CN111563549A - Medical image clustering method based on multitask evolutionary algorithm - Google Patents

Abstract

The invention discloses a medical image clustering method based on a multitask evolutionary algorithm, which comprises the following steps of: s1, extracting ROI feature description data of the medical image; s2, reading ROI feature description data of the extracted medical image, and obtaining a plurality of clustering results under a multi-task framework by optimizing a plurality of clustering internal indexes by applying an NMP clustering rule; and S3, selecting an optimal result from the results by using expert knowledge of the doctor. The method can fully express the connotation of the medical image, can simultaneously optimize one population to obtain a plurality of clustering results, is easier to converge to global optimum through cross-domain communication, and has more obvious clustering effect.

Description

Medical Image Clustering Method Based on Multi-task Evolutionary Algorithm

technical field

The invention relates to the two fields of medical technology and intelligent computing, in particular to a medical image clustering method based on a multi-task evolutionary algorithm.

Background technique

In the past 20 years, medical imaging technology has become one of the rapidly developing fields in medical technology, and medical images are more and more easily obtained and stored. Among the image data, medical image data occupies a large proportion. In the medical system, medical images play an important role. Through the observation of medical images, the lesion can be judged more directly and clearly, so as to obtain more accurate diagnosis results. Accurate clustering of medical images can better provide scientific reference for medical staff to judge and diagnose the cause of the disease, thereby greatly reducing the misdiagnosis rate caused by insufficient human vision resolution or insufficient clinical experience of medical staff. Further improve the utilization of medical images.

At present, for medical image clustering, although traditional clustering algorithms have been transplanted to medical images at home and abroad, such as the method based on K-means and its many variants, but these all have "sensitivity to the initial center and Sensitive to the value of K, the density-based methods are mostly sensitive to the value of the field radius and MinPtr, and are greatly affected by noise, and the grid-based methods are lacking in accuracy”.

In addition, there is also the use of evolutionary algorithms for clustering. Commonly used are genetic algorithms or differential algorithms, but they are based on single-task frameworks. Only a single target can be optimized in one operation to obtain a single optimization result.

What's more, for medical staff, not every pixel in a medical image is worth observing, and doctors will pay more attention to those distinctive pixel areas, which are called ROI (region of interest). In previous image studies, researchers extracted ROIs based on traditional image features such as color, texture, and shape. But this does not apply to medical images. Compared with ordinary image data, medical images have many characteristics.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the deficiencies of the prior art, to provide a method that can more fully express the connotation of medical images, can simultaneously optimize a population to obtain multiple clustering results, and is easier to converge to the global optimum through cross-domain communication. The medical image clustering method based on multi-task evolutionary algorithm with more obvious effect,

For achieving the above object, the technical scheme provided by the present invention is:

The medical image clustering method based on multi-task evolutionary algorithm includes the following steps:

S1. Extract the ROI feature description data of the medical image;

S2. Read in and extract the ROI feature description data of the medical image, and use the NMP clustering rules to obtain multiple clustering results under the multi-task framework by optimizing multiple clustering internal indicators;

S3. Use the expert knowledge of the doctor to select an optimal result.

Further, the ROI feature description data of the medical image extracted in the step S1 includes relative grayscale s1, relative area s2, relative centroid coordinates s3, circularity s4, angle s5, and symmetry s6.

Further, the specific process of extracting the ROI feature description data of the medical image in the step S1 is as follows:

S1-1, read in medical images;

S1-2. Scan the medical image to obtain the pixel number, length, width and maximum gray value of the image;

S1-3, detect the symmetry s6 of the medical image;

S1-4, extract the ROI area of the medical image according to the grayscale range;

S1-5, extracting the mean gray value, the number of pixels, the longest axis, the shortest axis, the coordinates of the centroid and the angle from the ROI region obtained in step S1-4;

S1-6, calculate the feature description data of the ROI area.

