Breast Cancer Mass Diagnosis System Based on Dynamic Integration Selection Algorithm

Abstract

The present invention provides a breast cancer mass diagnosis system based on a dynamic ensemble selection algorithm, which relates to the field of medical technology. In the present invention, a training module is used to obtain a base classifier pool according to a training set by combining homogeneous classifier generation and heterogeneous classifier generation; a generation module is used to obtain a dynamic ensemble data set corresponding to a test set according to the test set and the base classifier pool; an acquisition module is used to obtain the dynamic ability region of any sample in the dynamic ensemble data set according to a dynamic selection set and the dynamic ensemble data set; a classification module is used to select a preset number of classifiers according to the base classifier pool and the dynamic ability region by using a complementary index, and obtain the diagnostic classification results corresponding to the current sample in the dynamic ensemble data set in each selected classifier. A base classifier pool of the DES algorithm is generated by combining homogeneous classifier generation and heterogeneous classifier generation; and by proposing a dynamic ability region (DROC) algorithm, the diagnostic accuracy of breast cancer masses is improved.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and particularly to a breast cancer mass diagnosis system based on a dynamic ensemble selection algorithm. Background Art

[0002] In recent years, using machine learning to assist doctors in diagnosing the conditions of breast cancer patients has become a research hotspot. Among them, the combination of breast cancer diagnosis and ensemble learning is the top priority of research. This is because the number of malignant breast cancer mass data is less than that of benign mass data, and there is a problem of data imbalance in breast cancer data. At the same time, for different breast cancer data sets, different classifiers yield different results, and no single classifier algorithm can be adapted to all breast cancer data sets. Ensemble learning can combine the results of different weak heterogeneous classifiers to obtain a strong classifier, which often achieves better results and stronger stability than a single classifier, and can also improve the generalization ability of the classifier.

[0003] Dynamic ensemble selection (DES) is a relatively advanced ensemble learning algorithm. The existing ways to generate classifiers by dynamic ensemble mainly include two types: homogeneous classifier generation and heterogeneous classifier generation. Homogeneous classifier generation means that in the process of generating classifiers, the same classifier algorithm is used, but for the training set, data is randomly sampled to form a new data set, and different base classifiers are trained. Heterogeneous classifier generation means that for the same training set, different classifier algorithms are used, and finally the results of each classifier are integrated to obtain the final prediction result.

[0004] However, the basic requirement for enhancing the generalization performance of the ensemble system is that the individual classifiers are as irrelevant as possible. However, homogeneous classifier generation uses the same algorithm, and even if the data sets are different or the algorithm parameters are different, it is impossible to ensure that all individual classifiers are irrelevant. Using homogeneous classifiers will limit the performance of the DES algorithm to a certain extent; while heterogeneous classifiers use different algorithms to generate a classifier pool. In order to ensure accuracy requirements, the classifier pool needs a certain number of classifiers in DES. If a large number of different classifier algorithms are selected to ensure accuracy, this will make the DES process complex and the classification algorithms too many will make the decision boundary too complex. In summary, the diagnostic accuracy of the existing breast cancer diagnosis models needs to be improved. Summary of the Invention

[0005] (1) Technical Problems to be Solved

[0006] Aiming at the deficiencies of the prior art, the present invention provides a breast cancer mass diagnosis system based on a dynamic ensemble selection algorithm, which solves the technical problem of the low diagnostic accuracy of the existing breast cancer diagnosis models.

[0007] (2) Technical Solutions

[0008] To achieve the above object, the present invention is realized through the following technical solutions:

[0009] A breast cancer mass diagnosis system based on a dynamic ensemble selection algorithm, comprising:

[0010] A preprocessing module, configured to obtain and preprocess a breast cancer mass diagnosis data set, and divide it into a training set, a test set, and a dynamic selection set;

[0011] A training module, configured to obtain a base classifier pool by combining the generation of homogeneous classifiers and the generation of heterogeneous classifiers according to the training set;

[0012] A generation module, configured to obtain a dynamic ensemble data set corresponding to the test set according to the test set and the base classifier pool;

[0013] An acquisition module, configured to obtain a dynamic ability region of any sample in the dynamic ensemble data set according to the dynamic selection set and the dynamic ensemble data set;

[0014] A classification module, configured to select a preset number of classifiers according to the complementary index according to the base classifier pool and the dynamic ability region, and obtain the diagnostic classification results corresponding to the current sample in the selected classifiers in the dynamic ensemble data set;

[0015] An integration module, configured to integrate each of the diagnostic classification results, obtain the diagnostic classification result of the current sample in the dynamic ensemble data set, and finally obtain the diagnostic classification results of all samples in the dynamic ensemble data set.

[0016] Preferably, the acquisition module includes a first acquisition sub-module, configured to use an adaptive mutual k-nearest neighbor algorithm to obtain a first ability region of any sample in the dynamic ensemble data set in the dynamic selection set; or configured to obtain the dynamic ability region;

[0017] The first acquisition sub-module is specifically configured to:

[0018] S411. For any sample in the dynamic ensemble data set, calculate the adaptive distance between the current sample in the dynamic ensemble data set and each sample in the dynamic selection set;

[0019]

[0020] R i = min d E (x dsel，i′ , x dsel，i ) + ε, y(x dsel，i′ ) ≠ y(x dsel，i ), x dsel，i′ , x dsel，i∈X dsel

[0021]

[0022]

[0023] where d ada (x dkes，j , x dsel，i ) is the adaptive distance between the j-th sample x dkes，j in the dynamic integrated dataset and any sample x dsel in the dynamic selection set X dsel，i ;

[0024] R i represents the minimum distance between any sample x dsel in the dynamic selection set X dsel，i and any other different-class sample x dsel in the dynamic selection set X dsel，i′ ;

[0025] d E (x dsel，i′ , x dsel，i ) represents the Euclidean distance between any sample x dsel in the dynamic selection set X dsel，i and any other different-class sample x dsel in the dynamic selection set X dsel，i′ ;

[0026] d E (x dkes，j , x dsel，i ) represents the Euclidean distance between the j-th sample x dkes，j in the dynamic integrated dataset and any sample x dsel in the dynamic selection set X dsel，i ;

[0027] where x dsel，i′，l , x dkes，j，l or x dsel，i，l represents the value of the corresponding sample on the preset feature l of the breast cancer mass diagnosis dataset, and the total number of features is M;

[0028] y(x dsel，i′ ) represents the diagnosis class corresponding to x dsel，i′ ;

[0029] y(x dsel，i ) represents the diagnosis class corresponding to x dsel，i ;

[0030] ε represents an arbitrarily small number;

[0031] S412. Arrange in non - decreasing order according to each of the adaptive distance magnitudes, and select the first K samples with smaller distances to form the third ability region of the current sample in the dynamic integrated dataset;

[0032] S413. Sequentially remove single samples in the third ability region from the dynamic selection set in the above arrangement order, and add them to the current sample in the dynamic integrated dataset to obtain a new dynamic selection set;

[0033] S414. Calculate the adaptive distances between the single samples removed from the third ability region and each sample in the new dynamic selection set, and select the K samples with smaller distances to form the fourth ability region of the removed sample;

[0034] S415. Determine whether the current sample in the dynamic integrated dataset is within the fourth ability region; if so, it is determined that the current sample in the dynamic integrated dataset and the single sample removed from the third ability region satisfy the mutual adaptive K - nearest neighbor relationship;

[0035] S416. Summarize the judgment results of all samples in the third ability region, and each sample that meets the requirements forms the first ability region of the current sample in the dynamic integrated dataset in the dynamic selection set;

[0036] Or if the number of samples in the first ability region obtained above is less than the preset quantity threshold, select the first preset quantity threshold of samples in the third ability region to form the final dynamic ability region.

