CN107330869A - Extraordinary image vegetarian refreshments reconstruct after overlapping cell segmentation - Google Patents
Extraordinary image vegetarian refreshments reconstruct after overlapping cell segmentation Download PDFInfo
- Publication number
- CN107330869A CN107330869A CN201710504878.2A CN201710504878A CN107330869A CN 107330869 A CN107330869 A CN 107330869A CN 201710504878 A CN201710504878 A CN 201710504878A CN 107330869 A CN107330869 A CN 107330869A
- Authority
- CN
- China
- Prior art keywords
- cell
- formula
- model
- value
- parameter
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 230000011218 segmentation Effects 0.000 title claims abstract description 29
- 238000000034 method Methods 0.000 claims abstract description 32
- 230000003287 optical effect Effects 0.000 claims abstract description 10
- 230000002159 abnormal effect Effects 0.000 claims abstract description 8
- 238000012549 training Methods 0.000 claims description 19
- 239000011159 matrix material Substances 0.000 claims description 15
- 238000004364 calculation method Methods 0.000 claims description 13
- 230000008439 repair process Effects 0.000 claims description 10
- 230000003044 adaptive effect Effects 0.000 claims description 7
- 230000008859 change Effects 0.000 claims description 7
- 238000001514 detection method Methods 0.000 claims description 7
- 230000000694 effects Effects 0.000 claims description 6
- 238000007476 Maximum Likelihood Methods 0.000 claims description 2
- 238000000205 computational method Methods 0.000 claims 3
- 238000003745 diagnosis Methods 0.000 abstract description 7
- 238000005516 engineering process Methods 0.000 abstract description 5
- 238000005259 measurement Methods 0.000 abstract description 3
- 238000002474 experimental method Methods 0.000 abstract description 2
- 230000009467 reduction Effects 0.000 abstract description 2
- 206010028980 Neoplasm Diseases 0.000 abstract 1
- 201000011510 cancer Diseases 0.000 abstract 1
- 210000004027 cell Anatomy 0.000 description 69
- 238000004458 analytical method Methods 0.000 description 4
- 210000004698 lymphocyte Anatomy 0.000 description 4
- 208000006105 Uterine Cervical Neoplasms Diseases 0.000 description 3
- 201000010881 cervical cancer Diseases 0.000 description 3
- 208000019065 cervical carcinoma Diseases 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 201000010099 disease Diseases 0.000 description 2
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 2
- 238000009499 grossing Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000007704 transition Effects 0.000 description 2
- 210000004881 tumor cell Anatomy 0.000 description 2
- 206010008342 Cervix carcinoma Diseases 0.000 description 1
- 241001269238 Data Species 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 210000002919 epithelial cell Anatomy 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 239000012535 impurity Substances 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012758 nuclear staining Methods 0.000 description 1
- 230000001575 pathological effect Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30096—Tumor; Lesion
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
- Investigating Or Analysing Biological Materials (AREA)
Abstract
Extraordinary image vegetarian refreshments reconstruct after overlapping cell segmentation, the present invention relates to the problem of the pixel in DNA ploidy body analytical technology, occurred after overlapping cell segmentation exception.DNA ploidy body analytical technology measures the relative amount of cell DNA by image processing techniques, is had a wide range of applications in terms of cancer diagnosis.But the pixel after the segmentation of overlapping cell is abnormal, causes the features such as texture, gray scale and the most important optical density of cell to occur abnormal, easily occur mistaken diagnosis in diagnosis.To improve this problem, the present invention proposes a kind of cell overlap area pixel reconstructing method based on GMM UBM models.Experiment shows that this method can effectively adjust the characteristic values such as texture, gray scale, the optical density of cell, reduces the error of DNA content measurement, influence of the reduction extraordinary image vegetarian refreshments to grader discrimination.The present invention is applied to the extraordinary image vegetarian refreshments reconstruct after overlapping cell segmentation.
Description
Technical field
Extraordinary image vegetarian refreshments reconstruct after the overlapping cell segmentation of present invention design.