Further, the process of detecting the symmetry s6 of the medical image in the steps S1-3 is:

The medical image is folded in half, and the grayscale threshold is used for binarization. If the remaining pixels are less than the set value, the medical image is judged to be symmetrical, otherwise it is asymmetrical;

The specific process of calculating the feature description data of the ROI region in the step S1-6 is as follows:

其中ROI.gray为ROI区域的平均灰度，IMAGE.gray为整个图像的平均灰度；Relative grayscale s1:

Among them, ROI.gray is the average gray level of the ROI area, and IMAGE.gray is the average gray level of the entire image;

Relative area s2: s2=ROI.area/IMAGE.area, ROI.area is the number of pixels in the ROI area, and IMAGE.area is the number of pixels in the entire image;

Relative centroid coordinate s3: s3=(ROI.x/IMAGE.length, ROI.y/IMAGE.height), ROI.x is the abscissa of the ROI centroid, IMAGE.length is the length of the original image, and ROI.y is the ROI centroid , IMAGE.height is the height of the original image;

Circular s4: s4=4×π×ROI.area/ROI.perimeter ² , ROI.area is the number of pixels in the ROI area, and ROI.perimeter is the number of pixels around the ROI area;

Angle s5: s5=(Orientation+90)/180, and Orientation is the angle between the long axis of the ROI and the X axis.

Further, the specific process of the step S2 is as follows:

S2-1. Read in and extract the ROI feature description data of the medical image;

S2-2, initialize the population and the maximum number of iterations n, k=0;

S2-3, perform clustering in combination with NMP clustering rules and use the clustering index to evaluate;

S2-4, calculate the skill factor τ of each individual;

S2-5. Generate offspring; randomly select two individuals a and b from the population as the parent, and generate a random number rand greater than 0 and less than 1. If rand is less than the algorithm parameter rmp or the skill factor τ of the two individuals is equal, Then perform the simulated binary crossover, otherwise perform mutation on the two individuals respectively; repeat the steps of crossover or mutation until the number of offspring is equal to the number of individuals in the population, stop, and enter step S2-6;

S2-6, calculating the fitness value of the generated offspring under each clustering index optimization task;

S2-7, merge the parent and child to form a new population, and recalculate and update the skill factor τ and scalar fitness value φ of all individuals in it according to the fitness value after clustering;

S2-8. Sort the individuals in the population according to the scalar fitness value φ, and then select individuals from superior to inferior to enter the next-generation population, and individuals with larger φ are preferentially selected into the next-generation population; k=k+1;

S2-9, if k<n, return to step S2-3, otherwise stop the iteration;

S2-10, find out the individual whose skill factor τ is equal to 1, and record the individual as the current optimal individual for each task.

Further, in the step S2-2, the population is initialized by the following coding:

d为数据维度数，m_ij为簇中心的坐标向量，T_ij(j＝1,...,K_max)为得到类别质心点的激活阈值；Set the maximum value K _max of the number of cluster categories, and the individual coding of the population is a vector of K _max +K _max *d dimension

d is the number of data dimensions, m _ij is the coordinate vector of the cluster center, and T _ij (j=1,...,K _max ) is the activation threshold for obtaining the class centroid point;

The definition of the activated centroid point is as follows:

If T _ij is greater than 0.5, the centroid point m _ij of the corresponding cluster is activated, otherwise it is not activated; if T _ij is greater than 1 or a negative number, it is reset to 1 or 0, and if the number of centroids obtained is less than the minimum number of categories set, then Several activation threshold activations are randomly selected to satisfy the minimum number of classes.

