CN111583194A - High-Dimensional Feature Selection Algorithm Based on Bayesian Rough Sets and Cuckoo Algorithm - Google Patents

Abstract

The invention discloses a high-dimensional feature selection algorithm based on a Bayesian rough set and a cuckoo algorithm, which comprises the following steps: acquiring a lung tumor image, and performing target contour segmentation to obtain a segmented ROI image; extracting high-dimensional characteristic components of the segmented ROI image, and constructing a decision information table containing characteristic attributes based on the characteristic components; and reducing the original feature space by adopting a BRSGA algorithm to obtain an optimal feature subset, optimizing a penalty factor and a kernel function parameter of the SVM by utilizing a CS algorithm, and inputting the reduced feature subset into the optimized SVM to obtain a classification recognition result. According to the method, the optimal feature subset is generated through the genetic algorithm and the BRS, the feature dimension is reduced on the premise of not reducing the classification accuracy, the constraint of manual parameter setting is eliminated, and the time consumption is reduced. The CS is used for carrying out global optimization on SVM parameters, so that the method has the advantages of more effective exploration of search space, enriched population diversity, good robustness and stronger global search capability.

Description

High-Dimensional Feature Selection Algorithm Based on Bayesian Rough Set and Cuckoo Algorithm

technical field

The invention relates to the technical field of medical image recognition, in particular to a high-dimensional feature selection algorithm based on Bayesian rough set and cuckoo algorithm.

Background technique

With the development of computer aided diagnosis (CAD) research, medical image processing technology has developed rapidly. However, the multi-modality, grayscale ambiguity and uncertainty of medical images make the rate of missed diagnosis and misdiagnosis remain high in the process of single-modality medical imaging diagnosis. Therefore, different modalities of medical image processing technology emerge as the times require, which are divided into pixel level, feature level and decision level according to different levels. The feature-level processing can compress the amount of information on the basis of retaining important information, and the processing speed is faster. In the process of feature-level processing of medical images, the redundancy and correlation between features make the "curse of dimensionality" an NP-hard problem. Feature selection is an effective measure to solve this problem, which can effectively reduce the size of the feature space. dimension, reducing time complexity.

The problems existing in the high-dimensional feature selection process include how to generate optimal feature subsets, how to evaluate the effect, the selection of classifiers used for evaluation, and the optimization of classifier parameters. In recent years, experts and scholars have proposed many algorithms. First, the proposal of variable precision rough set (VPRS) can effectively overcome the limitation that rough set (RS) can only handle accurate classification data. By introducing the classification error rate β, the lower approximation of RS is changed from "Full containment" is relaxed to "partial containment", which improves robustness and generalization of results in noisy datasets. The core of VPRS research is the selection of classification error rate β. The main research areas include three aspects: first, without considering the details of β selection, a variety of extended VPRS models are proposed, such as: variable-precision fuzzy rough set, variable-precision multi-granularity Rough set, generalized VPRS, extended VPRS based on β-tolerance relationship and Bavarian distance, etc.; second, the value of β is obtained through different calculation methods, such as the average inclusion degree as the threshold for selecting upper and lower approximations; third, introducing Probabilistic formulas propose many probabilistic RS models, such as VPRS, game rough set, decision rough set, Bayesian rough set (BRS), 0.5 probability rough set, etc. There is a certain correlation among various methods in probability rough sets, and the difference is reflected in the calculation of probability formula and the way of parameter design. Among them, BRS introduces a priori probability on the basis of VPRS, and replaces the classification error rate β in VPRS with prior probability. It does not need to manually set parameters, which not only overcomes the complete and accurate division of RS pair approximation, but also avoids the parameters in VPRS. The effect of β on the upper and lower approximations. Most of the research on BRS is still in the stage of theoretical analysis, lacking mature and independent models, and has not been combined with other algorithms to deal with the problem of high-dimensional feature selection in medical images.