Background technology
Cervical carcinoma is occurred frequently in recent years, has become the social concern for threatening women's life.It is annual new in women worldwide
The people of cervix cancer about 52.76 ten thousand is diagnosed, nearly 26.50 ten thousand people dies from the disease, wherein 90 % the dead are located at under-developed area.It is Chinese every
Year new discovery cervix cancer number of cases about 7.5 ten thousand, accounts for global cervical carcinoma newly sends out total number of persons 1/7,3.5 ten thousand people and dies from this disease.At present
Effective screening methods of cervical cancer is cervical exfoliated cell smear pathologic finding.This method needs veteran pathologist to exist
Diagnosis is made after Microscopic observation sick cell again, this will expend substantial amounts of manpower and materials, it is difficult to adapt to the demand of reality.Computer
Auxiliary diagosis technology can effectively solve this problem, typically there is the analysis of DNA ploidy body.The technology is first with Feulgen to cell
Nuclear staining, is then recognized the various types of cells and rubbish impurity on image and picks out epithelial cell, finally surveyed using image
The method of amount determines DNA content in nucleus and is used as the foundation for judging abnormal cell.
The purpose of cell segmentation is to divide the image into for single cell, and is analyzed in units of cell, and this is
The premise of DNA ploidy body analysis.Some overlapping cells inevitably occur in cell picture, and after the segmentation of overlapping cell point
Pixel abnormal problem occurs.This will cause the features such as texture, gray scale and the most important optical density of cell deviation occur,
Reduce the accuracy rate of cell DNA measurement accuracy and diagnosis.No matter which kind of image repair method, its basic principle is using existing
The priori for having sample goes to repair the pixel of affected area.Whether therefore whether priori is important, and can be by fine profit
With this directly affects the quality repaired.But in cell reconstitution, an effective pixel of cell is very limited, cause instruction
White silk data are insufficient, and this causes current method to be all difficult to obtain preferable effect.GMM-UBM is based on set forth herein one kind
（Gaussian Mixture Models, GMM Universal Background Model, UBM）The overlapping region weight of model
Structure method, effectively to solve this problem.GMM, which is used to be distributed each cell pixel, to be modeled, and instruction is needed containing quantity of parameters
Practice.Less data volume can not meet this requirement.GMM-UBM trains a general UBM model with mass data, recycles
Each cell one specific GMM model of distinctive information self-adapting.The model has been obtained extensively in Speaker Identification field
Application, adaptability is stronger in terms of low volume data.
The content of the invention
Occur pixel abnormal problem after segmentation point the invention aims to solve overlapping cell, and propose
A kind of overlapping region reconstructing method based on GMM-UBM models.
Foregoing invention purpose is mainly achieved through the following technical solutions.
S1,1000 unicellular data for randomly selecting same sample, do training data.
S2, utilize the general gauss hybrid models UBM of EM Algorithm for Training.
The probability density formula of mixed Gauss model is as follows：
(1)
In formulamSpecified for degree of mixing generally according to experimental result, hereinm=3,π k Represent thekThe weight of individual single Gauss model,N
(x, μ k , C) for thekThe probability density of individual single Gauss model, wherein；The probability density formula of single Gauss is as follows：
(2)
Whereinμ k For mathematic expectaion,CFor covariance matrix.
EM algorithms are a kind of maximal possibility estimation sides that probabilistic model parameter is solved from the data set that there is implicit variable
Method；The training of gauss hybrid models is exactly what EM algorithms were realized, and implicit variable therein isπ k ；GMM model is understood by formula 1,2
It needs to be determined that parameter haveπ k 、μ k 、C；EM algorithms estimation gauss hybrid models have two steps；EStep, by initialization or
The parameter of each Gaussian component, goes to estimate the weights of each Gaussian component, weights known to the result of person's previous stepSuch as following formula
It is shown；MStep, the weights based on estimation, then go to determine the parameter of Gaussian component, thekThe expectation of individual Gaussian componentμ k With association side
Poor matrixC k It is shown below.The two steps are repeated, until fluctuation very little, extreme value are reached approximately；
(3)
It is the in formulaiIndividual sample iskProbability in individual single Gauss model；
(4)
(5)
(6)
In formulaN k Expression belongs toKThe number of samples of individual Gaussian component.
S3, randomly select segmentation after cell normal segments information, utilize MAP algorithm adaptive GMMs GMM.