Further, in step S2-3, the NMP clustering rules are specifically as follows:

Given a data set with a number of N, the data is represented as X=(x ₁ ,...,x _N ), the centroids of K clusters are C={C ₁ ,...,C _K }, and D represents the distance , then, in the NMP rule, the distance between the sample point x _i =(i=0,...,N) and a certain clustering category C _h (h=1,...,K) is defined as follows:

D(x _i ,C _h )=min{D(x _i ,x _j ),D(x _i ,m _h )|x _j ∈C _h }

That is, the distance from the sample to the cluster category is the distance from the sample to the points of the cluster with the smallest distance;

Each sample is assigned to the nearest cluster, and all samples assigned to the same cluster form a candidate sample set, which is called an undetermined cluster of these samples; then, for each cluster, from the candidate sample set Select a nearest sample, and the nearest sample points of all clusters are combined as the nearest sample set; finally, find a sample in the nearest sample set, the distance between this sample and the undetermined cluster where it is located is the smallest in the nearest sample set. The sample is assigned to its undetermined cluster; the above steps are repeated until all samples have been assigned.

Further, in step S2-3, the clustering index includes CH index, Dunn index, and SIL index, respectively corresponding to an optimization task; each index is as follows:

CH indicator:

表示类别间离差矩阵的迹，m表示整个数据集的平均值向量；In the above formula,

Represents the trace of the inter-class dispersion matrix, and m represents the mean vector of the entire dataset;

Dunn Metrics:

δ(C_i)为这个类别的两个最远距离点间距离：

In the above formula, D(C _i , C _j ) represents that the distance between different categories is the distance between the two closest data points. The formula is expressed as follows:

δ(C _i ) is the distance between the two most distant points of this category:

SIL indicator:

In the above formula, s _j =(b _j -a _j )/max(a _j ,b _j ) represents the outline width of the data point x _{j ; the average distances a j and} The formula for calculating the minimum distance b _j to other categories of data points is as follows:

Further, in step S2-5, the operation of simulating binary crossover is as follows:

Given two parents x _a =[x _a (1),...,x _a (d)] and x _b =[x _b (1),...,x _b (d)], d is The dimension of the data, first calculate the distribution factor c(j):

Among them, β is a system parameter greater than 0, and r is a random number greater than 0 and less than 1 in each dimension;

Get the offspring as:

x _e (j)=[(1+c(j))x _a (j)+(1-c(j))x _b (j)]/2

_xf (j)=[(1+c(j))xb(j)+(1- _c (j)) _xa (j)]/2.

Compared with the prior art, the principle and advantages of this scheme are as follows:

1. Through the multi-factor evolution algorithm, a population is simultaneously optimized for the internal indicators of the cluster such as CH index, Dunn index, and SIL index to obtain multiple clustering results, and it is easier to converge to the global optimum through cross-domain communication.

2. The traditional image feature mining method based on color, texture and shape often ignores the information carried by the unique attributes of the medical image ROI. The ROI attributes used in the present invention are relative grayscale, relative area, relative centroid coordinates, circularity, Angle and symmetry, characteristics specific to medical images are also taken into account when considering traditional image features.

3. Clustering rules based on NMP (nearest multiple prototypes): When the traditional particle swarm optimization method is applied to clustering optimization, the effect is not good for the clustering problem of non-circular cluster shape. It is also a distance-based clustering rule. NMP rule is a dynamic process, which is more flexible and usually has better clustering effect.

Description of drawings

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

Fig. 1 is the overall flow chart of the medical image clustering method based on multi-task evolutionary algorithm of the present invention;

Fig. 2 is the flow chart of medical image ROI data extraction;

FIG. 3 is a schematic diagram of coding of multi-task cluster evolution algorithm.

Detailed ways

Below in conjunction with specific embodiment, the present invention will be further described:

As shown in Figure 1, the medical image clustering method based on the multi-task evolutionary algorithm described in this embodiment includes the following steps:

S1. Extract ROI feature description data of the medical image, including relative grayscale s1, relative area s2, relative centroid coordinates s3, circularity s4, angle s5, and symmetry s6.