Secondly, the performance of the classifier is the basis for evaluating the high-dimensional feature selection algorithm. The support vector machine (SVM) is a commonly used binary classification algorithm. The introduction of the kernel function further broadens its application scope. The commonly used kernel functions include Polynomial kernel function, radial basis kernel function (RBF) and Sigmoid kernel function, among which the calculation speed of the polynomial kernel function is slow, which seriously affects its effect and is less used; Only the kernel matrix needs to be calculated, the time complexity is small, the manual parameter setting workload is large and the time is long, and the final parameters are not necessarily optimal, and the parameter selection needs to be transformed into an optimization problem for analysis.

Therefore, how to provide a high-dimensional feature selection algorithm with low time complexity and better robustness based on Bayesian rough set and cuckoo algorithm is an urgent problem for those skilled in the art to solve.

SUMMARY OF THE INVENTION

In view of this, the present invention provides a high-dimensional feature selection algorithm based on Bayesian rough set and cuckoo algorithm, combined with BRS, GA, CS and SVM algorithms, and proposes a high-dimensional feature selection algorithm based on BRSGA and CS two-stage optimization Feature selection algorithm. In the first stage of optimization, the BRSGA algorithm is used to reduce the original feature space to obtain the optimal feature subset. SVM classification model to identify lung tumor images.

In order to achieve the above object, the present invention provides the following technical solutions:

A high-dimensional feature selection algorithm based on Bayesian rough set and cuckoo algorithm, including the following steps:

S1. Acquire a lung tumor image, and perform target contour segmentation to obtain a segmented ROI image;

S2, extracting the high-dimensional feature components of the segmented ROI image, and constructing a decision information table containing feature attributes based on the high-dimensional feature components, the feature attributes and the features of different dimensions in the high-dimensional feature components correspond;

S3. Based on the Bayesian rough set model, a fitness objective function is constructed by using the weighted summation of the global relative gain function, the attribute reduction length and the genetic coding weight function, and the characteristic attribute is reduced by combining the genetic operator combination, Get the reduced feature subset;

S4, using the cuckoo algorithm to optimize the penalty factor and kernel function of the SVM, and input the reduced feature subset into the optimized SVM to obtain a classification and identification result.

Preferably, the high-dimensional feature components in S2 include shape features, texture features and grayscale features of the lung tumor image.

Preferably, the S3 specifically includes the following steps:

S31. Construct a fitness objective function:

采用全局相对增益衡量信息系统S的属性重要度；The first objective function is the global relative gain function of the equivalence relation E relative to the feature attribute D:

The global relative gain is used to measure the attribute importance of the information system S;

The second objective function is the attribute reduction length:

Among them, |C| is the number of conditional attributes, and L _r is the number of 1 genes in the r chromosome;

The third objective function is the gene encoding weight function:

Among them, the numerator is the product sum of genes other than 0 and 1, and the denominator is the length of the chromosome;

Construct a fitness objective function F(x)=-ω1×target1-ω2×target2+ω3×target3 to perform feature attribute reduction on the feature attribute;

S32, optimize the genetic operator according to the fitness objective function:

According to the fitness objective function, the fitness value of the feature attribute is calculated, and it is judged whether the termination condition is met. If so, the reduced feature subset is obtained; otherwise, the feature attribute is randomly selected without replacement residue, uniform crossover and Gaussian transformation. Constitute the genetic algorithm operation, and re-execute S32.

Preferably, the specific steps of optimizing the SVM parameters by the cuckoo algorithm in the S4 include:

S41. Initialization settings: including the probability P _a , the number of iterations N, the number of nests n, the upper and lower limits, the penalty factor c of the SVM and the RBF kernel function parameter σ;

S42, initialize n bird nest positions, calculate the fitness values of all bird nests, and save the current optimal position and fitness value;

S43. Update the bird's nest position according to the formula, and compare it with the bird's nest fitness value of the corresponding position of the previous generation, and retain the bird's nest position and the fitness value with the smallest fitness value as the optimal bird's nest;

S44, generating a random number r, discarding the poor bird's nest with a given probability P _a , if r > P _a , updating the bird's nest, otherwise not updating;

S45. Recalculate the fitness value of the bird's nest, replace the bird's nest with a low fitness value with a bird's nest with a high fitness value, and generate a new set of bird nest positions;

S46, determine whether the number of iterations is completed, if so, stop the search to obtain the global optimal fitness value and the corresponding optimal bird's nest, if the stopping condition is not satisfied, then skip to S43 to continue the optimization;

S47, construct an SVM prediction model according to the optimal parameters c and σ corresponding to the optimal bird's nest position.

