CN106845551A - A kind of histopathology image-recognizing method - Google Patents
A kind of histopathology image-recognizing method Download PDFInfo
- Publication number
- CN106845551A CN106845551A CN201710059300.0A CN201710059300A CN106845551A CN 106845551 A CN106845551 A CN 106845551A CN 201710059300 A CN201710059300 A CN 201710059300A CN 106845551 A CN106845551 A CN 106845551A
- Authority
- CN
- China
- Prior art keywords
- dictionary
- ill
- disease
- free
- overbar
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/03—Recognition of patterns in medical or anatomical images
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of histopathology image-recognizing method, comprise the following steps:Choose disease-free and ill training sample, disease-free and ill test sample;With reference to disease-free training sample and ill training sample, set up disease-free dictionary learning model and ill dictionary learning model, alternating iteration optimizes two object functions, until reaching maximum iteration untill, study obtains disease-free dictionary and ill dictionary;Using disease-free dictionary and ill dictionary, rarefaction representation is carried out to test sample, sparse reconstructed error vector of the test sample under disease-free dictionary and ill dictionary is calculated respectively;Classification statistic is obtained by sparse reconstructed error vector, the classification of test sample is determined by the comparing of classification statistic and threshold value.Application of the present invention to dictionary learning in histopathology image classification proposes new model and method, and the band category dictionary for learning has robustness in preferably sparse reconstitution and class to similar sample, has identification between preferable class to non-similar sample.
Description
Technical field
The present invention relates to a kind of histopathology image-recognizing method.
Background technology
With the development of computer-aided diagnosis technology, the research of " digital pathology " is also gradually subject to vast researcher
Concern, wherein, how accurately to automatically extract and hide identification feature in the picture, be follow-up histopathology graphical analysis
Or classification is submitted necessary information, so as to quick and precisely provide disease grade with classification, it has also become great in " digital pathology " to choose
One of research topic of war property.
Traditional feature extraction mode is broadly divided into two categories below:First major class is the spy based on special domain or particular task
Levy, such as the size of biological cell and morphological feature, the gray scale of image or colour information, texture;Second major class is main with space
Based on structure and Analysis On Multi-scale Features, such as morphological feature, drawing method, scale invariant feature, wavelet character.Above-mentioned traditional characteristic
Extracting mode is generally Pixel-level feature or manual feature, is typically only suitable for specific data object, and its range of application is restricted,
And feature redundancy is high, identification is low.
In the last few years, rarefaction representation obtained greatly pass because of its outstanding behaviours in numerous computer vision problems
Note.Its basic thought is that a primary signal is expressed as with one group of excessively complete dictionary to be the sparse signal of base.Rarefaction representation exists
Howling success is all obtained in the fields such as image denoising and recovery, recognition of face, image classification.And with the development of technology, such as
What study turns into the dictionary of scholars' focus of attention, i.e., to the dictionary suitable for particular problem (such as image classification)
Learning theory framework.
Whether the dictionary that it is critical only that construction of dictionary learning has preferable reconstitution and identification.This class is asked
Topic, Zhang etc. propose a kind of identification K-SVD (Discriminative K-SVD, DK-SVD) dictionary learning method.
Jiang etc. proposes the dictionary learning method based on the consistent K-SVD of category (Label Consistent K-SVD, LC-KSVD).
Yang etc. proposes identification dictionary learning (Fisher Discrimination Dictionary using Fisher criterions
Learning, FDDL) method, lift the differentiation performance of dictionary indirectly by constraining rarefaction representation coefficient.Vu etc. proposes one kind
Towards identification feature dictionary learning (Discriminative Feature-oriented Dictionary Learning,
DFDL) method, and it is applied to histopathology image classification.The above method, can obtain very good in image classification
Classifying quality.
However, because the feature that different types of histopathology image is presented is different, same type of histopathology image
Middle cellular morphology is larger with geometry changing features, and pathological characters also show variation, and this causes similar pathological image sample
The feature difference of this is more than the feature difference between non-similar pathology image pattern so that the ill dictionary of above method study with
Disease-free dictionary similarity degree is higher, still relatively low with the identification of ill sample to disease-free sample, what its classification performance still had
Treat in raising.
The content of the invention
In order to solve the above-mentioned technical problem, the present invention provides a kind of histopathology image that accuracy rate is high, robustness is high and knows
Other method.
