CN107103334A - It is a kind of based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study - Google Patents
It is a kind of based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study Download PDFInfo
- Publication number
- CN107103334A CN107103334A CN201710280338.0A CN201710280338A CN107103334A CN 107103334 A CN107103334 A CN 107103334A CN 201710280338 A CN201710280338 A CN 201710280338A CN 107103334 A CN107103334 A CN 107103334A
- Authority
- CN
- China
- Prior art keywords
- msub
- mrow
- dictionary
- convolutional neural
- neural networks
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
- 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
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/194—Segmentation; Edge detection involving foreground-background 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
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10081—Computed x-ray tomography [CT]
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- 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
- G06T2207/30061—Lung
- G06T2207/30064—Lung nodule
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study, dictionary is combined by this method to the classification layer of learning algorithm as convolutional neural networks with convolutional neural networks, referred to as CNN DPC.The algorithm includes the training of convolutional neural networks parameter and Lung neoplasm is classified two stages.Training data is pre-processed first, pulmonary parenchyma part is obtained；Then pre-training is carried out to convolutional neural networks；Then the classification layer of convolutional neural networks is replaced to learning algorithm with dictionary, and to the network parameter and dictionary after pre-training to optimizing；Finally treat grouped data to be pre-processed, obtain pulmonary parenchyma part, and realize by the disaggregated model trained the classification of Lung neoplasm.The present invention is realized simply, it can be readily appreciated that by the way that convolutional neural networks are combined with dictionary to the Dominant Facies of learning algorithm, can comparatively fast and preferably realize the good pernicious classification of Lung neoplasm.
Description
Technical field
The present invention relates to Lung neoplasm classification, and in particular to a kind of based on the Lung neoplasm of convolutional neural networks and dictionary to study
Sorting technique.
Background technology
The early screening of lung cancer is significant for incidence and the death rate for reducing lung cancer, and lung CAD system is answered
Work load is reduced with help doctor, and diagnosis speed is fast, diagnostic result is objective., can be true for lung CAD system
The type for examining Lung neoplasm is to judge the major criterion of its performance.Traditional sorting technique is imitated for large-scale medical image processing
Fruit is simultaneously bad, and convolutional neural networks achieve relatively good application effect in terms of image recognition and target detection, but works as
The preceding grader being used in combination with convolutional neural networks, such as SVM, sigmoid functions etc., it is impossible to excavating depth feature well
Labyrinth.The dictionary occurred for 2014 can not only realize the classification of sample data to learning algorithm, and can be fine
Express the feature of sample data in ground.But classification is carried out by artificial acquisition feature using this method, and for different samples
This needs to extract different features could preferably treatment classification task.In summary, in order to improve lung CAD system Diagnosis of pulmonary
The accuracy rate of tubercle, convolutional neural networks are combined with dictionary to learning method.
Dictionary includes dictionary learning and two stages of classification to learning algorithm.
(1) the dictionary learning stage
If X=[X1,…,Xk,…XK](Xk∈Rd×n_k, n_k represents the number of kth class training sample) and represent dictionary to learning
The corresponding feature of input sample of algorithm is practised, this feature is corresponding with the input picture from K classes.So, dictionary is to learning algorithm
Target be to find a structured analysis dictionary P=[P1,…,Pk,…PK](Pk∈Rm×d) it is combined to dictionary D with a structure
=[D1,…,Dk,…DK](Dk∈Rd×m), the coding to feature X and reconstruct are realized, m represents the number of dictionary atom.For instruction
Practice the feature X of kth class sample in sample setk, to stator dictionary to { Dk, PkOn the premise of, order, with XkCorresponding code coefficient
It is expressed as PkXk.Then, dictionary is defined as to learning algorithm DPL object function：
Wherein, Y represents class label matrix corresponding with feature X, and φ (P, D, X, Y) is one and sentenced to strengthen dictionary D and P
The differentiation of other property.
In dictionary in learning algorithm, it is desirable to sub- dictionary PkSample of the mapping from other class i (i ≠ k)(sample set
Except X in XkOther samples in addition) to one close to kernel, i.e. PkXi≈0,Then problem (1) is written as
Wherein,It is a constraint information, for avoiding PkTrivial solution is produced when=0, λ is a weight coefficient.
Due to the object function non-convex that problem (2) is represented, a matrix variables A presentation code coefficient matrix is introduced, and will
Problem (2) loosens as follows
Wherein, τ is a weight coefficient.