Compared with the prior art, the advantages of a high-dimensional feature selection algorithm based on Bayesian rough set and cuckoo algorithm designed by the present invention are:

The attribute importance is analyzed from the perspective of the global relative gain function, and the fitness function is constructed by combining the weighted sum of the attribute reduction length and the gene encoding weight function, and the optimal feature subset is generated through genetic operations such as selection, crossover and mutation. On the premise of reducing the classification accuracy, the feature dimension is reduced, and the constraints of manual parameter setting are freed, which greatly reduces the time consumption. Using CS to optimize the parameters of support vector machine (SVM) globally, the global search in the CS algorithm has infinite mean and variance, which can explore the search space more effectively than the algorithm using the standard Gaussian process, broaden the search field, enrich the It has good robustness and strong global search ability. It is feasible and effective to combine BRS with intelligent optimization algorithm for feature selection, and use CS to optimize the parameters of SVM.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only It is an embodiment of the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to the provided drawings without creative work.

Fig. 1 is the flow chart of the high-dimensional feature selection algorithm based on Bayesian rough set and cuckoo algorithm provided by the present invention;

Fig. 2 is the contrast diagram before and after the ROI region is segmented by utilizing Otsu algorithm provided by the embodiment of the present invention;

FIG. 3 is a flowchart for generating an optimal feature subset provided by an embodiment of the present invention;

Fig. 4 is the CS optimization SVM parameter flow chart provided by the embodiment of the present invention;

5 is a schematic diagram of a change in a fitness function during a certain feature subset generation process provided by an embodiment of the present invention;

FIG. 6 is a comparison diagram of the results of different classification algorithms based on the BRSGA selection algorithm provided by an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The embodiment of the present invention discloses a high-dimensional feature selection algorithm based on Bayesian rough set and cuckoo algorithm. The flowchart is shown in FIG. 1 , including data acquisition, data preprocessing, image segmentation, feature extraction, and attribute reduction. In the process of feature reduction, the GA hybrid BRS algorithm is used to optimize the original feature subset, and the optimized feature subset after reduction is obtained. The function parameters and penalty parameters are optimized to obtain the parameters with the best model fit and build the model according to the parameters. Finally, a two-stage optimized high-dimensional feature selection algorithm is used to classify and identify lung tumor CT images. The specific execution process is as follows:

S1. Acquire a lung tumor image, and perform target contour segmentation to obtain a segmented ROI image.

Before target contour segmentation, it includes image acquisition and preprocessing:

Taking into account the popularity of common lung tumor imaging methods, the acceptance of doctors and patients, the cost, and to avoid the influence of lung tumor CT images by the specifications, models, and environment of the inspection equipment, lung tumors with definite diagnostic conclusions were collected. A total of 3000 cases of tumor CT images were obtained, including 1500 cases of malignant (benign) tumors. The sub-images with strong discriminative ability are intercepted from the obtained images as ROI regions, and all ROI regions are normalized to experimental images with a size of 50*50 pixels.

Target contour segmentation process:

Segmenting target contours (including lesion contours) from the intercepted ROI region plays a crucial role in various clinical applications. However, the method of manual annotation by radiologists is still used in clinical practice at present, and a large number of intensive manual operations are prone to errors. Therefore, accurate segmentation using computer technology has great practical value. This embodiment adopts the Otsu threshold segmentation method, the core idea of which is to segment the image into two groups, and when the inter-group variance between the two groups reaches the maximum, the obtained value is the optimal segmentation threshold. The basic principle of Otsu algorithm is as follows:

Assuming that the size of an image is m×n, and the gray level of the image is l, the gray level range is [0,l-1], and n _i represents the number of occurrences of gray level i, then gray level i is present in all The frequency of occurrence in a pixel is p _i =n _i /(m×n). Assume that the pixels whose gray level is less than q constitute class A ₁ , that is, the gray level range of A ₁ is [0, q], and the pixel point whose gray level range is [q+1, l-1] is A ₂ . P ₁ (q), P ₂ (q) represent the probability of occurrence of A ₁ class and A ₂ class, respectively, u ₁ (q), u ₂ (q) represent the average value of the gray level of A ₁ class and A ₂ class, but:

The between-group variance σ _b (q) of the image is expressed as:

When the inter-group variance between the two groups reaches the maximum, the obtained value is the optimal segmentation threshold, that is, the pixel segmentation threshold is:

In this paper, Otsu is used to segment the ROI area. As shown in Figure 2, an example of the ROI image before and after segmentation using the Otsu algorithm is given. Figure 2(a) is the ROI image before segmentation, and Figure 2(b) is the ROI image after segmentation.