MAP algorithms are divided into two steps：The first step is identical with the E steps of EM algorithms, calculates each Gaussian component as shown in Equation 3
Weights, and be calculated as follows shown parameter：
(7)
(8)
Second step byN k Modifying factor is obtained, carrys out new and old UBM parameter, for Gaussian componentk, the expression formula that parameter is calculated
It is as follows respectively：
(9)
(10)
(11)
Wherein,γAllow all mixed weight-value and be 1,β k π ,β k μ ,β k C For the weight of Gaussian component, mean vector, covariance
The modifying factor of matrix, its effect is the new and old parameter for balancing GMM model, and its value is bigger, illustrates that data are more abundant, also
It is that new parameter is more credible；And if it is smaller, just illustrate the negligible amounts of data, more inaccurate is also become to GMM model estimation,
Its definition is as follows：
(12)
In formulaλ ρ It is constraint modifying factor for relationship factorβ k ρ Change yardstick（ρ∈ π,μ,C}）,λGenerally take 16.
S4, overlapping cell segmentation method is utilized to split overlapping cell.
S5, generated at random using GMM model and meet the gray value of constraints to change unusual part.
The diagnosis principle of the DNA ploidy body analysis system of area of computer aided diagosis is：The picture of suspected tumor cells is enumerated
Out, that is, DI is found>2.5 cell picture；DI calculation formula is as shown in Equation 13, because the DNA content of lymphocyte is relative
Stablize therefore be used as Standard values by the use of the IOD averages of sample medium size lymphocyte；The calculation formula of IOD values is as follows：
(13)
(14)
Wherein, be corresponding pixel points OD value it is as follows：
(15)
Wherein, the average gray value of background is represented, the gray value of pixel is represented；For lap gray value abnormal problem,
The gray value of required generation will make it not influence the IOD values of initial cell, then will limit its value in an effective codomain Y
It is interior；The OD value of i.e. newly-generated gray value will in the both sides of the OD value average of initial cell normal segments, codomain Y's
Calculation formula is as follows：
(16)
In formulaOD m For the mean optical density of nucleus normal segments,OD s For the optical density standard deviation of nucleus normal segments, by
Calculate and understand codomain。
S6, utilize the newly-generated part of median filter smoothness of image.
S7, repair using FMM algorithms newly-generated part and normal segments be connected edge.
The selection of repairing area is by taking two cell overlaps as an example, if the edge contour point set of the two cells is combined intoC 1、C 2 ,
The set of the lap profile point of two overlapping cellsM；Set is asked for firstM, it is made up of two parts, i.e., in profileC 1It is interior
ProfileC 2 On point and in profileC 2 Interior profileC 1On point.Then will setMExpand obtained regionDIt is exactly to be repaired
Region；FMM algorithm principles：Assuming that restoring areaDInpThe gray value of point, with pointpCentered on choose a small neighbourhoodβε,qFor it
In a bit, pointpThe calculation formula of modified values is as follows：
(17)
In formula△I(q) beqThe luminance gradient value of point,w(p, q)=dir(p, q)·dst(p, q)·lev(p, q), wherein
dir(p, q) be direction factor, dst (p, q) be the geometric distance factor, lev (p, q) for level set apart from the factor.
Invention effect
The invention provides a kind of overlapping region reconstructing method based on GMM-UBM models.This algorithm is chosen a large amount of slender first
Born of the same parents' image data trains a UBM, for the gray value modeling to all cells.Then the normal segments with each cell are adaptive
A GMM is answered to be used for the distribution modeling to the cell gray value.To the lap of each cell, ash is generated at random with its GMM
Angle value is filled into overlapping region, and to prevent random value unbalance, adds strict gray value limitation.Utilize the method for medium filtering
The rough grain of smooth newly-generated part.In order to solve newly-generated part and the unsmooth problem of transition of background and normal segments,
Proposition utilizes FMM（Fast Matching Method, FMM）Image repair algorithm repairs rank come the information according to linking edge
Socket part point.Experiment shows that this method can effectively adjust the characteristic values such as texture, gray scale, the optical density of cell, reduces DNA and contains
The error of measurement, influence of the reduction extraordinary image vegetarian refreshments to grader discrimination.
Brief description of the drawings
The overlapping cell separation process diagrams of Fig. 1；
The realization procedure chart of Fig. 2 this paper algorithms；
Design sketch after a variety of overlapping cell reconstitutions of Fig. 3.