Initialization P and D is unit Frobenius norms, and the solution of problem (3) can be real by being alternately performed following three step
It is existing：
(i) fixed { D, P, X }, the approximate solution represented with formula (4) updates A：
(ii) fixed { A, X }, the approximate solution represented with formula (5) updates P：
Wherein, γ is a weight coefficient.
(iii) it is fixed { A, X }, a variable S is introduced, problem (3) is converted into
So D value is updated by formula (7)：
Wherein, ρ is a weight parameter.
When the difference for the target function value of iteration twice in iterative process, occur is less than the situation of some threshold value, stop
Iteration.
(2) sorting phase
If sample data set includes K species, K sub- dictionaries can be obtained after the dictionary learning stage to { Dk,
Pk(k=1,2 ..., K).A test image is given, its feature is extracted, is designated as x, make x on the basis of k-th of sub- dictionary pair
Reconstruction error is E (x；Dk,Pk), then dictionary is to the grader layer DPC rules classified to test image：
That is, if k-th sub- dictionary is minimum to the reconstruction error for reconstructing x, then the corresponding sample images of feature x just belong to
Kth class.
The content of the invention
It is a kind of based on convolutional neural networks and dictionary it is an object of the invention to overcome the defect that prior art is present to provide
The Lung neoplasm sorting algorithm being combined to learning algorithm, improves the accuracy rate of Lung neoplasm classification.
The present invention is adopted the following technical scheme that：
It is a kind of based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study, including convolutional neural networks parameter
Training and Lung neoplasm classification be two stages of training stage and sorting phase：
1) training stage:
A. the low dosage thin layer scanning CT images as training sample are inputted；
B. the CT images of training sample are pre-processed；
C. train classification models CNN-DPC parameter；Comprise the following steps：
C1. the parameter to convolutional neural networks carries out pre-training；In the pre-training stage, classical convolutional neural networks are used for reference
LeNet-5 network structure, based on training dataset, the convolutional neural networks structure of one 9 layers of design, the convolutional Neural
Network includes input layer, three parts of hidden layer and output layer, and wherein hidden layer includes three convolutional layers, three maximum pond layers
With a full articulamentum；Input layer number is 128 × 128, and each node in hidden layer is different, and output layer nodes are
2, output layer is classification layer, and the function of classification is sigmoid functions, and the activation primitive of hidden layer is ReLU (Rectified
Linear Units)；Forward-propagating and ginseng of the error backpropagation algorithm to this 9 layers of convolutional neural networks using standard
Number carries out pre-training；
C2. after pre-training terminates, the classification layer of convolutional neural networks is replaced to learning algorithm with dictionary, then carries out parameter
Optimization；The optimization of parameter includes the renewal of dictionary pair and the fine setting of network parameter, specifically includes following steps：
C21. the sigmoid functions of convolutional neural networks classification layer are replaced to learning algorithm with dictionary；Assuming that XkRepresent instruction
Practice the output of the full articulamentum of stage kth class training data correspondence convolutional neural networks, C represents target letter of the dictionary to learning algorithm
Number, CkRepresent the corresponding dictionary of kth class training data to the object function of learning algorithm, { Pk,DkRepresent kth class training data
Corresponding structuring dictionary pair, due in dictionary in learning algorithm, the renewal of dictionary pair is broken down into following subproblem：
So in parameter optimisation step, orderThen C is to XkPartial derivative be：
Because our purpose is that Lung neoplasm is divided into two classes, so in above formula, k=1,2, K=2；
C22. obtaining allAfterwards, as the gradient in the back-propagation algorithm of standard to convolutional Neural net
The parameter of network is finely adjusted, while using the feature of the Network Capture training dataset after fine setting to dictionary to being updated；
2) sorting phase, the classification of data to be sorted；Sorting phase includes following three step：
A. low dosage thin layer scanning CT images to be sorted are inputted；
B. CT images to be sorted are pre-processed；
C. image to be classified is classified using the disaggregated model CNN-DPC for training network parameter.After optimization
Convolutional neural networks obtain the feature that all training datas concentrate pulmonary parenchyma images, to object function of the dictionary to learning algorithm
Solve, 2 sub- dictionaries can be obtained to { (D1,P1), (D2,P2)}.A test image is given, its feature is extracted, is designated as x, is made
Reconstruction errors of the x on the basis of the individual sub- dictionary pair of kth (k=1,2) is E (x；Dk,Pk), then disaggregated model CNN-DPC is to test chart
As the rule classified is：
That is, if k-th sub- dictionary is minimum to the reconstruction error for reconstructing x, then the corresponding sample images of feature x just belong to
Kth class.
