CN110110634A - Pathological image polychromatophilia color separation method based on deep learning - Google Patents
Pathological image polychromatophilia color separation method based on deep learning Download PDFInfo
- Publication number
- CN110110634A CN110110634A CN201910347578.7A CN201910347578A CN110110634A CN 110110634 A CN110110634 A CN 110110634A CN 201910347578 A CN201910347578 A CN 201910347578A CN 110110634 A CN110110634 A CN 110110634A
- Authority
- CN
- China
- Prior art keywords
- matrix
- image
- model
- dyeing
- prediction
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/69—Microscopic objects, e.g. biological cells or cellular parts
- G06V20/698—Matching; Classification
Abstract
The pathological image polychromatophilia color separation method based on deep learning that the invention discloses a kind of, comprising the following steps: (1) optical density transformation is carried out to pathological staining image, obtain the optical density matrix of original pathological staining image;(2) the optical density matrix for obtaining step (1) constructs ResU-Net model;(3) the ResU-Net model obtained to step (2) is trained;(4) image dyeing separation is carried out by the ResU-Net model after step (3) training.This method can carry out Pixel-level analysis, same class loading in better separate picture, to improve dyeing separating property to original image.
Description
Technical field
The present invention relates to technical field of image information processing, especially a kind of pathological image polychromatophilia color based on deep learning
Separation method.
Background technique
Dyeing separation is a kind of preconditioning technique automatically analyzed for aid in tissue pathological image.With specific structure knot
The identification of histological types can be enhanced in the chemical reagent of conjunction.In histopathology, most widely used coloring agent is
H and E (H&E).Hematoxylin makes nucleus that navy blue or purple be presented in conjunction with nucleic acid, and Yihong is adhered to tissue
In protein on, make cellular matrix present pink.Traditional dyeing divides method, such as color deconvolution (CD) and independent element point
Analysis (ICA) is intended to find optimum dyeing matrix and dyeing concentration matrix, and the predefined value for dyeing matrix is inaccurate.And
In practical applications, the hematoxylin and eosin dyeing effect observed will receive many factors influence, for example number used is swept
Retouch instrument type and Staining Protocol.In recent years, compared with traditional computer vision methods, deep learning is significant to improve many groups
Knit it is achievable in pathology task as a result, but research be concentrated mainly on classification and segmentation on, not yet dedicated for dyeing separation
Task.Convolutional neural networks (CNN) extract color and texture information, and can obtain higher level by stacking convolutional layer
Feature.When in conjunction with ICA, texture information, which has been displayed, can improve dyeing separating resulting.
Color deconvolution (CD) algorithm of Ruifrok and Johnston is first application for dyeing separation field.In coloured silk
In color deconvolution frame, need for image to be transferred in light intensity spatial, and need dyed substrates for dye separation.
Ruifrok and Johnston provides a small amount of example dyed substrates, but if dyeing condition changes, then these matrix will not
It is applicable in again.In order to meet the needs for being directed to specific image optimization dyed substrates, develop several based on color deconvolution
Method.In early stage, Macenko proposes a kind of automatic staining Matrix Prediction Method, one as dyeing method for normalizing
Part, and estimate to dye vector using singular value decomposition (SVD).Gavrilovic assumes that perceiving similar color will connect each other
Closely, and it was found that pixel is expected appear in in corresponding group of each dyeing in Maxwell's colorimetric plane.It then will be every
A dyeing vector is estimated as the average value of its corresponding Gaussian Profile.Kather suggests the best table that dyeing is obtained using PCA
Show.This is by projecting to the first two PCA component in the plane by using the dyeing vector creation for estimating dyeing Matrix Estimation
Come what is realized.However, PCA assumes that there are orthogonalities between main component, but not such was the case with for situation, especially in H and E etc.
In correlation dyeing.Trahearn, which is recently proposed, a kind of carries out the method deconvoluted of dyeing using ICA variant.This method be based on
Lower hypothesis: dyeing vector can be modeled as by isolated component according to ICA model.As application ICA, it is contemplated that the pixel of identical stain
By the main shaft distribution approximately along an isolated component, and the pixel of different stains will be along different main shaft distributions.However,
Trahearn shows that in some cases original isolated component cannot provide enough deconvolution.Therefore, using aligning step
To adjust the isolated component of estimation.This is found by minimizing the average value of the distance between each pixel and its nearest vector
Group optimum dyeing vector, the stopping when reaching convergence.Rabinovic compares two kinds of color Deconvolution Methods, nonnegative matrix because
Son decomposes (NMF) and independent component analysis (ICA).Although they show that NMF is performed better than, both methods has all been not enough to
Deconvolute image entirely.