Specific implementation method
To make the purpose, technical scheme and advantage of the embodiment of the present invention clearer, below in conjunction with the embodiment of the present invention
Accompanying drawing, the technical scheme in the embodiment of the present invention is clearly and completely described, it is clear that described embodiment is this hair
Bright a part of embodiment, rather than whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art are not having
There is the every other embodiment made and obtained under the premise of creative work, belong to the scope of protection of the invention.
Embodiment 1：
As shown in Figure 1 provided herein is overlapping cell segmentation after extraordinary image vegetarian refreshments reconstructing method, include model training and cell
Reconstruct：
The model training includes step：
S1,1000 unicellular data for randomly selecting same sample, do training data；
S2, utilize the general gauss hybrid models UBM of EM Algorithm for Training；
S3, randomly select segmentation after cell normal segments information, utilize MAP algorithm adaptive GMMs GMM；
The cell reconstitution includes step：
S4, overlapping cell segmentation method is utilized to split overlapping cell；
S5, generated at random using GMM model and meet the gray value of constraints to change unusual part；
S6, utilize the newly-generated part of median filter smoothness of image；
S7, repair using FMM algorithms newly-generated part and normal segments be connected edge.
The mass data that the embodiment of the present invention chooses same sample in the model training stage is calculated as characteristic vector using EM
Method iteration obtains a general UBM model；Cell after being split again by overlapping cell segmentation algorithm, and choose
The normal segments information of cell after segmentation, utilizes the distinctive GMM model of the adaptive cell of MAP algorithms；In cell reconstitution
Stage goes out to meet the gray value of constraints using the model prediction, and by newly-generated part smoothing techniques；Finally utilize FMM
Algorithm repairs the obvious marginal portion of transition.
The embodiment of the present invention is described in detail below：
The embodiment of the present invention realizes the reconstruct of cell after segmentation using inventive algorithm using the cell of 6 different reduplicative formses
Implement down.
Model training as shown in Figure 1 includes step.
, randomly select 1000 unicellular data of same sample, do training data.
Recognize and be partitioned into 1000 unicellular pictures of same sample first, then from the nucleus of each picture with
Machine chooses 300 gray values, finally inputs 300,000 gradation datas as characteristic vector.
S2, utilize the general gauss hybrid models UBM of EM Algorithm for Training；The weighted sum shape that GMM passes through several Gauss models
Into probability distribution data category is modeled mixed Gauss model probability density formula it is as follows：
(1)
In formulamSpecified for degree of mixing generally according to experimental result, hereinm=3,π k Represent thekThe weight of individual single Gauss model,N
(x, μ k , C) for thekThe probability density of individual single Gauss model, wherein；The probability density formula of single Gauss is as follows：
(2)
Whereinμ k For mathematic expectaion,CFor covariance matrix.
EM algorithms are a kind of maximal possibility estimation sides that probabilistic model parameter is solved from the data set that there is implicit variable
Method.The training of gauss hybrid models is exactly what EM algorithms were realized, and implicit variable therein isπ k .GMM model is understood by formula 1,2
It needs to be determined that parameter haveπ k 、μ k 、C.EM algorithms estimation gauss hybrid models have two steps.EStep, by initialization or
The parameter of each Gaussian component, goes to estimate the weights of each Gaussian component, weights are shown below known to the result of person's previous step；MStep, the weights based on estimation, then go to determine the parameter of Gaussian component, thekThe expectation of individual Gaussian componentμ k And covariance matrixC k It is shown below.The two steps are repeated, until fluctuation very little, extreme value are reached approximately；
(3)
It is the in formulaiIndividual sample iskProbability in individual single Gauss model；
(4)
(5)
(6)
In formulaN k Expression belongs toKThe number of samples of individual Gaussian component.
, randomly select segmentation after cell normal segments information, utilize MAP algorithm adaptive GMMs GMM.