S2. Extract the high-dimensional feature components of the segmented ROI image, including a total of 104-dimensional features including shape features, texture features, and grayscale features. The specific features are shown in Table 1. And based on the high-dimensional feature components, a decision information table containing feature attributes is constructed, and the feature attributes correspond to the features of different dimensions in the high-dimensional feature components. The size of the constructed decision information table is 3000*105, and the fuzzy The C-means clustering algorithm discretizes the decision information table. After discretization, a numerical label is assigned to the tumor feature, that is, the feature attribute, which represents the benign and malignant tumor, and is located in the last column of the decision information table.

Table 1 Lung tumor CT image feature set

S3. Based on the Bayesian rough set model, a fitness objective function is constructed by using the weighted summation of the global relative gain function, the attribute reduction length and the genetic coding weight function, and the characteristic attribute is reduced by combining the genetic operator combination, A reduced feature subset is obtained; this embodiment combines the BRS and GA algorithms to perform attribute reduction, reduces the time complexity and space complexity of the classifier, and improves the classification performance.

As shown in Figure 3, the reduction specifically includes the following steps:

S31. Establish a BRS model:

1) Parameter setting: a sequence consisting of 0 and 1 chromosomes, the length of which is equal to the number N of conditional attributes, the crossover probability P _c , the mutation probability P _m , the maximum number of iterations K=150, the initial population M=20, the fitness function is F(x);

2) Encoding: encoding in binary mode, its length is equal to the number of conditional attributes, "0" means that the feature is not selected, and "1" means that the feature is selected;

3) Characteristic attributes, that is, the generation of the initial population: randomly generate M chromosome strings whose length is equal to the number of conditional attributes to form the initial population;

4) Genetic operators: Genetic operators include selection, crossover and mutation operators. The combination of genetic operators is random remainder selection without playback, uniform crossover and Gaussian mutation.

S32. Construct a fitness objective function: comprehensively consider three aspects of global relative gain function, attribute reduction length and gene encoding weight function, construct a fitness function framework through weighted sum, carry out the optimization process of genetic algorithm, and find the most suitable fitness function. Discriminative feature subsets.

采用全局相对增益衡量信息系统S的属性重要度；The global relative gain function of the equivalence relation E with respect to the feature attribute D is:

The global relative gain is used to measure the attribute importance of the information system S;

In the BRS model, the process of attribute reduction algorithm with global relative gain as heuristic information is as follows:

S321: Calculate the core attribute set γ of the conditional attribute in the information system S=(U, A, V, f), and calculate the dependence degree R _C (D) of the decision attribute on the condition attribute;

计算

的值，构成集合M；S322: Calculate the degree of dependence R _γ (D) of the decision attribute on the core attribute, if R _γ (D)=R _C (D) is established, then go to S324 to obtain R reduction, otherwise let C=C-γ, for

calculate

The value of , constitutes a set M;

转到S322继续计算；S323: Sort the elements in the set M in ascending order, and add the maximum value to the set γ, that is, γ=γ∪C _i ,

Go to S322 to continue the calculation;

S324: The final result is an R reduction of BRS.

The attribute reduction length is

Among them, |C| is the number of conditional attributes, and L _r is the number of 1 genes in the r chromosome;

The gene encoding weight function is

The value of the locus can only be 0 and 1, otherwise it will be punished. Since there will be genes greater than 1 or less than 0 or less than -1 in the chromosome, for this situation, the gene encoding weight function is constructed as target3, and the molecule is calculated. If the product sum of non-0, 1 is not 0, if the gene on the locus i is 0, then 0×(0-1)=0, if the locus is 1, then 1×(1-1)=0, so only calculated The sum of products other than 0 and 1, and the denominator is expressed as the length of the chromosome.