As shown in Figure 2, the specific process of extraction is as follows:

S1-1, read in medical images;

S1-2. Scan the medical image to obtain the pixel number, length, width and maximum gray value of the image;

S1-3. Detect the symmetry of the medical image s6:

The medical image is folded in half, and the grayscale threshold is used for binarization. If the remaining pixels are less than the set value, the medical image is judged to be symmetrical, otherwise it is asymmetrical;

S1-4, extract the ROI area of the medical image according to the grayscale range;

S1-5, extracting the mean gray value, the number of pixels, the longest axis, the shortest axis, the coordinates of the centroid and the angle from the ROI region obtained in step S1-4;

S1-6. Calculate the feature description data of the ROI area:

其中ROI.gray为ROI区域的平均灰度，IMAGE.gray为整个图像的平均灰度；Relative grayscale s1:

Among them, ROI.gray is the average gray level of the ROI area, and IMAGE.gray is the average gray level of the entire image;

Relative area s2: s2=ROI.area/IMAGE.area, ROI.area is the number of pixels in the ROI area, and IMAGE.area is the number of pixels in the entire image;

Relative centroid coordinate s3: s3=(ROI.x/IMAGE.length, ROI.y/IMAGE.height), ROI.x is the abscissa of the ROI centroid, IMAGE.length is the length of the original image, and ROI.y is the ROI centroid , IMAGE.height is the height of the original image;

Circular s4: s4=4×π×ROI.area/ROI.perimeter ² , ROI.area is the number of pixels in the ROI area, and ROI.perimeter is the number of pixels around the ROI area;

Angle s5: s5=(Orientation+90)/180, and Orientation is the angle between the long axis of the ROI and the X axis.

S2. After extracting the ROI feature description data of the medical image, read in the data, use the NMP clustering rules, and obtain multiple clustering results under the multi-task framework by optimizing the internal indicators of multiple clusters; the specific process is as follows:

S2-1. Read in and extract the ROI feature description data of the medical image;

S2-2, initialize the population and the maximum number of iterations n, k=0; wherein, this step completes the initialization through the following coding:

d为数据维度数，m_ij为簇中心的坐标向量，T_ij(j＝1,...,K_max)为得到类别质心点的激活阈值；Set the maximum value K _max of the number of cluster categories, and the individual coding of the population is a vector of K _max +K _max *d dimension

d is the number of data dimensions, m _ij is the coordinate vector of the cluster center, and T _ij (j=1,...,K _max ) is the activation threshold for obtaining the class centroid point;

The definition of the activated centroid point is as follows:

If T _ij is greater than 0.5, the centroid point m _ij of the corresponding cluster is activated, otherwise it is not activated; if T _ij is greater than 1 or a negative number, it is reset to 1 or 0, and if the number of centroids obtained is less than the minimum number of categories set, then Several activation threshold activations are randomly selected to satisfy the minimum number of classes. As shown in Figure 3 of the accompanying drawings, an individual whose dimension d is 2 and the maximum number of cluster categories Kmax is 4, the front is the activation threshold, and the second position 0.4 is less than 0.5, so its corresponding second centroid is Inactive state, and so on, the threshold value of other positions is greater than 0.5 to be activated. Therefore, the number of clusters for this individual is 3. In practice, d is set to 6.

S2-3, perform clustering in combination with NMP clustering rules and use the clustering index to evaluate;

The NMP clustering rules described in this step are as follows:

Given a data set with a number of N, the data is represented as X=(x ₁ ,...,x _N ), the centroids of K clusters are C={C ₁ ,...,C _K }, and D represents the distance , then, in the NMP rule, the distance between the sample point x _i =(i=0,...,N) and a certain clustering category C _h (h=1,...,K) is defined as follows:

D(x _i ,C _h )=min{D(x _i ,x _j ),D(x _i ,m _h )|x _j ∈C _h }

That is, the distance from the sample to the cluster category is the distance from the sample to the points of the cluster with the smallest distance;

Each sample is assigned to the nearest cluster, and all samples assigned to the same cluster form a candidate sample set, which is called an undetermined cluster of these samples; then, for each cluster, from the candidate sample set Select a nearest sample, and the nearest sample points of all clusters are combined as the nearest sample set; finally, find a sample in the nearest sample set, the distance between this sample and the undetermined cluster where it is located is the smallest in the nearest sample set. The sample is assigned to its undetermined cluster; the above steps are repeated until all samples have been assigned.