Maximum a posteriori（Maximum a posteriori, MAP）Rule of thumb data are obtained to being difficult to observe method of estimation
The point estimation of amount, the prior distribution for the amount of being estimated is fused to wherein, so the maximum likelihood that it can be regarded as regularization is estimated
Meter.All normal segments gray values of cell after segmentation are chosen first, and training dataset isX={x 1,x 2,…,x t }.Then utilize
MAP algorithms calculate the parameter of specific mixed Gaussian.MAP algorithms are divided into two steps：The first step is identical with the E steps of EM algorithms, meter
The weights of each Gaussian component as shown in Equation 3 are calculated, and are calculated as follows shown parameter：
(7)
(8)
Second step byN k Modifying factor is obtained, carrys out new and old UBM parameter, for Gaussian componentk, the expression formula that parameter is calculated
It is as follows respectively：
(9)
(10)
(11)
Wherein,γAllow all mixed weight-value and be 1,β k π ,β k μ ,β k C For the weight of Gaussian component, mean vector, covariance
The modifying factor of matrix, its effect is the new and old parameter for balancing GMM model, and its value is bigger, illustrates that data are more abundant, also
It is that new parameter is more credible；And if it is smaller, just illustrate the negligible amounts of data, more inaccurate is also become to GMM model estimation,
Its definition is as follows：
(12)
In formulaλ ρ It is constraint modifying factor for relationship factorβ k ρ Change yardstick（ρ∈ π,μ,C}）,λGenerally take 16.
The cell reconstitution includes step：
S4, overlapping cell segmentation method is utilized to split overlapping cell；
Author proposes the overlapping nucleus splitting scheme based on identification in a paper, and overlapping nucleus segmentation key is to look for
To lap position and the defiber tallied with the actual situation.If there is overlapping phenomenon between cell, then the cell one after overlapping
Surely there can be corresponding concave point.Thus asked the concave point that solves whole profile is converted to the problem of finding lap position
Topic.This method uses the concave point detection method based on curvature estimation.But it is due to that the uneven slip of profile result in many false concave points
Appearance, in order to distinguish true and false concave point, this method identifies the classification of overlapping nucleus to determine concave point number by grader,
Foundation is provided for concave point detection.The concave point position of the overlapping nucleus of N number of series connection is sinking degree most deep individual points of preceding 2 (N-1);N
The concave point position of individual nucleus in parallel is sinking degree most deep top n point.Detailed process such as Fig. 2, is partitioned into cell wheel first
Exterior feature identifies the classification of the cell as shown in Figure 2 a and using grader, and accurate concave point position is found further according to nucleus classification
As shown in Figure 2 b, sub- profile is then partitioned into as shown in Figure 2 c according to the sub- contours segmentation scheme of a variety of reduplicative formses, finally
With ellipse fitting Reduced separating line and it is added on sub- profile, obtains complete daughter nuclei profile as shown in Figure 2 d.
, generated at random using GMM model and meet the gray value of constraints to change unusual part；
After overlapping cell segmentation, obtain it is unicellular in include lap, i.e., unicellular matrixC=A∪N, wherein matrixATable
Show lap in cell, matrixNRepresent non-folded part in cell.Utilize the gauss hybrid models and constraints of each cellY, generate some gray value modification matrixes for meeting Gaussian ProfileA.Generation method utilizes the principle of production model, first with
It is uniformly distributed the numerical value between generation (0,1).Then judge which Gaussian component it belongs to according to the weight of each single Gauss.
A random gray value finally is generated using the Gaussian component, but the gray value must be in constraintsYInside otherwise will
Regenerate.Aforesaid operations are repeated until by matrixAValue modification finish.ConstraintsYCalculation it is as follows.
The diagnosis principle of the DNA ploidy body analysis system of area of computer aided diagosis is：The picture of suspected tumor cells is enumerated
Out, that is, DI is found>2.5 cell picture.DI calculation formula is as shown in Equation 13, because the DNA content of lymphocyte is relative
Stablize therefore be used as Standard values by the use of the IOD averages of sample medium size lymphocyte.The calculation formula of IOD values is as follows：
(13)
(14)
Wherein,For corresponding pixel pointsOD value, it is as follows：
(15)
Wherein, the average gray value of background is represented,Represent pixelGray value.For lap gray scale
It is worth abnormal problem, the gray value of required generation will make it not influence the IOD values of initial cell, then will limit its value at one
Effective codomainYInterior, i.e., the OD value of newly-generated gray value will be in the OD value average of initial cell normal segments
Both sides, codomain Y calculation formula is as follows：
(16)
In formulaOD m For the mean optical density of nucleus normal segments,OD s For the optical density standard deviation of nucleus normal segments.By
Calculate and understand codomain。
, utilize the newly-generated part of median filter smoothness of image；
Because the coarse texture of newly-generated part is not consistent with the texture of cell normal segments.The mode of intermediate value rate ripple is utilized herein
Smoothing matrix obtains matrixAA。
, repair using FMM algorithms newly-generated part and normal segments be connected edge；
Its marginal portion is excessively protruded after newly-generated part merges with archaeocyte., herein will be newly-generated in order to solve this problem
The mark image that partial fringe region is regarded, is carried out using the gray value for the surrounding pixel point for marking image by FMM algorithms
Repairing.
