesions in medical images. By combining microscopic texture features and macroscopic morphological features, the accuracy of lesion identification is significantly improved. Through multi-level feature extraction technology, this method can capture detailed information of lesions from different scales, making the feature representation more comprehensive and accurate. The design of the feature fusion layer enables the deep combination of microscopic and macroscopic features, thereby achieving optimal weight distribution in the feature representation and enhancing the model's ability to capture lesion features.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for extracting and identifying features of microlesions in medical images, the specific steps of the method are: S1, image preprocessing: input medical images Perform preprocessing, the preprocessing including denoising and contrast enhancement, to obtain a preprocessed image ,in is the image pixel coordinate; S2, Micro texture feature extraction: The preprocessed medical image is analyzed by wavelet transform-based texture analysis. Processing, through multi-scale wavelet decomposition, decompose the image into low-frequency and high-frequency sub-bands, through the wavelet function The wavelet coefficients in the sub-band are extracted to reflect the microscopic texture characteristics of the small lesions, and the statistics of the high-frequency sub-band energy and entropy are calculated as the descriptors of the microscopic texture characteristics, that is, through the wavelet transform function , decomposition scale , perform Wavelet coefficients are obtained by wavelet decomposition of the scale , and calculate the energy of the high frequency sub-band and entropy ,in is the wavelet coefficient The probability distribution function of Represents the high frequency sub-band region; The gray level co-occurrence matrix GLCM is used to further extract micro texture features and calculate the and distance The pixel gray level co-occurrence matrix is ​​used to obtain the texture feature parameters such as contrast, correlation, energy and entropy based on GLCM, that is, For direction and distance Grayscale on the image, calculate the contrast , correlation ,energy ,entropy ,in Respectively about The mean and standard deviation in the direction combine the micro-texture features into a micro-texture feature vector ; S3, macroscopic morphological feature extraction: using morphological processing technology, dilate the medical image to extract the macroscopic morphological features of the edge shape, size and contour of the tiny lesions, and use the edge perimeter, area, and circularity geometric parameters as macroscopic morphological feature descriptors, that is, using the morphological dilation function right Processing, define the structural element as , then the expanded image is , and based on the macroscopic shape after expansion, the edge detection algorithm Obtain the edge contour of the lesion , calculate the perimeter of the edge ,area ,in The edge contour Enclosed area, circularity ; Perform region growing algorithm on the lesion area Processing, determine the scope of the lesion, and calculate its volume and center of gravity position characteristics to further enrich the macroscopic morphological feature information, that is, set the seed point to The growth threshold is , the lesion area after region growing is , calculate its volume , center of gravity position ,in , combining the macroscopic morphological features into a macroscopic morphological feature vector ; S4, micro-macro feature fusion: The extracted micro texture features and macroscopic morphological characteristics The input is sent to the fusion layer, and a weighted fusion method is used to assign corresponding weights to the importance of the vectors for micro-lesion recognition according to the different characteristics of micro-texture and macro-morphology. and , and perform weighted summation of the micro texture feature vector and the macro morphology feature vector to obtain the fused feature vector ; The fused feature vector is analyzed by using the principal component analysis (PCA) dimensionality reduction technique. Perform dimensionality reduction processing to remove redundant information and retain representative feature information; S5, microlesion recognition: The fusion feature vector after dimension reduction is Input to the trained classifier In the classifier According to the characteristic information of the feature vector, the microlesions are classified and identified to obtain the identification results .

[0007] Furthermore, the S3 region growing algorithm Middle seed point The selection method is: perform grayscale threshold segmentation on the image and set the threshold to , get the binary image , in the binary image Select the central pixel with the largest connected area as the seed point .

[0008] Furthermore, the growth threshold in S3 , determined based on the grayscale of the image and the grayscale changes within the microlesion, that is, using the grayscale standard deviation of the image , estimated value of grayscale variation inside small lesions , then the growth threshold ,in is the average gray value of the image, and is the empirical coefficient.

[0009] Furthermore, the S4 converts the micro texture feature vector and the macroscopic morphological feature vector Input to the fusion layer and set the weight of the micro texture feature vector to , the weight of the macroscopic morphological feature vector is ,and , the fused feature vector , using principal component analysis The fused feature vector Perform dimensionality reduction processing, that is The transformation matrix is , the feature vector after dimensionality reduction .

[0010] Furthermore, the S4 uses principal component analysis The fused feature vector The process of dimensionality reduction is as follows: The fused feature vector The sample set is split into ,in is the number of samples, for the fused feature vector , whose dimensions for , the covariance matrix Elements for: ,in , Indicates The first fusion feature vector elements, is the first of all samples The mean of the features, ; Covariance matrix Perform eigenvalue decomposition and find a new set of coordinate axes represented by eigenvectors so that the data variance represented by the eigenvalues ​​on the coordinate axes is maximized, thereby obtaining the eigenvalues and the corresponding eigenvector ; Set the reduced dimension to , the latitude is based on the ratio of the sum of the retained eigenvalues ​​to the sum of the total eigenvalues ,Right now , select the eigenvalue feature vectors; Before The eigenvectors corresponding to the largest eigenvalues ​​form the transformation matrix , the feature vector after dimensionality reduction pass Get the feature vector after fusion Projected onto In the new subspace determined, dimensionality reduction is achieved.

[0011] Furthermore, the S5 classifier A neural network structure classifier is used, wherein the neural network structure includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is equal to the feature vector after dimensionality reduction. The hidden layer adopts a multi-layer structure, and the number of nodes in each layer depends on the number of training samples. Number of categories with minimal lesions To determine, specifically the number of hidden layer nodes , the number of nodes in the output layer and the number of categories of minimal lesions The neural network is equal to each other, so that the neural network can output the results corresponding to different categories of microlesions. The weights and thresholds of the neural network are trained through the back propagation algorithm, and the loss function is set. ,in Representative The true category labels of the training samples, Represents the neural network for The prediction results of the training samples are calculated. During the training process, the weights of the neural network are adjusted according to the loss function. and threshold Find the partial derivative, As the hidden layer The input of the node, For its output, for the output layer, , for the hidden layer , according to the partial derivatives, the weights and thresholds are updated using gradient descent, that is, , , is the learning rate, and it is continuously updated until the loss function converges, so that the neural network can accurately classify and identify tiny lesions.

[0012] Furthermore, the weight of the S4 micro-texture feature vector and the weight of the macro-morphological feature vector Through cross-validation method, the training sample set is divided into Fold each time The first fold is used as the training set, and the remaining fold is used as the validation set. The loss function on the validation set is: ,in Indicates The sample index set of the validation set, For the The true category labels of samples, is the classifier for The prediction results of samples, the total loss function , when calculating partial derivatives, for : , and similarly we get , according to the total loss function and Calculate partial derivatives and update weights, repeat the update until the total loss function converges, and then obtain the optimal weights and .

[0013] Furthermore, the medical image includes one or a combination of CT images, MRI images, and PET images.

[0014] Compared with the prior art, this method for extracting and identifying features of microlesions in medical images has the following beneficial effects: 1. The method of the present invention realizes a comprehensive characterization of the characteristics of tiny lesions through a joint feature extraction strategy of microtexture and macromorphology. The microtexture features can deeply reflect the fine structural information such as cells and blood vessels inside the lesions, which is helpful for judging the pathological nature of the lesions, while the macromorphological features determine the position, range and approximate shape of the lesions as a whole, which is conducive to the positioning and overall evaluation of the lesions. The organic combination of the two avoids the one-sidedness of a single feature, makes the recognition results more accurate and reliable, and provides clinicians with more comprehensive and accurate diagnostic information, thereby significantly improving the accuracy of early diagnosis of the disease and improving the treatment effect and prognosis of patients.