The clustering index includes CH index, Dunn index, and SIL index, which correspond to an optimization task respectively. The specific indicators are as follows:

CH indicator:

表示类别间离差矩阵的迹，m表示整个数据集的平均值向量；In the above formula,

Represents the trace of the inter-class dispersion matrix, and m represents the mean vector of the entire dataset;

Dunn Metrics:

δ(C_i)为这个类别的两个最远距离点间距离：

In the above formula, D(C _i , C _j ) represents that the distance between different categories is the distance between the two closest data points. The formula is expressed as follows:

δ(C _i ) is the distance between the two most distant points of this category:

SIL indicator:

In the above formula, s _j =(b _j -a _j )/max(a _j ,b _j ) represents the outline width of the data point x _{j ; the average distances a j and} The formula for calculating the minimum distance b _j to other categories of data points is as follows:

S2-4. Calculate the skill factor τ of each individual, and this value records the optimal task that a single individual i performs best in all tasks. For example, if the individual i is the most advanced in the individual ranking of the jth task, then τ _i =j. When calculating the skill factor, after sorting all individuals in the population according to the size of the fitness value under a certain task, the corresponding serial number in the sequence is an individual i; the factorial level r _i ^{j corresponding to the task j} .

S2-5. Generate offspring; randomly select two individuals a and b from the population as the parent, and generate a random number rand greater than 0 and less than 1, if rand is less than the algorithm parameter rmp (random mating probability) or the two individuals If the skill factor τ is equal, the simulated binary crossover is performed, otherwise, the mutation is performed on the two individuals respectively; the steps of crossover or mutation are repeated until the number of offspring is equal to the number of individuals in the population, and the process goes to step S2-6;

Among them, the operation of simulating binary crossover is as follows:

Given two parents x _a =[x _a (1),...,x _a (d)] and x _b =[x _b (1),...,x _b (d)], d is The dimension of the data, first calculate the distribution factor (spread factor) c(j):

Among them, β is a system parameter greater than 0, and r is a random number greater than 0 and less than 1 in each dimension;

Get the offspring as:

x _e (j)=[(1+c(j))x _a (j)+(1-c(j))x _b (j)]/2

_xf (j)=[(1+c(j))xb(j)+(1- _c (j)) _xa (j)]/2.

S2-6, calculating the fitness value of the generated offspring under each clustering index optimization task;

那么该个体的标量适应值为

即该个体标量适应值由它在表现最好的那个任务中的阶乘等级决定；S2-7. Combine the parent and child to form a new population. After clustering, recalculate and update the skill factor τ and scalar fitness value φ of all individuals in it according to the fitness value; if the factorial of the i-th individual in all K tasks The grades are

Then the scalar fitness of the individual is

That is, the scalar fitness value of the individual is determined by its factorial level in the task that performs best;

S2-8. Sort the individuals in the population according to the scalar fitness value φ, and then select individuals from superior to inferior to enter the next-generation population, and individuals with larger φ are preferentially selected into the next-generation population; k=k+1;

S2-9, if k<n, return to step S2-3, otherwise stop the iteration;

S2-10, find out the individual whose skill factor τ is equal to 1, and record the individual as the current optimal individual for each task.

S3. Three clustering results are obtained after step S2, and an optimal result is selected from the expert knowledge of the doctor.

In the embodiment, the optimization of different indicators can be regarded as different tasks through multi-task clustering, and the optimal individuals under each task can be found respectively, and expert knowledge can be used to find out which clustering indicator is most suitable for this type of medical image. Clustering indicators can choose indicators suitable for the actual situation to achieve better results. This method utilizes the migration process of multi-task learning, and the communication between different learning tasks is conducive to breaking through the local optimum to get closer to the global optimum, and its optimization speed is faster than using the same method to run multiple times for different optimization objectives. Efficient.