The selection of repairing area is by taking two cell overlaps as an example, if the edge contour point set of the two cells is combined into C1, C2,
The set M of the lap profile point of two overlapping cells.Set M is asked for first, it is made up of two parts, i.e., in profile C1
Profile C2 on point and the point on profile C1 in profile C2.It is exactly to be repaired that then set M is expanded to obtained region D
Region.FMM algorithm principles：Assuming that in restoring area D p points gray value, with pointpCentered on choose a small neighbourhoodβε, q is
More therein, the calculation formula of point p modified values is as follows：
(17)
In formula △ I (q) be q points luminance gradient value, w (p, q)=dir (p, q) dst (p, q) lev (p, q), wherein
Dir (p, q) is direction factor, and dst (p, q) is the geometric distance factor, and lev (p, q) is level set apart from the factor.
Final realizes that effect such as Fig. 3 shows, the cell after as can be seen from the figure reconstructing can effectively eliminate extraordinary image
Vegetarian refreshments, contrast segmentation after cell closer to and archaeocyte.
The present invention can also have other various embodiments, in the case of without departing substantially from spirit of the invention and its essence, this area
Technical staff works as can make various corresponding changes and deformation according to the present invention, but these corresponding changes and deformation should all belong to
The protection domain of appended claims of the invention.
Claims (6)
1. the extraordinary image vegetarian refreshments reconstruct after overlapping cell segmentation, it is characterised in that include model training and cell reconstitution：
The model training includes step：
S1,1000 unicellular data for randomly selecting same sample, do training data；
S2, utilize the general gauss hybrid models UBM of EM Algorithm for Training；
Cell normal segments information after S3, selection segmentation, utilizes MAP algorithm adaptive GMMs GMM；
The cell reconstitution includes step：
S4, overlapping cell segmentation method is utilized to split overlapping cell；
S5, generated at random using GMM model and meet the gray value of constraints to change unusual part；
S6, utilize the newly-generated part of median filter smoothness of image；
S7, repair using FMM algorithms newly-generated part and normal segments be connected edge.
2. detection method as claimed in claim 1, it is characterised in that the principle of the gauss hybrid models described in step S2 and
The calculation formula of EM algorithms is as follows：
The probability density formula of mixed Gauss model is as follows：
(1)
In formulamSpecified for degree of mixing generally according to experimental result, hereinm=3,π k Represent thekThe weight of individual single Gauss model,N
(x, μ k , C) for thekThe probability density of individual single Gauss model, wherein；The probability density formula of single Gauss is as follows,
Whereinμ k For mathematic expectaion,CFor covariance matrix：
(2)
EM algorithms are a kind of maximum Likelihoods that probabilistic model parameter is solved from the data set that there is implicit variable；It is high
The training of this mixed model is exactly what EM algorithms were realized, and implicit variable therein isπ k ；Understand that GMM model needs by formula 1,2
The parameter of determination hasπ k 、μ k 、C；EM algorithms estimation gauss hybrid models have two steps；EStep, by initialization or on
The parameter of each Gaussian component, goes to estimate the weights of each Gaussian component, weights known to the result of one stepSuch as following formula institute
Show；MStep, the weights based on estimation, then go to determine the parameter of Gaussian component, thekThe expectation of individual Gaussian componentμ k And covariance
MatrixC k It is shown below；The two steps are repeated, until fluctuation very little, extreme value are reached approximately；
(3)
It is the in formulaiIndividual sample iskProbability in individual single Gauss model；
(4)
(5)
(6)
In formulaN k Expression belongs toKThe number of samples of individual Gaussian component.