[0037] Preferably, the acquisition module further includes a second acquisition sub - module, which is used to obtain the second ability region after removing samples far from the edge according to the first ability region by using the method of approaching the current sample with the center point of its sample set; or is used to obtain the dynamic ability region;

[0038] The second acquisition sub - module is specifically used for:

[0039] S421. Arrange each sample \(x\) in the first ability region \(AMDROC(x dkes，j |X dsel ) in non - decreasing order according to the adaptive distance from the current sample \(x\) in the dynamic integrated dataset, and sequentially calculate AMD，i , and the distance from the center point \(x dkes，j of the sample set where \(k'\in Z\), to obtain the set of distances from the center points of the sample sets where \(k'\) represents the \(k'\) - th sample in the first ability region, and the total number of samples is c，k′ , to obtain the set of distances from the center points of the sample sets

[0040] where \(k'\) represents the \(k'\) - th sample in the first ability region, and the total number of samples is Limk ad represents the preset quantity threshold; Z represents the set of integers;

[0041] The distance between the sample set and the center point x c，k′ The solution process includes:

[0042]

[0043] Among them, define x c，k′ to the sum of the Euclidean distances of other samples from AMDROC(x dkes，j |X dsel )[ : k'] is the smallest; d E (x i , x j ) represents the Euclidean distance between any sample x dkes，j |X dsel and sample x i in AMDROC(x j |X c，k′ S422. Calculate the Euclidean distance between each sample set and the current sample x dkes，j in the dynamic integrated data set, and record the k' corresponding to the minimum Euclidean distance as Conk. Then, obtain the second ability region SAMDROC(x dkes，j |X dsel ) after removing the samples at the distance edge, that is, SAMDROC(x dkes，j |X dsel ) = AMDROC(x dkes，j |X dsel

[0044] Or if the number of samples in the obtained second ability region is less than the preset quantity threshold, then take the samples of the first preset numerical threshold in the first ability region to form the final dynamic ability region.

[0045] Preferably, the obtaining module further includes a third obtaining sub-module, which is used to obtain the dynamic ability region after removing the edge category samples according to the second ability region; or is used to obtain the dynamic ability region;

[0047] The third obtaining sub-module specifically is used for:

[0048] Define the class corresponding to the i-th sample i in the second ability region SAMDROC(x SMAD，i ) as y = class(x dkes，j ) dsel ;

[0049] If Then, obtain the dynamic ability region TAMDROC(x after excluding the edge class samples dkes，j |X dsel ) = SAMDROC(x dkes，j |X dsel ) - x SMAD，i ;

[0050] Where, num(y i |SAMDROC(x dkes，j |X dsel )) represents the number of classes of the diagnostic class y i in the second ability region SAMDROC(x dkes，j |X dsel ); represents the number of samples of the second ability region SAMDROC(x dkes，j |X dsel ); marginrate represents the edge class sample rate;

[0051] Or if the number of samples in the obtained dynamic ability region is less than the preset number threshold, then take the samples of the first preset number threshold in the second ability region to form the final dynamic ability region.

[0052] Preferably, the pre-training module is specifically configured to:

[0053] S11. Use BI-RADS grading to represent the corresponding features and mass diagnosis levels of the initial breast cancer mass diagnosis dataset, calculate the interval number distance between different gradings and the minimum grade 3, and perform standardization processing on the obtained results. The features include boundary clarity, edge regularity, internal echo, calcification, and blood flow abundance;

[0054] S12. Use the oversampling method to balance the classes of the initial breast cancer mass diagnosis dataset;

[0055] S13. Divide the processed breast cancer mass diagnosis dataset into the training set, test set, and dynamic selection set.

[0056] Preferably, the training module is specifically configured to:

[0057] Select several different classifier algorithms to form a base classifier algorithm pool Use the Bagging algorithm to perform multiple samplings on the training set X train to generate different sample sets For the same classifier algorithm c m b By training different sample sets obtain different isomorphic classifiers cm b，o , and finally generate the base classifier pool

[0058] Preferably, the generating module is specifically configured to:

[0059] S31. For the test set X test and the j-th sample x to be tested in it test，j , use the base classifier pool For the current sample x in the dynamic integration dataset dkes，j perform classification to obtain the classification result set C(x test，j ) = {c 1 (x test，j ), c 2 (x test，j ),......, c T (x test，j )};

[0060] S32. According to the classification result set C(x test，j ), calculate the classification consensus degree of the sample x to be tested test，j ,

[0061]

[0062] where H(x test，j , C) represents the proportion of the maximum number of votes corresponding to the voting method in the classification result set C(x test，j ), where Voting(C(x test，j )) is the number of votes corresponding to each category in C(x test，j ); T represents the total number of classifiers in the base classifier pool;

[0063] S33. Predetermine the consensus screening criterion Cond. If H(x test，j , C) > Cond, add the sample x test，j to the dynamic integration dataset X dkes ;

[0064] If H(x test，j , C) ≤ Cond, obtain the diagnostic classification result of the sample x to be tested test，j

[0065] where argclass max(Voting(C(x test，j ))) represents the category with the maximum number of votes in C(x test，j ).

[0066] Preferably, the classification module is specifically configured to:

[0067] S51. Adopt a base classifier pool For the dynamic ability region TAMDROC(x dkes，j |X dsel ), obtain the accuracy rate set Place the classifier ranked first at the first position of the classifier order set U;

[0068] S52. Define C′ = C - U, and select classifiers from C′ one by one to calculate the overall complementary index with the classifier order set U, and obtain the overall complementary index set of classifiers

[0069] where T′ represents the number of classifiers in C′, and select the classifier corresponding to the maximum value of the overall complementary index as the classifier ranked i, that is C t ′ ∈ C is added to the classifier order set U;

[0070] The solution process of the overall complementary index TCI(c t ′) includes

[0071]

[0072]

[0073] where c i and c j respectively represent two different classifiers in the base classifier pool ; c i and c j For the dynamic ability region TAMDROC(x dkes，j |X dsel ), the misclassified sets are S i and S j respectively; D(S) represents the number of samples in S; the intersection of S i and S j is S i ∩S j , and the union is S i ∪S j ; G ij represents the complementary index of classifiers c i and c j ;

[0074] S53. Repeat the above operations until the classifier order set contains all the classifiers in the base classifier pool, that is Select the first η classifiers in the finally obtained U according to the preset parameter η;

[0075] S54. Obtain the current sample x in the dynamic integration dataset according to the selected top η classifiers dkes，j for the corresponding respective diagnostic classification results

[0076] Preferably, the integration module is used to obtain the diagnostic classification result of the current sample in the dynamic integration dataset, specifically including:

[0077] Classify the current sample x dkes，j to obtain the result {e 1 (x dkes，j ),......, e η (x dkes，j )}, and use the majority voting method to obtain the final diagnostic classification result arg class(max(Voting(e dkes，j (x i (x dkes，j )))), where i = 1,......, η

[0078] Preferably, the breast cancer mass diagnosis system further includes:

[0079] A merging module, configured to merge the diagnostic classification results corresponding to some samples in the test set obtained by the generating module with the diagnostic classification results corresponding to all samples in the dynamic integration dataset obtained by the integration module, to obtain the diagnostic classification result corresponding to each sample in the test set

[0080] (III) Beneficial effects

[0081] The present invention provides a breast cancer mass diagnosis system based on a dynamic integration selection algorithm. Compared with the prior art, it has the following beneficial effects:

[0082] In the present invention, a preprocessing module is used to obtain and preprocess a breast cancer mass diagnosis dataset, and divide it into a training set, a test set, and a dynamic selection set; a training module is used to obtain a base classifier pool by combining the generation of homogeneous classifiers and the generation of heterogeneous classifiers according to the training set; a generation module is used to obtain a dynamic integration dataset corresponding to the test set according to the test set and the base classifier pool; an acquisition module is used to obtain the dynamic ability region of any sample in the dynamic integration dataset according to the dynamic selection set and the dynamic integration dataset; a classification module is used to select a preset number of classifiers according to the base classifier pool and the dynamic ability region by using a complementary index, and obtain the diagnostic classification results corresponding to the current sample in the selected classifiers in the dynamic integration dataset; an integration module is used to integrate the diagnostic classification results to obtain the diagnostic classification result of the current sample in the dynamic integration dataset, and finally obtain the diagnostic classification results of all samples in the dynamic integration dataset. The base classifier pool of the DES algorithm is generated by combining the generation of homogeneous classifiers and the generation of heterogeneous classifiers; and by proposing the dynamic ability region (DROC) algorithm, the diagnostic accuracy of breast cancer masses is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to these drawings.