Example: Set chromosome r=[0 1 -2 3 1], (r-1)=[-1 0 -3 2 0], then:

r×(r-1)=[0 1 -2 3 1]×[-1 0 -3 2 0]=[0 0 6 6 0]

∑abs(r×(r-1))=12, and the chromosome length is 5, then target3=12/5=2.4.

Construct a fitness objective function F(x)=-ω1×target1-ω2×target2+ω3×target3 to perform feature attribute reduction on the feature attribute.

S32, optimize the genetic operator according to the fitness objective function:

Calculate the fitness value of the feature attribute according to the fitness objective function, and judge whether the termination condition is satisfied. The termination condition is the set value; if so, the reduced feature subset is obtained; otherwise, the feature attributes are sequentially performed without replacement The genetic algorithm operation composed of random selection of remainder, uniform crossover and Gaussian transformation, and re-execute S32.

S4, using the cuckoo algorithm to optimize the penalty factor and kernel function of the SVM, and input the reduced feature subset into the optimized SVM to obtain a classification and identification result.

Referring to Figure 4, the specific steps of optimizing SVM parameters by the cuckoo algorithm include:

S41. Initialization settings: including the probability P _a , the number of iterations N, the number of nests n, the upper and lower limits, the penalty factor c of the SVM and the RBF kernel function parameter σ;

S42: Initialize n bird nest positions, calculate the fitness values of all bird nests and save the current optimal positions and fitness values; a bird nest is a feasible solution, and calculating the fitness value of the bird nest is the n feasible solutions obtained by initialization The value obtained after the calculation of the objective function is brought in, and the optimal value (maximum or minimum can be selected according to specific needs) is reserved, that is, the optimal position and fitness value of the bird's nest are obtained;

S43, update the position of the bird's nest according to the position update formula, and compare it with the bird's nest fitness value of the corresponding position of the previous generation, and retain the bird's nest position and the fitness value with the smallest fitness value as the optimal bird's nest;

S44, the random number r is automatically generated according to the Gaussian random function, and the poor bird's nest is discarded with a given probability P _a , if r > P _a , the bird's nest is updated, otherwise it is not updated;

S45. Recalculate the fitness value of the bird's nest, replace the bird's nest with a low fitness value with a bird's nest with a high fitness value, and generate a new set of bird nest positions;

S46, determine whether the number of iterations is completed, if so, stop the search to obtain the global optimal fitness value and the corresponding optimal bird's nest, if the stopping condition is not satisfied, then skip to S43 to continue the optimization;

S47, construct an SVM prediction model according to the optimal parameters c and σ corresponding to the optimal bird's nest position.

The search path of the cuckoo CS algorithm is levy flight. In this form of walking, short-distance exploration and occasional longer-distance walking are alternated, which can expand the search range, increase the diversity of the population, and avoid falling into local optimum. The relevant definitions are as follows:

The formula for the CS algorithm to search for the bird's nest location is:

为第i个鸟巢在第t代的鸟巢位置，α为步长控制量，一般取0.1，α用于确定随机搜索范围：where:

is the bird's nest position of the i-th bird's nest in the t-th generation, α is the step size control amount, generally 0.1, and α is used to determine the random search range:

Among them, α ₀ is a constant (α ₀ =0.01), and x _best represents the current optimal solution.

表示点对点乘积，L(λ)为随机搜索路径，Levy～u＝t^-λ,1＜λ≤3，服从levy分布，则相应的位置更新公式为：In the formula to search for the location of the bird's nest,

Represents the point-to-point product, L(λ) is the random search path, Levy～u=t ^-λ , 1<λ≤3, obeys the levy distribution, then the corresponding position update formula is:

where μ and ν both obey a normal distribution:

where Γ is the standard Gamma function.