3. detection method as claimed in claim 1, it is characterised in that the computational methods of the MAP algorithms described in step S3 are such as
Under：
Maximum a posteriori（Maximum a posteriori, MAP）Rule of thumb data are obtained to being difficult to observed quantity method of estimation
Point estimation, the prior distribution for the amount of being estimated is fused to wherein, so it can be regarded as the maximal possibility estimation of regularization；It is first
All normal segments gray values of cell after segmentation are first chosen, training dataset isX={x 1,x 2,…,x t }；Then calculated using MAP
Method calculates GMM parameter；MAP algorithms are divided into two steps：The first step is identical with the E steps of EM algorithms, calculates as shown in Equation 3 each
The weights of Gaussian component, and it is calculated as follows shown parameter：
(7)
(8)
Second step byN k Modifying factor is obtained, carrys out new and old UBM parameter, for Gaussian componentk, the expression formula that parameter is calculated
It is as follows respectively：
(9)
(10)
(11)
Wherein,γAllow all mixed weight-value and be 1,β k π ,β k μ ,β k C For the weight of Gaussian component, mean vector, covariance
The modifying factor of matrix, its effect is the new and old parameter for balancing GMM model, and its value is bigger, illustrates that data are more abundant, also
It is that new parameter is more credible；And if it is smaller, just illustrate the negligible amounts of data, more inaccurate is also become to GMM model estimation,
Its definition is as follows：
(12)
In formulaλ ρ It is constraint modifying factor for relationship factorβ k ρ Change yardstick（ρ∈ π,μ,C}）,λGenerally take 16.
4. detection method as claimed in claim 1, it is characterised in that the computational methods of the constraints described in step S5 are such as
Under：
The calculation formula of IOD values is as follows：
(14)
Wherein, be corresponding pixel points OD value it is as follows：
(15)
Wherein, the average gray value of background is represented, the gray value of pixel is represented；For lap gray value abnormal problem,
The gray value of required generation will make it not influence the IOD values of initial cell, then will limit its value in an effective codomain Y
It is interior；The OD value of i.e. newly-generated gray value will in the both sides of the OD value average of initial cell normal segments, codomain Y's
Calculation formula is as shown in Equation 16：
(16)
In formulaOD m For the mean optical density of nucleus normal segments,OD s For the optical density standard deviation of nucleus normal segments, by counting
Calculate and understand codomain。
5. detection method as claimed in claim 1, it is characterised in that the computational methods of the FMM algorithms described in step S7 are such as
Under：
Assuming that in restoring area D p points gray value, centered on point p choose a small neighbourhoodβ ε , q is more therein, point p
The calculation formula of modified values is as follows：
(17)
△ I (q) are the luminance gradient value of q points, w (p, q)=dir (p, q) dst (p, q) lev (p, q), wherein dir in formula
(p, q) is direction factor, and dst (p, q) is the geometric distance factor, and lev (p, q) is level set apart from the factor.
6. detection method as claimed in claim 1, it is characterised in that novelty of the invention is as follows：
The present invention is modeled using UBM model to the gray value of all cells；Then with the normal segments of each cell adaptive one
The distribution that individual GMM is used for the cell gray value is modeled；Generated at random with GMM again meet gray scale limitation gray value be filled into weight
Folded region；Linking part is finally repaired using image repair algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710504878.2A CN107330869B (en) | 2017-06-28 | 2017-06-28 | Reconstruction method of abnormal pixel points after segmentation of overlapped cells |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710504878.2A CN107330869B (en) | 2017-06-28 | 2017-06-28 | Reconstruction method of abnormal pixel points after segmentation of overlapped cells |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107330869A true CN107330869A (en) | 2017-11-07 |
| CN107330869B CN107330869B (en) | 2021-04-23 |
Family
ID=60197737
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710504878.2A Active CN107330869B (en) | 2017-06-28 | 2017-06-28 | Reconstruction method of abnormal pixel points after segmentation of overlapped cells |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107330869B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109992962A (en) * | 2019-03-08 | 2019-07-09 | 天津理工大学 | A kind of intrusion detection method based on Multi-Fuzzy-Neural Networks |