[0015] 2. The present invention adopts multi-level feature extraction technology and combines it with feature fusion strategy to fully explore the characteristic information of tiny lesions at different scales, improve the robustness and generalization of the recognition method, and effectively extract and recognize features when facing different types of medical images and tiny lesions of different individuals and parts, thereby reducing misjudgments caused by the diversity and complexity of image data, improving the applicability and effectiveness of medical imaging diagnosis in different clinical scenarios, and reducing excessive reliance on specific imaging equipment and professional experience.

[0016] Other advantages, objectives and features of the present invention will be set forth in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative work. Figure 1 It is a flowchart of a method for feature extraction and recognition of microlesions in medical images; Figure 2 This is a flowchart of S4 dimensionality reduction processing in a method for feature extraction and identification of tiny lesions in medical images. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] In the description of the present application, the term "for example" is used to indicate an example, illustration or description. Any embodiment described as "for example" in the present application is not necessarily interpreted as being more preferred or more advantageous than other embodiments. In order to enable any technician in the field to implement and use the present invention, the following description is given. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes will not be elaborated in detail to avoid unnecessary details that make the description of the present invention obscure. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest range of principles and features disclosed in the present application.

[0020] Embodiment 1 This embodiment describes in detail a feature extraction and identification method for tiny lung lesions. By processing lung CT images, including image preprocessing, micro-texture feature extraction, macro-morphological feature extraction, micro-macro feature fusion and tiny lesion identification, the sensitivity and accuracy of tiny lung lesion detection are effectively improved, providing strong support for the early diagnosis of lung diseases.

[0021] In the specific implementation, first, a CT image of the lungs is obtained. , in the image Represents pixel coordinates, using median filtering To denoise, replace the value of each pixel with the median value of its neighborhood pixel value, that is, the neighborhood window size is , for any pixel in the image , the pixel value after denoising , by Sort the pixel values ​​and take the middle value to get the pixel value after removing noise , while retaining the edge information of the image; then using the adaptive histogram equalization algorithm to enhance the image contrast, the image is divided into non-overlapping sub-regions, for each sub-region ), calculate its local histogram ( , is the grayscale level), set the pixel point in the image The sub-region is , whose gray value is , then when calculating the local histogram hour, The value of is added by 1, and each pixel is mapped according to the local histogram of the sub-region to obtain the enhanced pixel value , the specific mapping formula is: For sub-region The cumulative distribution function of the enhanced pixel value is (in For sub-region In this way, the contrast of lung tissue and tiny lesions in the image can be significantly improved, which is convenient for feature extraction.

[0022] Then, the texture analysis algorithm based on wavelet transform is used to analyze the preprocessed image. For processing, Haar wavelet basis function is selected to process the image Scale wavelet decomposition, wavelet transform function is , then through Perform the image Scaling wavelet decomposition to obtain wavelet coefficients , calculate the energy of the high frequency sub-band (in represents the high frequency subband area) and entropy is the wavelet coefficient Probability distribution function of ); calculate different directions and distance The pixel gray-level co-occurrence matrix on , for each pixel in the image , in the direction and distance Find the corresponding pixel point , count the number of occurrences of the gray value combination of these two pixels, and construct the gray level co-occurrence matrix ,based on Calculate contrast , Correlation ,energy and entropy ,in Respectively about The mean and standard deviation in the direction are calculated as follows: The size is ,but for The calculation will be and Interchange them to obtain the above micro texture features, and combine them into a micro texture feature vector .

[0023] Then, the morphological dilation function is applied right Processing, defining structural elements is a circular structural element, then the expanded image is The dilation operation will expand the boundaries of objects in the image outward, which helps to connect broken edges and fill small holes, thereby better extracting the overall morphology of the lesion and obtaining the edge contour of the lesion through the edge detection algorithm based on morphological gradient. , let the expanded image be , the image after corrosion is (The erosion operation is the same as the dilation operation, using the same structural element , the erosion function is defined as ), then the morphological gradient image , threshold processing is performed on the morphological gradient image to obtain the edge contour , set the threshold to ,but , calculate the perimeter of the edge , , Domain and circularity When calculating the perimeter, traverse the pixel points on the edge contour, and add 1 to the perimeter each time a pixel point is encountered. The area can be calculated by counting the area The number of pixels in the lesion area is obtained, and the region growing algorithm is performed on the lesion area. Processing, according to the seed point selection method, the image is segmented by grayscale threshold, and the threshold is set to , get the binary image , in the binary image Select the central pixel with the largest connected area as the seed point , let the growth threshold be , using the grayscale standard deviation of the image and the estimated value of grayscale variation within the microlesion ,but (in is the average gray value of the image, and is the empirical coefficient , the lesion area after region growing is , calculate the volume of the lesion area , Center of gravity ,in The above macroscopic morphological features are combined into a macroscopic morphological feature vector .

[0024] The micro texture feature vector and the macroscopic morphological feature vector Input to the fusion layer and set the weight of the micro texture feature vector to , the weight of the macroscopic morphological feature vector is ,and , optimize and adjust the weights through the cross-validation method, and divide the training sample set into Fold each time The first fold is used as the training set, and the remaining fold is used as the validation set. The loss function on the validation set is: ,in Indicates The sample index set of the validation set, For the The true category labels of samples, is the classifier for The prediction results of samples, the total loss function , when calculating partial derivatives, for : , and similarly we get , according to the total loss function and Calculate partial derivatives and update weights, repeat the update until the total loss function converges, and then obtain the optimal weights and , the micro and macro features are fused according to the optimal weights, and the fused feature vector , use principal component analysis (PCA) to analyze the fused feature vector Perform dimensionality reduction processing on the fused feature vector The sample set is processed and the covariance matrix is ​​calculated , whose elements (in is the sample size Indicates The first fusion feature vector elements, is the first of all samples The mean of the features, ), perform eigenvalue decomposition on the covariance matrix and obtain the eigenvalue and the corresponding eigenvector , according to the ratio of the sum of the retained eigenvalues ​​to the sum of the total eigenvalues Determine the dimensions after dimensionality reduction ,Right now , take the front The eigenvectors corresponding to the largest eigenvalues ​​form the transformation matrix , and get the feature vector after dimensionality reduction .

[0025] The fused feature vector after dimension reduction Input to the trained classifier In the classifier A neural network structure classifier is used. The neural network structure includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is equal to the feature vector after dimensionality reduction. The hidden layer adopts a multi-layer structure, and the number of nodes in each layer depends on the number of training samples. Number of categories with minimal lesions To determine, specifically the number of hidden layer nodes , the number of nodes in the output layer and the number of categories of minimal lesions The neural network is equal to each other, so that the neural network can output the results corresponding to different categories of microlesions. The weights and thresholds of the neural network are trained through the back propagation algorithm, and the loss function is set. (in Representative The true category labels of the training samples, Represents the neural network for The prediction results of training samples) are obtained. During the training process, the weights of the neural network are adjusted according to the loss function. and threshold Find the partial derivative, As the hidden layer The input of the node, For its output ( is the activation function), for the output layer, , for the hidden layer (Here represents the next layer node connected to the current hidden layer node), and the weights and thresholds are updated using gradient descent according to the partial derivatives, that is, (in is the learning rate), and the update is continued until the loss function converges, so that the neural network can accurately classify and identify small lesions.

[0026] In summary, in the entire process of detecting tiny lung lesions, starting from image acquisition and preprocessing, to micro-texture feature extraction, macro-morphological feature extraction, micro-macro feature fusion, and finally to tiny lesion identification, each step is closely linked and complements each other. Through this comprehensive processing flow, the sensitivity and accuracy of tiny lung lesion detection can be effectively improved, which is helpful for the early detection of lung diseases and provides an important basis for the timely treatment of patients.

[0027] Embodiment 2 This embodiment describes in detail the feature extraction and recognition method for tiny brain lesions. By integrating brain CT and MRI images, image preprocessing, micro-texture and macro-morphological feature extraction, feature fusion and tiny lesion recognition, the accuracy of diagnosis of tiny brain lesions is improved, laying the foundation for the early treatment of brain diseases.

[0028] First, obtain CT images of the brain and MRI images ,in is the image pixel coordinate, Denoising to get the denoised image , use the histogram equalization algorithm to enhance the image contrast and obtain the enhanced image ;right Denoising to get the denoised image , and at the same time perform contrast enhancement to obtain an enhanced image , the enhanced CT and MRI images are registered, and a registration method based on feature points is used to extract feature points using a scale-invariant feature transformation algorithm. The algorithm constructs the scale space of the image, detects extreme points at different scales, and accurately locates and describes the features of the extreme points, thereby obtaining a feature point set with scale and rotation invariance, namely for The feature point set in for The feature point set in the calculation is used to calculate the matching relationship between the feature point sets, and the nearest neighbor distance ratio is used for matching. Each feature point in ,exist Find its nearest neighbor feature point and the next nearest feature point , calculate them with The characteristic distance and ,like ( Threshold , then it is believed that and The match is successful, and the matching point pair set is obtained , according to the matching point pair set Calculate the transformation model between images and set the transformation matrix to , for matching point pairs (in ,satisfy , and get the transformation matrix The optimal estimate of the transformation matrix is ​​used to minimize the error of all matching points after transformation. Transform the MRI image to achieve image registration, that is , so that the tiny lesions in the two images correspond in spatial position, facilitating unified feature extraction and recognition operations.

[0029] Then, the registered images and The texture analysis algorithm using wavelet transform decomposition scale is , set the wavelet transform function to , then the image is Scaling wavelet decomposition to obtain wavelet coefficients and , calculate the energy of the high frequency sub-band (in represents the high-frequency sub-band area of ​​CT image), (in represents the high-frequency subband area of ​​MRI images) and entropy , , and The probability distribution functions of the wavelet coefficients of CT and MRI images are calculated in different directions. and distance The pixel gray-level co-occurrence matrix on and For each pixel in the image, find its corresponding pixel in the specified direction and distance, count the number of occurrences of gray value combinations to construct a gray level co-occurrence matrix, based on and Calculate the contrast separately , , correlation , ,energy , and entropy , ,in and CT and MRI images are Directional mean and standard deviation.

[0030] Use morphological erosion and dilation functions to process the registered images. For CT images, define the structural element is a rectangular structural element, the image after corrosion (Corrosion function Defined as ), the expanded image (Expansion function Defined as , for the MRI image, the same corrosion and expansion operations are performed to obtain and , the edge contour of the lesion is obtained by the Canny edge detection algorithm. For CT images, , calculate the perimeter of the edge ,area ,(in domain) and circularity , for MRI images, , calculate the corresponding perimeter ,area and circularity .

[0031] Perform region growing algorithm on the lesion area For CT images, the seed point selection method is as follows: grayscale threshold segmentation is performed on the image, and the threshold is set to , get the binary image , select the central pixel with the largest connected area in the binary image as the seed point For MRI images, the same method is used to set the appropriate threshold , get the seed point , for CT images, the growth threshold is , according to the grayscale standard deviation of the image and the estimated value of grayscale variation range within microlesions ,but (in is the average gray value of the image, and is the empirical coefficient), for MRI images, the growth threshold can be obtained by the same logic , the lesion area after region growing is and , calculate the volume of the lesion area , Center of gravity (in , for MRI images, calculate the volume and center of gravity The calculation method is the same as that of CT images. The above macroscopic morphological features are combined into macroscopic morphological feature vectors and .

[0032] The micro texture feature vector , and the macroscopic morphological feature vector , Input to the fusion layer, and determine the weight of the micro texture feature vector through brain lesion sample training and the weight of the macro-morphological feature vector and , the fused feature vector , use principal component analysis to analyze the fused feature vector Perform dimensionality reduction to obtain the eigenvector after dimensionality reduction , the fusion feature vector after dimension reduction Input to the trained neural network classifier In this way, the neural network can accurately classify and identify tiny lesions.

[0033] In summary, in the process of detecting microscopic brain lesions, the integration of multimodal images (CT and MRI) fully utilizes the information advantages of different image modalities, helps to distinguish different types of microscopic lesions, reduces misdiagnosis and missed diagnosis, and provides a reliable basis for the early treatment of brain diseases.

[0034] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A method for extracting and identifying features of microlesions in medical images, characterized in that: The specific steps of this method are: S1, image preprocessing: input medical images Perform preprocessing, the preprocessing including denoising and contrast enhancement, to obtain a preprocessed image ,in is the image pixel coordinate; S2, Micro texture feature extraction: The preprocessed medical image is analyzed using texture analysis based on wavelet transform. Processing, through multi-scale wavelet decomposition, decompose the image into low-frequency and high-frequency sub-bands, through the wavelet function The wavelet coefficients in the sub-band are extracted to reflect the microscopic texture characteristics of the small lesions, and the statistics of the high-frequency sub-band energy and entropy are calculated as the descriptors of the microscopic texture characteristics, that is, through the wavelet transform function , decomposition scale , perform Wavelet coefficients are obtained by wavelet decomposition of the scale , and calculate the energy of the high frequency sub-band and entropy ,in is the wavelet coefficient The probability distribution function of Represents the high frequency sub-band region; The gray level co-occurrence matrix GLCM is used to further extract micro texture features and calculate the and distance The pixel gray level co-occurrence matrix is ​​used to obtain the texture feature parameters such as contrast, correlation, energy and entropy based on GLCM, that is, For direction and distance Grayscale on the image, calculate the contrast , correlation ,energy ,entropy ,in Respectively about The mean and standard deviation in the direction combine the micro-texture features into a micro-texture feature vector ; S3, macroscopic morphological feature extraction: using morphological processing technology, dilate the medical image to extract the macroscopic morphological features of the edge shape, size and contour of the tiny lesions, and use the edge perimeter, area, and circularity geometric parameters as macroscopic morphological feature descriptors, that is, using the morphological dilation function right Processing, define the structural element as , then the expanded image is , and based on the macroscopic shape after expansion, the edge detection algorithm Obtain the edge contour of the lesion , calculate the perimeter of the edge ,area ,in The edge contour Enclosed area, circularity ; Perform region growing algorithm on the lesion area Processing, determine the scope of the lesion, and calculate its volume and center of gravity position characteristics to further enrich the macroscopic morphological feature information, that is, set the seed point to The growth threshold is , the lesion area after region growing is , calculate its volume , center of gravity position ,in , combining the macroscopic morphological features into a macroscopic morphological feature vector ; S4, micro-macro feature fusion: The extracted micro texture features and macroscopic morphological characteristics The input is sent to the fusion layer, and a weighted fusion method is used to assign corresponding weights to the importance of the vectors for micro-lesion recognition according to the different characteristics of micro-texture and macro-morphology. and , and perform weighted summation of the micro texture feature vector and the macro morphology feature vector to obtain the fused feature vector ; The fused feature vector is analyzed by using the principal component analysis (PCA) dimensionality reduction technique. Perform dimensionality reduction processing to remove redundant information and retain representative feature information; S5, microlesion recognition: The fusion feature vector after dimension reduction is Input to the trained classifier In the classifier According to the characteristic information of the feature vector, the microlesions are classified and identified to obtain the identification results .

2. According to claim 1, a method for extracting and identifying features of microlesions in medical images, characterized in that: The S3 region growing algorithm Middle seed point The selection method is: perform grayscale threshold segmentation on the image and set the threshold to , get the binary image , in the binary image Select the central pixel with the largest connected area as the seed point .

3. The method for extracting and identifying features of microlesions in medical images according to claim 1, characterized in that: The growth threshold in S3 , determined based on the grayscale of the image and the grayscale changes within the microlesion, that is, using the grayscale standard deviation of the image , estimated value of grayscale variation inside small lesions , then the growth threshold ,in is the average gray value of the image, and is the empirical coefficient.

4. The method for extracting and identifying features of microlesions in medical images according to claim 1, characterized in that: The S4 transforms the micro texture feature vector and the macroscopic morphological feature vector Input to the fusion layer and set the weight of the micro texture feature vector to , the weight of the macroscopic morphological feature vector is ,and , the fused feature vector , using principal component analysis The fused feature vector Perform dimensionality reduction processing, that is The transformation matrix is , the feature vector after dimensionality reduction .

5. The method for extracting and identifying features of microlesions in medical images according to claim 4, characterized in that: The S4 uses principal component analysis The fused feature vector The process of dimensionality reduction is as follows: The fused feature vector The sample set is split into ,in is the number of samples, for the fused feature vector , whose dimensions for , the covariance matrix Elements for: ,in , Indicates The first fusion feature vector elements, is the first of all samples The mean of the features, ; Covariance matrix Perform eigenvalue decomposition and find a new set of coordinate axes represented by eigenvectors so that the data variance represented by the eigenvalues ​​on the coordinate axes is maximized, thereby obtaining the eigenvalues and the corresponding eigenvector ; Set the reduced dimension to , the latitude is based on the ratio of the sum of the retained eigenvalues ​​to the sum of the total eigenvalues ,Right now , select the eigenvalue feature vectors; Before The eigenvectors corresponding to the largest eigenvalues ​​form the transformation matrix , the feature vector after dimensionality reduction pass Get the feature vector after fusion Projected onto In the new subspace determined, dimensionality reduction is achieved.

6. The method for extracting and identifying features of microlesions in medical images according to claim 1, characterized in that: The S5 classifier A neural network structure classifier is used, wherein the neural network structure includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is equal to the feature vector after dimensionality reduction. The hidden layer adopts a multi-layer structure, and the number of nodes in each layer depends on the number of training samples. Number of categories with minimal lesions To determine, specifically the number of hidden layer nodes , the number of nodes in the output layer and the number of categories of minimal lesions The neural network is equal to each other, so that the neural network can output the results corresponding to different categories of microlesions. The weights and thresholds of the neural network are trained through the back propagation algorithm, and the loss function is set. ,in Representative The true category labels of the training samples, Represents the neural network for The prediction results of the training samples are calculated. During the training process, the weights of the neural network are adjusted according to the loss function. and threshold Find the partial derivative, As the hidden layer The input of the node, For its output, for the output layer, , for the hidden layer , according to the partial derivatives, the weights and thresholds are updated using gradient descent, that is, , , is the learning rate, and it is continuously updated until the loss function converges, so that the neural network can accurately classify and identify tiny lesions.

7. The method for extracting and identifying features of microlesions in medical images according to claim 1, characterized in that: The weight of the S4 microtexture feature vector and the weight of the macro-morphological feature vector Through cross-validation method, the training sample set is divided into Fold each time The first fold is used as the training set, and the remaining fold is used as the validation set. The loss function on the validation set is: ,in Indicates The sample index set of the validation set, For the The true category labels of samples, is the classifier for The prediction results of samples, the total loss function , when calculating partial derivatives, for : , and similarly we get , according to the total loss function and Calculate partial derivatives and update weights, repeat the update until the total loss function converges, and then obtain the optimal weights and .

8. The method for extracting and identifying features of microlesions in medical images according to claim 1, characterized in that: The medical image includes one or a combination of CT images, MRI images, and PET images.

Images