The performance evaluation of medical image recognition results includes two indicators, sensitivity and specificity, but these two indicators are difficult to comprehensively describe the overall performance of the classifier. Therefore, in this embodiment, evaluation indicators are respectively set in the two stages of feature selection and classification and identification, and the feature selection stage includes reduction length, attribute importance, and time. The classification and identification stage includes Accuracy, Sensitivity, Specificity, F value, Matthews Correlation Coefficient (MCC), Balanced F Score (F1Score), Youden Index (YI) and Time (Time), the calculation formula is as follows:

YI=Sensitivity+Specificity-1

Among them, TP refers to the number of successfully identified malignant tumor target contours; FP refers to the number of incorrectly identified malignant tumor target contours; TN refers to the number of successfully identified benign tumor target contours; FN refers to the number of incorrectly identified benign tumor target contours.

In order to verify the feasibility and effectiveness of the technical solution of the present invention, two sets of comparative experiments are designed. The first experiment is to verify the feasibility and effectiveness of the BRSGA feature selection algorithm. The GS algorithm is fixed to optimize the parameters of the SVM, and the BRSGA algorithm is compared with different The pros and cons of VPRSGA in the β case at different stages. Experiment 2 fixed the feature selection algorithm on the basis of Experiment 1, and compared the advantages and disadvantages of CS-SVM and GS-SVM, GA-SVM and PSO-SVM in the classification stage.

Experiment 1: Comparison of experimental results of different feature selection algorithms based on the same classification algorithm

The fixed classification and recognition algorithm is GS-SVM, and the advantages and disadvantages of BRSGA and VPRS algorithms with different parameters in the two stages of feature selection and classification and recognition are compared, and the VPRS parameter β is set to 0.1, 0.2, 0.3 and 0.4 respectively. The specific results are shown in Table 3, Figure 5 and Table 4. In the optimal feature subset generation stage, each parameter combination is reduced 5 times, and the reduction length, attribute importance, and time are obtained respectively, and the average of the 5 reduction results and indicators for each parameter is obtained as the experimental result under this parameter. . In the classification and identification stage, the results of each parameter reduction are five-fold crossover using LIBSVM (that is, 300 cases of benign tumors and 300 cases of malignant tumors are selected as the test set each time, and the remaining data are used as the training set) for classification and identification. Each parameter Five sets of identification results are obtained, including: accuracy, sensitivity, specificity, F value, MCC, F1Score, YI, and time, and the average value of each index of the five-fold cross is calculated as the reduced classification result under this parameter. The average value of each index of the five reductions is used as the reduction and classification result under this parameter combination.

Table 3 Comparison of attribute reduction results of different feature selection algorithms

It can be seen from Table 3 that in the present invention, when the classification error rate β does not need to be manually set, the reduction length is 7.8 dimensions, which is between the reduction lengths of the VPRSGA algorithm under different β values, compared with the reduction length of β=0.1 The decrease is more significant, with a slight increase in the reduction length compared to β=0.2 and β=0.4. The fitness value is only slightly higher than the VPRSGA algorithm with β=0.4. The importance is 0.0002 lower than that of VPRSGA with β=0.4, which is higher than the VPRSGA model with other parameters. Compared with the VPRSGA model whose reduction time is higher than β=0.2, the reduction time of VPRSGA under other parameter values is reduced by 16.54-419.35 seconds, and the time is shortened by 2.7 times compared with β=0.1. It can be seen from Figure 5 that the algorithm in this paper does not have premature convergence in the reduction stage, and the VPRSGA algorithm has different degrees of premature convergence under the condition of different β values. More serious premature convergence phenomenon. Therefore, compared with the VPRSGA algorithm, the present invention not only gets rid of the constraints of manual parameter setting, but also achieves a relatively ideal effect.

Table 4 Comparison of classification and recognition results of different feature selection algorithms

It can be seen from Table 4: Compared with the VPRSGA algorithm with β=0.1, the present invention reduces the accuracy, specificity, MCC, F1Score, and YI by 0.07%, 0.43%, 0.0015, 0.0006, and 0.0013, respectively, but the sensitivity increases by 0.3%, and the classification time β The VPRSGA algorithm of =0.1 is 3.4 times that of the BRSGA algorithm. Although the BRS algorithm reduces the accuracy within an acceptable range, it greatly reduces the time consumption. Considering the accuracy and time consumption, the overall performance of the BRS algorithm is better than the VPRS algorithm with β=0.1; the BRSGA algorithm Compared with the VPRSGA algorithm with parameters β=0.2, 0.3 and 0.4, the time is reduced, and the other indexes are improved to different degrees, among which the indexes of the VPRSGA algorithm with β=0.2 are improved significantly. It can be seen from the classification results that compared with the VPRSGA model, the BRSGA model not only gets rid of the constraints of parameters, but also improves the classification performance of the model.