| CN112784922A (en) * | 2021-02-07 | 2021-05-11 | 重庆渝抗医药科技有限公司 | Extraction and classification method of intelligent cloud medical images |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080195230A1 (en) * | 2007-02-09 | 2008-08-14 | Quijano Rodolfo C | Pericardial tissue sheet |
| CN102663416A (en) * | 2012-03-20 | 2012-09-12 | 苏州迪凯尔医疗科技有限公司 | Segmentation method of viscera and internal blood vessels thereof in surgical planning system |
| CN103984957A (en) * | 2014-05-04 | 2014-08-13 | 中国科学院深圳先进技术研究院 | Automatic early warning system for suspicious lesion area of capsule endoscope image |
-
2017
- 2017-06-28 CN CN201710504878.2A patent/CN107330869B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080195230A1 (en) * | 2007-02-09 | 2008-08-14 | Quijano Rodolfo C | Pericardial tissue sheet |
| CN102663416A (en) * | 2012-03-20 | 2012-09-12 | 苏州迪凯尔医疗科技有限公司 | Segmentation method of viscera and internal blood vessels thereof in surgical planning system |
| CN103984957A (en) * | 2014-05-04 | 2014-08-13 | 中国科学院深圳先进技术研究院 | Automatic early warning system for suspicious lesion area of capsule endoscope image |
Non-Patent Citations (2)
| Title |
|---|
| MARIE-HELENE ROY CARDINAL 等: "《Intravascular ultrasound image segmentation:A three-dimensional fast-marching method based on gray level distributions》", 《IEEE TRANSACTIONS ON MEDICAL IMAGING》 * |
| 霍春宝 等: "《基于GMM-UBM的说话人确认系统的研究》", 《辽宁工业大学学报(自然科学版)》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109992962A (en) * | 2019-03-08 | 2019-07-09 | 天津理工大学 | A kind of intrusion detection method based on Multi-Fuzzy-Neural Networks |
| CN109992962B (en) * | 2019-03-08 | 2023-01-31 | 天津理工大学 | Intrusion detection method based on multistage fuzzy neural network |
| CN112784922A (en) * | 2021-02-07 | 2021-05-11 | 重庆渝抗医药科技有限公司 | Extraction and classification method of intelligent cloud medical images |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107330869B (en) | 2021-04-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106803253B (en) | A kind of three-dimensional rock image crack identification method | |
| CN109359684A (en) | Fine granularity model recognizing method based on Weakly supervised positioning and subclass similarity measurement | |
| CN109191471A (en) | Based on the pancreatic cell image partition method for improving U-Net network | |
| CN110189255A (en) | Method for detecting human face based on hierarchical detection | |
| CN108090830B (en) | Credit risk rating method and device based on facial portrait | |
| CN108122008A (en) | SAR image recognition methods based on rarefaction representation and multiple features decision level fusion | |
| CN107392900B (en) | Multi-scale enhancement method for lung nodule image | |
| CN104182763A (en) | Plant type identification system based on flower characteristics | |
| CN105046694A (en) | Quick point cloud registration method based on curved surface fitting coefficient features | |
| CN104268600B (en) | A kind of mineral flotation foam image texture analysis and operating mode's switch method based on Minkowski distances | |
| CN108509843A (en) | A kind of face identification method of the Huber constraint sparse codings based on weighting | |
| CN112215847B (en) | Method for automatically segmenting overlapped chromosomes based on counterstudy multi-scale features | |
| CN114742758A (en) | Cell nucleus classification method in full-field digital slice histopathology picture | |
| CN108830842A (en) | A kind of medical image processing method based on Corner Detection | |
| CN112215842A (en) | Malignant nodule edge detection image processing method based on benign thyroid template | |
| CN111639878A (en) | Landslide risk prediction method and system based on knowledge graph construction | |
| CN110349159A (en) | 3D shape dividing method and system based on the distribution of weight energy self-adaptation | |
| CN106600613A (en) | Embedded GPU-based improved LBP infrared target detection method | |
| CN107330869A (en) | Extraordinary image vegetarian refreshments reconstruct after overlapping cell segmentation | |
| Zafari et al. | Resolving overlapping convex objects in silhouette images by concavity analysis and Gaussian process | |
| CN105160666B (en) | SAR image change detection based on Non-Stationary Analysis and condition random field | |
| CN116311387B (en) | Cross-modal pedestrian re-identification method based on feature intersection | |
| CN106980878B (en) | Method and device for determining geometric style of three-dimensional model | |
| CN106204564A (en) | A kind of laser photocentre extracting method | |
| CN108776809A (en) | A kind of dual sampling Ensemble classifier model based on Fisher cores |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |