Feature classification method, computer device and storage medium

Abstract

The present application relates to a feature classification method, a computer device and a storage medium. The feature classification result of the target breast is obtained by obtaining a medical image of the contralateral breast of the target breast and inputting the medical image of the contralateral breast into a preset machine learning model, wherein the medical image includes the breast glandular tissue of the contralateral breast and the background parenchyma enhanced tissue; the method simplifies the breast feature analysis method, making the classification result obtained by the breast feature analysis more convenient; at the same time, the extracted features are ensured to be more accurate, and the efficiency and accuracy of the feature classification result are also improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of medical technology, and in particular, to a feature classification method, a computer device, and a storage medium. Background Art

[0002] Breast cancer is a phenomenon in which mammary epithelial cells proliferate uncontrollably under the influence of multiple carcinogenic factors. It is the most common malignant tumor in middle-aged and elderly women, and its incidence and mortality are increasing year by year.

[0003] Early detection and prediction of breast cancer can effectively improve the cure rate of breast cancer. In the related art, when analyzing the breast characteristics of the breast to determine whether breast cancer exists, it is usually determined by analyzing the characteristics of the tumor to determine whether the tumor is benign or malignant, or based on tumor staging and typing, and based on biomarkers that characterize tumor deterioration.

[0004] However, breast feature analysis in related technologies is based on tumor features themselves, resulting in a relatively complex and single breast analysis method. Summary of the invention

[0005] Based on this, it is necessary to provide a feature classification method, computer equipment and storage medium to address the above technical problems, which can simplify the breast feature analysis method and improve the diversity of the feature analysis method.

[0006] In a first aspect, an embodiment of the present application provides a feature classification method, the method comprising:

[0007] Acquire a medical image of the contralateral breast of the target breast; the medical image includes breast glandular tissue of the contralateral breast and background parenchymal enhancement tissue;

[0008] The medical image of the contralateral breast is input into the preset machine learning model to obtain the feature classification results of the target breast.

[0009] In one embodiment, the machine learning model includes a feature extractor and a preset combination model;

[0010] The above method inputs the medical image of the contralateral breast into the preset machine learning model to obtain the feature classification results of the target breast, including:

[0011] The available radiomic features of the contralateral breast are extracted from the breast glandular tissue and the background parenchymal enhancement tissue by a feature extractor;

[0012] The available radiomics features were input into the combined model to obtain the feature classification results of the target breast.

[0013] In one embodiment, the combined model includes a feature selector and a classifier;

[0014] The above-mentioned available radiomics features are input into the combined model to obtain the characteristic classification results of the target breast, including:

[0015] Through the feature selector, the collinear features among the available radiomics features were screened out;

[0016] The available radiomics features after filtering out the collinear features are input into the classifier to obtain the feature classification results of the target breast.

[0017] In one embodiment, the extracting of usable radiomic features of the contralateral breast from the breast glandular tissue and the background parenchymal enhancement tissue by the feature extractor includes:

[0018] The tissue radiomics features of the contralateral breast are extracted from the breast glandular tissue and the background parenchymal enhancement tissue by the feature extractor;

[0019] The tissue radiomics features were screened to obtain the available radiomics features.

[0020] In one embodiment, the above-mentioned screening of tissue radiomics features to obtain usable radiomics features includes:

[0021] Standardization of tissue radiomics features;

[0022] The variance of each feature in the standardized tissue radiomics features was obtained, and the features with variance greater than the threshold were identified as usable radiomics features.

[0023] In one embodiment, the process of constructing the above-mentioned combined model includes:

[0024] Obtain sample medical images of breasts of multiple samples; the sample medical images include medical images of the contralateral breast of the breast affected by breast cancer when only one breast is affected by breast cancer; the sample medical images include sample breast glandular tissue and sample background parenchymal enhancement tissue;

[0025] Based on the sample breast glandular tissue and the sample background parenchymal enhancement tissue in each sample medical image, the training radiomics features of each sample medical image are obtained;

[0026] According to the features of each training radiomics and the preset gold standard, a preset number of initial combination models are trained to obtain candidate combination models corresponding to each initial combination model;

[0027] According to the prediction performance indicators of each candidate combination model, a combination model is determined from the candidate combination models.

[0028] In one embodiment, the above-mentioned training of a preset number of initial combination models according to the training radiomics features to obtain candidate combination models corresponding to each initial combination model includes:

[0029] By executing the training step, each of the preset number of initial combination models is trained to obtain a candidate combination model corresponding to each initial combination model;

[0030] The training step includes:

[0031] For any initial combination model:

[0032] Input each training radiomics feature into the feature selector in the initial combination model, and use the feature selector to filter out the collinearity features in each training radiomics feature;

[0033] The training radiomics features after screening out the collinear features are input into the initial classifier in the initial combination model, and the initial classifier is trained according to the gold standard to obtain the candidate combination model corresponding to the initial combination model.

[0034] In one embodiment, the above-mentioned prediction performance indicators include area under the curve and relative standard deviation;

[0035] Then, according to the prediction performance index of each candidate combination model, the combination model is determined from each candidate combination model, including:

[0036] Obtaining the area under the curve of each candidate combination model and the relative standard deviation of the area under the curve of each candidate combination model;

[0037] The candidate combination model whose area under the curve and relative standard deviation both meet the preset conditions is determined as the combination model.

[0038] In a second aspect, an embodiment of the present application provides a feature classification device, the device comprising:

[0039] An acquisition module is used to acquire a medical image of the contralateral breast of the target breast; the medical image includes the breast glandular tissue of the contralateral breast and the background parenchymal enhancement tissue;

[0040] The classification module is used to input the medical image of the contralateral breast into a preset machine learning model to obtain the feature classification results of the target breast.

[0041] In a third aspect, an embodiment of the present application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the method steps of any one of the embodiments of the first aspect when executing the computer program.

[0042] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method steps of any one of the embodiments of the first aspect above.