It can be seen from Experiment 1 that when the classification algorithm is fixed (that is, the SVM parameters are optimized by using the grid optimization algorithm), the BRSGA feature selection algorithm is free from the constraints of manual parameter setting compared with the VPRSGA algorithm, and it can be used in both attribute reduction and classification stages. It shows an ideal effect. Therefore, when verifying the effectiveness of CS algorithm for SVM parameter optimization, the fixed feature selection algorithm is BRSGA.

Experiment 2: Comparison of experimental results of different classification algorithms based on the same feature selection algorithm

The fixed optimal feature subset generation algorithm is BRSGA, and the CS algorithm is used to optimize the SVM parameters and compare with GS-SVM, GA-SVM and PSO-SVM. That is to classify and identify the results obtained by the five reductions of the BRSGA algorithm in experiment 1, and obtain the classification results by five-fold cross-validation for each reduction, including accuracy, sensitivity, specificity, F value, MCC, F1Score, YI and time. , and the average value of each index of the five-fold intersection is taken as the classification result after this reduction, and the average value of the five reductions is the final result of the classification model. In order to quantitatively describe whether the identification accuracy of the present invention and the comparison algorithm is statistically significant, a paired t test is used to perform hypothesis testing. The statistical hypothesis test is based on five indicators that comprehensively describe the accuracy of accuracy, F value, MCC, F1Score, and YI. , the significance level was set at p<0.05. The null hypothesis is that the difference between the average value of the same evaluation index between the present invention and the comparison algorithm is 0. For each evaluation index, the average and standard deviation of the five reduction classification and identification results are given.

Table 5 Comparison of results of different classification algorithms based on BRSGA selection algorithm

*Indicates that the marked result is significantly different from the corresponding index of the algorithm in this paper (CS-SVM) when the significance level is 0.05.

It can be seen from Table 5 that the quantitative analysis results show that the present invention is superior to the other three comparison algorithms in all five evaluation indexes, and all have statistically significant differences. From Figure 6 as a whole, the classification results of BRSGA after five reductions show a fluctuating trend. The fourth reduction is relatively good for most of the classification indicators, and the third reduction is relatively low except for the over-classification time. Because the initial population of the genetic algorithm is randomly generated in the process of generating the optimal feature subset, the results of each reduction are different. Therefore, in this paper, each parameter combination is reduced 5 times, and the classification adopts the method of five-fold crossover. The overall performance of the model is evaluated by the average value of each index of the classification result, which can effectively avoid one-sided evaluation.

Figure 6 shows the comparison charts of the classification results in the following classification and recognition stages: (a) accuracy; (b) classification time; (c) F value; (d) sensitivity; (e) specificity; (f) MCC; (g) F1Score; (h) Youden. In the five reduction process, the CS-SVM algorithm is higher than the GS-SVM, GA-SVM and PSO-SVM algorithms in the six evaluation indicators such as accuracy, F value, sensitivity, MCC, F1Score and Youden index, and the classification time slightly higher than GS-SVM. The classification time of the PSO-SVM algorithm is much higher than that of the other three algorithms. Except for oversensitivity, the other six evaluation indicators are higher than those of the GS-SVM and GA-SVM algorithms in the third reduction. lower than GS-SVM, GA-SVM and CS-SVM. It can be seen from Figure 6a, c, f, g and h that the CS-SVM algorithm in the five reductions is higher than GS-SVM, GA-SVM and PSO-SVM in all comprehensive evaluation indicators, and has certain robustness and Greater promotional value.

The reasons why the classification time of GS-SVM is lower than that of CS-SVM are as follows: First, due to the difficulty in obtaining medical image data, the test set data in this paper is only 600 cases, and the time complexity is lower than that of the CS algorithm, but the real clinical data It is massive, increasing sharply every day, or even increasing exponentially. When the number of samples increases, the time complexity of the GS algorithm will increase significantly, which cannot meet the needs of clinical applications. Second, the GS algorithm is based on experience. The search range has a certain randomness, and there is no guarantee that the optimal parameters are obtained. The CS algorithm is a swarm intelligence search algorithm, which has both local and global search capabilities, broadens the search field, enriches the diversity of the population, and has good robustness. Compared with the GS algorithm, it can effectively avoid the problems caused by experience. randomness.

The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments can be referred to each other. As for the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part can be referred to the description of the method.

The above description of the disclosed embodiments enables any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (4)

1. A high-dimensional feature selection algorithm based on a Bayesian rough set and a cuckoo algorithm is characterized by comprising the following steps:

s1, obtaining a lung tumor image, and performing target contour segmentation to obtain a segmented ROI image;

s2, extracting high-dimensional feature components of the segmented ROI image, and constructing a decision information table containing feature attributes based on the high-dimensional feature components, wherein the feature attributes correspond to features of different dimensions in the high-dimensional feature components;

s3, based on a Bayes rough set model, constructing a fitness objective function by utilizing the weighted summation of a global relative gain function, an attribute reduction length and a gene coding weight function, and reducing the feature attributes by combining genetic operator combination to obtain a reduced feature subset;

s4, optimizing the penalty factor and the kernel function of the SVM by using a cuckoo algorithm, and inputting the reduced feature subset into the optimized SVM to obtain a classification recognition result.

2. The high-dimensional feature selection algorithm based on Bayesian-coarse-set and Cuckoo-distribution algorithm as recited in claim 1, wherein the high-dimensional feature components in S2 comprise shape features, texture features and gray scale features of lung tumor images.

3. The high-dimensional feature selection algorithm based on bayesian rough set and cuckoo algorithm as claimed in claim 1, wherein said S3 specifically comprises the following steps:

s31, constructing a fitness objective function:

the first objective function is a global relative gain function of the equivalence relation E with respect to the feature attribute D:

measuring the attribute importance of the information system S by adopting global relative gain;

the second objective function is attribute reduction length:

wherein | C | is the number of conditional attributes, L_rThe number of genes in chromosome r is 1;

the third objective function is a gene coding weight function:

wherein, the numerator is the product sum of genes with the length of non-0 and 1, and the denominator is the length of the chromosome;

constructing a fitness objective function F (x) — ω 1 × target1- ω 2 × target2+ ω 3 × target3 to perform feature attribute reduction on the feature attributes;

s32, optimizing the genetic operator according to the fitness objective function:

calculating the fitness value of the characteristic attribute according to the fitness objective function, judging whether a termination condition is met, and if so, obtaining a reduced characteristic subset; if not, the characteristic attributes are sequentially subjected to genetic algorithm operation consisting of non-return remainder random selection, uniform intersection and Gaussian transformation, and S32 is executed again.

4. The high-dimensional feature selection algorithm based on the bayesian rough set and the cuckoo algorithm as claimed in claim 1, wherein the step of optimizing the SVM parameters by the cuckoo algorithm in S4 comprises:

s41, initialization setting: including probability P_aIteration times N, bird nest number N, upper and lower limits, penalty factor c of SVM and RBF kernel function parameter sigma;

s42, initializing n bird nest positions, calculating the fitness values of all bird nests, and storing the current optimal positions and the fitness values;

s43, updating the position of the bird nest according to a formula, comparing the position with the adaptability value of the bird nest at the corresponding position of the previous generation, and keeping the position of the bird nest with the minimum adaptability value and the adaptability value as the optimal bird nest;

s44, generating a random number r with a given probability P_aDiscarding bad bird nest if r > P_aIf not, updating the bird nest;

s45, recalculating the fitness value of the bird nest, replacing the bird nest with a high fitness value with the bird nest with a low fitness value, and generating a group of new bird nest positions;

s46, judging whether iteration times are finished, if so, stopping searching to obtain a global optimal fitness value and a corresponding optimal bird nest, and if not, jumping to S43 to continue optimizing;

and S47, constructing an SVM prediction model according to the optimal parameters c and sigma corresponding to the optimal bird nest position.

Images