Technical proposal that the invention solves the above-mentioned problems is:A kind of histopathology image-recognizing method, comprises the following steps:
Step one, chooses some image blocks as disease-free and have respectively from disease-free and ill two kinds of images of a certain tissue
Sick training sample, disease-free and ill test sample;
Step 2, Optimization Learning is disease-free dictionary:With reference to disease-free training sample and ill training sample, disease-free dictionary is set up
Model is practised, object function is minimized by the optimal way of two step alternating iterations, study obtains disease-free dictionary;
Step 3, the ill dictionary of Optimization Learning:With reference to ill training sample and disease-free training sample, ill dictionary is set up
Model is practised, object function is minimized by the optimal way of two step alternating iterations, study obtains ill dictionary;
Step 4, judges whether to reach maximum iteration, if so, then enter step 5, if it is not, then return to step
Two;
Step 5, obtains the reconstructed error vector of test sample:Using the disease-free dictionary and ill dictionary that obtain, to test
Sample carries out rarefaction representation, then calculate respectively sparse reconstructed error of the test sample under disease-free dictionary and ill dictionary to
Amount;
Step 6:Obtain the classification results of test sample:Classification statistic is obtained by sparse reconstructed error vector, then
The classification of test sample is determined by the comparing of classification statistic and threshold value.
Above-mentioned histopathology image-recognizing method, the step one is concretely comprised the following steps, disease-free and ill two from a certain tissue
The image block of equivalent amount is chosen in kind image respectively, each image block is then divided into RGB triple channels, by the pixel of triple channel
Series connection obtains characteristic vector after value is converted into column vector, finally using characteristic vector side by side as disease-free and ill training sample Y,Similarly obtain test sample.
Above-mentioned histopathology image-recognizing method, the step 2 is concretely comprised the following steps
2-1:Randomly selected respectively from disease-free and ill training sample n column vectors as initialization disease-free dictionary D and
Ill dictionary
2-2:Disease-free dictionary learning model is set up, model is as follows:
Wherein, argmin represents variate-value when making object function take minimum value, Y,Represent respectively disease-free with ill training
Sample, X,Represent the disease-free rarefaction representation coefficient with ill training sample respectively, N andDisease-free and ill image is represented respectively
The quantity of characteristic vector, L1The coding degree of rarefication for being disease-free sample and ill sample under disease-free dictionary, ρ is regularization parameter,
And ρ>0;In formulaThe sparse reconstructed error of disease-free dictionary and disease-free training sample is represented,Represent nothing
The reconstructed error of sick dictionary and ill training sample, F represents norm, and Ψ (D) is the Fisher criterion bound terms of disease-free dictionary, its
Expression formula is:Wherein m is the average of all atoms in disease-free dictionary D, and M is disease-free
The matrix of the atom average m compositions of dictionary D,It is ill dictionaryIn all atoms average, α, β represent spacing in class respectively
The penalty coefficient of spacing, α, β between class>0;
2-3:Fixed disease-free dictionary D, updates sparse coding coefficient, and object function now is as follows:
Make training sampleCode coefficient matrixL1It is disease-free sample and ill sample disease-free
Coding degree of rarefication under dictionary, optimal sparse solution isThe then solution of object function is divided into disease-free training sample
The step iteration of rarefaction representation two of rarefaction representation under disease-free dictionary D with ill training sample under disease-free dictionary D is completed, unified
Be simplified as:
Using the OMP algorithms in SPAMS tool boxes, training sample is solved respectively in disease-free dictionary D sparse solutions
2-4:Fixed sparse coding coefficient, updates disease-free dictionary D, and object function now is as follows:
Obtained by abbreviation:
Wherein, the mark of tr representing matrixs
Disease-free dictionary D optimal solutions are obtained using coordinate gradient descent method.
