CN101923712A - Particle swarm optimization-based gene chip image segmenting method of K-means clustering algorithm - Google Patents

Abstract

The invention discloses a particle swarm optimization-based gene chip image segmenting method of a K-means clustering algorithm. The particle swarm optimization-based gene chip image segmenting method is characterized by comprising the following steps of: firstly, classifying all pixels of a gene chip image into K types according to the K-means clustering algorithm and searching a local optimal position by each particle in a particle swarm according to a fitness function; and secondly, updating the speed and the position value per se by the particles in the particle swarm according to an individual extreme value and the optimal value. After multiple iterations, the subgroup at the global optimal position is the clustering classified result. The invention has the advantages of simple and definite algorithm process so as to effectively avoid the situation of involving in the local optimization or empty class, high convergence rate, strong global optimal search capability, fewer parameters required to be set and adjusted, accurate and quick result classification and no interference from human factors, and is suitable for segmenting large-scale gene chip images.

Description

Genetic chip image segmentation based on the K-means clustering algorithm of particle group optimizing

Technical field

The present invention relates to a kind of division processing method of image, be specifically related to the method that a kind of K-means clustering algorithm that utilizes particle group optimizing is cut apart automatically to the genetic chip image.

Background technology

Genetic chip (claiming DNA chip or biochip again) is the biotechnology of a kind of novel practical that last century, the mid-80 grew up, and has become one of focus of international life science at present.Biochip technology is based on the hybridization principle, micro-fabrication technique and Protocols in Molecular Biology in conjunction with semi-conductor industry, with the oligonucleotides of enormous amount or cDNA as probe, mode by high speed robot's point sample, sequentially or arrangement mode be fixed on a silicon chip that area is minimum, on the substrate such as slide or nylon membrane, after the dna sequence dna on fluorescently-labeled sample and the chip is by the hybridization of base pairing principle, utilize the laser co-focusing fluorescence detecting system to obtain fluorescence signal, fluorescence signal intensity has been reacted the expression of mRNA in the sample in the different samples.By handling and analyzing gene chip hybridization detected image, can analyze a large amount of gene information in biological cell or the tissue.This technology has the advantages such as concurrency, diversity, miniature property and robotization of height, can be in a large amount of biomolecule of very short time inner analysis, and obtain biological information in the sample rapidly and accurately, thus improved detection efficiency greatly.Biochip technology has become efficiently, has obtained fast, on a large scale the important means of associated biomolecule information.

The researchist is in research and application process to image, often only to some part in the image (zone specific, that have peculiar property in the general correspondence image) interested, this part zone is referred to as target or prospect usually, and other parts are referred to as background.Only on the basis of image segmentation, could carry out feature extraction and parameter measurement, make that more high-rise graphical analysis and understanding becomes possibility target.

It is important step in the genetic chip application process that the image segmentation of genetic chip is handled, the process of genetic chip image segmentation is exactly to determine the process of echo signal (target spot) and background signal, just in background, identify the process of target spot signal, its objective is the monochrome information that will extract target spot in image, personnel for deliberation further explore and research.The analysis result of this process will be directly used in definite result who detects and follow-up research.At present, occurred some both at home and abroad and be specifically designed to the software product of processing and analyzing gene chip image, but needed artificial participation mostly, and existed analysis precision not reach shortcomings such as requirement.

Cluster analysis has in pattern-recognition and image processing field widely uses, its fundamental purpose is given things is distinguished and to be classified by the similarity between things, make the element in each class have identical characteristic as far as possible, the characteristic difference between the different polymeric type is big as much as possible.Image segmentation and object extraction are the main application facet of cluster analysis.K-means clustering algorithm (J.B.MacQueen, 1967) is as the simplest in the clustering algorithm, and is the most a kind of, has in a large number and widely to use.

Find that in actual applications the K-means clustering algorithm has characteristics clearly.Because computing method, extendability and the efficient of K-means clustering algorithm when big data quantity is all more satisfactory, is applicable to that the image segmentation of genetic chip is handled.But when the initialization at class center,, might be absorbed in the situation of local optimum or generation empty class if selection is improper.Simultaneously, different cluster results be may produce, the analysis and the use of cluster result are unfavorable for for different initialization.In addition, the K-means clustering algorithm is subjected to noise and unusual several influence is bigger.

Summary of the invention

The object of the invention provides a kind of genetic chip image segmentation of the K-means clustering algorithm based on particle group optimizing, by introducing particle cluster algorithm optimization, obtain a kind of few, affected by noise little image partition method of parameter that is provided with, the image segmentation that helps genetic chip is handled.

For achieving the above object, the present invention at first is divided into the K class according to all pixels of K-means clustering algorithm genetic chip image, and each particle in the population is searched the local optimum position according to fitness function; Particle in the population upgrades speed and the positional value of oneself according to its individual extreme value and optimal location then.Through after the iteration repeatedly, subgroup, place, global optimum position is produced is the Cluster Classification result.This algorithm exchanges the information between the particle by to the cluster of population, and the information of having utilized multiparticle more to comprise in the iteration searching process, and the global convergence of algorithm is stronger.

The concrete technical scheme that adopts is: a kind of genetic chip image segmentation of the K-means clustering algorithm based on particle group optimizing comprises the following steps:

(1) imports the genetic chip image, and the genetic chip image is carried out pre-service;

Described pre-service comprises, the genetic chip image is converted into monochromatic gray level image, by the method for mathematical morphology this monochrome gray level image is carried out Filtering Processing, and connected component is less than the image section elimination of n pixel, and wherein, n gets the integer between 15～50;

(2) image after step (1) is handled carries out grid location, obtains a plurality of genetic chip images target area, and a target spot and the background area thereof of each image target area after by grid location constitutes;

(3) respectively image segmentation being carried out in each image target area handles, described image segmentation is treated to, a pixel is by a data vector representation, the horizontal ordinate of the horizontal ordinate of data vector and ordinate corresponding pixel points and ordinate, the gray-scale value of data vector value corresponding pixel points, adopt the K-means clustering algorithm based on particle group optimizing to carry out to the division of data vector, all pixels are divided into target and background two classes the most at last, realize cutting apart of genetic chip image; Described K-means clustering algorithm based on particle group optimizing is, at first according to the K-means clustering algorithm all pixels of genetic chip image is divided into the K class, K=2, and each particle in the population is searched the local optimum position according to fitness function; Particle in the population upgrades speed and the positional value of oneself according to its individual extreme value and optimal location then; Through after the iteration repeatedly, the subgroup at place, global optimum position is the Cluster Classification result.

In the technique scheme, a particle i represents potential separating of genetic chip image segmentation, and the good and bad degree of separating is weighed by fitness function.Each particle i comprises following information: x _i(t) be the current location of particle; v _i(t) be the current flight speed of particle, i.e. the distance that moves of particle; y _i(t) be the individual optimal location of particle.The individual optimal location of a particle is the optimal location that particle searches up to now, just can produce maximum fitness value.

The K-means clustering algorithm step of described particle group optimizing is as follows:

1) initialization: configure population N and maximum iteration time t _Max, generate each particle position vector x at random _i(t) and velocity vector v _i(t), generate cluster centre m at random _Ic, i=1,2 ..., N, cluster numbers is set to K, thus c=1 ..., K;

2) to each particle i, do down column operations:

A) calculating pixel point z _p(z _pPixel gray-scale value vector for image to be split) to its cluster centre m _IcEuclidean distance

By minimal distance principle

Pixel is reassigned to each cluster C _Ij, j=1 ..., K;

B) each cluster C _IjAccording to

Calculate cluster centre m _Ic, n _jFor belonging to cluster C _IjThe pixel number;

C) calculate fitness function

3) calculate local optimum position this moment

With the global optimum position

Wherein

4) according to v _i(t+1)=ω v _i(t)+c ₁r ₁(t) (y _l(t)-x _i(t))+c ₂r ₂(t) (y _g(t)-x _i(t)) upgrade particle's velocity, if v _i(t+1)＞v _MaxV then _i(t+1)=v _MaxThen according to x _i(t+1)=x _i(t)+v _i(t+1) to upgrading particle position;

Step 2), 3), 4) circulation t _MaxInferior, the position y of global optimum _g(t) the Cluster Classification result that produced of subgroup, place is net result.

In the technique scheme, described each particle is searched the local optimum position according to fitness function, and the good and bad degree of each particle is determined by following fitness function:

$f(m_{\mathrm{ic}},Z)={\mathrm{\ω}}_1{\stackrel{\‾}{d}}_{\max }(Z,m_{\mathrm{ic}})+{\mathrm{\ω}}_2(z_{\max }-d_{\min }\left(m_{\mathrm{ic}}\right))$

Wherein, z _MaxBe grey scale pixel value maximum in the gradation of image value vector; Z is the matrix of display pixel distribution condition, as a certain element z wherein _IjpRemarked pixel point z _pThe C that whether belongs to particle i _IjCluster; ω ₁And ω ₂Be the given positive constant of user, different initial values can cause different first search schemes;

Be the maximum average range in the corresponding particle i cluster;

Be the Euclidean distance of pixel p to the cluster average;

Be the minor increment between cluster; | C _Ij| for belonging to cluster C _IjThe pixel number.

By constantly searching for the minimum value of fitness function, minimize distance and maximization between class distance in the class, thereby hunt out optimum classification schemes.

Because the technique scheme utilization, the present invention compared with prior art has following advantage:

1. the present invention makes the K-means clustering algorithm not need to treat the prior distribution knowledge of grouped data by introducing particle cluster algorithm optimization, also lessly is subjected to the influence that initial solution selects and obtains suboptimal solution, and the image segmentation that helps genetic chip is handled;

2. when the present invention was cut apart the genetic chip image, the algorithmic procedure of employing was simply clear, was easy to realize; Cluster centre produces at random, is not subjected to interference from human factor; The situation of effectively avoiding being absorbed in local optimum or producing empty class; Fast convergence rate, search global optimum ability is strong; The parameter that needs setting and adjust is few.

Description of drawings

Fig. 1 is first width of cloth original color genetic chip image in the embodiment of the invention.

Fig. 2 is the gray level image of Fig. 1.

Fig. 3 is the genetic chip image after the pre-service of Fig. 2 process.

Fig. 4 for Fig. 3 through the image after the grid location.

Fig. 5 is the genetic chip bianry image of Fig. 4 after cutting apart through the K-means clustering algorithm of particle group optimizing of the present invention.

Fig. 6 is the genetic chip bianry image of Fig. 4 after cutting apart through the K-means clustering algorithm.

Fig. 7 is the result of target edge stack among Fig. 1 and Fig. 5.

Fig. 8 is finally cut apart the gained result for what Fig. 1 utilized the method for the invention.

Fig. 9 utilizes the segmentation result of K-means clustering method for Fig. 1.

Figure 10 is second width of cloth original color genetic chip image.

Figure 11 utilizes the segmentation result of K-means clustering method for Figure 10.

Figure 12 is finally cut apart the gained result for what Figure 10 utilized the method for the invention.

Embodiment

Below in conjunction with drawings and Examples the present invention is further described:

Embodiment:

A kind of genetic chip image segmentation of the K-means clustering algorithm based on particle group optimizing comprises following steps:

(1) pre-service of genetic chip image.Because obtaining in the technical process of genetic chip image, be subjected to the influence of substrate impurity, LASER Light Source and scanning process etc. easily, various noises can appear in image, as shown in Figure 1, directly influence accuracy and the result of experiment that detects.Therefore in this step, at first the color fluorescence genetic chip image with rgb format is converted into monochromatic gray level image, as shown in Figure 2, has simplified the processing procedure of computing machine greatly; By the method for mathematical morphology (JeanSerra, 1986) the genetic chip image is carried out Filtering Processing then.Fig. 3 is that at first to utilize radius be the ψ of 10 pixels _TopHatOperator strengthens Fig. 2, then uses the mathematical morphology area to open the image section elimination that filtering is less than connected component 20 pixels.

(2) grid location of genetic chip image.Grid location is exactly in the fluoroscopic image zone of genetic chip, determines each target area and target position, and obtains the process of related chip data.Do not need artificial participation, draw the target that lattice are genetic chip grid location automatically by computing machine fully, can accelerate the speed of Computer Processing like this, the error that can avoid artificial participation to bring again.Utilize the level of image and vertical projection signal that the genetic chip image is carried out grid location based on the method for the automatic division grid of mathematical morphology method (R.Hi rata, J.Barrera, R.F.Hash i mot o, and D.0.Dantas, 2001).Fig. 4 is that Fig. 3 is through the image after the grid location.

(3) the genetic chip image cuts apart.By the processing of above-mentioned steps, can obtain independent genetic chip image target area (target spot and its background area).In this step, image is cut apart automatically by the K-means clustering algorithm of particle group optimizing.

The purpose of K-means clustering algorithm is that some data vectors are divided in the known cluster of classification number.For image classification or image segmentation, a pixel of a data vector representative image.Each pixel all is divided to nearest average or cluster centre; When all pixels all are classified, the average of each cluster is calculated according to the pixel of repartitioning; Till the average of each cluster does not obviously change.The K-means clustering algorithm may be summarized to be as follows:

A) at random for each cluster produces an initial cluster centre, there be K cluster centre in K cluster;

B) each sample is assigned to some in K the cluster according to minimal distance principle;

C) with the average of all samples of each cluster as new cluster centre;

D) the c repeating step b if cluster centre changes)) till marked change no longer takes place in cluster centre;

A resulting K cluster centre is exactly K the division arrival square error minimum that clustering result K-meanss clustering algorithm is determined.When cluster be intensive and class and class between difference when obvious, effect is better.For handling large data sets, this algorithm is scalable relatively and efficiently.But the selection to initial cluster center in algorithm has bigger influence to cluster result, in case the initial value selection is bad, possibly can't obtain effective cluster result, and this also becomes a subject matter of K-means algorithm.Owing to the reason of algorithm " greed " itself, cluster result may be absorbed in the situation of local optimum in addition.

At the shortcoming of K-means clustering algorithm, the present invention introduces particle swarm optimization algorithm.In particle cluster algorithm, each individuality is called one " particle ", potential separating of each particle representing optimized problem, and the good and bad degree of separating is weighed by fitness function.

Each particle i comprises following information: x _i(t) be the current location of particle; v _i(t) be the current flight speed of particle, i.e. the distance that moves of particle; y _i(t) be the individual optimal location of particle.。The individual optimal location of a particle is the optimal location that particle searches up to now, just can produce maximum fitness value.In each iteration, the individual optimal location of particle upgrades according to following formula:

$y_i(t+1)=\left\{\begin{array}{ccc}{y}_i\left(t\right)& \mathrm{if}& f\left(x_i(t+1)\right)\&GreaterEqual;f\left(y_i\left(t\right)\right)\\ x_i(t+1)& \mathrm{if}& f\left(x_i(t+1)\right)<f\left(y_i\left(t\right)\right)\end{array}\right.---\left(1\right)$

For this particle, current local optimum position is

$y_l(t+1)=\min \{y_i(t+1),\&ForAll;t\}---\left(2\right)$

For whole population, current whole optimal locations are

$y_g(t+1)=\min \{y_l(t+1),\&ForAll;t\}---\left(3\right)$

And each particle is according to following formula renewal speed and position:

v _i(t+1)＝ωv _i(t)+c ₁r ₁(t)(y _l(t)-x _i(t))c ₂r ₂(t)(y _g(t)-x _i(t))(4)

x _i(t+1)＝x _i(t)+v _i(t+1) (5)

Wherein, ω is an inertia weight, plays balance local optimum and global optimum; r ₁(t) and r ₂(t) for being evenly distributed on the random number between (0,1), be used for keeping the diversity of colony; c ₁With c ₂Be the study factor, make particle can the oneself sum up and in colony outstanding particle learn, thereby can be close to local optimum and global optimum, regulate these two parameters and can jump out local minimum and quickening speed of convergence; The maximal value v of the flying speed of particle _MaxMust be controlled, otherwise may be caused too early convergence.

In the present invention, each particle is represented K cluster centre.Therefore, population is represented candidate's image classification result's set.To classification schemes evaluation is the key that the optimizing application algorithm carries out cluster.The present invention is determined by following fitness function the good and bad degree of each particle:

$f(m_{\mathrm{ic}},Z)={\mathrm{\&omega;}}_1{\stackrel{\&OverBar;}{d}}_{\max }(Z,m_{\mathrm{ic}})+{\mathrm{\&omega;}}_2(z_{\max }-d_{\min }\left(m_{\mathrm{ic}}\right))---\left(6\right)$

Wherein, z _MaxBe grey scale pixel value maximum in the gradation of image value vector; Z is the matrix of display pixel distribution condition, as a certain element z wherein _IjpRemarked pixel point z _pThe C that whether belongs to particle i _IjCluster; ω ₁And ω ₂Be the given positive constant of user, different initial values can cause different first search schemes;

Be the maximum average range in the corresponding particle i cluster;

Be the Euclidean distance of pixel p to the cluster average;

Be the minor increment between cluster; | C _Ij| for belonging to cluster C _IjThe pixel number.

By the minimum value of continuous search fitness function, can minimize distance and maximization between class distance in the class, thereby hunt out optimum classification schemes.

Concrete dividing method is as follows:

1) initialization.Configure population N, set maximum iteration time t _Max, generate each particle position vector x at random _i(t) and velocity vector v _i(t), generate cluster centre m at random _Ic, i=1,2 ..., N, cluster numbers is set to K, thus c=1 ..., K;

2) to each particle i, do down column operations:

A) calculating each pixel arrives and cluster centre m _IcEuclidean distance d (z _p, m _Ic), by minimal distance principle

Pixel is reassigned to each cluster C _Ij, j=1 ..., K;

B) each cluster C _IjAccording to

Calculate cluster centre m _Ic, n _jFor belonging to cluster C _IjThe pixel number;

C) calculate fitness function f (m according to formula (6) _Ic, Z);

3) calculate local optimum position y this moment respectively according to formula (2) and formula (3) _l(t) and the position y of global optimum _g(t);

4) upgrade particle's velocity according to formula (4), if v _i(t+1)＞v _MaxV then _i(t+1)=v _MaxUpgrade particle position according to formula (5) then.

Step 2), 3), 4) circulation t _MaxInferior.Final Cluster Classification result is the position y of global optimum _g(t) the Cluster Classification result that produced of subgroup, place.

Fig. 5 is the binary map of the genetic chip after utilizing the method for the invention to cut apart at Fig. 4; Fig. 6 is for utilizing the binary map of the genetic chip after the K-means clustering method is cut apart at Fig. 4; Fig. 7 is the result of target edge stack among original gene chip image and Fig. 5; Fig. 8 be utilize the method for the invention finally cut apart gained result (population N _kGet 30, iterations t _MaxGet 30, it is target and background two classes that cluster numbers K gets 2, inertia weight and study factor ω=c ₁=c ₂=0.5, the maximal value v of the flying speed of particle _Max=2, ω ₁=ω ₂=0.45); Fig. 9 is the segmentation result that utilizes the K-means clustering method.Contrast by Fig. 5 and Fig. 6 can obviously be found out, utilizes the binary map of the genetic chip after the K-means clustering method is cut apart to keep a lot of noises, and these will impact final result, and the contrast of Fig. 9 and Fig. 8 has just demonstrated this point.Fig. 9 has obviously been Duoed several noise spots than Fig. 8, and the method for the invention is with respect to utilizing the K-means clustering method that certain advantage is arranged.Figure 10 is second width of cloth original color genetic chip image; Figure 11 utilizes the segmentation result of K-means clustering method for Figure 10; Figure 12 is finally cut apart the gained result for what Figure 10 utilized the method for the invention.By contrast as can be seen, the method for the invention is with respect to the K-means clustering method, has incomparable advantage at the aspect such as smooth of the removal of the removal of ground unrest, target internal noise, fuzzy edge.

The present invention utilizes particle group optimizing K-means clustering algorithm to the genetic chip image is cut apart automatically, algorithmic procedure is simply clear, be not subjected to interference from human factor, that effectively avoids may occurring in the K-means clustering algorithm is absorbed in local optimum or produces the situation of empty class, fast convergence rate, it is strong to search for overall ability, keeps extendability and the efficient of K-means clustering algorithm when big data quantity simultaneously.