The above-mentioned embodiments are only preferred embodiments of the present invention, and are not intended to limit the scope of implementation of the present invention. Therefore, any changes made according to the shape and principle of the present invention should be included within the protection scope of the present invention.

Claims (9)

1. The medical image clustering method based on the multitask evolutionary algorithm is characterized by comprising the following steps of:

s1, extracting ROI feature description data of the medical image;

s2, reading ROI feature description data of the extracted medical image, and obtaining a plurality of clustering results under a multi-task framework by optimizing a plurality of clustering internal indexes by applying an NMP clustering rule;

and S3, selecting an optimal result from the results by using expert knowledge of the doctor.

2. The medical image clustering method based on the multi-task evolutionary algorithm as claimed in claim 1, wherein the ROI feature description data of the medical image extracted in step S1 comprises relative gray S1, relative area S2, relative centroid coordinates S3, circularity S4, angle S5 and symmetry S6.

3. The medical image clustering method based on the multitask evolutionary algorithm according to claim 2, wherein the specific process of extracting the ROI feature description data of the medical image in the step S1 is as follows:

s1-1, reading in a medical image;

s1-2, scanning the medical image to obtain the maximum value of the pixel number, the length, the width and the gray level of the image;

s1-3, detecting the symmetry of the medical image S6;

s1-4, extracting an ROI (region of interest) of the medical image according to the gray scale range;

s1-5, extracting a gray average value, the number of pixels, a longest axis, a shortest axis, a centroid coordinate and an angle from the ROI obtained in the step S1-4;

and S1-6, calculating the feature description data of the ROI area.

4. The medical image clustering method based on the multitask evolutionary algorithm according to the claim 3, wherein the step S1-3 is to detect the symmetry S6 of the medical image by:

folding the medical image for difference, performing binarization processing by using a gray threshold value, if the remaining pixel points are less than a set value, judging that the medical image is symmetrical, otherwise, judging that the medical image is asymmetrical;

the specific process of calculating the feature description data of the ROI region in step S1-6 is as follows:

relative gray s 1:

wherein, ROI.gray is the average gray of the ROI area, and IMAGE.gray is the average gray of the whole image;

relative area s 2: s2 is roi.area, which is the number of pixels in the ROI region, and image.area is the number of pixels in the entire image;

relative centroid coordinate s 3: s3 ═ roi.x/image.length, roi.y/image.height, roi.x is the abscissa of the ROI centroid, image.length is the length of the original image, roi.y is the ordinate of the ROI centroid, and image.height is the height of the original image;

circularity s 4: s4 ═ 4 × pi × roi²ROI is the number of pixels in the ROI region, and ROI is the number of pixels around the ROI region;

angle s 5: s5 (Orientation +90)/180, Orientation being the angle from the long axis of the ROI to the X-axis.

5. The medical image clustering method based on the multitask evolutionary algorithm according to the claim 1, wherein the specific process of the step S2 is as follows:

s2-1, reading ROI feature description data of the extracted medical image;

s2-2, initializing a population and setting the maximum iteration number n, wherein k is 0;

s2-3, combining the NMP clustering rule to execute clustering and evaluating the clustering indexes;

s2-4, calculating the skill factor tau of each individual;

s2-5, generating offspring; randomly selecting two individuals a and b from the population as parents, generating a random number rand which is greater than 0 and less than 1, if rand is less than an algorithm parameter rmp or skill factors tau of the two individuals are equal, executing analog binary crossing, otherwise, respectively executing variation on the two individuals; repeating the step of crossing or mutation until the number of filial generations is equal to the number of population individuals, and then entering the step S2-6;

s2-6, calculating the adaptive value of the generated filial generation under each clustering index optimization task;

s2-7, merging the parents and the offspring to form a new population, and recalculating and updating the skill factors tau and the scalar fitness values phi of all the individuals according to the fitness values after clustering;

s2-8, sorting individuals in the population according to the scalar fitness value phi, then sequentially selecting the individuals from good to bad to enter the next generation of population, and preferentially selecting the individuals with larger phi to enter the next generation of population; k is k + 1;

s2-9, if k is less than n, returning to the step S2-3, otherwise, stopping iteration;

and S2-10, finding out the individuals with the skill factor tau equal to 1, and recording the individuals as the individuals with the optimal current tasks.

6. The medical image clustering method based on the multitask evolution algorithm according to the claim 5, wherein in the step S2-2, the population is initialized by the following codes:

setting the maximum value K of the cluster category number_maxThe individual code of the population is K_max+K_maxVector of dimension x d

d is the number of dimensions of the data, m_ijCoordinate vector of cluster center, T_ij(j＝1,...,K_max) To obtain an activation threshold for a class centroid point;

the definition of the activation-derived centroid point is specifically as follows:

if T_ijIf the mass center point m is larger than 0.5, the corresponding cluster_ijActivated, otherwise not activated; if T_ijGreater than 1 or negative, reset to 1 or 0, if soAnd if the number of the centers of mass is less than the set minimum number of categories, randomly selecting a plurality of activation threshold values to activate so as to meet the requirement of the minimum number of categories.

7. The medical image clustering method based on the multitask evolutionary algorithm according to the claim 5, wherein in the step S2-3, the NMP clustering rule is as follows:

given a number N of data sets, the data is represented as X ═ X (X)₁,...,x_N) The centroid of the K clusters is C ═ C₁,...,C_KD denotes distance, then sample point x in the NMP rule_iA certain cluster class C and (i ═ 0.·, N)_hThe distance of (h ═ 1.., K) is defined as follows:

D(x_i,C_h)＝min{D(x_i,x_j),D(x_i,m_h)|x_j∈C_h}

that is, the distance from the sample to the cluster category is the distance from the sample to the cluster point with the smallest distance;

each sample is assigned to the nearest cluster, and all samples assigned to the same cluster constitute a candidate sample set, and the cluster is called an undetermined cluster of the samples; then, for each cluster, selecting a nearest sample from the candidate sample set, and merging the nearest sample points of all clusters to be called a nearest sample set; finally, finding a sample in the nearest sample set, wherein the distance between the sample and the undetermined cluster in which the sample is located is the smallest in the nearest sample set, and distributing the sample to the undetermined cluster; the above steps are repeated until all samples are allocated.

8. The medical image clustering method based on the multitask evolution algorithm according to the claim 5, wherein in the step S2-3, the clustering index comprises a CH index, a Dunn index and a SIL index, which respectively correspond to an optimization task; the indexes are as follows:

CH index:

in the above formula, the first and second carbon atoms are,

the trace of the inter-class dispersion matrix is represented, and m represents the mean vector of the whole data set;

dunn index:

in the above formula, D (C)_i,C_j) Representing the distance between the different classes as the distance between the two closest data points, the formula is expressed as follows:

(C_i) The two farthest distance points for this category are separated:

SIL index:

in the above formula, s_j＝(b_j-a_j)/max(a_j,b_j) Representing data point x_jThe width of the profile of (a); data point x_jAverage distance a to other data points of the class to which it belongs_jAnd minimum distance b to other category data points_jThe calculation formula of (a) is as follows:

9. the medical image clustering method based on the multitask evolutionary algorithm according to the claim 5, wherein in the step S2-5, the operation of simulating the binary crossing is as follows:

is provided with two parents x_a＝[x_a(1),...,x_a(d)]And x_b＝[x_b(1),...,x_b(d)]D is the dimension of the data, the distribution factor c (j) is first calculated:

wherein, beta is a system parameter larger than 0, and r is a random number larger than 0 and smaller than 1 in each dimension;

the obtained offspring is:

x_e(j)＝[(1+c(j))x_a(j)+(1-c(j))x_b(j)]/2

x_f(j)＝[(1+c(j))x_b(j)+(1-c(j))x_a(j)]/2。

Images