[0043] The embodiments of the present application provide a feature classification method, computer device and storage medium. The feature classification result of the target breast is obtained by acquiring the medical image of the contralateral breast of the target breast and inputting the medical image of the contralateral breast into a preset machine learning model. The medical image includes the mammary glandular tissue of the contralateral breast and the background parenchymal enhancement tissue. In this method, on the one hand, based on the correlation between the bilateral breasts of the human body, the information in the medical image of the contralateral breast is used to determine the feature classification result of the target breast, providing a new way to analyze breast features to determine the feature classification result, which increases the diversity of breast feature analysis methods. On the other hand, this method is based on the mammary gland tissue in the breast. The characteristics of the tissue and the background parenchymal enhanced tissue are used as the basis for analysis. Compared with the related art that uses the characteristics of the lesion itself as the basis for analysis, it also provides a new way to analyze breast features to determine the feature classification results, further increasing the diversity of breast feature analysis methods; and the breast glandular tissue and the background parenchymal enhanced tissue area in the breast are relatively large, and the feature extraction method is relatively convenient, thereby simplifying the breast feature analysis method and making the classification results obtained by the breast feature analysis more convenient; the larger area can also ensure that the extracted features are more accurate, thus making the classification results obtained by the breast feature analysis more accurate. Furthermore, in this method, a machine learning model is used to perform feature analysis to obtain feature classification results. The machine learning model is pre-constructed and can be directly called to improve the efficiency and accuracy of the feature classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 An application environment diagram of a feature classification method provided in one embodiment;

[0045] Figure 2 A schematic diagram of a flow chart of a feature classification method provided in one embodiment;

[0046] Figure 3 A schematic flow chart of a feature classification method provided in another embodiment;

[0047] Figure 4 A schematic flow chart of a feature classification method provided in another embodiment;

[0048] Figure 5 A schematic flow chart of a feature classification method provided in another embodiment;

[0049] Figure 6A schematic flow chart of a feature classification method provided in another embodiment;

[0050] Figure 7 A schematic flow chart of a feature classification method provided in another embodiment;

[0051] Figure 8 A schematic diagram of combined model verification and test results provided in one embodiment;

[0052] Fig. 9 A schematic flow chart of a feature classification method provided in another embodiment;

[0053] Fig.10 is a structural block diagram of a feature classification device provided in one embodiment;

[0054] Fig.11 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0056] The feature classification method provided in this application can be applied to computer devices. The computer devices include but are not limited to various personal computers, laptops, smart phones, tablet computers, portable wearable devices, etc. Among them, the internal structure of the computer device is as follows Figure 1 As shown, the processor in the internal structure is used to provide computing and control capabilities. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database is used to store relevant data of the feature classification process. The network interface is used to communicate with other external devices through a network connection. When the computer program is executed by the processor, a feature classification method is implemented.

[0057] The embodiments of the present application provide a feature classification method, a computer device, and a storage medium, which can simplify the feature analysis method and improve the diversity of the feature analysis method. The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems will be described in detail below through embodiments and in combination with the accompanying drawings. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. It should be noted that when the feature classification method provided in the embodiment of the present application is described below, the execution subject is a computer device. The technical solution in the embodiment of the present application will be clearly and completely described below in combination with the drawings in the embodiment of the present application. Obviously, the described embodiment is a part of the embodiment of the present application, not all of the embodiments.

[0058] In one embodiment, Figure 2 As shown, a feature classification method is provided, and the embodiment includes the following steps:

[0059] S101, obtaining a medical image of a contralateral breast of a target breast; the medical image includes breast glandular tissue of the contralateral breast and background parenchymal enhancement tissue.

[0060] If one of the two breasts is the target breast, the other is the opposite breast of the target breast. The target breast refers to the breast that needs to be feature classified at present. It should be noted that the breast in the embodiment of the present application includes breast and gland.

[0061] The feature classification in the embodiment of the present application refers to the classification of the features of breast glandular tissue and background parenchyma enhancement tissue in the breast. Among them, breast glandular tissue (fibroglandular tissues, FGT) and background parenchyma enhancement (background parenchyma enhancement, BPE) tissue are tissues other than adipose tissue in the breast, and BPE refers to the contrast enhancement of normal breast fibroglandular tissue, which is shown as progressive enhancement during dynamic enhancement.

[0062] Based on this, it is necessary to obtain medical images including FGT and BPE to extract the features of FGT and BPE. The medical images include but are not limited to computed tomography (CT) images, magnetic resonance imaging (MRI) images, positron emission computed tomography-magnetic resonance imaging (PET-MRI) images, dynamic contrast enhancement magnetic resonance imaging (DCE-MRI) images, etc., which are not limited in this embodiment.

[0063] In practical applications, the changes between the two breasts of a human body are related. If the internal morphological characteristics of one breast change due to malignant transformation, the internal morphological characteristics of the other breast will also be affected. Based on this, in the embodiment of the present application, when it is necessary to perform feature classification on the target breast, the medical image of the contralateral breast of the target breast is used as the analysis basis.

[0064] Taking DCE-MRI images as an example, DCE-MRI uses a continuous, repeated, and rapid imaging method to obtain images before and after the injection of contrast agents, and obtains semi-quantitative or quantitative parameters through a series of calculations and analyses. DCE-MRI images can be used to obtain information on the morphological characteristics of lesions and reflect the physiological changes of lesion tissues.

[0065] Therefore, in the embodiment of the present application, the medical image of the contralateral breast of the target breast is obtained, and what can be obtained is the DCE-MRI image of the contralateral breast of the target breast, and the DCE-MRI image needs to include the FGT and BPE of the contralateral breast.

[0066] It should be noted that the DCE-MRI images of the contralateral breast obtained are images before and after the contrast agent is injected into the contralateral breast, and the DCE-MRI images of the contralateral breast are not limited to images of any phase after the contrast agent is injected, and the embodiments of the present application do not limit this.

[0067] Based on the acquired DCE-MRI images, FGT is extracted from the images acquired at each time phase, and then the FGT in the image before contrast agent injection is subtracted from the FGT in the image after contrast agent injection to obtain BPE, which reflects the enhancement of glandular tissue. Therefore, based on the acquired DCE-MRI images of the contralateral breast, FGT and BPE can be obtained, and then by analyzing FGT and BPE, features reflecting the morphological information of FGT and BPE can be obtained and classified.

[0068] For example, in actual application, the method for the embodiments of the present application to obtain the medical image of the contralateral breast of the target breast includes but is not limited to obtaining the medical image of the contralateral breast that meets the above requirements from a database pre-stored therein; it may also be downloading the medical image of the contralateral breast that meets the requirements from a network platform; it may also be generating the corresponding medical image of the contralateral breast according to demand, etc. The embodiments of the present application do not limit the method for obtaining the medical image of the contralateral breast.