Above-mentioned histopathology image-recognizing method, the step 3 is concretely comprised the following steps
3-1:Randomly selected respectively from disease-free and ill training sample n column vectors as initialization disease-free dictionary D and
Ill dictionary
3-2:Ill dictionary learning model is set up, model is as follows:
Wherein, Y,Represent respectively it is disease-free with ill training sample, X,Represent respectively disease-free dilute with ill training sample
Dredge represent coefficient, N andThe quantity of disease-free and ill image feature vector, L are represented respectively2It is that disease-free sample and ill sample exist
Coding degree of rarefication under ill dictionary, ρ is regularization parameter, and ρ>0;In formulaIll dictionary is represented with ill sample
This sparse reconstructed error,The reconstructed error of ill dictionary and disease-free sample is represented,It is ill dictionary
Fisher criterion bound terms, its expression formula is:Wherein m is institute in disease-free dictionary D
There is the average of atom,It is ill dictionaryIn all atoms average, M be ill dictionaryIn all atoms averageGroup
Into matrix;
3-3:The ill dictionary of fixationSparse coding coefficient is updated, object function now is as follows:
Make training sampleCode coefficient matrixL2It is disease-free sample and ill sample ill
Coding degree of rarefication under dictionary, optimal sparse solution isThe then solution of object function is divided into disease-free training sample
In ill dictionaryUnder rarefaction representation with ill training sample in ill dictionaryUnder the step iteration of rarefaction representation two complete, system
One is simplified as:
Using the OMP algorithms in SPAMS tool boxes, training sample is solved respectively in ill dictionarySparse solution
3-4:Fixed sparse coding coefficient, updates ill dictionaryObject function now is as follows:
Obtained by abbreviation:
Wherein,
Ill dictionary is obtained using coordinate gradient descent methodOptimal solution.
Above-mentioned histopathology image-recognizing method, the step 5 is concretely comprised the following steps
5-1, by test sample image piecemeal, each segment is considered as a column vector h, and u segment composition matrix H is taken at random
As test sample, utilizeTest sample H is tried to achieve in band category dictionaryUnder sparse coding
5-2, calculates test sample in disease-free dictionary D and ill dictionaryUnder sparse reconstructed error vector, i.e. δ1=diag
((H-DX)(H-DX)T),Wherein, the unit on diag () representing matrix leading diagonal
Element.
Above-mentioned histopathology image-recognizing method, the step 6 is concretely comprised the following steps
6-1, definition vectorNtIt is the number of test sample;
6-2, classification statistic S is obtained by vectorial C:
When classification statistic S is more than or equal to threshold value Th, test sample is disease-free sample;Conversely, working as classification statistic S
Less than threshold value Th, then test sample is ill sample.
The beneficial effects of the present invention are:Step of the invention includes:Concentrated from histopathology view data first and distinguished
Some image blocks are randomly selected as training sample and test sample;Then different types of training sample is input to model
In, model is solved using the method for alternating iteration, object function is continued to optimize, study obtains band category dictionary;Finally
Rarefaction representation is carried out to test set matrix based on the band category dictionary for obtaining, is determined by the contrast of reconstructed error vector sum threshold value
The classification of this test set matrix.Application of the present invention to dictionary learning in histopathology image classification propose new model and
Method, the band category dictionary for learning has robustness in preferably sparse reconstitution and class to similar sample, to non-similar sample
This has identification between preferable class, can effectively improve histopathology image classification performance.
Brief description of the drawings
Fig. 1 is flow chart of the invention.
Fig. 2 be ADL databases in lung, spleen, the histopathology schematic diagram of kidney, wherein (a) from left to right be respectively lung,
The disease-free image of spleen, kidney, (b) is respectively lung, spleen, the ill image of kidney from left to right.
Fig. 3 is the histopathology schematic diagram of adenopathy and lobate cancer in BreaKHis databases, wherein the tissue of (a) for adenopathy
Pathological image, (b) is the histopathology image of lobate cancer.
Specific embodiment
The present invention is further illustrated with reference to the accompanying drawings and examples.
As shown in figure 1, the present invention is comprised the following steps:
Step one:Some image blocks are chosen respectively from disease-free and ill two kinds of images of a certain tissue as disease-free and have
Sick training sample, disease-free and ill test sample.Concretely comprise the following steps:
40 images are randomly selected respectively from disease-free and ill two kinds of images of a certain tissue, are carried at random from every image
Take 250 segments, the size of block is 20 × 20, then 10000 colored segments altogether, each colored segment is then divided into RGB
Triple channel, is converted into the pixel value of triple channel series connection after column vector and obtains characteristic vector, finally using characteristic vector it is arranged side by side as
Training sample, then Y,R1200×10000The size of representing matrix, respectively from remaining a certain organization chart picture with
Machine chooses disease-free and ill each 110 of two kinds of images as test set.
Step 2, Optimization Learning is disease-free dictionary:With reference to disease-free training sample and ill training sample, disease-free dictionary is set up
Model is practised, object function is minimized by the optimal way of two step alternating iterations, study obtains disease-free dictionary.Concretely comprise the following steps:
2-1:Randomly selected respectively from disease-free and ill training sample n column vectors as initialization disease-free dictionary D and
Ill dictionary
2-2:Disease-free dictionary learning model is set up, model is as follows:
Wherein, argmin represents variate-value when making object function take minimum value, Y,Represent respectively disease-free with ill training
Sample, X,Represent the disease-free rarefaction representation coefficient with ill training sample respectively, N andDisease-free and ill image is represented respectively
The quantity of characteristic vector, L1The coding degree of rarefication for being disease-free sample and ill sample under disease-free dictionary, ρ is regularization parameter,
And ρ>0;In formulaThe sparse reconstructed error of disease-free dictionary and disease-free training sample is represented,Represent nothing
The reconstructed error of sick dictionary and ill training sample, F represents norm, and Ψ (D) is the Fisher criterion bound terms of disease-free dictionary, its
Expression formula is:Wherein m is the average of all atoms in disease-free dictionary D, and M is disease-free
The matrix of the atom average m compositions of dictionary D,It is ill dictionaryIn all atoms average, α, β represent spacing in class respectively
The penalty coefficient of spacing, α, β between class>0;Purpose of model is simultaneously to maximize the 2nd simultaneously by minimizing the 1st and the 3rd,
What is then learnt is preferable to the reconstruction property of similar sample with category dictionary, poor for non-similar sample reconstruction property, or even nothing
Method is reconstructed, and has stronger resolving ability between the dictionary of study, so as to obtain with identification feature so that further can be more
Good classification;
2-3:Fixed disease-free dictionary D, updates sparse coding coefficient, and object function now is as follows:
Make training sampleCode coefficient matrixL1It is disease-free sample and ill sample disease-free
Coding degree of rarefication under dictionary, optimal sparse solution isThe then solution of object function is divided into disease-free training sample
The step iteration of rarefaction representation two of rarefaction representation under disease-free dictionary D with ill training sample under disease-free dictionary D is completed, unified
Be simplified as:
Using the OMP algorithms in SPAMS tool boxes, training sample is solved respectively in disease-free dictionary D sparse solutions
2-4:Fixed sparse coding coefficient, updates disease-free dictionary D, and object function now is as follows:
Obtained by abbreviation:
Wherein, the mark of tr representing matrixs
Above-mentioned function is convex function, and disease-free dictionary D optimal solutions are obtained using coordinate gradient descent method.
Step 3, the ill dictionary of Optimization Learning:With reference to ill training sample and disease-free training sample, ill dictionary is set up
Model is practised, object function is minimized by the optimal way of two step alternating iterations, study obtains ill dictionary.Concretely comprise the following steps:
3-1:Randomly selected respectively from disease-free and ill training sample n column vectors as initialization disease-free dictionary D and
Ill dictionary
3-2:Ill dictionary learning model is set up, model is as follows:
Wherein, Y,Represent respectively it is disease-free with ill training sample, X,Represent respectively disease-free dilute with ill training sample
Dredge represent coefficient, N andThe quantity of disease-free and ill image feature vector, L are represented respectively2It is that disease-free sample and ill sample exist
Coding degree of rarefication under ill dictionary, ρ is regularization parameter, and ρ>0;In formulaIll dictionary is represented with ill sample
This sparse reconstructed error,The reconstructed error of ill dictionary and disease-free sample is represented,It is ill dictionary
Fisher criterion bound terms, its expression formula is:Wherein m is institute in disease-free dictionary D
There is the average of atom,It is ill dictionaryIn all atoms average, M be ill dictionaryIn all atoms averageComposition
Matrix;Purpose of model is by minimizing the 1st and the 3rd and maximizing the 2nd, then the band category dictionary pair for learning simultaneously
The reconstruction property of similar sample is preferably, poor for non-similar sample reconstruction property, or even cannot reconstruct, and between the dictionary of study
With stronger resolving ability, so as to obtain with identification feature so as to further can preferably classify.
3-3:The ill dictionary of fixationSparse coding coefficient is updated, object function now is as follows:
Make training sampleCode coefficient matrixL2It is disease-free sample and ill sample ill
Coding degree of rarefication under dictionary, optimal sparse solution isThe then solution of object function is divided into disease-free training sample
In ill dictionaryUnder rarefaction representation with ill training sample in ill dictionaryUnder the step iteration of rarefaction representation two complete, system
One is simplified as:
Using the OMP algorithms in SPAMS tool boxes, training sample is solved respectively in ill dictionarySparse solution
3-4:Fixed sparse coding coefficient, updates ill dictionaryObject function now is as follows:
Obtained by abbreviation:
Wherein,
Ill dictionary is obtained using coordinate gradient descent methodOptimal solution;
3-5:The process of return to step two, Optimization Learning is disease-free dictionary and the ill dictionary of Optimization Learning alternately, until
Stop when reaching maximum iteration.
Step 4, judges whether to reach maximum iteration, if so, then enter step 5, if it is not, then return to step
Two.
Step 5, obtains the reconstructed error vector of test sample:Using the disease-free dictionary and ill dictionary that obtain, to test
Sample carries out rarefaction representation, then calculate respectively sparse reconstructed error of the test sample under disease-free dictionary and ill dictionary to
Amount.Concretely comprise the following steps:
5-1, by test sample image piecemeal, each segment is considered as a column vector h, and 250 segment composition squares are taken at random
Battle array H is utilized as test sampleTest sample H is tried to achieve in band class marking-up
Allusion quotationUnder sparse coding
5-2, calculates test sample in disease-free dictionary D and ill dictionaryUnder sparse reconstructed error vector, i.e. δ1=diag
((H-DX)(H-DX)T),Wherein, the unit on diag () representing matrix leading diagonal
Element.
Step 6:Obtain the classification results of test sample:Classification statistic is obtained by sparse reconstructed error vector, then
The classification of test sample is determined by the comparing of classification statistic and threshold value.Concretely comprise the following steps:
6-1, definition vectorNtIt is the number of test sample;
6-2, classification statistic S is obtained by vectorial C:
When classification statistic S is more than or equal to threshold value Th, test sample is disease-free sample;Conversely, working as classification statistic S
Less than threshold value Th, then test sample is ill sample.
Table 1 is the present invention and the classification results contrast table of other approach applications to the lung images in ADL databases.
Table 1
Table 2 is the present invention and the classification results contrast table of other approach applications to the spleen image in ADL databases.
Table 2
Table 2 is the present invention and the classification results contrast table of other approach applications to the renal image in ADL databases.
Table 3
By table 1, table 2, table 3 it is recognised that model proposed by the present invention is obvious to the medical diagnosis on disease effect of this three organoid
Other method is better than, a just point rate all increases under disease-free sample and ill sample.Especially, lung's classification knot of table 1
Fruit becomes apparent, and compared with DFDL, the nicety of grading of context of methods improves 2~3%.As shown in Figure 2, disease-free lung images
In comprising the larger alveolar of volume, and alveolar volume is smaller in ill lung images, and it is thin to be covered with the inflammation of bluish violet
Born of the same parents, and texture is increasingly complex, the disease-free otherness between ill lung images is significantly greater than spleen and renal image.Meanwhile,
It is disease-free high with structural similarity with ill spleen image texture, but because color distortion is larger, two class image discriminatings take second place, its
Classification performance takes second place;It is disease-free that not only texture is similar to structure high but also color similarity is high, identification to ill renal image
Worst, its classification performance is most weak.Experimental result is consistent completely with Fig. 1 in table, and the effective of model proposed by the present invention is illustrated again
Property.
In order to verify the universality of the identification feature learning framework of the histopathology image of present invention structure, particularly,
The model that the present invention is carried is applied to the diagnosis of disease type in BreaKHis data sets.
Table 4 is the present invention and other approach applications to classification results contrast table in BreaKHis databases.
Table 4
Table 4 gives classification results of the distinct methods on BreaKHis databases, test result indicate that, the present invention is proposed
Model also showed that preferable classification of diseases performance for two kinds of benign breast cancer images in Fig. 3, the result shows this
Invention has preferably effect with category dictionary for effectively improving to the reconstitution and robustness of the rarefaction representation of similar sample,
It is poor for non-similar sample identification also to solve the problems, such as simultaneously.
Claims (6)
1. a kind of histopathology image-recognizing method, comprises the following steps:
Step one, chooses some image blocks as disease-free and ill instruction respectively from disease-free and ill two kinds of images of a certain tissue
Practice sample, disease-free and ill test sample;
Step 2, Optimization Learning is disease-free dictionary:With reference to disease-free training sample and ill training sample, disease-free dictionary learning mould is set up
Type, object function is minimized by the optimal way of two step alternating iterations, and study obtains disease-free dictionary;
Step 3, the ill dictionary of Optimization Learning:With reference to ill training sample and disease-free training sample, ill dictionary learning mould is set up
Type, object function is minimized by the optimal way of two step alternating iterations, and study obtains ill dictionary;
Step 4, judges whether to reach maximum iteration, if so, then enter step 5, if it is not, then return to step two;
Step 5, obtains the reconstructed error vector of test sample:Using the disease-free dictionary and ill dictionary that obtain, to test sample
Rarefaction representation is carried out, sparse reconstructed error vector of the test sample under disease-free dictionary and ill dictionary is then calculated respectively;
Step 6:Obtain the classification results of test sample:Classification statistic is obtained by sparse reconstructed error vector, is then passed through
Classification statistic determines the classification of test sample with the comparing of threshold value.
2. histopathology image-recognizing method according to claim 1, it is characterised in that:Step one specific steps
The image block of equivalent amount to be chosen respectively from a certain disease-free and ill two kinds of images of tissue, then by each image block point
It is RGB triple channels, series connection obtains characteristic vector after the pixel value of triple channel is converted into column vector, finally that characteristic vector is arranged side by side
As disease-free and ill training sample Y,Similarly obtain test sample.
3. histopathology image-recognizing method according to claim 2, it is characterised in that:The specific steps of the step 2
For
2-1:Randomly select disease-free dictionary D and ill of the n column vectors as initialization respectively from disease-free and ill training sample
Dictionary
2-2:Disease-free dictionary learning model is set up, model is as follows:
Wherein, argmin represents variate-value when making object function take minimum value, Y,Represent respectively disease-free with ill training sample
This, X,Represent the disease-free rarefaction representation coefficient with ill training sample respectively, N andIt is special that disease-free and ill image is represented respectively
Levy the quantity of vector, L1The coding degree of rarefication for being disease-free sample and ill sample under disease-free dictionary, ρ is regularization parameter, and ρ
>0;In formulaThe sparse reconstructed error of disease-free dictionary and disease-free training sample is represented,Represent disease-free
The reconstructed error of dictionary and ill training sample, F represents norm, and Ψ (D) is the Fisher criterion bound terms of disease-free dictionary, its table
It is up to formula:
Wherein m is the average of all atoms in disease-free dictionary D, and M is disease-free dictionary D
The matrix of atom average m compositions,It is ill dictionaryIn all atoms average, between α, β are represented in class between spacing and class respectively
Away from penalty coefficient, α, β>0;
2-3:Fixed disease-free dictionary D, updates sparse coding coefficient, and object function now is as follows:
Make training sampleCode coefficient matrixL1It is disease-free sample and ill sample in disease-free dictionary
Under coding degree of rarefication, optimal sparse solution isThe then solution of object function is divided into disease-free training sample in nothing
The step iteration of rarefaction representation two of rarefaction representation under sick dictionary D with ill training sample under disease-free dictionary D is completed, unified letter
Change as follows:
Using the OMP algorithms in SPAMS tool boxes, training sample is solved respectively in disease-free dictionary D sparse solutions
2-4:Fixed sparse coding coefficient, updates disease-free dictionary D, and object function now is as follows:
Obtained by abbreviation:
Wherein, the mark of tr representing matrixs
Disease-free dictionary D optimal solutions are obtained using coordinate gradient descent method.
4. histopathology image-recognizing method according to claim 3, it is characterised in that:The specific steps of the step 3
For
3-1:Randomly select disease-free dictionary D and ill of the n column vectors as initialization respectively from disease-free and ill training sample
Dictionary
3-2:Ill dictionary learning model is set up, model is as follows:
Wherein, Y,Represent respectively it is disease-free with ill training sample, X,The disease-free sparse table with ill training sample is represented respectively
Show coefficient, N and N represents the quantity of disease-free and ill image feature vector, L respectively2It is disease-free sample and ill sample ill
Coding degree of rarefication under dictionary, ρ is regularization parameter, and ρ>0;In formulaIll dictionary is represented with ill sample
Sparse reconstructed error,The reconstructed error of ill dictionary and disease-free sample is represented,For the Fisher of ill dictionary is accurate
Then bound term, its expression formula is:Wherein m is all atoms in disease-free dictionary D
Average,It is ill dictionaryIn all atoms average, M be ill dictionaryIn all atoms averageThe matrix of composition;
3-3:The ill dictionary of fixationSparse coding coefficient is updated, object function now is as follows:
Make training sampleCode coefficient matrixL2It is disease-free sample and ill sample in ill dictionary
Under coding degree of rarefication, optimal sparse solution isThe then solution of object function is divided into disease-free training sample and is having
Sick dictionaryUnder rarefaction representation with ill training sample in ill dictionaryUnder the step iteration of rarefaction representation two complete, it is unified
It is simplified as:
Using the OMP algorithms in SPAMS tool boxes, training sample is solved respectively in ill dictionarySparse solution
3-4:Fixed sparse coding coefficient, updates ill dictionaryObject function now is as follows:
Obtained by abbreviation:
Wherein,
Ill dictionary is obtained using coordinate gradient descent methodOptimal solution.
5. histopathology image-recognizing method according to claim 4, it is characterised in that:The specific steps of the step 5
For
5-1, by test sample image piecemeal, each segment is considered as a column vector h, and u segment composition matrix H conduct is taken at random
Test sample, utilizesTest sample H is tried to achieve in band category dictionaryUnder
Sparse coding
5-2, calculates test sample in disease-free dictionary D and ill dictionaryUnder sparse reconstructed error vector, i.e. δ1=diag ((H-
DX)(H-DX)T),Wherein, the element on diag () representing matrix leading diagonal.
6. histopathology image-recognizing method according to claim 5, it is characterised in that the specific steps of the step 6
For
6-1, definition vectorNtIt is the number of test sample;
6-2, classification statistic S is obtained by vectorial C:
When classification statistic S is more than or equal to threshold value Th, test sample is disease-free sample;Conversely, when classification statistic S is less than
Threshold value Th, then test sample is ill sample.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710059300.0A CN106845551B (en) | 2017-01-24 | 2017-01-24 | Tissue pathology image identification method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710059300.0A CN106845551B (en) | 2017-01-24 | 2017-01-24 | Tissue pathology image identification method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106845551A true CN106845551A (en) | 2017-06-13 |
CN106845551B CN106845551B (en) | 2020-08-11 |
Family
ID=59122438
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710059300.0A Active CN106845551B (en) | 2017-01-24 | 2017-01-24 | Tissue pathology image identification method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106845551B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107832786A (en) * | 2017-10-31 | 2018-03-23 | 济南大学 | A kind of recognition of face sorting technique based on dictionary learning |
CN109063766A (en) * | 2018-07-31 | 2018-12-21 | 湘潭大学 | A kind of image classification method based on identification prediction sparse decomposition model |
CN109308485A (en) * | 2018-08-02 | 2019-02-05 | 中国矿业大学 | A kind of migration sparse coding image classification method adapted to based on dictionary domain |
CN109376802A (en) * | 2018-12-12 | 2019-02-22 | 浙江工业大学 | A kind of gastroscope organ classes method dictionary-based learning |
CN111027594A (en) * | 2019-11-18 | 2020-04-17 | 西北工业大学 | Step-by-step anomaly detection method based on dictionary representation |
CN113627556A (en) * | 2021-08-18 | 2021-11-09 | 广东电网有限责任公司 | Method and device for realizing image classification, electronic equipment and storage medium |
CN113793319A (en) * | 2021-09-13 | 2021-12-14 | 浙江理工大学 | Fabric image flaw detection method and system based on class constraint dictionary learning model |
CN114428873A (en) * | 2022-04-07 | 2022-05-03 | 源利腾达(西安)科技有限公司 | Thoracic surgery examination data sorting method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104866810A (en) * | 2015-04-10 | 2015-08-26 | 北京工业大学 | Face recognition method of deep convolutional neural network |
CN105844223A (en) * | 2016-03-18 | 2016-08-10 | 常州大学 | Face expression algorithm combining class characteristic dictionary learning and shared dictionary learning |
US20160307073A1 (en) * | 2015-04-20 | 2016-10-20 | Los Alamos National Security, Llc | Change detection and change monitoring of natural and man-made features in multispectral and hyperspectral satellite imagery |
-
2017
- 2017-01-24 CN CN201710059300.0A patent/CN106845551B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104866810A (en) * | 2015-04-10 | 2015-08-26 | 北京工业大学 | Face recognition method of deep convolutional neural network |
US20160307073A1 (en) * | 2015-04-20 | 2016-10-20 | Los Alamos National Security, Llc | Change detection and change monitoring of natural and man-made features in multispectral and hyperspectral satellite imagery |
CN105844223A (en) * | 2016-03-18 | 2016-08-10 | 常州大学 | Face expression algorithm combining class characteristic dictionary learning and shared dictionary learning |
Non-Patent Citations (2)
Title |
---|
TIEP HUU VU 等: "Histopathological Image Classification Using Discriminative Feature-Oriented Dictionary Learning", 《IEEE TRANSACTIONS ON MEDICAL IMAGING》 * |
ZHUOLIN JIANG 等: "Label Consistent K-SVD: Learning a Discriminative Dictionary for Recognition", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107832786A (en) * | 2017-10-31 | 2018-03-23 | 济南大学 | A kind of recognition of face sorting technique based on dictionary learning |
CN107832786B (en) * | 2017-10-31 | 2019-10-25 | 济南大学 | A kind of recognition of face classification method dictionary-based learning |
CN109063766A (en) * | 2018-07-31 | 2018-12-21 | 湘潭大学 | A kind of image classification method based on identification prediction sparse decomposition model |
CN109308485A (en) * | 2018-08-02 | 2019-02-05 | 中国矿业大学 | A kind of migration sparse coding image classification method adapted to based on dictionary domain |
CN109308485B (en) * | 2018-08-02 | 2022-11-29 | 中国矿业大学 | Migrating sparse coding image classification method based on dictionary field adaptation |
CN109376802B (en) * | 2018-12-12 | 2021-08-03 | 浙江工业大学 | Gastroscope organ classification method based on dictionary learning |
CN109376802A (en) * | 2018-12-12 | 2019-02-22 | 浙江工业大学 | A kind of gastroscope organ classes method dictionary-based learning |
CN111027594A (en) * | 2019-11-18 | 2020-04-17 | 西北工业大学 | Step-by-step anomaly detection method based on dictionary representation |
CN111027594B (en) * | 2019-11-18 | 2022-08-12 | 西北工业大学 | Step-by-step anomaly detection method based on dictionary representation |
CN113627556A (en) * | 2021-08-18 | 2021-11-09 | 广东电网有限责任公司 | Method and device for realizing image classification, electronic equipment and storage medium |
CN113793319A (en) * | 2021-09-13 | 2021-12-14 | 浙江理工大学 | Fabric image flaw detection method and system based on class constraint dictionary learning model |
CN113793319B (en) * | 2021-09-13 | 2023-08-25 | 浙江理工大学 | Fabric image flaw detection method and system based on category constraint dictionary learning model |
CN114428873A (en) * | 2022-04-07 | 2022-05-03 | 源利腾达(西安)科技有限公司 | Thoracic surgery examination data sorting method |
Also Published As
Publication number | Publication date |
---|---|
CN106845551B (en) | 2020-08-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106845551A (en) | A kind of histopathology image-recognizing method | |
CN107220980B (en) | A kind of MRI image brain tumor automatic division method based on full convolutional network | |
EP3432263B1 (en) | Semantic segmentation for cancer detection in digital breast tomosynthesis | |
Kooi et al. | Large scale deep learning for computer aided detection of mammographic lesions | |
CN109886273B (en) | CMR image segmentation and classification system | |
CN110889852B (en) | Liver segmentation method based on residual error-attention deep neural network | |
CN107506761A (en) | Brain image dividing method and system based on notable inquiry learning convolutional neural networks | |
WO2019001208A1 (en) | Segmentation algorithm for choroidal neovascularization in oct image | |
CN112001928B (en) | Retina blood vessel segmentation method and system | |
CN110705555A (en) | Abdomen multi-organ nuclear magnetic resonance image segmentation method, system and medium based on FCN | |
CN110265141B (en) | Computer-aided diagnosis method for liver tumor CT image | |
CN108491849A (en) | Hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks | |
CN109785300A (en) | A kind of cancer medical image processing method, system, device and storage medium | |
CN111145181A (en) | Skeleton CT image three-dimensional segmentation method based on multi-view separation convolutional neural network | |
CN109767459B (en) | Novel fundus image registration method | |
CN110675411A (en) | Cervical squamous intraepithelial lesion recognition algorithm based on deep learning | |
CN106021330A (en) | A three-dimensional model retrieval method used for mixed contour line views | |
CN104484886A (en) | Segmentation method and device for MR image | |
CN111476794B (en) | Cervical pathological tissue segmentation method based on UNET | |
Wang et al. | EAR-U-Net: EfficientNet and attention-based residual U-Net for automatic liver segmentation in CT | |
CN115965602A (en) | Abnormal cell detection method based on improved YOLOv7 and Swin-Unet | |
CN109685830A (en) | Method for tracking target, device and equipment and computer storage medium | |
You et al. | EG-Trans3DUNet: a single-staged transformer-based model for accurate vertebrae segmentation from spinal CT images | |
CN112419335B (en) | Shape loss calculation method of cell nucleus segmentation network | |
CN108280417A (en) | A kind of finger vena method for quickly identifying |
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 |