Summary of the invention
The pathological image polychromatophilia color separation method based on deep learning that the object of the present invention is to provide a kind of, is utilized CNN
The various features that can be extracted, remain the structure of organization object, while also having separated the dye that other methods often cannot be distinguished
Color spot.
To achieve the above object, the technical solution adopted by the present invention are as follows:
A kind of pathological image polychromatophilia color separation method based on deep learning, comprising the following steps:
(1) optical density transformation is carried out to pathological staining image, obtains the optical density matrix of original pathological staining image;
(2) the optical density matrix for obtaining step (1) constructs ResU-Net model;
(3) the ResU-Net model obtained to step (2) is trained;
(4) image dyeing separation is carried out by the ResU-Net model after step (3) training.
The specific steps of the step (1) are as follows: each pixel R, G, B tri- are calculated to original pathological staining image matrix first
Optical density corresponding to channel obtains the optical density matrix of original pathological staining image;R, each channel institute in tri- channels G, B
Corresponding optical density OD is expressed as follows:
Wherein, D is image in the space optical density OD, I0It is incident intensity, I is transmitted intensity.
The specific steps of the step (2) are as follows: optical density matrix is three row matrix D=[dr, dg, db]T, wherein each row pair
Ying Yuyi Color Channel;Using optical density matrix as the input of model, constructs ResU-Net framework and carry out multitask dyeing point
From;ResU-Net framework is supported by constricted path, bridge joint, path expander three parts, to complete the prediction of linear coloring color matrix,
Non-linear dyed color Matrix prediction and dyeing concentration prediction;Wherein, constricted path is used to reduce the Spatial Dimension of characteristic pattern, together
When successively increase the quantity of characteristic pattern, input picture is extracted as compact feature;Bridging part is for connecting constricted path and expansion
Path is opened, and realizes linear coloring color matrix forecast function;Path expander is for the gradually details of recovery target and accordingly
The path expander of Spatial Dimension, right side will up-sample, and every up-sampling is primary, and just port number corresponding with constricted path is identical
Scale fusion is respectively used to dyed color matrix and dyeing concentration Matrix prediction will export.
In the step (2), constricted path has several residual blocks, in each residual block, by deconvolution, by feature
Mapping reduces half;Correspondingly, path expander is also made of corresponding residual block;Before each residual block, exist from compared with
The cascade of the up-sampling of the Feature Mapping of low level and the Feature Mapping from corresponding encoded path.
In the step (2), residual block is explained are as follows: it is assumed that the input of a neural network unit is x, desired output is H
(x), it in addition defines a residual error and maps F (x)=H (x)-x, if x is directly passed to output, which will be learned
The target of habit is exactly residual error mapping F (x)=H (x)-x, and residual error unit is made of a series of convolutional layers and a shortcut, defeated
Enter the output that x passes to residual error unit by this shortcut, then the output of residual unit is z=F (x)+x, and z is to the inclined of x
It leads as follows:
In formula, z is greater than 1 to x local derviation;Furthermore include batch standardization BN (Batch in each residual block
) and line rectification function ReLU (Rectified Linear Unit) normalization.
In the step (3), ResU-Net model includes prior model, posterior model and concentration model;Prior model is logical
The single dyed color matrix for predicting each image is crossed, solves Nonnegative matrix factorization in combination with dyeing concentration matrix
(NMF) problem, and posterior model predicts the dyed color matrix combination concentration prediction model of each pixel based on Pixel-level, it can
Accurately every kind of dyeing independent characteristic is arrived in study;Given priori dyes matrix, poststaining matrix and shared concentration matrix, pass through by
Concentration matrix dyes matrix with priori and posteriority dyeing matrix combines to form two reconstructions;By minimum input picture and each
Reconstruction loss between reconstruction carrys out training pattern;Library is added between the distribution that prior model and posterior model are predicted and strangles Bark-
Leibler (Kullback-Leibler) bound term, ensures that the prediction of posterior model is only offset slightly from prior model;Wherein, two
KL divergence such as following formula between a gaussian variable:
In formula, σ1、σ2、μ1、μ2Respectively indicate the normal distribution standard difference of prior model prediction and posterior model prediction with it is equal
Value;
The number of pixels that N is each image is defined, M is every wheel picture number, and K indicates dyeing type, and C indicates image channel
Number, then obtaining bound term:
In formula, μm,n,k,c、Prior model prediction is respectively indicated to predict with posterior model
The mean value and standard deviation of normal distribution.Subscript m, n, k, c respectively indicate every wheel picture number, image pixel number, dyeing type
Number and image channel number serial number;
Loss function is defined as follows in order to examine its separating effect for dyeing separation task:
In formula,Indicate the nth pixel of m-th of image,Indicate corresponding forecast image pixel.