[0084] Figure 1 It is a structural block diagram of a breast cancer mass diagnosis system based on a dynamic integration selection algorithm provided by an embodiment of the present invention;

[0085] Figure 2 It is an example of a sample away from the edge provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0086] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described below. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0087] By providing a breast cancer mass diagnosis system based on a dynamic integration selection algorithm in the embodiments of the present application, the technical problem of low diagnostic accuracy of existing breast cancer diagnosis models is solved.

[0088] The technical solution in the embodiments of this application is to solve the above technical problems, and the general idea is as follows:

[0089] In the embodiments of the present invention, a preprocessing module is used to obtain and preprocess a breast cancer mass diagnosis dataset, and divide it into a training set, a test set, and a dynamic selection set; a training module is used to obtain a base classifier pool by combining homogeneous classifier generation and heterogeneous classifier generation according to the training set; a generation module is used to obtain a dynamic integration dataset corresponding to the test set according to the test set and the base classifier pool; an acquisition module is used to obtain the dynamic ability region of any sample in the dynamic integration dataset according to the dynamic selection set and the dynamic integration dataset; a classification module is used to select a preset number of classifiers according to the base classifier pool and the dynamic ability region by using a complementary index, and obtain the diagnostic classification results corresponding to the current sample in the selected classifiers in the dynamic integration dataset; an integration module is used to integrate each of the diagnostic classification results to obtain the diagnostic classification result of the current sample in the dynamic integration dataset, and finally obtain the diagnostic classification results of all samples in the dynamic integration dataset. A base classifier pool of the DES algorithm is generated by combining homogeneous classifier generation and heterogeneous classifier generation; and by proposing a dynamic ability region (DROC) algorithm, the diagnostic accuracy of breast cancer masses is improved.

[0090] To better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments.

[0091] Embodiment:

[0092] As Figure 1 shown, the embodiments of the present invention provide a breast cancer mass diagnosis system based on a dynamic integration selection algorithm, including:

[0093] A preprocessing module is used to obtain and preprocess a breast cancer mass diagnosis dataset, and divide it into a training set, a test set, and a dynamic selection set;

[0094] A training module is used to obtain a base classifier pool by combining homogeneous classifier generation and heterogeneous classifier generation according to the training set;

[0095] A generation module is used to obtain a dynamic integration dataset corresponding to the test set according to the test set and the base classifier pool;

[0096] An acquisition module is used to obtain the dynamic ability region of any sample in the dynamic integration dataset according to the dynamic selection set and the dynamic integration dataset;

[0097] A classification module, configured to select a preset number of classifiers according to the base classifier pool and the dynamic ability region by using a complementary index, and obtain the diagnostic classification results corresponding to the current sample in the dynamic integration dataset in each of the selected classifiers;

[0098] An integration module, configured to integrate each of the diagnostic classification results, obtain the diagnostic classification result of the current sample in the dynamic integration dataset, and finally obtain the diagnostic classification results of all samples in the dynamic integration dataset.

[0099] In the embodiment of the present invention, a base classifier pool of the DES algorithm is generated by combining homogeneous classifier generation and heterogeneous classifier generation; and by proposing a dynamic ability region (DROC) algorithm, the diagnostic accuracy of breast cancer masses is improved.

[0100] The following will introduce each step of the above solution in detail with specific examples:

[0101] A preprocessing module, configured to obtain and preprocess a breast cancer mass diagnosis dataset, and divide it into a training set, a test set, and a dynamic selection set; specifically:

[0102] S11. Use BI-RADS grading to represent the corresponding features and mass diagnosis levels of the initial breast cancer mass diagnosis dataset, calculate the interval number distance between different grades and the minimum grade 3, and perform normalization processing on the obtained results. The features include boundary clarity, edge regularity, internal echo, calcification, and blood flow abundance.

[0103] In the embodiment of the present invention, historical diagnosis reports of four ultrasound doctors in the ultrasound room of a certain tertiary hospital in Hefei, Anhui are collected and sorted from information systems such as the electronic medical record system and the report system (partial data are shown in Tables 1 and 2). Then, the breast cancer mass diagnosis dataset in the historical diagnosis reports is standardized. First, the descriptions of the diseases of breast cancer patients by the ultrasound doctors are feature-extracted, and five features including boundary clarity, edge regularity, internal echo, calcification, and blood flow abundance and the mass diagnosis grade are extracted.

[0104] Table 1 Partial data display diagram

[0105]

[0106] A brief introduction to the breast cancer mass diagnosis data of four doctors is shown in Table 2.

[0107] Table 2 Brief introduction to the diagnostic data of four doctors

[0108]

[0109]

[0110] Both the features and the mass diagnosis grades are represented by BI-RADS grades. BI-RADS is a commonly used clinical diagnosis grading standard for breast cancer masses. Each grade can be converted into an interval number, indicating the degree to which doctors suspect the mass is malignant, as shown in Table 1.

[0111] Table 3 Correspondence Table between BI-RADS Grades and Interval Numbers

[0112] Diagnostic level Probability interval Conversion value 3 (0，2％] 0 4A [3％，30％] 0.1798 4B [31％，60％] 0.4692 4C [61％，94％] 0.7988 5 [95％，100％] 1

[0113] First, assume that two interval numbers a and b are given. υ = a ∩ b = [υ - , u + represents the intersection of the two interval numbers W(a) = b + - b - , W(b) = b + - b - , W(u) = u + - υ - , the distance d IN (a, b) of the two interval numbers is calculated as shown in Equation (1).

[0114]

[0115] Combining Table 1 and Equation 1, calculate the interval number distances between different grades and the minimum grade 3. Convert the grade data {3, 4A, 4B, 4C, 5} corresponding to the five features of boundary clarity, edge regularity, internal echo, calcification, and blood flow abundance into [0, 0.1735, 0.4528, 0.7709, 0.9651]. On this basis, perform standardization processing on the data and convert it into [0, 0.1798, 0.4692, 0.7988, 1]. Through the above process, the standardization of the data is achieved, and all feature data values are converted into numerical values between [0, 1] to obtain the dataset X d , d = 1, 2, 3, 4, where d represents different doctors.

[0116] Part of the converted data is shown in Table 4.

[0117] Table 4 Part of the Converted Data

[0118]

[0119] S12. Use the oversampling method to balance the categories of the initial breast cancer mass diagnosis dataset.

[0120] The imbalance of the dataset means that the number of samples with different class labels varies greatly. When using a classifier algorithm for classification, the imbalance of the dataset will make the established classifier more inclined to the class with a larger number of samples, affecting the performance of the classifier.

[0121] Since the proportion of severe cases in life is much smaller than that of mild cases, the imbalance is more serious in medical data. The model proposed in the embodiments of the present invention takes the breast cancer mass diagnosis dataset as the research object, and there is an imbalance problem. The oversampling method, such as the BorderlineSMOTE method, is used to balance the classes of the breast cancer mass diagnosis dataset X.

[0122] S13. Divide the processed breast cancer mass diagnosis dataset into the training set, the test set, and the dynamic selection set.

[0123] The processed data is divided into the training set X train , the dynamic selection set X dsel and the test set X test . Specifically, the processed dataset of each doctor is divided into X d train , X d dsel and X d test in the ratio of 5∶3∶2. Considering that the number of diagnosis categories is 5, the minimum threshold of the ability region of the input sample to be measured can be set as Limk ad =5, the number of dynamic selection classifiers η = 30, and the initial number of samples in the ability region K = 25.

[0124] The training module is used to obtain a base classifier pool by combining the generation of homogeneous classifiers and the generation of heterogeneous classifiers according to the training set; specifically for:

[0125] Select several different classifier algorithms to form a base classifier algorithm pool Use the Bagging algorithm, such as Bootstrap sampling, to sample the training set X train multiple times to generate different sample sets The same classifier algorithm c m b By training different sample sets obtain different homogeneous classifiers c m b，o , and finally generate the base classifier pool

[0126] For example, the classifier algorithms at least include LB, GNB, SVM, MLP, LDA or AB algorithms. That is, in the embodiments of the present invention, six algorithms such as logistic regression (LR), Gaussian Naive Bayes (GNB), Support vector machine (SVM), Multi-layer Perceptron classifier (MLP), Linear Discriminant Analysis (LDA), and AdaBoost (AB) are selected as the base classifier algorithms. Some parameter settings of the base classifier are shown in Table 5. The parameters not shown in Table 5 use the default parameters of the scikit-learn package in Python.

[0127] Table 5 Parameter Settings Table of Base Classifier Algorithms

[0128]

[0129] The reason for choosing the above method of combining homogeneous classifier generation and heterogeneous classifier generation is that for the analysis of the generalization error of the ensemble classifier, the basic requirement for enhancing the generalization performance of the ensemble system is that the individual classifiers are as irrelevant as possible. However, homogeneous classifier generation uses the same algorithm, and even if the data sets are different or the algorithm parameters are different, it is impossible to ensure that all individual classifiers are irrelevant. Using homogeneous classifiers will limit the performance of the DES algorithm to a certain extent.

[0130] Heterogeneous classifiers use different algorithms to generate a classifier pool. In order to ensure the accuracy requirement of DES, the classifier pool needs a certain number of classifiers. However, if a large number of different classifier algorithms are selected to ensure accuracy, this will make the DES process complex and too many classification algorithms will make the decision boundary too complex.

[0131] Therefore, in view of the problems that the correlation between the classifiers generated by the existing DES algorithm using homogeneous classifiers is strong and the generation of heterogeneous classifiers requires more algorithms, the embodiments of the present invention adopt a method of combining two methods of homogeneous classifier generation and heterogeneous classifier generation.

[0132] In addition, in the existing DES algorithm, the methods for determining the region of competence (ROC) of the sample to be tested, whether it is the K-nearest neighbor or the natural K-nearest neighbor method, are composed of fixed-value samples. However, due to the different characteristic situations of the sample to be tested and its dynamically selected set, for different input samples, selecting the same K value to form the region of competence cannot ensure that the similarity between the samples in the region of competence and the sample to be tested can achieve the purpose of selecting a classifier. For example, some samples to be tested can search for more nearest neighbors within a short distance, but some samples to be tested must expand the search distance for nearest neighbors in order to reach the K value, resulting in too large a distance between the nearest neighbor samples, and the formed region of competence cannot achieve the expected effect.

[0133] Determining the region of competence is the basis for selecting a classifier in the DES algorithm. If there are problems in the selection of the region of competence, a classifier with high performance cannot be selected, ultimately affecting the performance of the DES algorithm. Based on this, the embodiments of the present invention provide the following method: obtaining a dynamic region of competence through the generating module and the obtaining module. Specifically, it includes:

[0134] The generating module is used to obtain the dynamic integrated data set corresponding to the test set according to the test set and the base classifier pool; specifically used for:

[0135] S31. For the test set X test the jth sample x test，j to be tested in it, use the base classifier pool to classify the current sample x dkes，j in the dynamic integrated data set, and obtain the classification result set C(x test，j ) = {c 1 (x test，j ), c 2 (x test，j ),......, c T (x test，j )}.

[0136] S32. According to the classification result set C(x test，j ), calculate the classification consensus degree of the sample x test，j to be tested according to formula (2),

[0137]

[0138] where H(x test，j , C) represents the proportion of the maximum number of votes corresponding to the voting method in the classification result set C(x test，j ), where Voting(C(x test，j )) is C(x test，j) The number of votes corresponding to each category; T represents the total number of classifiers in the base classifier pool.

[0139] S33. Predetermine the consensus screening criterion Cond. If H(x test，j , C) > Cond, add the sample x test，j to the dynamic integration dataset X dkes ;

[0140] If H(x test，j , C) ≤ Cond, obtain the diagnostic classification result of the sample x test，j to be measured

[0141] Among them, argclass max(Voting(C(x test，j ))) represents the category with the maximum number of votes in C(x test，j ). In the embodiment of the present invention, the consensus screening criterion Cond is taken as 0.9.

[0142] It can be understood that when the sample to be measured in the test sample is not selected into the dynamic integration dataset, the diagnostic classification result of this part of the sample to be measured can be directly determined by the above method.

[0143] An acquisition module, configured to acquire the dynamic ability region of any sample in the dynamic integration dataset according to the dynamic selection set and the dynamic integration dataset.

[0144] It is emphasized again that in existing research, the determination method of the ability region is composed of fixed values. However, this determination method only selects samples unidirectionally according to the distance. This selection method is only unidirectional and very one-sided. Because the ability region obtained in this way is only the nearest neighbor relationship considered from the perspective of the sample to be measured, without considering the characteristic situation of different samples and the distribution situation of the dynamic selection set data, it may expand the similarity search interval in order to reach the number of nearest neighbors and select some marginal samples.

[0145] For example, as Figure 2 shown, given that the number of samples in the ROC is 5, the ability region of x 1 includes x 2 , but obviously, x 2 is farther away from x 1 compared with other samples in the ROC and does not meet the requirements of the samples constituting the ability region. Therefore, x 2 needs to be excluded.

[0146] Correspondingly, the acquisition module includes a first acquisition sub-module, configured to use the adaptive mutual K-nearest neighbor algorithm to acquire the first ability region of any sample in the dynamic integration dataset in the dynamic selection set; or configured to acquire the dynamic ability region;

[0147] The first acquisition sub-module is specifically configured to:

[0148] S411. For any sample in the dynamic integration dataset, calculate the adaptive distance between the current sample in the dynamic integration dataset and each sample in the dynamic selection set;

[0149]

[0150] R i = min d E (x dsel，i′ , x dsel，i ) + ε, y(x dsel，i′ ) ≠ y(x dsel，i ), x dsel，i′ , x dsel，i ∈X dsel (4)

[0151]

[0152]

[0153] where d ada (x dkes，j , x dsel，i ) is the adaptive distance between the jth sample x dkes，j in the dynamic integration dataset and any sample x dsel in the dynamic selection set X dsel，i ;

[0154] R i represents the minimum distance between any sample x dsel in the dynamic selection set X dsel，i and any other different-class sample x dsel in the dynamic selection set X dsel，i′ ;

[0155] d E (x dkes，j , x dsel，i ) represents the Euclidean distance between the jth sample x dkes，j in the dynamic integration dataset and any sample x dsel in the dynamic selection set X dsel，i ;

[0156] d E (x dsel，i′ , x dsel，i ) represents the Euclidean distance between any sample x dsel in the dynamic selection set X dsel，i and any sample x dselAny other different category sample x among them dsel，i′ The Euclidean distance between;

[0157] where x dsel，i′ , x dkes，j，l or x dsel，i，l represents the value of the corresponding sample on the preset feature l of the breast cancer mass diagnosis dataset, and the total number of features is M; based on the initial data processed by the above preprocessing module, it is known that M = 5;

[0158] y(x dkes，j ) represents the diagnostic category corresponding to x dkes，j ;

[0159] y(x dsel，i ) represents the diagnostic category corresponding to x dsel，i ;

[0160] ε represents an arbitrarily small number, and in the embodiment of the present invention, ε = 0.00001 is taken.