Claims (4)

1. the genetic chip image segmentation based on the K-means clustering algorithm of particle group optimizing is characterized in that, comprises the following steps:

(1) imports the genetic chip image, and the genetic chip image is carried out pre-service;

(2) image after step (1) is handled carries out grid location, obtains a plurality of genetic chip images target area, and a target spot and the background area thereof of each image target area after by grid location constitutes;

(3) respectively image segmentation being carried out in each image target area handles, described image segmentation is treated to, a pixel is by a data vector representation, the horizontal ordinate of the horizontal ordinate of data vector and ordinate corresponding pixel points and ordinate, the gray-scale value of data vector value corresponding pixel points, adopt the K-means clustering algorithm based on particle group optimizing to carry out to the division of data vector, all pixels are divided into target and background two classes the most at last, realize cutting apart of genetic chip image; Described K-means clustering algorithm based on particle group optimizing is, at first according to the K-means clustering algorithm all pixels of genetic chip image is divided into the K class, K=2, and each particle in the population is searched the local optimum position according to fitness function; Particle in the population upgrades speed and the positional value of oneself according to its individual extreme value and optimal location then; Through after the iteration repeatedly, the subgroup at place, global optimum position is the Cluster Classification result.

2. the genetic chip image segmentation of the K-means clustering algorithm based on particle group optimizing according to claim 1, it is characterized in that: pre-service comprises described in the step (1), the genetic chip image is converted into monochromatic gray level image, method by mathematical morphology is carried out Filtering Processing to this monochrome gray level image, connected component is less than the image section elimination of n pixel, wherein, n gets the integer between 15～50.

3. the genetic chip image segmentation of the K-means clustering algorithm based on particle group optimizing according to claim 1 is characterized in that:

The K-means clustering algorithm step of described particle group optimizing is as follows:

1) initialization: configure population N and maximum iteration time t _Max, generate each particle position vector x at random _i(t) and velocity vector v _i(t), generate cluster centre vector m at random _Ic, i=1,2 ..., N, cluster numbers is set to K, thus c=1 ..., K;

2) to each particle i, do down column operations:

A) calculating pixel point z _p(z _pPixel gray-scale value vector for image to be split) to its cluster centre m _IcEuclidean distance

By minimal distance principle Pixel is reassigned to each cluster C _Ij, j=1 ..., K;

B) each cluster C _IjAccording to

Calculate cluster centre m _Ic, n _jFor belonging to cluster C _IjThe pixel number;

C) calculate fitness function

Wherein, z _MaxBe grey scale pixel value maximum in the gradation of image value vector; Z is the matrix of display pixel distribution condition, as a certain element z wherein _IjpRemarked pixel point z _pThe C that whether belongs to particle i _IjCluster; ω ₁And ω ₂Be the given positive constant of user, different initial values can cause different first search schemes;

Be the maximum average range in the corresponding particle i cluster;

Be the Euclidean distance of pixel p to the cluster average;

Be the minor increment between cluster; | C _Ij| for belonging to cluster C _IjThe pixel number;

3) calculate local optimum position this moment

With the global optimum position

Wherein

4) according to v _i(t+1)=ω v _i(t)+c ₁r ₁(t) (y _l(t)-x _i(t))+c ₂r ₂(t) (y _g(t)-x _i(t)) upgrade particle's velocity, if v _i(t+1)＞v _MaxV then _i(t+1)=v _MaxThen according to x _i(t+1)=x _i(t)+v _i(t+1) upgrade particle position;

Step 2), 3), 4) circulation t _MaxInferior, the position y of global optimum _g(t) the Cluster Classification result that produced of subgroup, place is net result.

4. the genetic chip image segmentation of the K-means clustering algorithm based on particle group optimizing according to claim 1, it is characterized in that: described each particle is searched the local optimum position according to fitness function, and the good and bad degree of each particle is determined by following fitness function:

$f(m_{\mathrm{ic}},Z)={\mathrm{\&omega;}}_1{\stackrel{\&OverBar;}{d}}_{\max }(Z,m_{\mathrm{ic}})+{\mathrm{\&omega;}}_2(z_{\max }-d_{\min }\left(m_{\mathrm{ic}}\right))$

Wherein, z _MaxBe grey scale pixel value maximum in the gradation of image value vector; Z is the matrix of display pixel distribution condition, as a certain element z wherein _IjpRemarked pixel point z _pThe C that whether belongs to particle i _IjCluster; ω ₁And ω ₂Be the given positive constant of user, different initial values can cause different first search schemes;

Be the maximum average range in the corresponding particle i cluster;

Be the Euclidean distance of pixel p to the cluster average;

Be the minor increment between cluster; | C _Ij| for belonging to cluster C _IjThe pixel number.

By constantly searching for the minimum value of fitness function, minimize distance and maximization between class distance in the class, thereby hunt out optimum classification schemes.

Images