In the step (4), feature is extracted from image optical density by ResU-net model, the feature of extraction is passed
To each submodule, a series of ginseng of Gaussian Profiles that these submodules are predicted the dyeing concentration of each pixel and therefrom sampled
Number, parameter includes mean value and variance;For pixel each in image, to hematoxylin coloured portions, eosin stains part, background
It is predicted in each RGB color channel of 3 kinds of part situation;The affiliated area of each pixel is intuitively expressed by probability distribution
Domain, to achieve the purpose that dyeing separation.
In the step (4), while realizing three kinds of staining conditions separation, when occurring stained region in image, by table
Now it is the low probability event in the case of three kinds, spot can be recognized accurately at this time.
The utility model has the advantages that being used for polychromatophilia color separation (bush the invention proposes a kind of new unsupervised deep learning method
Essence, Yihong, background and spot).Inspiration of this method by Nonnegative matrix factorization (NMF), and input picture is decomposed into
Dyed color matrix and dyeing concentration matrix.Method of the invention predicts the Gaussian Profile about dyed color, from Gauss point
Available each pixel belongs to the probability of each pigmented section in cloth, is statistically analyzed sufficiently to learn various dyeing
Independent characteristic, can be to dyeing Accurate classification in image.By analyzing the Gaussian Profile of dyed color, converge to model as early as possible
Optimum state and make model have good generalization ability.Can precisely identify hematoxylin in model, Yihong, background these three
Under staining conditions, when occurring stained region in image, it will appear as the low probability event under three cases above, at this time can
Spot is recognized accurately.In deep learning network, the dyeing matrix of image level is fixed, but is directed to real image, same dye
Color region (i.e. similar tissue) will lead to same pigmented section due to the extraneous factors such as artificially colored, and dyeing matrix is also different
's.This method can carry out Pixel-level analysis, same class loading in better separate picture to original image, to improve dyeing point
From performance.
Compared with prior art, the invention has the following advantages that
1. predicting the Gaussian Profile about dyed color, available each pixel belongs to each from Gaussian Profile
The probability of pigmented section is statistically analyzed sufficiently to learn the independent characteristic of various dyeing, can be accurate to dyeing in image
Classification.
2. making model converge to optimum state as early as possible by the Gaussian Profile of analysis dyed color and there is model very
Good generalization ability.Hematoxylin can be precisely identified in model, Yihong, under these three staining conditions of background, when occurring in image
Stained region will appear as the low probability event under three cases above, and spot can be recognized accurately at this time.
3. conventional method has limitation to the feature that single image learns, pass through mass data using deep learning
The model that the feature practised is trained has good robustness in terms of dyeing separation.
Detailed description of the invention
Fig. 1 is image dyeing separation overall flow figure;
Fig. 2 is the detailed structure view of ResU-Net model;
Fig. 3 is prior model schematic diagram;
Fig. 4 is overall model schematic diagram;
Fig. 5 is the color Gaussian Profile figure obtained to a certain area sampling;
Fig. 6 is training front and back distribution of color exemplary diagram;
Fig. 7 is dyeing separating effect figure;
Fig. 8 is and existing method separation effect comparison figure;
Fig. 9 is the effect picture using method detection spot of the invention.
Specific embodiment
Further explanation is done to the present invention with reference to the accompanying drawing.
A kind of pathological image polychromatophilia color separation method based on deep learning of the invention, comprising the following steps:
(1) optical density transformation is carried out to pathological staining image, obtains the optical density matrix of original pathological staining image;
Dyeing separation frame leads to as shown in Figure 1, calculating each pixel R, G, B tri- to original pathological staining image matrix first
Optical density corresponding to road obtains the optical density matrix of original pathological staining image;R, each channel institute in tri- channels G, B is right
The optical density OD answered is expressed as follows:
Wherein, D is image in the space optical density OD, I0It is incident intensity, I is transmitted intensity.
(2) the optical density matrix for obtaining step (1) constructs ResU-Net model;
Optical density matrix is three row matrix D=[dr, dg, db]T, wherein each row is corresponding to a Color Channel;By optical density
Input of the matrix as model, construction ResU-Net framework carry out multitask dyeing separation, as shown in Figure 2;ResU-Net framework
By constricted path, bridge joint, path expander three parts support, to complete the prediction of linear coloring color matrix, non-linear dyed color
Matrix prediction and dyeing concentration prediction;Wherein, constricted path is used to reduce the Spatial Dimension of characteristic pattern, while successively increasing feature
Input picture is extracted as compact feature by the quantity of figure;Bridging part realizes line for connecting constricted path and path expander
Property dyed color Matrix prediction function;Path expander is used for the gradually details of recovery target and corresponding Spatial Dimension, right side
Path expander will up-sample, and every up-sampling is primary, and just port number same scale corresponding with constricted path is merged with will be defeated
It is respectively used to dyed color matrix and dyeing concentration Matrix prediction out.
Wherein, constricted path has several residual blocks, and in each residual block, by deconvolution, Feature Mapping is reduced
Half;Correspondingly, path expander is also made of corresponding residual block;Before each residual block, exist other from lower level
The cascade of the up-sampling of Feature Mapping and the Feature Mapping from corresponding encoded path.Residual block is explained are as follows: it is assumed that a mind
Input through network unit is x, and desired output is H (x), in addition defines a residual error and maps F (x)=H (x)-x, if x is direct
Output is passed to, then the neural network unit target to be learnt is exactly residual error mapping F (x)=H (x)-x, residual error unit
It is made of a series of convolutional layers and a shortcut, input x passes to the output of residual error unit by this shortcut, then residual error
The output of unit is z=F (x)+x, and z is as follows to the local derviation of x:
In formula, z is greater than 1 to x local derviation, and the problem of gradient disappears is effectively prevented in back-propagation process;Furthermore each
All comprising batch standardization BN (Batch normalization) and line rectification function ReLU (Rectified in residual block
Linear Unit), effectively accelerate convergence rate.
(3) the ResU-Net model obtained to step (2) is trained;
The ResU-Net model of formation is as shown in figure 3, include prior model, posterior model and concentration model;Prior model
By predicting the single dyed color matrix of each image, Nonnegative matrix factorization is solved in combination with dyeing concentration matrix
(NMF) problem, and posterior model predicts the dyed color matrix combination concentration prediction model of each pixel based on Pixel-level, it can
Accurately every kind of dyeing independent characteristic is arrived in study;Given priori dyes matrix, poststaining matrix and shared concentration matrix, pass through by
Concentration matrix dyes matrix with priori and posteriority dyeing matrix combines to form two reconstructions;By minimum input picture and each
Reconstruction loss between reconstruction carrys out training pattern;Library is added between the distribution that prior model and posterior model are predicted and strangles Bark-
Leibler (Kullback-Leibler) bound term, ensures that the prediction of posterior model is only offset slightly from prior model;Wherein, two
KL divergence such as following formula between a gaussian variable:
In formula, σ1、σ2、μ1、μ2Respectively indicate the normal distribution standard difference of prior model prediction and posterior model prediction with it is equal
Value;
The number of pixels that N is each image is defined, M is every wheel picture number, and K indicates dyeing type, and C indicates image channel
Number, then obtaining bound term:
In formula, μm,n,k,c、Prior model prediction is respectively indicated to predict with posterior model
The mean value and standard deviation of normal distribution.Subscript m, n, k, c respectively indicate every wheel picture number, image pixel number, dyeing type
Number and image channel number serial number;
Loss function is defined as follows in order to examine its separating effect for dyeing separation task:
In formula,Indicate the nth pixel of m-th of image,Indicate corresponding forecast image pixel.
(4) image dyeing separation is carried out by the ResU-Net model after step (3) training;
Feature is extracted from image optical density by ResU-net model, the feature of extraction is passed to each submodule,
A series of parameter of Gaussian Profiles that these submodules are predicted the dyeing concentration of each pixel and therefrom sampled, parameter include equal
Value and variance;For pixel each in image, to hematoxylin coloured portions, eosin stains part, 3 kinds of situations of background parts
It is predicted in each RGB color channel;Each pixel affiliated area is expressed, intuitively by probability distribution to reach dyeing point
From purpose.While realizing three kinds of staining conditions separation, when occurring stained region in image, it will appear as in the case of three kinds
Low probability event, spot can be recognized accurately at this time.
Below with reference to embodiment, the present invention will be further described.
Embodiment
Original image progress light intensity spatial conversion is input in the ResU-Net model of 12 level frameworks and carries out multitask
Dyeing separation.ResU-Net model is supported by constricted path, bridge joint, path expander three parts to complete linear coloring color matrix
Prediction, non-linear dyed color Matrix prediction and dyeing concentration prediction.Constricted path is made of preceding 1-4 grades of network, for reducing
The Spatial Dimension of characteristic pattern, while successively increasing the quantity of characteristic pattern, input picture is extracted as compact feature.5th grade is bridge
Socket part point connection is shunk and path expander, and realizes linear coloring color matrix forecast function.Path expander is by 6-9 grades of network structures
At, it being up-sampled for gradually restoring details and corresponding Spatial Dimension, the path expander of target, every up-sampling is primary,
Just port number same scale fusion corresponding with constricted path is respectively used to dyed color matrix and dyeing concentration square will export
Battle array prediction.During linear prediction, Z5 feature is flattened into first as vector, deploys two full articulamentums, intermediate node is
500, output node 9 respectively indicates the dyed color matrix in three kinds of each channels staining conditions R, G, B.9 grades of Z9 feature is used for
Two subtasks: non-linear dyed color prediction and dyeing concentration prediction.Non-linear dyed color is predicted by 10 grades and 11 grades
Matrix, and 12 grades will realize that dyeing concentration is predicted to individual element.All ranks are all constructed using residual unit, and every level-one includes
Two 3 × 3 convolution blocks and an identity map.Identity map connection unit is output and input.Include in each residual block
BN and ReLU prevents gradient from disappearing and accelerates convergence rate.Design parameter and output size are as shown in table 1 in network.
Table 1
It in training process, for linear model, will a little be sampled to being randomly choosed in image Gaussian Profile, to be formed
About the estimated value of field color a certain in image distribution, and for nonlinear model, to pixel Gaussian Profile stochastical sampling, with
The dyed color matrix of prediction pixel grade.As shown in Figure 5.The process is repeated to each case, these distributions are combined to form
The dyeing matrix of estimation.The mean value of each distribution indicates the value of model distribution maximum probability, and standard deviation describes model
Accuracy.To further illustrate: assuming that distribution represents the red value in Yihong, if standard deviation is very low, the value sampled very may be used
0.5 can be connect.If the true red value in Yihong is close to 0.5, sampled value should cause to rebuild loss reduction;Therefore, if really
Red value far from 0.5, then sampled value will lead to very high reconstruction loss.If the big standard deviation of model prediction, is adopted
Sample value will change it is very big, therefore even if average value correctly can generate big reconstruction loss.Therefore, in order to find optimum value, often
The mean value of a distribution must be close to true value, and standard deviation must be low as far as possible.Network weight W uses Gaussian Profile random initializtion.
Learning rate is originally defined as 0.001, and when training reaches 250 wheel, training stops.Fig. 6 gives these distributions of training front and back
Example.
The sampling from a variety of different organization types and two scanners (Philips and Aperio), it means that each
The contrast of image and staining power difference.The model proposed is realized based on PyTorch deep learning frame, and is used
NVidia GTX 1080Ti GPU is trained.Data set organizes picture group by 22000 RGB having a size of 128x128 pixel
At.Dyeing disjunctive model is formed by 8 with the convolution feature extraction layer of ResU-net structural arrangement and 3 lesser convolutional layers,
To export image mode and pixel-by-pixel dyeing matrix and dyeing concentration matrix.Using batch size be 64 trining mock up about
Need to obtain for 30 minutes good as a result, although can realize good result with less achievement.Optimized using ADAM
Device is gradually decreased at the end of each epoch using the initial learning rate of 1-e3.The loss of standard mean square error, which is used as to rebuild, punishes
Penalize item.The result shows that the background of hematoxylin and eosin dyeing and hybrid RGB image can be had successfully been isolated, while retaining tissue
Structure.The second row of Fig. 7 shows coloring and isolated hematoxylin dyeing.
Method of the invention is compared with many tradition and state-of-the-art method, Fig. 8 illustrates comparison result.
The method of Mikto is used for cell segmentation, therefore can be only used for the separation of H dyeing.Following NMF method is shown using conventional method
Solve the result of NMF;Although H dyeing is quite obvious, cytoplasmic structure is seriously reduced, and any spot all may cause separation
Effect is poor.Color deconvolution (CD) is the isolated classical way of dyeing, and optimum dyeing matrix could be calculated by requiring manual intervention.With
CD method can storage configuration well, it can however not rationally separating background color.SDSA is a kind of nearest strategy, it makes
Spot is separated with the statistical analysis of multiresolution dyeing data.It can be seen that SDSA is successfully partitioned into H dyeing, but when figure
When as in more than two kinds of dyeing, this method failure.Fig. 9 shows the result of method spot detection of the invention.Belong to spot
Pixel has low probability in obtained probability distribution;It is less than the pixel of predefined thresholds λ by select probability, it can be to these
Region is identified and is split.
The invention proposes a kind of unsupervised deep learning methods for polychromatophilia color separation.It is image in this frame
Dyeing has been deployed separately different subtasks.In order to overcome the variation of dyeing, linear-nonlinear model is constructed, can be provided
The dyeing of Pixel-level samples.In order to make minimization of loss, carry out training pattern using Gaussian Profile, and also using the distribution come into
Row dyeing Matrix prediction and spot separation.
The above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications are also answered
It is considered as protection scope of the present invention.
Claims (10)
1. a kind of pathological image polychromatophilia color separation method based on deep learning, it is characterised in that: the following steps are included:
(1) optical density transformation is carried out to pathological staining image, obtains the optical density matrix of original pathological staining image;
(2) the optical density matrix for obtaining step (1) constructs ResU-Net model;
(3) the ResU-Net model obtained to step (2) is trained;
(4) image dyeing separation is carried out by the ResU-Net model after step (3) training.
2. the pathological image polychromatophilia color separation method according to claim 1 based on deep learning, it is characterised in that: described
The specific steps of step (1) are as follows: original pathological staining image matrix is calculated corresponding to tri- channels each pixel R, G, B first
Optical density obtains the optical density matrix of original pathological staining image;R, optical density corresponding to each channel in tri- channels G, B
OD is expressed as follows:
Wherein, D is image in the space optical density OD, I0It is incident intensity, I is transmitted intensity;
The specific steps of the step (2) are as follows: optical density matrix is three row matrix D=[dr, dg, db]T, wherein each row is corresponding to one
A Color Channel;Using optical density matrix as the input of model, constructs ResU-Net framework and carry out multitask dyeing separation;
ResU-Net framework is predicted by constricted path, bridge joint, the support of path expander three parts with completing linear coloring color matrix, non-
The prediction of linear coloring color matrix and dyeing concentration prediction;Wherein, constricted path is used to reduce the Spatial Dimension of characteristic pattern, simultaneously
The quantity for successively increasing characteristic pattern, is extracted as compact feature for input picture;Bridging part is for connecting constricted path and expansion
Path, and realize linear coloring color matrix forecast function;Path expander is for gradually restoring the details and sky accordingly of target
Between dimension, the path expander on right side will up-sample, and every up-sampling is primary, just the identical ruler of port number corresponding with constricted path
Degree fusion is respectively used to dyed color matrix and dyeing concentration Matrix prediction will export.
3. the pathological image polychromatophilia color separation method according to claim 2 based on deep learning, it is characterised in that: described
In step (2), constricted path has several residual blocks, and in each residual block, by deconvolution, Feature Mapping is reduced one
Half;Correspondingly, path expander is also made of corresponding residual block;Before each residual block, exist from the other spy of lower level
Levy the up-sampling of mapping and the cascade of the Feature Mapping from corresponding encoded path.
4. the pathological image polychromatophilia color separation method according to claim 2 based on deep learning, it is characterised in that: described
In step (2), residual block is explained are as follows: it is assumed that the input of a neural network unit is x, desired output is H (x), is in addition defined
One residual error maps F (x)=H (x)-x, if x is directly passed to output, which is exactly
Residual error maps F (x)=H (x)-x, and residual error unit is made of a series of convolutional layers and a shortcut, and input x passes through this victory
Diameter passes to the output of residual error unit, then the output of residual unit is z=F (x)+x, and z is as follows to the local derviation of x:
In formula, z is greater than 1 to x local derviation;Furthermore comprising including batch standardization and line rectification function in each residual block.
5. the pathological image polychromatophilia color separation method according to claim 2 based on deep learning, it is characterised in that: described
In step (2), constricted path is made of preceding 1-4 grades of network, for reducing the Spatial Dimension of characteristic pattern, while successively increasing feature
Input picture is extracted as compact feature by the quantity of figure;5th grade is bridging part connection contraction and path expander, and realizes line
Property dyed color Matrix prediction function;The path expander is made of 6-9 grades of networks, for gradually restoring the details and phase of target
The Spatial Dimension answered, path expander will up-sample, and every up-sampling is primary, just the identical ruler of port number corresponding with constricted path
Degree fusion is respectively used to dyed color matrix and dyeing concentration Matrix prediction will export;During linear prediction, Z5 feature
It is flattened into first as vector, deploys two full articulamentums, intermediate node 500, output node 9 respectively indicates three kinds of dyes
Pornographic condition R, the dyed color matrix in each channel G, B;9 grades of Z9 feature is used for two subtasks: non-linear dyed color prediction
It is predicted with dyeing concentration;Non-linear dyed color matrix is predicted by 10 grades and 11 grades, and 12 grades will realize dye to individual element
Colour saturation prediction;All ranks are all constructed using residual unit, and every level-one includes two 3 × 3 convolution blocks and an identity map;
Identity map connection unit is output and input.
6. the pathological image polychromatophilia color separation method according to claim 1 based on deep learning, it is characterised in that: described
In step (3), ResU-Net model includes prior model, posterior model and concentration model;Prior model is by predicting each figure
The single dyed color matrix of picture, Nonnegative matrix factorization (NMF) is solved the problems, such as in combination with dyeing concentration matrix, then
The dyed color matrix combination concentration prediction model that model predicts each pixel based on Pixel-level is tested, can accurately be learnt to every
Kind dyeing independent characteristic;Given priori dyes matrix, poststaining matrix and shared concentration matrix, by by concentration matrix and priori
Dyeing matrix and posteriority dyeing matrix combine to form two reconstructions;By minimizing the reconstruction between input picture and each reconstruction
Loss carrys out training pattern;Library is added between the distribution that prior model and posterior model are predicted and strangles Bark-leibler bound term,
Ensure that the prediction of posterior model is only offset slightly from prior model;Wherein, the KL divergence such as following formula between two gaussian variables:
In formula, σ1、σ2、μ1、μ2Respectively indicate the normal distribution standard difference and mean value of prior model prediction and posterior model prediction;
The number of pixels that N is each image is defined, M is every wheel picture number, and K indicates dyeing type, and C indicates image channel number,
So obtain bound term:
In formula, μm,n,k,c、It respectively indicates prior model prediction and predicts normal state with posterior model
The mean value and standard deviation of distribution.Subscript m, n, k, c respectively indicate every wheel picture number, image pixel number, dyeing type number
And image channel number serial number;
Loss function is defined as follows in order to examine its separating effect for dyeing separation task:
In formula,Indicate the nth pixel of m-th of image,Indicate corresponding forecast image pixel.
7. the pathological image polychromatophilia color separation method according to claim 1 based on deep learning, it is characterised in that: described
In step (3), in training process, for linear model, will a little it be sampled to being randomly choosed in image Gaussian Profile, with shape
At the estimated value being distributed about field color a certain in image, and for nonlinear model, to pixel Gaussian Profile stochastical sampling,
With the dyed color matrix of prediction pixel grade;The process is repeated to each case, these distributions are combined to form the dye of estimation
Colour moment battle array;The mean value of each distribution indicates the value of model distribution maximum probability, and standard deviation describes the accuracy of model.
8. the pathological image polychromatophilia color separation method according to claim 7 based on deep learning, it is characterised in that: described
In step (3), to find optimum value, the mean value of each distribution must be close to true value, and standard deviation must be low as far as possible.
9. the pathological image polychromatophilia color separation method according to claim 1 based on deep learning, it is characterised in that: described
In step (4), feature is extracted from image optical density by ResU-net model, the feature of extraction is passed to each submodule
Block, a series of parameter of Gaussian Profiles that these submodules are predicted the dyeing concentration of each pixel and therefrom sampled, parameter packet
Include mean value and variance;For pixel each in image, to hematoxylin coloured portions, eosin stains part, 3 kinds of feelings of background parts
It is predicted in each RGB color channel of condition;Each pixel affiliated area is expressed, intuitively by probability distribution to reach dye
The purpose of color separation.
10. the pathological image polychromatophilia color separation method according to claim 9 based on deep learning, it is characterised in that: institute
It states in step (4), while realizing three kinds of staining conditions separation, when occurring stained region in image, will appear as three kinds of feelings
Spot can be recognized accurately in low probability event under condition at this time.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910347578.7A CN110110634B (en) | 2019-04-28 | 2019-04-28 | Pathological image multi-staining separation method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910347578.7A CN110110634B (en) | 2019-04-28 | 2019-04-28 | Pathological image multi-staining separation method based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110110634A true CN110110634A (en) | 2019-08-09 |
CN110110634B CN110110634B (en) | 2023-04-07 |
Family
ID=67487107
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910347578.7A Active CN110110634B (en) | 2019-04-28 | 2019-04-28 | Pathological image multi-staining separation method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110110634B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111507992A (en) * | 2020-04-21 | 2020-08-07 | 南通大学 | Low-differentiation gland segmentation method based on internal and external stresses |
CN111539883A (en) * | 2020-04-20 | 2020-08-14 | 福建帝视信息科技有限公司 | Digital pathological image H & E dyeing restoration method based on strong reversible countermeasure network |