[0161] S412. Arrange in non - decreasing order according to the size of each of the adaptive distances, and take the first K samples with smaller distances to form the third ability region of the current sample in the dynamic integration dataset; that is, the initial ability region ADROC(x dkes，j |X dsel );

[0162] S413. Sequentially remove a single sample in the third ability region from the dynamic selection set in the above arrangement order, and add it to the current sample in the dynamic integration dataset to obtain a new dynamic selection set.

[0163] That is, remove the k - th sample point x dkes，j |X dsel ) in ADROC(x dkes，j，k A , k = 1,......, K, from the dynamic selection set X dsel Remove x dkes，j，k A , and add it to x dkes，j The sample points form x dkes，j k The new dynamic selection set, that is, X dsel ′ = {(X dsel -x dkes，j，k A ) ∪ x dkes，j};

[0164] S414. Similarly referring to the processing ideas provided by formulas (3) - (6), calculate the adaptive distances of the single sample removed in the third ability region from each sample in the new dynamic selection set, and take the K samples with smaller distances to form the fourth ability region of the removed sample.

[0165] That is, calculate x dkes，j，k A In X dsel ′, and take K samples with smaller distance to form x dkes，j，k A X-based dsel 'The fourth capability region ADROC(x dkes，j，k A |X dsel ),

[0166] S415. Determine whether the current sample in the dynamic integrated data set is within the fourth capability region; if so, determine that the current sample in the dynamic integrated data set and the single sample removed from the third capability region satisfy a mutually adaptive K-nearest neighbor relationship.

[0167] S416, summarizing the judgment results of all samples in the third capability area, and each sample that meets the requirements constitutes the first capability area of ​​the current sample in the dynamic integrated data set in the dynamic selection set;

[0168] Alternatively, if the number of samples in the first capability region obtained above is less than a preset number threshold, samples before the preset number threshold of the third capability region are taken to form the final dynamic capability region.

[0169] That is, if x dkes，j ∈ADROC(x dkes，j，k A |X dsel ′), indicating that x dkes，j With x dkes，j，k A In X dsel In the adaptive K-nearest neighbor relationship, x dkes，j，k A Add to the first capability region AMDROC(x dkes，j |X dsel )middle;

[0170] Considering the first capability region AMDROC(x dkes，j |X dsel ) is the number of samples in It may be 0, and you need to set a preset quantity threshold Limk ad .like This means that AMDROC(x dkes，j |X dsel ) The number of samples does not meet the requirements of the composition capability area, and the third capability area ADROC (x dkes，j |X dsel ) ex Limkad The values form AMDROC(x dkes，j |X dsel ), that is, AMDROC(x dkes，j |X dsel ) = ADROC(x dkes，j )[:Limk ad , and it is used as the final dynamic capability region mentioned above.

[0171] When the first acquisition sub-module belongs to the case of acquiring the first capability region; the acquisition module further includes a second acquisition sub-module, configured to, according to the first capability region, adopt the method of approaching the current sample by the distance from the center point of its sample set, acquire a second capability region after removing samples far from the edge; or acquire the dynamic capability region.

[0172] It should be noted that in the embodiments of the present invention, it is considered that the case where the sample set of the first capability region is completely coincident with the center point x c (AMDROC) and x dkes，j is the most ideal case (the detailed content about the distance from the center point of the sample set will be introduced in detail later), but there may be some samples x dkes，j in AMDROC(x dsel |X AMD，i (x AMD，i ∈AMDROC(x dkes，j |X dsel )) such that the distance between x c (AMDROC) and x dkes，j is too large. At this time, it is considered that x AMD，i satisfying the above situation is a point far from the edge, and the existence of points far from the edge increases the overall difference between x dkes，j and AMDROC(x dkes，j |X dsel ). Therefore, in order to obtain a capability region with smaller difference from x dkes，j , it is necessary to remove points far from the edge.

[0173] Correspondingly, the second acquisition sub-module is specifically configured to:

[0174] S421. Arrange each sample x dkes，j in AMDROC(x dsel |X AMD，i ) in non-decreasing order according to the adaptive distance from the current sample x dkes，j in the dynamic integration dataset, and calculate and k'∈Z for the distance from the center point x c，k′ of the sample set, and obtain the set of distances from the center point of the sample set

[0175] Among them, k′ represents the k′-th sample in the first ability region, and the total number of samples is Limk ad represents the preset quantity threshold; Z represents the set of integers;

[0176] The distance between the sample set and the center point x c，k′ The solution process includes:

[0177]

[0178] Among them, it is defined that x c，k′ to the Euclidean distance sum of other samples in AMDROC(x dkes，j |X dsel )[:k′] is the smallest; d E (x i , x j ) represents the Euclidean distance between any sample x dkes，j |X dsel )[:k′] and the sample x i and the sample x j , and the total number of samples is k′.

[0179] S422. Calculate the Euclidean distance between each sample set and the current sample x c，k′ in the dynamic integrated data set, and record the k′ corresponding to the minimum Euclidean distance as Conk. Then, obtain the second ability region SAMDROC(x dkes，j |X dkes，j |X dsel ) = AMDROC(x dkes，j |X dsel )[:Conk];

[0180] Or if the number of samples in the obtained second ability region is less than the preset quantity threshold, then take the samples of the first preset numerical threshold in the first ability region to form the final dynamic ability region.

[0181] Considering that the number of samples in the second ability region SAMDROC(x dkes，j |X dsel ) obtained by the above process is and may not reach the preset quantity threshold Limk ad for forming the ability region. If then SAMDROC(x dkes，j |X dsel ) takes the first Limk samples in the first ability region AMDROC(x dkes，j |X dsel ) advalue, i.e., SAMDROC(x dkes，j |X dsel ) = AMDROC(x dkes，j |X dsel )[:Limk ad , as the final dynamic capability region described above.

[0182] When the second acquisition sub-module belongs to the case of acquiring the second capability region; the acquisition module further includes a third acquisition sub-module, configured to acquire the dynamic capability region after removing the marginal class samples according to the second capability region; or configured to acquire the dynamic capability region.

[0183] The reason for introducing the third acquisition sub-module is that considering the second capability region SAMDROC(x dkes，j |X dsel ), there may still be such a situation. As can be seen from Table 6, the distance between x dkes，j ' and all samples in SAMDROC(x dkes，j |X dsel ) is 0, but the diagnostic results of SAMDROC(x dkes，j |X dsel ) are quite different. However, for any classifier algorithm, there can only be one prediction result for all SAMDROC(x dkes，j |X dsel ) samples. The samples in SAMDROC(x dkes，j |X dsel ) are all the neighbor samples of x dkes，j , which means that there should not be a large gap in their classes. However, the classes of the four samples 1, 8, 10, and 12 in SAMDROC(x dkes，j |X dsel ) are different from those of most samples and account for a small proportion. They are defined as marginal class samples. The present invention believes that the existence of marginal class samples will to a certain extent reduce the accuracy of the classifier in SAMDROC(x dkes，j |X dsel ). However, accuracy is an important indicator for classifier selection and may lead to the inability to select a suitable classifier in the dynamic selection process of the classifier, affecting the performance of the algorithm. Therefore, it is necessary to eliminate the influence of marginal class samples.

[0184] Table 6 Example of SAMDROC(x dkes，j |X dsel )

[0185] Input sample Feature 1 Feature 2 Feature 3 Feature 4 Feature 5 Diagnostic result <![CDATA[x dkes，j ′]]> 0.1798 0.7988 0.4692 0.4692 0.4692 SAMDROC Feature 1 Feature 2 Feature 3 Feature 4 Feature 5 Diagnostic result 1 0.1798 0.7988 0.4692 0.4692 0.4692 4 2 0.1798 0.7988 0.4692 0.4692 0.4692 3 3 0.1798 0.7988 0.4692 0.4692 0.4692 3 4 0.1798 0.7988 0.4692 0.4692 0.4692 3 5 0.1798 0.7988 0.4692 0.4692 0.4692 3 6 0.1798 0.7988 0.4692 0.4692 0.4692 3 7 0.1798 0.7988 0.4692 0.4692 0.4692 3 8 0.1798 0.7988 0.4692 0.4692 0.4692 5 9 0.1798 0.7988 0.4692 0.4692 0.4692 3 10 0.1798 0.7988 0.4692 0.4692 0.4692 5 11 0.1798 0.7988 0.4692 0.4692 0.4692 3 12 0.1798 0.7988 0.4692 0.4692 0.4692 1

[0186] Correspondingly, the third acquisition sub-module is specifically configured to:

[0187] Define the $i$-th sample in the second ability region $SAMDROC(x$ dkes，j $|X$ dsel ) as corresponding to the category $y$ i $= class(x$ SMAD，i )$;

[0188] If then obtain the dynamic ability region $TAMDROC(x$ dkes，j $|X$ dsel ) after removing the marginal category samples, where $TAMDROC(x$ dkes，j $|X$ dsel ) = $SAMDROC(x$ SMAD，i ) - $x$

[0189] where $num(y$ i $|SAMDROC(x$ dkes，j $|X$ dsel )) represents the number of categories of the diagnostic category $y$ i in the second ability region $SAMDROC(x$ dkes，j $|X$ dsel ); represents the number of samples in the second ability region $SAMDROC(x$ dkes，j $|X$ dsel ); marginrate represents the marginal category sample rate, and in the embodiments of the present invention, $marginrate = 0.2$;

[0190] Or if the number of samples in the obtained dynamic ability region is less than the preset number threshold, then take the first preset number threshold of samples in the second ability region to form the final dynamic ability region.

[0191] Considering that the number of samples in the dynamic ability region $TAMDROC(x$ dkes，j $|X$ dsel ) obtained in the above process is It may not reach the preset number threshold $Limk$ for forming the ability region ad . If then the dynamic ability region $TAMDROC(x$ dkes，j $|X$ dsel ) takes the first $Limk$ dkes，j $|X$ dsel values of $SAMDROC(x$ ad , that is, $TAMDROC(x$ dkes，j $|X$ dsel ) = $SAMDROC(x$ dkes，j $|X$ dsel )[:$Limk$ ad, and used as the final dynamic capability area.

[0192] As can be seen from the above, the DES-DROC-CIIH algorithm (CEIH represents the classifier ensemble algorithm generated by heterogeneous and homogeneous) proposed in the embodiment of the present invention uses a three-step strategy to determine the dynamic capability area DROC of the sample to be tested. Generally speaking

[0193] The first step: Combine the adaptive K-nearest neighbor algorithm with the mutual K-nearest neighbor algorithm, and propose an adaptive K-nearest neighbor algorithm to determine the initial DROC of the sample to be tested.

[0194] The second step: On the basis of the first step, use the method of the distance from the sample to be tested approaching the center point of the DROC sample set to eliminate the samples far from the edge in the DROC.

[0195] The third step: On the basis of the first two steps, aiming at the problem of possible edge class samples in the DROC of the sample to be tested, eliminate the edge samples to finally obtain the DROC of the sample to be tested.

[0196] Through the above process, the dynamic integration dataset X dkes The sample x dkes，j in X dsel The dynamic capability domain TAMDROC(x dkes，j |X dsel ).

[0197] The classification module is used to select a preset number of classifiers according to the base classifier pool and the dynamic capability area, and obtain the diagnostic classification results corresponding to the current sample in the dynamic integration dataset in each selected classifier; specifically used for:

[0198] S51. Use the base classifier pool For the dynamic capability area TAMDROC(x dkes，j |X dsel ), obtain the accuracy rate set Put the classifier ranked first At the first place in the classifier order set U;

[0199] S52. Define C′ = C - U, and select classifiers from C′ one by one to calculate the overall complementary index with the classifier order set U, and obtain the overall complementary index set of classifiers

[0200] Among them, T′ represents the number of classifiers in C′, and select the classifier corresponding to the maximum value of the overall complementary index as the classifier ranked i, that is c t ′ ∈ C is added to the classifier order set U;

[0201] The solution process of the overall complementary index TCI(c t ) includes,

[0202]

[0203]

[0204] wherein, c i and c j respectively represent two different classifiers in the base classifier pool ; c i and c j divide the set of misclassified samples of the dynamic ability region TAMDROC(x dkes，j |X dsel ) into S i and S j respectively; D(S) represents the number of samples in S; the intersection of S i and S j is S i ∩S j , and the union is S i ∪S j ; G ij represents the complementary index of classifiers c i and c j ;

[0205] S53. Repeat the above operations until the classifier order set contains all the classifiers in the base classifier pool, that is select the first η classifiers in the finally obtained U according to the preset parameter η;

[0206] S54. According to the selected first η classifiers, obtain the respective diagnostic classification results corresponding to the current sample x dkes，j in the dynamic integration dataset.

[0207] An integration module, configured to integrate the respective diagnostic classification results to obtain the diagnostic classification result of the current sample in the dynamic integration dataset, and finally obtain the diagnostic classification results of all samples in the dynamic integration dataset.

[0208] Among them, the integration module is used to obtain the diagnostic classification result of the current sample in the dynamic integration dataset, specifically including:

[0209] Classify the current sample x dkes，j to obtain the result as {e 1 (x dkes，j ),......, e η (x dkes，j )}, and use the majority voting method to obtain the current sample x dkes，jThe final diagnostic classification result of arg class(max(Voting(e i (x dkes，j ))), i = 1,......, η.

[0210] For the dynamic integrated dataset X dkes All samples in are repeatedly substituted into the acquisition module, classification module, and integration module to execute the above steps, obtaining where num(X dkes ) represents the number of samples in X dkes . The classification results are combined with those in step (4) to obtain the classification results of the test set X test That is, the category with the largest number of votes is taken as the classification result. That is, the category with the largest number of votes is taken as the classification result.

[0211] To obtain the diagnostic classification results of each sample to be tested in the test set, the breast cancer mass diagnosis system provided by the embodiments of the present invention further includes:

[0212] A merging module, configured to merge the diagnostic classification results corresponding to some samples in the test set obtained by the generating module with the diagnostic classification results corresponding to all samples of the dynamic integrated dataset obtained by the integration module, to obtain the diagnostic classification result corresponding to each sample of the test set

[0213] The embodiments of the present invention consider that the BI-RADS diagnostic grade given by a doctor essentially indicates the degree to which the doctor suspects that the mass is malignant. For example, if 3 is misdiagnosed as 4A, and if 3 is misdiagnosed as 5, obviously the cost of the latter is greater. In order to better show the difference between the costs of misdiagnosis, the embodiments of the present invention combine the diagnostic grade with the diagnostic accuracy rate (IAcc) defined in formula (10),

[0214]

[0215] where y(x i ) is the BI-RADS diagnostic grade given by the doctor, is the diagnostic level number finally given by the merging model, I(Y) represents converting Y into the corresponding interval number, and M represents the number of samples to be tested.

[0216] For R 1 , R 2 , R 3 , R 4 The diagnostic datasets X 1 , X 2 , X 3 , X 4 of four doctors are used for experiments.

[0217] To ensure the effectiveness of the model proposed in the present invention, the mean of the results of 10 runs is obtained by measuring the effectiveness of the diagnostic accuracy IAcc method given by Equation (10). The experimental results are shown in Table 7 as follows:

[0218] Table 7 Experimental Results

[0219] Doctor Accuracy rate <![CDATA[R 1 > 0.8286 <![CDATA[R 2 > 0.8468 <![CDATA[R 3 > 0.8709 <![CDATA[R 4 > 0.8552

[0220] It can be seen from this that the breast cancer mass diagnosis system based on the dynamic ensemble selection algorithm provided by the embodiments of the present invention can obtain a relatively accurate diagnostic classification result of breast cancer masses.

[0221] In addition, in order to verify the superiority of the proposed DES-DROC-IHCI algorithm of the model, the embodiments of the present invention adopt the static adaptive ability region and the classifier dynamic ensemble selection algorithm DES-ROC-IHCI generated by homogeneous and heterogeneous structures, the classifier dynamic ensemble selection algorithm DES-DROC-ICI based on the dynamic ability region and homogeneous generation, the classifier dynamic ensemble selection algorithm DES-DROC-HCI based on the dynamic ability region and heterogeneous generation, and the classifier ensemble algorithm CEIH generated by homogeneous and heterogeneous structures.

[0222] Among them, the algorithms such as LB, GNB, SVM, MLP, LDA, and AB refer to the experimental parameters in Table 5, and the parameters of other algorithms are shown in Table 8 below. The parameters not specified are taken as the default parameters in the sklearns and deslib algorithm packages in Python.

[0223] Table 8 Experimental Parameters

[0224]

[0225]

[0226] The results of the comparative experiment are shown in Table 9.

[0227] Table 9 Comparison of Comparative Experiment Results

[0228]

[0229] From Table 9, by observing the diagnostic results of the models obtained by various algorithms, it can be seen that among the 5 algorithms used in the comparative experiment, the diagnostic result of DES-DROC-HCI is the worst. The main reason may be that this algorithm uses a heterogeneous classifier to generate a classifier pool for dynamic ensemble selection, and the generated classifier pool only has 6 classifiers. During the process of ensemble selection, the classifiers with higher classification accuracy will be restricted by the classifiers with lower accuracy, resulting in poor performance of the obtained diagnostic model.

[0230] By comparing the experimental results of DES-ROC-IHCI and CEIH, it is found that DES-ROC-IHCI using dynamic classifier selection has better performance than CEIH. The main reason is that the former selects a group of the most suitable classifiers for integration instead of integrating all classifiers for different samples to be tested, avoiding the influence of classifiers with poor performance in the base classifiers.

[0231] By comparing the experimental results of DES-DROC-IHCI, DES-DROC-ICI-AB, and DES-DROC-HCI, it can be seen that the DES-DROC-IHCI algorithm combines the advantages of both and achieves better classification results for all doctors. The main reason is that the DES-DROC-IHCI algorithm uses a method of combining homogeneous and heterogeneous classifiers to generate a classifier pool, which not only effectively combines the advantages of different heterogeneous classifiers, but also ensures the diversity of different classifiers in the integrated system, and avoids the problem of restricting the performance of the DES algorithm due to insufficient number of classifiers in the classifier pool.

[0232] By comparing DES-DROC-IHCI and DES-ROC-IHCI, compared with DES-ROC-IHCI, the classification performance of DES-DROC-IHCI on the diagnostic data set of each doctor has been significantly improved, among which the improvement of R 4 is the highest, reaching 1.27%. The reason why the DES-ROC-IHCI algorithm proposed in the present invention has better prediction effect is that the DES-DROC-IHCI algorithm, based on the DES-ROC-IHCI algorithm, adopts a three-step strategy for different samples to be tested to determine a more representative dynamic ability region, so that in the dynamic selection process, compared with the DES-ROC-IHCI algorithm, more suitable classifiers can be selected, making the grading prediction results of breast cancer masses more accurate, and finally improving the diagnostic accuracy of the breast cancer mass diagnosis model.

[0233] In summary, compared with the prior art, the following beneficial effects are achieved:

[0234] The embodiment of the present invention uses a combination of homogeneous classifier generation and heterogeneous classifier generation to generate the base classifier pool of the DES algorithm; and by proposing the dynamic ability region (DROC) algorithm, the diagnostic accuracy of breast cancer masses is improved.

[0235] It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the existence of additional identical elements in the process, method, article or device comprising the said element.

[0236] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A breast cancer mass diagnosis system based on a dynamic ensemble selection algorithm, characterized in that, it includes: A preprocessing module for obtaining and preprocessing a breast cancer mass diagnosis data set and dividing it into a training set, a test set, and a dynamic selection set; A training module for obtaining a base classifier pool by combining the use of homogeneous classifiers and heterogeneous classifiers according to the training set; A generation module for obtaining a dynamic ensemble data set corresponding to the test set according to the test set and the base classifier pool; An acquisition module for obtaining the dynamic ability region of any sample in the dynamic ensemble data set according to the dynamic selection set and the dynamic ensemble data set; A classification module for selecting a preset number of classifiers according to the complementary index according to the base classifier pool and the dynamic ability region, and obtaining the diagnostic classification results corresponding to the current sample in the dynamic ensemble data set in each selected classifier; An integration module for integrating each of the diagnostic classification results to obtain the diagnostic classification result of the current sample in the dynamic ensemble data set, and finally obtaining the diagnostic classification results of all samples in the dynamic ensemble data set; The acquisition module includes a first acquisition sub-module for obtaining the first ability region of any sample in the dynamic ensemble data set in the dynamic selection set by using an adaptive mutual K-nearest neighbor algorithm; or for obtaining the dynamic ability region; The first acquisition sub-module is specifically used for: S411. For any sample in the dynamic ensemble data set, calculate the adaptive distance between the current sample in the dynamic ensemble data set and each sample in the dynamic selection set; R i = min d E (x dsel，i′ ,x dsel，i ) + ε, y(x dsel，i′ ) ≠ y(x dsel，i ),x dsel，i′ ,x dsel，i ∈ X dsel where d ada (x dkes，j , x dsel，i ) is the adaptive distance between the j-th sample x dkes，j in the dynamic integrated dataset and any sample x dsel in the dynamic selection set X dsel，i ; R i represents any sample x dsel in the dynamic selection set X dsel，i and any other sample x of a different class dsel in the dynamic selection set X dsel，i′ the minimum distance therebetween; d E (x dsel，i′ ,x dsel，i ) represents any sample x dsel in the dynamic selection set X dsel，i and any other different-class sample x dsel in the dynamic selection set X dsel，i′ is the Euclidean distance between them; d E (x dkes，j ,x dsel，i ) represents the i-th sample x in the dynamic integrated dataset dkes，j and any sample x in the dynamic selection set X dsel The Euclidean distance between them; dsel，i ​ where x dsel，i′，l , x dkes，j，l or x dsel，i，l represents the value of the corresponding sample on the preset feature l of the breast cancer mass diagnosis dataset, and the total number of features is M; y(x dsel，i′ ) represents x dsel，i′ corresponding diagnostic category; y(x dsel，i ) represents x dsel，i corresponding diagnostic category; ε represents an arbitrarily small number; S412. Arrange in non-decreasing order according to the magnitudes of the respective adaptive distances, and take the first K samples with smaller distances to form the third ability region of the current sample in the dynamic ensemble data set; S413. Sequentially remove the individual samples in the third ability region from the dynamic selection set in the above arrangement order, and add the current sample in the dynamic ensemble data set to obtain a new dynamic selection set; S414. Calculate the adaptive distances between the individual samples removed from the third ability region and each sample in the new dynamic selection set, and take the K samples with smaller distances to form the fourth ability region of the removed sample; S415. Determine whether the current sample in the dynamic ensemble data set is within the fourth ability region; if so, it is determined that the current sample in the dynamic ensemble data set and the individual sample removed from the third ability region satisfy the mutual adaptive K-nearest neighbor relationship; S416. Summarize the judgment results of all samples in the third ability region, and each sample that meets the requirements forms the first ability region of the current sample in the dynamic ensemble data set in the dynamic selection set; Or if the number of samples in the first ability region obtained above is less than a preset number threshold, then take the first preset number threshold of samples in the third ability region to form the final dynamic ability region; The obtaining module further includes a second obtaining sub-module, which is configured to obtain a second ability region after removing samples far from the edge according to the first ability region by using the method of approaching the current sample with the distance from the center point of its sample set; or to obtain the dynamic ability region; specifically, the second obtaining sub-module is configured to: S421. Arrange each sample \(x\) in the first ability region \(AMDROC(x dkes，j |X dsel ) in non - decreasing order according to the adaptive distance from the current sample \(x\) in the dynamic integrated dataset AMD，i , and calculate the distance from the sample set with dkes，j \(k'\in Z\) to the center point \(x\) in turn , obtaining the set of distances from the sample sets to the center points c，k′ , that is, the set of distances from the sample sets to the center points Among them, k' represents the k'-th sample in the first ability region, and the total number of samples is Limk ad represents the preset quantity threshold; Z represents the set of integers; The distance between the sample set and the center point x c，k′ The solution process includes: In which, define x c，k′ to AMDROC(x dkes，j |X dsel )[:k′] The sum of the Euclidean distances of other samples is the smallest; d E (x i , x j ) represents AMDROC(x dkes，j |X dsel )[:k′] for any sample x i and sample x j The Euclidean distance between them, the total number of samples is k′; S422. Calculate the distance between each sample set and the center point x c，k′ and the current sample x in the dynamic integration dataset dkes，j to obtain the Euclidean distance. Denote the k' corresponding to the minimum Euclidean distance as Conk. Then, obtain the second ability region SAMDROC(x dkes，j |X dsel ) = AMDROC(x dkes，j |X dsel )[ : Conk]; Alternatively, if the number of samples in the obtained second ability region is less than the preset number threshold, then take the samples of the first preset numerical threshold in the first ability region to form the final dynamic ability region; The obtaining module further includes a third obtaining sub-module, which is configured to obtain the dynamic ability region after removing edge category samples according to the second ability region; or to obtain the dynamic ability region; Specifically, the third obtaining sub-module is configured to: Define the $i$-th sample in the second ability region $SAMDROC(x dkes，j |X dsel )$ whose corresponding class is $y i = class(x SMAD，i )$; If then obtain the dynamic ability region TAMDROC(x after removing the edge category samples dkes，j |X dsel ) = SAMDROC(x dkes，j |X dsel ) - x SMAD，i ; where num(y i |SAMDROC(x dkes，j |X dsel )) represents the diagnostic category y i in the second ability region SAMDROC(x dkes，j |X dsel ); represents the number of categories in the second ability region SAMDROC(x dkes，j |X dsel ); marginrate represents the marginal category sample rate; Alternatively, if the number of samples in the obtained dynamic ability region is less than the preset number threshold, then take the samples of the first preset number threshold in the second ability region to form the final dynamic ability region.

2. The breast cancer mass diagnosis system according to claim 1, wherein, the preprocessing module is specifically configured to: S11. Represent the features corresponding to the initial breast cancer mass diagnosis data set and the mass diagnosis level by using BI-RADS grading, calculate the interval number distance between different gradings and the minimum grade 3, and perform standardization processing on the obtained results. The features include boundary clarity, edge regularity, internal echo, calcification, and blood flow abundance; S12. Use the oversampling method to balance the categories of the initial breast cancer mass diagnosis data set; S13. Divide the breast cancer mass diagnosis data set processed as above into the training set, test set, and dynamic selection set.

3. The breast cancer mass diagnosis system according to claim 2, wherein, the training module is specifically configured to: Select several different classifier algorithms to form a base classifier algorithm pool Use the Bagging algorithm for the training set X train Conduct multiple samplings to generate different sample sets The same classifier algorithm c m b By training different sample sets Obtain different isomorphic classifiers c m b，o Finally, generate the base classifier pool 4. The breast cancer mass diagnosis system according to claim 2, wherein, the generating module is specifically configured to: S31. For the test set X test For the i-th sample x to be tested in it test，j , use the base classifier pool For the current sample x in the dynamic integration dataset dkes，j Perform classification to obtain the classification result set C(x test，j ) = {c 1 (x test，j ), c 2 (x test，j ), ……, c T (x test，j )}; S32. According to the classification result set C(x test，j ), calculate the classification consensus degree of the sample x to be measured test，j . Among them, H(x test，j , C) represents the proportion of the maximum vote count corresponding to the voting method in the classification result set C(x test，j ), where Voting(C(x test，j )) is the number of votes corresponding to each category in C(x test，j ); T represents the total number of classifiers in the base classifier pool; S33. Predetermine a consensus screening criterion Cond. If H(x test，j , C) > Cond, add the sample x test，j to the dynamic integration dataset X dkes ; If H(x test，j , C) ≤ Cond, obtain the diagnostic classification result of the sample x to be measured test，j ​ Among them, argclass max(Voting(C(x test，j ))) represents the class with the maximum number of votes in C(x test，j ).

5. The breast cancer mass diagnosis system according to claim 4, wherein, the classification module is specifically configured to: S51. Adopt a base classifier pool For the dynamic capability region TAMDROC(x dkes，j |X dsel ), obtain an accuracy set Place the classifier ranked first at the first position of the classifier order set U; S52. Define C′ = C - U, and select classifiers one by one from C′ to calculate the overall complementary index with the overall classifier order set U, and obtain the overall complementary index set of classifiers where T′ represents the number of classifiers in C′, and select the classifier corresponding to the maximum value of the overall complementary index as the classifier ranked i, that is Add it to the classifier order set U; The solving process of the overall complementary index TCI(c, ′) includes, Among them, c i and c j respectively represent two different classifiers in the base classifier pool ; c i and c j make the sets of misclassified samples of the dynamic ability region TAMDROC(x dkes，j |X dsel ) be S i and S j respectively; D(S) represents the number of samples in S; the intersection of S i and S j is S i ∩S j , and the union is S i ∪S j ; G ij represents the complementary index of classifiers c i and c j ; S53. Repeat the above operations until all classifiers in the base classifier pool are included in the classifier order set, i.e., Select the first η classifiers in the finally obtained U according to the preset parameter η; S54. Obtain the current sample x in the dynamic ensemble dataset according to the selected top η classifiers dkes，j The corresponding diagnostic classification results 6. The breast cancer mass diagnosis system according to claim 5, wherein, the integration module is used to obtain the diagnostic classification result of the current sample in the dynamic integration data set, specifically including: For the current sample x dkes，j classify it to get the results as {e 1 (x dkes，j ), ……, eη(x dkes，j )}, and use the majority voting method to obtain the final diagnosis classification result arg class(max(Voting(e dkes，j )(x i ))), where i = 1, ……, η. dkes，j ​ 7. The breast cancer mass diagnosis system according to claim 4, wherein, further includes: A merging module, which is configured to merge the diagnostic classification results corresponding to some samples in the test set obtained by the generating module with the diagnostic classification results corresponding to all samples in the dynamic integration data set obtained by the integration module to obtain the diagnostic classification result corresponding to each sample in the test set.

Images