[0069] S102, inputting the medical image of the contralateral breast into a preset machine learning model to obtain a feature classification result of the target breast.

[0070] According to the medical image of the contralateral breast obtained above, the medical image of the contralateral breast is input into a preset machine learning model, and the FGT and BPE in the medical image of the contralateral breast are analyzed by the machine learning model to obtain the characteristics of FGT and BPE, and the characteristics are classified to obtain a classification result, which is the characteristic classification result of the target breast. After obtaining the characteristic classification result of the target breast, according to the characteristic classification result of the target breast, further information processing or analysis can be performed, and applied in different scenarios according to the processing result or analysis result.

[0071] The feature classification result may be the result of classifying the features according to different types of information; or the result of classifying according to different degrees of lesions; or the result of classifying according to normal tissue features and abnormal tissue features, etc. The present application embodiment does not limit the classification conditions in the feature classification result.

[0072] The preset machine learning model in the embodiments of the present application can be implemented in different ways.

[0073] In one embodiment, the machine learning model is a holistic, separate, end-to-end neural network model, that is, the medical image of the contralateral breast is used as the input of the neural network model, and after passing through the neural network model, the output is the feature classification result of the target breast.

[0074] In another embodiment, the machine learning model includes multiple different sub-neural network models or algorithm models, that is, the medical image of the contralateral breast is used as one of the algorithm models in the machine learning model. After being processed by the algorithm model, it enters the next algorithm model or sub-neural network model until all models are passed and the feature classification result of the target breast is finally output.

[0075] Regardless of which of the above methods is used to implement the machine learning model, the function of determining the characteristic classification results of the target breast from the medical image of the contralateral breast can be achieved. The embodiments of the present application do not limit the implementation method of the machine learning model.

[0076] The feature classification method provided in the embodiment of the present application obtains a medical image of the contralateral breast of the target breast and inputs the medical image of the contralateral breast into a preset machine learning model to obtain a feature classification result of the target breast, wherein the medical image includes the breast glandular tissue of the contralateral breast and the background parenchymal enhancement tissue; in this method, on the one hand, based on the correlation between the bilateral breasts of the human body, the information in the medical image of the contralateral breast is used to determine the feature classification result of the target breast, providing a new way to analyze breast features to determine the feature classification result, thereby increasing the diversity of breast feature analysis methods; on the other hand, in this method, The characteristics of the breast glandular tissue and the background parenchymal enhanced tissue in the breast are used as the basis for analysis. Compared with the analysis based on the characteristics of the lesion itself in the related art, a new way to analyze breast characteristics is provided to determine the feature classification results, which further increases the diversity of breast feature analysis methods; and the breast glandular tissue and the background parenchymal enhanced tissue area in the breast are relatively large, and the feature extraction method is relatively convenient, thereby simplifying the breast feature analysis method and making the classification results obtained by the breast feature analysis more convenient; the larger area can also ensure that the extracted features are more accurate, thus making the classification results obtained by the breast feature analysis more accurate. Furthermore, in this method, a machine learning model is used to perform feature analysis to obtain feature classification results. The machine learning model is pre-constructed and can be directly called to improve the efficiency and accuracy of the feature classification results.

[0077] Based on the above embodiment, the process of the above step S102 is described by taking the machine learning model including multiple different sub-neural network models or algorithm models as an example. In one embodiment, the above machine learning model includes a feature extractor and a preset combination model; based on this, Figure 3 As shown, the above S102 includes the following steps:

[0078] S201, extracting usable radiomic features of the contralateral breast from breast glandular tissue and background parenchymal enhancement tissue through a feature extractor.

[0079] The feature extractor can be implemented by any feature extraction algorithm, for example, the Pyradiomics toolkit is used to extract radiomic features in FGT and BPE tissues based on medical images of the contralateral breast; pyradiomics is an open source python package that can be used for radiomic feature extraction of medical images. Specifically, in practical applications, the pyradiomics toolkit can be used to write code to generate a usable feature extractor based on actual needs, and then the feature extractor can be used to extract usable radiomic features of the contralateral breast from breast glandular tissue and background parenchymal enhancement tissue.

[0080] Optionally, the radiomics features include at least one of first-order statistical features, shape features, grayscale co-occurrence matrix, grayscale level length matrix, grayscale and band matrix, neighborhood grayscale difference matrix and grayscale independent matrix. Among them, the first-order statistical features reflect the grayscale distribution of the lesion, including: energy, entropy, skewness, kurtosis, variance, mean absolute deviation, etc.; the shape features reflect the overall shape of the lesion, including: compactness, spherical imbalance, sphericity, volume, surface area, etc.; the texture features describe the spatial distribution of the grayscale intensity of the lesion, which is composed of grayscale co-occurrence matrix features and grayscale run matrix features; for example, a total of 103 types of features can be extracted.

[0081] The available radiomics features refer to a part of the radiomics features selected from the radiomics features. For example, the remaining features after removing irrelevant features and redundant features from the radiomics features are determined as available radiomics features.

[0082] In one embodiment, a medical image of the contralateral breast including breast glandular tissue and background parenchymal enhancement tissue is used as an input to a feature extractor. After passing through the feature extractor, available imaging omics features of the contralateral breast are directly obtained.

[0083] In another embodiment, a feature extractor may be used to first extract unscreened imaging omics features of breast glandular tissue and background parenchymal enhancement tissue from a medical image of the contralateral breast including breast glandular tissue and background parenchymal enhancement tissue, and then remove irrelevant features and redundant features from the unscreened imaging omics features, and determine the remaining features after removing the irrelevant features and redundant features as usable imaging omics features.

[0084] It should be noted that, in one embodiment, this step can be used as a preprocessing process, and the preprocessing process is placed before the machine learning model of the previous embodiment. That is, after obtaining the medical image of the contralateral breast, this step is first performed from the medical image of the contralateral breast to extract the available imaging genomics features, and then the available imaging genomics features are input into the above-mentioned machine learning model to obtain the feature classification results of the target breast.

[0085] S202, inputting the available radiomics features into the combined model to obtain the feature classification results of the target breast.

[0086] Based on the available radiomic features of the contralateral breast extracted from breast glandular tissue and background parenchymal enhancement tissue, the available radiomic features are input into the combined model to obtain the feature classification results of the target breast.

[0087] The combined model can be a model composed of a combination of multiple different algorithm models. After the multiple algorithm models in the combined model are combined, its performance is higher than the performance of the combination of other algorithm models. For example, the combined model includes algorithm model A1, algorithm model B1 and algorithm model C1, then the performance of the combined model is compared with the performance of the combination of algorithm model A2, algorithm model B1 and algorithm model C2. Performance refers to the accuracy, precision, sensitivity, stability, specificity and other performance indicators of the output feature classification results of the target breast.

[0088] In this embodiment, the available radiomics features of the contralateral breast are extracted from the breast glandular tissue and the background parenchymal enhancement tissue by a feature extractor, and then the available radiomics features are input into the combined model to obtain the feature classification results of the target breast. Because the available radiomics features are the features after the redundant features and irrelevant features are eliminated, the feature classification results of the target breast are determined based on the available radiomics features, which can improve the accuracy of the feature classification results.

[0089] In one embodiment, taking the above-mentioned combined model including a feature selector and a classifier as an example, Figure 4 As shown, the above S202 includes the following steps:

[0090] S301, screening out collinear features from available radiomics features through a feature selector.

[0091] Collinearity refers to the characteristics of the independent variables due to the high correlation. Since collinearity will cause some highly correlated but important features to be excluded during feature selection, it is necessary to screen out the collinearity in the available imaging features. After the collinearity is screened out, the available imaging features are equivalent to screening out the highly correlated features, which will make the back-end combined model more interpretable. Interpretability can also better determine the cause of the characteristic classification results of the target breast and guide subsequent applications.

[0092] Among them, the feature selector may include but is not limited to the Least Absolute Shrinkage and Selection Operator (LASSO), Recursive Feature Elimination (RFE) and Random Forest (RF), etc., which is not limited in the embodiments of the present application.

[0093] Optionally, in one embodiment, before the available radiomics features are screened out for collinear features using a feature selector, the available radiomics features can be further optimized to select the best features. For example, by using a univariate feature selection method, the correlation between each feature in the available radiomics features and the benign and malignant nature of the tumor is measured, and features with insignificant differences between the two groups of samples of benign tumors and malignant tumors are eliminated, and features with a p-value (assumed value) <0.05 are retained as features after the available radiomics features are optimized. Then, based on the optimized features, the step of screening out collinear features can be performed. In this way, the best features can be further selected from the available radiomics features by calculating the statistical measure of the univariate, so that the features are more accurate.

[0094] S302, inputting the available radiomics features after filtering out the collinear features into a classifier to obtain a feature classification result of the target breast.

[0095] A classifier is a classification model that can map given data to one of the given categories, so that it can be applied to data prediction. Its implementation can adopt algorithms such as decision tree, logistic regression, naive Bayes, neural network, etc. For example, the classifier includes but is not limited to any one of support vector machine, logistic regression, decision tree, gradient boosting decision tree and extreme gradient.

[0096] Taking the constructed classifier as a gradient boosting decision tree as an example, the available imaging omics features after screening out the collinear features are input into the gradient boosting decision tree, and the gradient boosting decision tree can determine the characteristic classification result of the target breast according to the input data.

[0097] In one embodiment, the combined model is a model composed of LASSO + Gradient Boosting Decision Tree (GBDT); in another embodiment, the combined model is a model composed of LASSO + eXtreme Gradient Boosting (XGBoost); this embodiment of the present application is not limited to this. Among them, GBDT refers to all gradient boosting tree algorithms, including XGBoost, that is, XG Boost is also a variant of GBDT. In actual application, the difference between the two is the selection of different scenarios. When applying the embodiment of the present application, any one of the two can be selected as the combined model.

[0098] In the embodiment of the present application, the collinearity features in the available imaging omics features are screened out by a feature selector, and then the available imaging omics features after the collinearity features are screened out are input into a classifier to obtain the feature classification results of the target breast. By combining the feature selector and the classifier, and the combined model formed after the combination is a combination method with the best performance, the feature classification results of the target breast outputted in the end can be made more accurate and effective.

[0099] As mentioned above, the available radiomics features of the contralateral breast are the radiomics features after irrelevant features and redundant features are screened out; and, based on this, an embodiment is provided for obtaining the available radiomics features. Figure 5 As shown, in one embodiment, the above S201 includes the following steps:

[0100] S401, extracting tissue imaging features of the contralateral breast from breast glandular tissue and background parenchymal enhancement tissue through a feature extractor.

[0101] As described above, the feature extractor can be a feature extractor generated after the code is written according to the actual situation. For details, please refer to the description of the previous embodiment. Based on the feature extractor, the tissue imaging omics features of the contralateral breast are extracted from the breast glandular tissue and the background parenchymal enhanced tissue, that is, the tissue imaging omics features of the breast glandular tissue and the background parenchymal enhanced tissue of the contralateral breast are directly extracted from the medical image of the contralateral breast. The tissue imaging omics features here refer to the imaging omics features without screening out irrelevant features and redundant features.

[0102] S402, performing feature screening on tissue radiomics features to obtain usable radiomics features.

[0103] Then, the tissue imaging omics features are screened to remove irrelevant and redundant features, and the remaining ones are the usable imaging omics features.

[0104] In one embodiment, irrelevant features and redundant features in tissue radiomics features can be directly screened out through an algorithm model to obtain usable radiomics features.

[0105] In another embodiment, irrelevant features and redundant features in tissue radiomics features can be screened out by data preprocessing. Optionally, feature screening is performed on tissue radiomics features to obtain usable radiomics features, including: standardizing the tissue radiomics features; obtaining the variance of each feature in the standardized tissue radiomics features, and determining the features with variance greater than a threshold as usable radiomics features.

[0106] All features in the tissue imaging genomics features are normalized. Data normalization is to scale the data so that it falls into a small specific interval. That is, all feature values ​​in the tissue imaging genomics features are processed into the interval [0,1], and then the variance selection method is used to calculate the variance of each feature value after normalization, and the variance threshold is set for screening; taking the variance threshold of 0.1 as an example, the low-variance features below and equal to the variance threshold of 0.1 are eliminated, and all feature values ​​with variance greater than the variance threshold of 0.1 are retained. All retained features are usable imaging genomics features.

[0107] In the embodiment of the present application, the tissue imaging omics features of the contralateral breast are extracted from the breast glandular tissue and the background parenchymal enhancement tissue by a feature extractor, and the tissue imaging omics features are feature screened to obtain usable imaging omics features. Irrelevant features and redundant features in the tissue imaging omics features are screened out, and only better features are retained as usable imaging omics features. The usable imaging omics features used are optimized so that the usable imaging omics features can accurately determine the feature classification results of the target breast after entering the combined model.

[0108] The following is an explanation of the construction process of the above combination model, specifically how to determine the combination model that is the best combination of feature selection and classifier. Figure 6 As shown, in one embodiment, the construction process of the above combined model includes the following steps:

[0109] S501, obtaining sample medical images of breasts of multiple samples; the sample medical images include medical images of the contralateral breast of the breast affected by breast cancer when only one breast is affected by breast cancer; the sample medical images include sample breast glandular tissue and sample background parenchymal enhancement tissue.

[0110] This step is the process of obtaining model training data, that is, it is necessary to obtain a training data set. In order to make the training data set cover all situations as much as possible and improve the diversity of the training data set, it is necessary to obtain sample medical images of the breast of multiple samples. Among them, samples refer to different subjects, that is, one subject corresponds to one sample, and each sample can also select multiple sample medical images of the breast. For example, 100 sample medical images of the breast can be obtained, and the sample medical images obtained for each sample include three situations:

[0111] The first is that only one of the bilateral breasts of the sample has breast cancer; the second is that both of the bilateral breasts of the sample have breast cancer; the third is that neither of the bilateral breasts of the sample has breast cancer.

[0112] For the first type, when obtaining sample medical images, what needs to be obtained is the sample medical image of the contralateral breast of the breast suffering from breast cancer. For the second type, when obtaining sample medical images, what needs to be obtained is the sample medical image of either breast on both sides. For the third type, when obtaining sample medical images, what needs to be obtained is also the sample medical image of either breast on both sides. And no matter which one, the sample medical image obtained must include sample breast glandular tissue and sample background parenchymal enhanced tissue. It can be understood that, taking 100 samples as an example, each sample can obtain at least one of the above three types of sample medical images, and there is no limitation on this.

[0113] For example, the method for obtaining a sample medical image of the breast of the sample may be to select a medical image that meets the requirements from a pre-stored database; or to download a medical image that meets the requirements from a network platform; or to make a simulated medical image according to the requirements, etc. The embodiment of the present application does not limit the method for obtaining the sample medical image.

[0114] S502, based on the sample breast glandular tissue and the sample background parenchymal enhancement tissue in each sample medical image, obtaining the training radiomics feature of each sample medical image.

[0115] For the above-mentioned combined model, its input is the available imaging omics features of the breast, so the training imaging omics features of the combined model (i.e., training data) are also the available imaging omics features of the sample breast glandular tissue and the sample background parenchymal enhancement tissue extracted from the sample medical images of the above-mentioned samples. They are extracted from the sample medical images and used to train the combined model, so they are called training imaging omics features.

[0116] Among them, the training imaging genomics features of the sample breast glandular tissue and the sample background parenchymal enhanced tissue are obtained from each sample medical image, which can be extracted by using the feature extractor in the aforementioned embodiment, or by using the pre-processing and then extraction method in the aforementioned embodiment, and the embodiments of the present application are not limited to this.

[0117] For example, for the sample FGT and sample BPE tissues in the sample medical images of 100 samples, 1037 radiomics features were first extracted, including 14 shape features, 18 first-order features, 75 second-order features (texture features) and 930 filter features, as shown in Table 1 below:

[0118] Table 1

[0119] Category (Quantity) feature Shape characteristics(14) Volume, maximum diameter, surface area, major axis length, etc. First-order features (18) Mean, median, standard deviation, energy, entropy, etc. Texture features(75) GLCM, GLRLM, NGTDM, GLDM Filter Characteristics(930) Laplace-Gaussian filtering, wavelet

[0120] Among them, the Gray-level Co-occurrence Matrix (GLCM) in Table 1 is also called the Gray-level Co-occurrence Matrix, which is the texture of the image formed by the repeated appearance of pixel gray in spatial position. GLCM describes the joint distribution of two pixel grays with a certain spatial position relationship; Gray-Level Run-Length Matrix (GLRLM) is a matrix composed of the length of the gray value parade; Neighborhood gray-tone difference matrix (NGTDM) reflects the relationship between a pixel and its surrounding pixel values; Gray-level dependence matrix (GLDM) is a texture feature extraction based on the gray-level co-occurrence matrix. It can be understood that the features in Table 1 above are only an example and are not limited.

[0121] After extracting the features in Table 1 above, feature selection is required to eliminate irrelevant features, redundant features, and collinear features. For example, all features are first standardized so that all feature values ​​are in the interval [0,1], and then the variance selection method is used to calculate the variance of each feature value. Taking the variance threshold of 0.1 as the standard, all feature values ​​with variance greater than the threshold are retained, and all feature values ​​with variance less than or equal to the threshold are retained and eliminated, and the training radiomics features of each sample medical image can be obtained.

[0122] S503 , training a preset number of initial combination models according to the training radiomics features and the preset gold standard, to obtain candidate combination models corresponding to the initial combination models.

[0123] According to the training imaging features of each sample medical image and the preset gold standard, a preset number of initial combination models are trained to obtain candidate combination models corresponding to each initial combination model. Wherein, the gold standard is a preset standard value. For example, in the embodiment of the present application, the gold standard is the characteristic classification result of the target side breast. During the training process, the gold standard is used as a supervisory role to determine whether the initial combination model is trained and converged. Wherein, the gold standard can be obtained by taking the sample medical images of the breast of each sample as an example of a sample with only one breast suffering from breast cancer, and the characteristic classification of the breast of the breast cancer side of the sample is used as the gold standard. In this way, during training, the sample medical images of the breast of the breast cancer side of the sample on the opposite side are used as training data, and the gold standard for comparison is the real characteristic classification of the breast of the breast cancer side of the sample. In this way, the characteristic classification of the breast of the breast cancer side of the sample is used as the gold standard to supervise the output result of the training, so that the trained combination model can accurately determine the breast characteristic classification result of the target side breast based on the imaging features in the sample medical images of the contralateral breast.