CN113538422A (en) * | 2021-09-13 | 2021-10-22 | 之江实验室 | Pathological image automatic classification method based on dyeing intensity matrix |
CN114627010A (en) * | 2022-03-04 | 2022-06-14 | 透彻影像(北京)科技有限公司 | Dyeing space migration method based on dyeing density map |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104200428A (en) * | 2014-08-18 | 2014-12-10 | 南京信息工程大学 | Microscopic image color convolution removal method and cutting method based on non-negative matrix factorization (NMF) |
CN106780498A (en) * | 2016-11-30 | 2017-05-31 | 南京信息工程大学 | Based on point depth convolutional network epithelium and matrix organization's automatic division method pixel-by-pixel |
CN108470337A (en) * | 2018-04-02 | 2018-08-31 | 江门市中心医院 | A kind of sub- reality Lung neoplasm quantitative analysis method and system based on picture depth feature |
CN108830813A (en) * | 2018-06-12 | 2018-11-16 | 福建帝视信息科技有限公司 | A kind of image super-resolution Enhancement Method of knowledge based distillation |
-
2019
- 2019-04-28 CN CN201910347578.7A patent/CN110110634B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104200428A (en) * | 2014-08-18 | 2014-12-10 | 南京信息工程大学 | Microscopic image color convolution removal method and cutting method based on non-negative matrix factorization (NMF) |
CN106780498A (en) * | 2016-11-30 | 2017-05-31 | 南京信息工程大学 | Based on point depth convolutional network epithelium and matrix organization's automatic division method pixel-by-pixel |
CN108470337A (en) * | 2018-04-02 | 2018-08-31 | 江门市中心医院 | A kind of sub- reality Lung neoplasm quantitative analysis method and system based on picture depth feature |
CN108830813A (en) * | 2018-06-12 | 2018-11-16 | 福建帝视信息科技有限公司 | A kind of image super-resolution Enhancement Method of knowledge based distillation |
Non-Patent Citations (1)
Title |
---|
HARIYANTI MOHD SALEH等: "Overlapping Chromosome Segmentation using U-Net: Convolutional Networks with Test Time Augmentation", 《RESEARCHGATE》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111539883A (en) * | 2020-04-20 | 2020-08-14 | 福建帝视信息科技有限公司 | Digital pathological image H & E dyeing restoration method based on strong reversible countermeasure network |
CN111539883B (en) * | 2020-04-20 | 2023-04-14 | 福建帝视信息科技有限公司 | Digital pathological image H & E dyeing restoration method based on strong reversible countermeasure network |
CN111507992A (en) * | 2020-04-21 | 2020-08-07 | 南通大学 | Low-differentiation gland segmentation method based on internal and external stresses |
CN113538422A (en) * | 2021-09-13 | 2021-10-22 | 之江实验室 | Pathological image automatic classification method based on dyeing intensity matrix |
CN114627010A (en) * | 2022-03-04 | 2022-06-14 | 透彻影像(北京)科技有限公司 | Dyeing space migration method based on dyeing density map |
Also Published As
Publication number | Publication date |
---|---|
CN110110634B (en) | 2023-04-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110110634A (en) | Pathological image polychromatophilia color separation method based on deep learning | |
CN109740413A (en) | Pedestrian recognition methods, device, computer equipment and computer storage medium again | |
CN109784386A (en) | A method of it is detected with semantic segmentation helpers | |
CN110059586B (en) | Iris positioning and segmenting system based on cavity residual error attention structure | |
CN107563999A (en) | A kind of chip defect recognition methods based on convolutional neural networks | |
US11593607B2 (en) | Method and system for predicting content of multiple components in rare earth extraction process | |
US11605163B2 (en) | Automatic abnormal cell recognition method based on image splicing | |
CN106156781A (en) | Sequence convolutional neural networks construction method and image processing method and device | |
CN108846835A (en) | The image change detection method of convolutional network is separated based on depth | |
CN111523521A (en) | Remote sensing image classification method for double-branch fusion multi-scale attention neural network | |
CN115063796B (en) | Cell classification method and device based on signal point content constraint | |
CN115705637A (en) | Improved YOLOv5 model-based spinning cake defect detection method | |
CN110826411B (en) | Vehicle target rapid identification method based on unmanned aerial vehicle image | |
CN109903339B (en) | Video group figure positioning detection method based on multi-dimensional fusion features | |
CN106023121A (en) | BGA position back bore manufacture method | |
Abousamra et al. | Weakly-supervised deep stain decomposition for multiplex IHC images | |
CN109376753A (en) | A kind of the three-dimensional space spectrum separation convolution depth network and construction method of dense connection | |
CN109410171A (en) | A kind of target conspicuousness detection method for rainy day image | |
Liang et al. | An improved DualGAN for near-infrared image colorization | |
CN115170427A (en) | Image mirror surface highlight removal method based on weak supervised learning | |
CN112633257A (en) | Potato disease identification method based on improved convolutional neural network | |
CN114511710A (en) | Image target detection method based on convolutional neural network | |
CN111179272B (en) | Rapid semantic segmentation method for road scene | |
CN115546620A (en) | Lightweight target detection network and method based on YOLO (YOLO) and electronic equipment | |
CN113837154B (en) | Open set filtering system and method based on multitask assistance |
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 |