[0124] Among them, the preset number of initial combination models refers to the pre-construction of multiple different initial combination models; for example, taking the combination model including feature selectors and classifiers as an example, and the feature selectors are LASS O, RFE, and RF; the classifiers are support vector machine, logistic regression, decision tree, gradient boosting decision tree and extreme gradient boosting; then the initial combination models are 15 combinations of these three feature selectors and five classifiers. Taking LASSO as an example, the initial combination models are: LASSO+support vector machine, LASSO+logistic regression, LASSO+decision tree, LASSO+gradient boosting decision tree, LASSO+extreme gradient boosting; other combinations are not listed here one by one. In short, the preset number is determined according to the number of feature selectors and classifiers.

[0125] After the initial combination model is determined, each initial combination model is trained according to the training radiomics features. During training, a five-fold cross-validation method can be used for training. For example, the training radiomics features of the above 100 samples of medical images are divided into five parts, each of which is the training radiomics features of medical images of 20 samples. During training, four of them are selected as training sets and one is used as a validation set. In this way, each initial combination model is trained with the training set and validated with the validation set; the next time, four different sets are re-selected as training sets, and the remaining one is used as a validation set, and each initial combination model is trained and validated; the training is repeated several times, and each time the selected training set and validation set are different. In this way, each initial combination model is trained, making each initial combination model more robust and generating more accurate breast feature classification results.

[0126] Each initial combination model is trained and supervised by the gold standard. When each initial combination model converges, a candidate combination model corresponding to each initial combination model is obtained.

[0127] It can be understood that the above-mentioned training imaging genomics features can also first use the univariate feature selection method to measure the correlation between the features and the benign and malignant nature of the tumor, eliminate the features with insignificant differences between the two groups of samples of benign tumors and malignant tumors, retain the features with p-value<0.05, and then train the initial combination model. The idea is the same as that of the previous embodiment and is not limited here.

[0128] S504: Determine a combination model from the candidate combination models according to the prediction performance indicators of the candidate combination models.

[0129] After the training of each initial combination model is completed, a corresponding candidate combination model is obtained, and the prediction performance indicators of different candidate combination models are analyzed. For example, the prediction performance indicator is the area under the curve (AUC) and the relative standard deviation (RSD). According to the prediction performance indicator, the combination model in the above embodiment is determined from each candidate combination model.

[0130] Wherein, AUC represents the area under the ROC curve and the coordinate axis. The ROC curve is the receiver operating characteristic curve (ROC). In combination with the embodiments of the present application, AUC represents the area under the curve generated by the sensitivity and specificity of the feature classification results output by each initial combination model in the training process; RSD represents the mean and standard deviation of the area under the curve of each initial combination model in the training process.

[0131] Based on this, Figure 7 As shown, according to the prediction performance indicators of each candidate combination model, an embodiment of determining a combination model from each candidate combination model includes the following steps:

[0132] S601, obtaining the area under the curve of each candidate combination model and the relative standard deviation of the area under the curve of each candidate combination model.

[0133] According to the sensitivity and specificity of the feature classification results output by the training process of each initial combination model, a corresponding ROC curve is generated, and the sensitivity and specificity of the feature classification results output by the training process can be directly obtained; according to the ROC curve, the corresponding AUC (i.e., the area under the curve) is determined; wherein each ROC curve corresponds to an AUC value, and the value of AUC is between 0 and 1. The closer the AUC is to 1, the higher the accuracy of the result output by the candidate combination model.

[0134] In addition, since the number of features used by different candidate combination models is not equal, RSD can be used as an auxiliary indicator to compare the stability of the classification results of the output features of each candidate combination model. RSD is defined as follows:

[0135]

[0136] Among them, mean AUC is the mean value of AUC; sd AUC is the standard deviation of AUC; and the smaller the RSD, the more stable the classification results of the output features of the candidate combination model.

[0137] S602, determining the candidate combination model whose area under the curve and relative standard deviation both meet preset conditions as the combination model.

[0138] According to the AUC (area under the curve) and RSD (relative standard deviation) of each candidate combination model obtained above, the candidate combination model whose AUC and RSD both meet the preset conditions can be determined as the combination model in the above embodiment. For example, the preset condition is that the AUC is closest to 1 and the RSD is the smallest, which means that the performance of the candidate combination model is the best. Then, the candidate combination model whose AUC is closest to 1 and the RSD is the smallest is determined as the final combination model, and the performance of the obtained combination model is also the best. In this way, it can be ensured that when the available imaging genomics features of the contralateral breast are predicted by the constructed combination model, an accurate and effective feature classification result of the target breast can be obtained.

[0139] For example, the following provides an example of the results of testing the above-mentioned candidate combination models using a test set, in which the performance is evaluated from the perspectives of accuracy, precision, sensitivity, specificity and AUC.

[0140] After testing each candidate combination model on each test set, the test results show:

[0141] (1) The performance of the candidate combination model LASSO+XGBoost (i.e., the extreme gradient boosting mentioned above) in the test process is consistent with the performance in the cross-validation during training, with an accuracy of 0.881, a precision of 0.873, and an AUC of 0.893. Its accuracy, precision, and AUC are all ranked first among all candidate combination models, while its sensitivity ranks seventh among all candidate combination models, and its specificity ranks second among all candidate combination models.

[0142] (2) The performance of the candidate combination model LASSO+GBDT (i.e., the above-mentioned gradient boosting decision tree) is slightly inferior to that of LASSO+XGBoost. The accuracy, precision, and AUC of the LASSO+GBDT combination model are 0.865, 0.869, and 0.878, ranking second among all candidate combination models. Its sensitivity is 0.816, ranking eighth among all candidate combination models. However, the LASSO+GBDT combination model has the highest specificity (0.850).

[0143] like Figure 8 As shown, Figure 8 Schematic diagram of ROC curves of each candidate combination model in the training process verification and test set test; specifically, Figure 8 (a) is a schematic diagram of the ROC curve of the combined model LASSO+XGBoos in the five-fold cross validation during the training process; Figure 8 (b) is a schematic diagram of the ROC curve of the combined model LASSO+XGBoos on the test set of this example; Figure 8 (c) is a schematic diagram of the ROC curve of the combined model LASSO+GBDT in the five-fold cross validation during the training process; Figure 8 (d) in the figure is a schematic diagram of the ROC curve of the combined model LA SSO+GBDT on the test set of this example.

[0144] Therefore, according to the performance of each candidate combination model in the test set, the combination model LASSO+GBDT and the combination model LASSO+XGBoost can achieve better performance in the task of determining the feature classification result of the target breast according to the medical image of the contralateral breast in the embodiment of the present application. In practical applications, both can be used as the final combination model, which can be determined according to the actual situation.

[0145] For the above-mentioned process of training a preset number of initial combination models according to each training imaging omics feature to obtain candidate combination models corresponding to each initial combination model, a training embodiment is provided, which includes: by executing a training step, each initial combination model in the preset number of initial combination models is trained to obtain a candidate combination model corresponding to each initial combination model; wherein the training step includes: for any initial combination model: inputting each training imaging omics feature into a feature selector in the initial combination model, and filtering out collinear features in each training imaging omics feature through the feature selector; inputting each training imaging omics feature after filtering out the collinear features into an initial classifier in the initial combination model, and training the initial classifier according to the gold standard to obtain a candidate combination model corresponding to the initial combination model.

[0146] When training is performed by the five-fold cross validation method, each initial candidate model is trained separately. Therefore, each initial candidate model is trained using the training steps to obtain a candidate combination model corresponding to each initial candidate model.

[0147] Taking one of the initial candidate models A as an example, the process of training the initial candidate model A using the training step includes: inputting each training radiomics feature into the feature selector in the initial combination model A, filtering out the collinearity features in each training radiomics feature through the feature selector, and after filtering out the collinearity features, inputting each training radiomics feature after filtering out the collinearity features into the initial classifier in the initial combination model A, and the initial classifier will output a feature classification result, repeating the execution, and each time the initial classifier will output a feature classification result, and each output feature classification result will be compared with the preset gold standard. If the similarity / closeness between the output feature classification result and the gold standard reaches a preset threshold, it is determined that the initial classifier training is completed, and at this time, the candidate combination model of the initial combination model A is obtained. In this way, the supervised training of the initial classifier according to the gold standard can make the classifier after iterative training more accurate, so that the candidate combination model corresponding to the initial combination model A can also obtain accurate and effective feature classification results of the target breast according to the available radiomics features of the contralateral breast.

[0148] It should be noted that the combined model provided in the embodiment of the present application, in addition to being used in the present application to generate characteristic classification results of the target breast, can also be used for other analyses based on breast tissue imaging genomics, such as predicting survival rate based on glandular tissue imaging genomics characteristics, etc., without limitation to this.

[0149] In addition, it is understandable that when applied to different imaging genomics analysis tasks, after the above training process, the final combination of feature selectors and classifiers selected may not be the same, that is, for different imaging genomics analysis tasks, the final combination model determined from the candidate combination models is also different, which depends on actual needs.

[0150] like Fig. 9 As shown, the embodiment of the present application also provides a feature classification method, which includes:

[0151] S1, obtain sample medical images of breasts of various samples.

[0152] S2, based on the sample breast glandular tissue and the sample background parenchymal enhancement tissue in each sample medical image, obtaining the training radiomics features of each sample medical image.

[0153] S3, training a preset number of initial combination models according to the features of each training imaging genomics model and a preset gold standard, and obtaining candidate combination models corresponding to each initial combination model.

[0154] S4, obtaining the area under the curve of each candidate combination model and the relative standard deviation of the area under the curve of each candidate combination model.

[0155] S5, determining the candidate combination model whose area under the curve and relative standard deviation both meet preset conditions as the combination model.

[0156] S6, obtaining a medical image of the contralateral breast of the target breast; the medical image includes breast glandular tissue of the contralateral breast and background parenchymal enhancement tissue.

[0157] S7, extracts tissue radiomic features of the contralateral breast from breast glandular tissue and background parenchymal enhancement tissue through a feature extractor.

[0158] S8, feature screening of tissue radiomics features to obtain usable radiomics features.

[0159] S9, through the feature selector, the collinear features in the available radiomics features are screened out.

[0160] S10, inputting the available radiomics features after filtering out the collinear features into the classifier to obtain the feature classification result of the target breast.

[0161] The feature classification method provided in this embodiment has similar implementation principles and technical effects as the feature classification methods in the aforementioned embodiments, and will not be described in detail here.

[0162] It should be understood that, although the various steps in the flow chart of the above-described embodiment are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flow chart of the above-described embodiment may include a plurality of steps or a plurality of stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of the steps or stages in other steps.

[0163] In one embodiment, Fig.10 As shown, the embodiment of the present application provides a feature classification device, which includes: an acquisition module 10 and a classification module 11; wherein,

[0164] An acquisition module 10 is used to acquire a medical image of the contralateral breast of the target breast; the medical image includes the breast glandular tissue of the contralateral breast and the background parenchymal enhancement tissue;

[0165] The classification module 11 is used to input the medical image of the contralateral breast into a preset machine learning model to obtain the feature classification result of the target breast.

[0166] In one embodiment, the machine learning model includes a feature extractor and a preset combination model; the classification module 11 includes: an extraction unit and a classification unit, wherein:

[0167] An extraction unit, used for extracting usable radiomics features of the contralateral breast from breast glandular tissue and background parenchymal enhancement tissue through a feature extractor;

[0168] The classification unit is used to input the available imaging omics features into the combined model to obtain the feature classification results of the target breast.

[0169] In one embodiment, the combined model includes a feature selector and a classifier; the above-mentioned classification unit is also used to screen out collinear features in available imaging genomics features through the feature selector; the available imaging genomics features after screening out the collinear features are input into the classifier to obtain the feature classification results of the target breast.

[0170] In one embodiment, the extraction unit comprises an extraction subunit and a screening subunit, wherein:

[0171] An extraction subunit, used for extracting tissue imaging features of the contralateral breast from breast glandular tissue and background parenchymal enhancement tissue through a feature extractor;

[0172] The screening subunit is used to screen out tissue radiomics features to obtain usable radiomics features.

[0173] In one embodiment, the above-mentioned screening subunit is also used to standardize the tissue imaging omics features; obtain the variance of each feature in the standardized tissue imaging omics features, and determine the features with variance greater than a threshold as available imaging omics features.

[0174] In one embodiment, the apparatus comprises:

[0175] A sample acquisition module is used to acquire sample medical images of breasts of multiple samples; the sample medical images include medical images of the contralateral breast of the breast affected by breast cancer when only one breast is affected by breast cancer; the sample medical images include sample breast glandular tissue and sample background parenchymal enhancement tissue;

[0176] A feature acquisition module, for acquiring training radiomics features of each sample medical image based on sample breast glandular tissue and sample background parenchymal enhancement tissue in each sample medical image;

[0177] A training module is used to train a preset number of initial combination models according to the characteristics of each training radiomics and a preset gold standard to obtain candidate combination models corresponding to each initial combination model;

[0178] The determination module is used to determine the combination model from the candidate combination models according to the prediction performance indicators of the candidate combination models.

[0179] In one embodiment, the training module includes: a training unit, the training unit being used to train each of a preset number of initial combination models by executing a training step to obtain a candidate combination model corresponding to each initial combination model;

[0180] The training step includes:

[0181] For any initial combination model: input each training radiomics feature into the feature selector in the initial combination model, and use the feature selector to screen out the collinear features in each training radiomics feature; input each training radiomics feature after screening out the collinear features into the initial classifier in the initial combination model, train the initial classifier according to the gold standard, and obtain the candidate combination model corresponding to the initial combination model.

[0182] In one embodiment, the predictive performance indicators include area under the curve and relative standard deviation;

[0183] The above-mentioned determination module includes: an indicator acquisition unit and a model determination unit, wherein:

[0184] An indicator acquisition unit, used to acquire the area under the curve of each candidate combination model and the relative standard deviation of the area under the curve of each candidate combination model;

[0185] The model determination unit is used to determine the candidate combination model whose area under the curve and relative standard deviation both meet preset conditions as the combination model.

[0186] For the specific definition of the feature classification device, please refer to the definition of the feature classification method above, which will not be repeated here. Each module in the above-mentioned feature classification device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0187] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Fig.11 As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a feature classification method is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0188] Those skilled in the art will understand that Fig.11 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0189] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0190] Acquire a medical image of the contralateral breast of the target breast; the medical image includes breast glandular tissue of the contralateral breast and background parenchymal enhancement tissue;

[0191] The medical image of the contralateral breast is input into the preset machine learning model to obtain the feature classification results of the target breast.

[0192] The above embodiment provides a computer device, whose implementation principle and technical effect are similar to those of the above method embodiment, and will not be repeated here.

[0193] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0194] Acquire a medical image of the contralateral breast of the target breast; the medical image includes breast glandular tissue of the contralateral breast and background parenchymal enhancement tissue;

[0195] The medical image of the contralateral breast is input into the preset machine learning model to obtain the feature classification results of the target breast.

[0196] The above embodiment provides a computer-readable storage medium, whose implementation principle and technical effect are similar to those of the above method embodiment, and will not be repeated here.

[0197] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0198] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0199] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A feature classification method, characterized in that: The method comprises: Acquire a medical image of the contralateral breast of the target breast; the medical image includes breast glandular tissue of the contralateral breast and background parenchymal enhancement tissue; Extracting tissue imaging features of the contralateral breast from the breast glandular tissue and the background parenchymal enhancement tissue through a feature extractor in a preset machine learning model; Standardizing the tissue radiomic features; Obtain the variance of each feature in the standardized tissue radiomics features, and determine the features with variance greater than a threshold as usable radiomics features; Screening out collinear features from the available radiomics features through a feature selector in the machine learning model; Inputting the available radiomics features after filtering out the collinear features into the classifier in the machine learning model to obtain the feature classification result of the target breast; The feature selector and classifier in the machine learning model form a combined model, and the construction process of the combined model includes: Obtaining sample medical images of breasts of multiple samples; the sample medical images include medical images of the contralateral breast of the breast affected by breast cancer when only one breast is affected by breast cancer; the sample medical images include sample breast glandular tissue and sample background parenchymal enhancement tissue; Based on the sample breast glandular tissue and the sample background parenchymal enhancement tissue in each of the sample medical images, obtaining the training radiomics features of each of the sample medical images; According to each of the training radiomics features and the preset gold standard, a preset number of initial combination models are trained to obtain candidate combination models corresponding to each of the initial combination models; The combination model is determined from the candidate combination models according to the prediction performance indicators of the candidate combination models.

2. The method according to claim 1, characterized in that The medical image of the contralateral breast is a DCE-MRI image.

3. The method according to claim 2, characterized in that The DCE-MRI images are images of the contralateral breast before and after contrast agent injection.

4. The method according to any one of claims 1 to 3, characterized in that: The background parenchymal enhancement tissue is obtained by subtracting the image before contrast agent injection and the image after contrast agent injection.

5. The method according to any one of claims 1 to 3, characterized in that: The mammary gland includes the breast and the gland.

6. The method according to any one of claims 1 to 3, characterized in that: The method of training a preset number of initial combination models according to the training imaging omics features and the preset gold standard to obtain candidate combination models corresponding to the initial combination models includes: By executing the training step, each of the preset number of initial combination models is trained to obtain a candidate combination model corresponding to each of the initial combination models; Wherein, the training step comprises: For any initial combination model: Inputting each of the training radiomics features into the feature selector in the initial combination model, and filtering out the collinearity features in each of the training radiomics features through the feature selector; Each of the training imaging omics features after screening out collinear features is input into the initial classifier in the initial combination model, and the initial classifier is trained according to the gold standard to obtain a candidate combination model corresponding to the initial combination model.

7. The method according to claim 6, characterized in that The prediction performance indicators include area under the curve and relative standard deviation; Then, determining the combination model from the candidate combination models according to the prediction performance indicators of the candidate combination models includes: Obtaining the area under the curve of each candidate combination model and the relative standard deviation of the area under the curve of each candidate combination model; The candidate combination model whose area under the curve and the relative standard deviation both meet preset conditions is determined as the combination model.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images