It is described based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study, in step B, specifically include with
Lower step：
1st step, binary conversion treatment is carried out to original CT image using Otsu threshold methods；
2nd step, to the result of the 1st step, region is carried out by seed point of first, upper left corner pixel and is increased, by background portion
Divide and separated with foreground part；
3rd step, to the result of the 2nd step, lathe part is first obtained using region growing algorithm；
4th step, the result of preceding 3 step is overlapped, pulmonary parenchyma part is obtained；
5th step, based on the result of the 4th step, region growing algorithm is respectively adopted to pulmo and obtains complete pulmo, so
After merge；
6th step, the result to the 5th step performs morphological operation, obtains pulmonary parenchyma mask；
7th step, the pulmonary parenchyma in original CT image is partitioned into according to pulmonary parenchyma mask；
8th step, 128 × 128 are adjusted to by the obtained pulmonary parenchyma image size in the 7th step by bilinear interpolation.
It is described based on convolutional neural networks and dictionary to the Lung neoplasm sorting technique of study, sorting phase is to be sorted
CT images are pre-processed, and specifically include following steps：
1st step, binary conversion treatment is carried out to original CT image using Otsu threshold methods；
2nd step, to the result of the 1st step, region is carried out by seed point of first, upper left corner pixel and is increased, by background portion
Divide and separated with foreground part；
3rd step, to the result of the 2nd step, lathe part is first obtained using region growing algorithm；
4th step, the result of preceding 3 step is overlapped, pulmonary parenchyma part is obtained；
5th step, based on the result of the 4th step, region growing algorithm is respectively adopted to pulmo and obtains complete pulmo, so
After merge；
6th step, the result to the 5th step performs morphological operation, obtains pulmonary parenchyma mask；
7th step, the pulmonary parenchyma in original CT image is partitioned into according to pulmonary parenchyma mask；
8th step, 128 × 128 are adjusted to by the obtained pulmonary parenchyma image size in the 7th step by bilinear interpolation.
Compared with prior art, the present invention has advantages below：
1. the present invention can overcome defect of the prior art, by convolutional neural networks and advantage of the dictionary to learning algorithm
Combine, realize simple, it is easy to understand；
Can be doctor 2. by using the technology of the present invention, species that can be accurately and efficiently to Lung neoplasm is diagnosed
Raw diagnosis provides reference.
Brief description of the drawings
Fig. 1 is the flow chart that the present invention carries out Lung neoplasm classification.
Fig. 2 is the procedure chart that the present invention is pre-processed to original CT image.
Embodiment
Below in conjunction with specific embodiment, the present invention is described in detail.
Reference picture 1, the present invention is based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study, including training rank
Section and sorting phase, are comprised the following steps that：
1) training stage:
A. reference picture 1, inputs the low dosage thin layer scanning CT images as training sample.
B. reference picture 2, are pre-processed to CT images, obtain the pulmonary parenchyma image that unified size is 128 × 128.Specifically
Step includes：
1st step, binary conversion treatment is carried out to original CT image using Otsu threshold methods；
2nd step, to the result of the 1st step, region is carried out by seed point of first, upper left corner pixel and is increased, by background portion
Divide and separated with foreground part；
3rd step, to the result of the 2nd step, lathe part is first obtained using region growing algorithm；
4th step, the result of preceding 3 step is overlapped, pulmonary parenchyma part is obtained；
5th step, based on the result of the 4th step, region growing algorithm is respectively adopted to pulmo and obtains complete pulmo, so
After merge；
6th step, the result to the 5th step performs morphological operation, obtains pulmonary parenchyma mask；
7th step, the pulmonary parenchyma in original CT image is partitioned into according to pulmonary parenchyma mask；
8th step, 128 × 128 are adjusted to by the obtained pulmonary parenchyma image size in the 7th step by bilinear interpolation.
C. train classification models CNN-DPC parameter；
Size, activation primitive and the output knot for each hidden layer correspondence core of convolutional neural networks that the present invention that table 1 provides is designed
Really.The convolutional neural networks include input layer, three parts of hidden layer and output layer, wherein hidden layer comprising three convolutional layers,
Three maximum pond layers and a full articulamentum.The activation of hidden layer is used as using ReLU (Rectified Linear Units)
Function, the then convolution operation that an input feature vector is mapped between f and convolution kernel h is defined as：
Wherein, fkWith hkInput mapping and the convolution kernel of kth layer are represented respectively, and * represents convolution operation, and b represents biasing, hk
Constantly updated in the training process with b.
Maximum pond Operation Definition is
Wherein, w illustrates that pond size is w × w, and CO represents that the feature of the previous convolutional layer output adjacent with pond layer is reflected
Penetrate, i and j are greater than being equal to 0 integer, but are to ensure that iw+m≤size (CO, 1) and jw+n≤size (CO, 2).
The output of maximum pond layer is changed into an one-dimensional vector by full articulamentum.
Size, activation primitive and the output result of each hidden layer correspondence core of the convolutional neural networks of table 1
C1. the parameter to convolutional neural networks carries out pre-training.In the pre-training stage, classical convolutional neural networks are used for reference
LeNet-5 network structure, based on training dataset, the convolutional neural networks structure of one 9 layers of design, the convolutional Neural
Network includes input layer, three parts of hidden layer and output layer, and wherein hidden layer includes three convolutional layers, three maximum pond layers
With a full articulamentum.Input layer number is 128 × 128, and each node in hidden layer is different, and output layer nodes are
2.When convolutional neural networks are used to classify, output layer is classification layer, and the function of classification is sigmoid functions.According to design
9 layers of good convolutional neural networks, the pre-training of parameter is carried out with error backpropagation algorithm using the forward-propagating of standard.
C2. after pre-training terminates, the classification layer of convolutional neural networks is replaced to learning algorithm with dictionary, then carries out parameter
Optimization.The optimization of parameter includes the renewal of dictionary pair and the fine setting of network parameter, specifically includes following steps：
C21. the sigmoid functions of convolutional neural networks classification layer are replaced to learning algorithm with dictionary.Assuming that XkRepresent instruction
Practice the output of the full articulamentum of stage kth class training data correspondence convolutional neural networks, C represents target letter of the dictionary to learning algorithm
Number, CkRepresent the corresponding dictionary of kth class training data to the object function of learning algorithm, { Pk,DkRepresent kth class training data
Corresponding structuring dictionary pair, due in dictionary in learning algorithm, the renewal of dictionary pair can be broken down into following son and ask
Topic：
So in parameter optimisation step, orderThen C is to XkPartial derivative be：
Because our purpose is that Lung neoplasm is divided into two classes, so in above formula, k=1,2, K=2.
C22. obtaining allAfterwards, as the gradient in the back-propagation algorithm of standard to convolutional Neural net
The parameter of network is finely adjusted, while using the feature of the Network Capture training dataset after fine setting to dictionary to being updated.
2) sorting phase, the classification of data to be sorted.Sorting phase mainly has following three step：
A. reference picture 1, inputs CT images to be sorted；
B. reference picture 2, are pre-processed to CT images to be sorted, obtain the pulmonary parenchyma figure that unified size is 128 × 128
Picture.Specific steps include：
1st step, binary conversion treatment is carried out to original CT image using Otsu threshold methods；
2nd step, to the result of the 1st step, region is carried out by seed point of first, upper left corner pixel and is increased, by background portion
Divide and separated with foreground part；
3rd step, to the result of the 2nd step, lathe part is first obtained using region growing algorithm；
4th step, the result of preceding 3 step is overlapped, pulmonary parenchyma part is obtained；
5th step, based on the result of the 4th step, region growing algorithm is respectively adopted to pulmo and obtains complete pulmo, so
After merge；
6th step, the result to the 5th step performs morphological operation, obtains pulmonary parenchyma mask；
7th step, the pulmonary parenchyma in original CT image is partitioned into according to pulmonary parenchyma mask；
8th step, 128 × 128 are adjusted to by the obtained pulmonary parenchyma image size in the 7th step by bilinear interpolation.
C. reference picture 1, is divided CT images to be sorted using the disaggregated model CNN-DPC for training network parameter
Class.The feature that all training datas concentrate pulmonary parenchyma image is obtained by the convolutional neural networks after optimization, to dictionary to study
The object function of algorithm is solved, and can obtain 2 sub- dictionaries to { (D1,P1), (D2,P2)}.A test image is given, is extracted
Its feature, is designated as x, and it is E (x to make reconstruction errors of the x on the basis of kth (k=1,2) individual sub- dictionary pair；Dk,Pk), then disaggregated model
The rule that CNN-DPC is classified to test image is：
That is, if k-th sub- dictionary is minimum to the reconstruction error for reconstructing x, then the corresponding sample images of feature x just belong to
Kth class.
Table 2 gives the present invention and is carrying out lung to the convolutional neural networks before learning algorithm, replacement classification layer with dictionary
Tubercle relevant accuracy of classifying, susceptibility, results contrast of specificity when classifying.The present invention is in Lung neoplasm as can be seen from Table 2
The application aspect of classification better than dictionary to learning algorithm, and classification accuracy, two aspects of susceptibility than replace classification layer it
Preceding convolutional neural networks improve 3%~6%.
The corresponding accuracy of 2 three kinds of algorithms of table, susceptibility, specificity Comparative result
It is proposed by the present invention based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study, fully with reference to convolution
Neutral net and dictionary be directly used as process object to the advantage of learning algorithm using pulmonary parenchyma image, it is to avoid Lung neoplasm
The complex operations of segmentation.The feature for the sample data extracted using convolutional neural networks, can more reflect the bottom of Lung neoplasm
Abstract characteristics, such depth characteristic is conducive to diagnosis of the computer to Lung neoplasm, and utilizes dictionary can not only to learning algorithm
The classification of Lung neoplasm is enough realized, and the complexity that can fully excavate representated by the depth characteristic that convolutional neural networks are extracted contains
Justice, this is beneficial to the accuracy rate of diagnosis for improving follow-up Lung neoplasm.
It should be appreciated that for those of ordinary skills, can according to the above description be improved or converted,
And all these modifications and variations should all belong to the protection domain of appended claims of the present invention.
Claims (3)
1. it is a kind of based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study, it is characterised in that including convolution god
Training and Lung neoplasm classification through network parameter are two stages of training stage and sorting phase：
1) training stage:
A. the low dosage thin layer scanning CT images as training sample are inputted；
B. the CT images of training sample are pre-processed；
C. train classification models CNN-DPC parameter；Comprise the following steps：
C1. the parameter to convolutional neural networks carries out pre-training；In the pre-training stage, classical convolutional neural networks LeNet- is used for reference
5 network structure, based on training dataset, the convolutional neural networks structure of one 9 layers of design, the convolutional neural networks bag
Input layer, three parts of hidden layer and output layer are included, wherein hidden layer includes three convolutional layers, three maximum pond layers and one
Full articulamentum；Input layer number is 128 × 128, and each node in hidden layer is different, and output layer nodes are 2, output
Layer is classification layer, and the function of classification is sigmoid functions, and the activation primitive of hidden layer is ReLU (Rectified
Linear Units)；Forward-propagating and ginseng of the error backpropagation algorithm to this 9 layers of convolutional neural networks using standard
Number carries out pre-training；
C2. after pre-training terminates, the classification layer of convolutional neural networks is replaced to learning algorithm with dictionary, then carries out the excellent of parameter
Change；The optimization of parameter includes the renewal of dictionary pair and the fine setting of network parameter, specifically includes following steps：
C21. the sigmoid functions of convolutional neural networks classification layer are replaced to learning algorithm with dictionary；Assuming that XkRepresent the training stage
The output of the full articulamentum of kth class training data correspondence convolutional neural networks, C represents object function of the dictionary to learning algorithm, Ck
Represent the corresponding dictionary of kth class training data to the object function of learning algorithm, { Pk,DkRepresent that kth class training data is corresponding
Structuring dictionary pair, due in dictionary in learning algorithm, the renewal of dictionary pair is broken down into following subproblem：
<mrow>
<mtable>
<mtr>
<mtd>
<mrow>
<mi>arg</mi>
<munder>
<mi>min</mi>
<mrow>
<msub>
<mi>P</mi>
<msup>
<mi>k</mi>
<mo>,</mo>
</msup>
</msub>
<msub>
<mi>D</mi>
<mi>k</mi>
</msub>
</mrow>
</munder>
<msub>
<mi>C</mi>
<mi>k</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>P</mi>
<mi>k</mi>
</msub>
<mo>,</mo>
<msub>
<mi>D</mi>
<mi>k</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
<mtd>
<mrow></mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mo>=</mo>
<mi>arg</mi>
<munder>
<mi>min</mi>
<mrow>
<msub>
<mi>P</mi>
<msup>
<mi>k</mi>
<mo>,</mo>
</msup>
</msub>
<msub>
<mi>D</mi>
<mi>k</mi>
</msub>
</mrow>
</munder>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>X</mi>
<mi>k</mi>
</msub>
<mo>-</mo>
<msub>
<mi>D</mi>
<mi>k</mi>
</msub>
<msub>
<mi>P</mi>
<mi>k</mi>
</msub>
<msub>
<mi>X</mi>
<mi>k</mi>
</msub>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>F</mi>
<mn>2</mn>
</msubsup>
<mo>+</mo>
<mi>&lambda;</mi>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>P</mi>
<mi>k</mi>
</msub>
<mover>
<msub>
<mi>X</mi>
<mi>k</mi>
</msub>
<mo>&OverBar;</mo>
</mover>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>F</mi>
<mn>2</mn>
</msubsup>
</mrow>
</mtd>
<mtd>
<mrow>
<mi>s</mi>
<mo>.</mo>
<mi>t</mi>
<mo>.</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>d</mi>
<mi>i</mi>
</msub>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mn>2</mn>
<mn>2</mn>
</msubsup>
<mo>&le;</mo>
<mn>1</mn>
</mrow>
</mtd>
</mtr>
</mtable>
<mo>;</mo>
</mrow>
So in parameter optimisation step, orderThen C is to XkPartial derivative be：
<mrow>
<mfrac>
<mrow>
<mo>&part;</mo>
<mi>C</mi>
</mrow>
<mrow>
<mo>&part;</mo>
<msub>
<mi>X</mi>
<mi>k</mi>
</msub>
</mrow>
</mfrac>
<mo>=</mo>
<mn>2</mn>
<mrow>
<mo>(</mo>
<mi>I</mi>
<mo>-</mo>
<msubsup>
<mi>P</mi>
<mi>k</mi>
<mi>T</mi>
</msubsup>
<msubsup>
<mi>D</mi>
<mi>k</mi>
<mi>T</mi>
</msubsup>
<mo>)</mo>
</mrow>
<mrow>
<mo>(</mo>
<msub>
<mi>X</mi>
<mi>k</mi>
</msub>
<mo>-</mo>
<msub>
<mi>D</mi>
<mi>k</mi>
</msub>
<msub>
<mi>P</mi>
<mi>k</mi>
</msub>
<msub>
<mi>X</mi>
<mi>k</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>+</mo>
<msub>
<mi>&Sigma;</mi>
<mrow>
<msup>
<mi>k</mi>
<mo>&prime;</mo>
</msup>
<mo>&NotEqual;</mo>
<mi>k</mi>
</mrow>
</msub>
<mn>2</mn>
<msubsup>
<mi>&lambda;P</mi>
<msup>
<mi>k</mi>
<mo>&prime;</mo>
</msup>
<mi>T</mi>
</msubsup>
<msub>
<mi>p</mi>
<msup>
<mi>k</mi>
<mo>&prime;</mo>
</msup>
</msub>
<msub>
<mi>X</mi>
<mi>k</mi>
</msub>
</mrow>
Because our purpose is that Lung neoplasm is divided into two classes, so in above formula, k=1,2, K=2；
C22. obtaining allAfterwards, as ginseng of the gradient in the back-propagation algorithm of standard to convolutional neural networks
Number is finely adjusted, while using the feature of the Network Capture training dataset after fine setting to dictionary to being updated；
2) sorting phase, the classification of data to be sorted；Sorting phase includes following three step：
A. low dosage thin layer scanning CT images to be sorted are inputted；
B. CT images to be sorted are pre-processed；
C. image to be classified is classified using the disaggregated model CNN-DPC for training network parameter；Pass through the volume after optimization
Product neutral net obtains the feature that all training datas concentrate pulmonary parenchyma image, and dictionary is asked the object function of learning algorithm
Solution, can obtain 2 sub- dictionaries to { (D1,P1), (D2,P2)}；A test image is given, its feature is extracted, is designated as x, make x
Reconstruction error on the basis of the individual sub- dictionary pair of kth (k=1,2) is E (x；Dk,Pk), then disaggregated model CNN-DPC is to test chart
As the rule classified is：
<mrow>
<mi>E</mi>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>;</mo>
<msub>
<mi>D</mi>
<mi>k</mi>
</msub>
<mo>,</mo>
<msub>
<mi>P</mi>
<mi>k</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>|</mo>
<mo>|</mo>
<mi>x</mi>
<mo>-</mo>
<msub>
<mi>D</mi>
<mi>k</mi>
</msub>
<msub>
<mi>P</mi>
<mi>k</mi>
</msub>
<mi>x</mi>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>F</mi>
<mn>2</mn>
</msubsup>
</mrow>
1
<mrow>
<mi>l</mi>
<mi>a</mi>
<mi>b</mi>
<mi>e</mi>
<mi>l</mi>
<mo>=</mo>
<mi>arg</mi>
<munder>
<mrow>
<mi>m</mi>
<mi>i</mi>
<mi>n</mi>
</mrow>
<mi>k</mi>
</munder>
<mi>E</mi>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>;</mo>
<msub>
<mi>D</mi>
<mi>k</mi>
</msub>
<mo>,</mo>
<msub>
<mi>P</mi>
<mi>k</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
That is, if k-th sub- dictionary is minimum to the reconstruction error for reconstructing x, then the corresponding sample images of feature x just belong to kth
Class.
2. as claimed in claim 1 based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study, its feature exists
In in step B, specifically including following steps：
1st step, binary conversion treatment is carried out to original CT image using Otsu threshold methods；
2nd step, to the result of the 1st step, by seed point of first, upper left corner pixel carry out region increase, by background parts with
Foreground part is separated；
3rd step, to the result of the 2nd step, lathe part is first obtained using region growing algorithm；
4th step, the result of preceding 3 step is overlapped, pulmonary parenchyma part is obtained；
5th step, based on the result of the 4th step, region growing algorithm is respectively adopted to pulmo and obtains complete pulmo, Ran Houjin
Row merges；
6th step, the result to the 5th step performs morphological operation, obtains pulmonary parenchyma mask；
7th step, the pulmonary parenchyma in original CT image is partitioned into according to pulmonary parenchyma mask；
8th step, 128 × 128 are adjusted to by the obtained pulmonary parenchyma image size in the 7th step by bilinear interpolation.
3. as claimed in claim 1 based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study, its feature exists
In sorting phase is pre-processed to CT images to be sorted, specifically includes following steps：
1st step, binary conversion treatment is carried out to original CT image using Otsu threshold methods；
2nd step, to the result of the 1st step, by seed point of first, upper left corner pixel carry out region increase, by background parts with
Foreground part is separated；
3rd step, to the result of the 2nd step, lathe part is first obtained using region growing algorithm；
4th step, the result of preceding 3 step is overlapped, pulmonary parenchyma part is obtained；
5th step, based on the result of the 4th step, region growing algorithm is respectively adopted to pulmo and obtains complete pulmo, Ran Houjin
Row merges；
6th step, the result to the 5th step performs morphological operation, obtains pulmonary parenchyma mask；
7th step, the pulmonary parenchyma in original CT image is partitioned into according to pulmonary parenchyma mask；
8th step, 128 × 128 are adjusted to by the obtained pulmonary parenchyma image size in the 7th step by bilinear interpolation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710280338.0A CN107103334A (en) | 2017-04-26 | 2017-04-26 | It is a kind of based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710280338.0A CN107103334A (en) | 2017-04-26 | 2017-04-26 | It is a kind of based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107103334A true CN107103334A (en) | 2017-08-29 |
Family
ID=59656989
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710280338.0A Pending CN107103334A (en) | 2017-04-26 | 2017-04-26 | It is a kind of based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107103334A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108078581A (en) * | 2017-12-12 | 2018-05-29 | 北京青燕祥云科技有限公司 | The good pernicious method of discrimination system of lung cancer and realization device based on convolutional neural networks |
| CN108198179A (en) * | 2018-01-03 | 2018-06-22 | 华南理工大学 | A kind of CT medical image pulmonary nodule detection methods for generating confrontation network improvement |
| CN109154989A (en) * | 2018-03-02 | 2019-01-04 | 香港应用科技研究院有限公司 | The classification performance that convolutional neural networks apply cancer cell screening is improved using mask |
| CN109544510A (en) * | 2018-10-24 | 2019-03-29 | 广州大学 | A kind of three-dimensional Lung neoplasm recognition methods based on convolutional neural networks |
| CN110223743A (en) * | 2019-06-14 | 2019-09-10 | 吉林大学第一医院 | A kind of structuring processing method and system for pulmonary cancer diagnosis record |
| CN111247556A (en) * | 2017-10-24 | 2020-06-05 | 爱克发医疗保健公司 | Training artificial neural networks on additional tasks while avoiding catastrophic interference |
| CN111340209A (en) * | 2020-02-18 | 2020-06-26 | 北京推想科技有限公司 | Network model training method, image segmentation method and focus positioning method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101763644A (en) * | 2010-03-10 | 2010-06-30 | 华中科技大学 | Pulmonary nodule three-dimensional segmentation and feature extraction method and system thereof |
| CN104268893A (en) * | 2014-10-16 | 2015-01-07 | 太原理工大学 | Method for segmenting and denoising lung parenchyma through lateral scanning and four-corner rotary scanning |
| CN104700118A (en) * | 2015-03-18 | 2015-06-10 | 中国科学院自动化研究所 | Pulmonary nodule benignity and malignancy predicting method based on convolutional neural networks |
| CN105160361A (en) * | 2015-09-30 | 2015-12-16 | 东软集团股份有限公司 | Image identification method and apparatus |
| CN106372390A (en) * | 2016-08-25 | 2017-02-01 | 姹ゅ钩 | Deep convolutional neural network-based lung cancer preventing self-service health cloud service system |
-
2017
- 2017-04-26 CN CN201710280338.0A patent/CN107103334A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101763644A (en) * | 2010-03-10 | 2010-06-30 | 华中科技大学 | Pulmonary nodule three-dimensional segmentation and feature extraction method and system thereof |
| CN104268893A (en) * | 2014-10-16 | 2015-01-07 | 太原理工大学 | Method for segmenting and denoising lung parenchyma through lateral scanning and four-corner rotary scanning |
| CN104700118A (en) * | 2015-03-18 | 2015-06-10 | 中国科学院自动化研究所 | Pulmonary nodule benignity and malignancy predicting method based on convolutional neural networks |
| CN105160361A (en) * | 2015-09-30 | 2015-12-16 | 东软集团股份有限公司 | Image identification method and apparatus |
| CN106372390A (en) * | 2016-08-25 | 2017-02-01 | 姹ゅ钩 | Deep convolutional neural network-based lung cancer preventing self-service health cloud service system |
Non-Patent Citations (4)
| Title |
|---|
| KEZE WANG ET AL.: "Dictionary Pair Classifier Driven Convolutional Neural Networks for Object Detection", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| XIAOJIAO XIAO ET AL.: "A Deep Learning Model of Automatic Detection of Pulmonary Nodules Based on Convolution Neural Networks (CNNs)", 《INTERNATIONAL CONFERENCE ON BIO-INSPIRED COMPUTING: THEORIES AND APPLICATIONS》 * |
| 纪国华: "基于PET-CT的孤立性肺结节分割及分类方法的研究", 《中国优秀硕士学位论文全文数据库 医药卫生科技辑》 * |
| 金奇樑: "基于CT图像的肺结节自动识别系统研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111247556A (en) * | 2017-10-24 | 2020-06-05 | 爱克发医疗保健公司 | Training artificial neural networks on additional tasks while avoiding catastrophic interference |
| CN108078581A (en) * | 2017-12-12 | 2018-05-29 | 北京青燕祥云科技有限公司 | The good pernicious method of discrimination system of lung cancer and realization device based on convolutional neural networks |
| CN108078581B (en) * | 2017-12-12 | 2019-02-12 | 北京青燕祥云科技有限公司 | The good pernicious judgement system of lung cancer and realization device based on convolutional neural networks |
| CN108198179A (en) * | 2018-01-03 | 2018-06-22 | 华南理工大学 | A kind of CT medical image pulmonary nodule detection methods for generating confrontation network improvement |
| CN109154989A (en) * | 2018-03-02 | 2019-01-04 | 香港应用科技研究院有限公司 | The classification performance that convolutional neural networks apply cancer cell screening is improved using mask |
| CN109154989B (en) * | 2018-03-02 | 2021-07-06 | 香港应用科技研究院有限公司 | Using masks to improve classification performance of convolutional neural networks for cancer cell screening applications |
| CN109544510A (en) * | 2018-10-24 | 2019-03-29 | 广州大学 | A kind of three-dimensional Lung neoplasm recognition methods based on convolutional neural networks |
| CN109544510B (en) * | 2018-10-24 | 2021-10-26 | 广州大学 | Three-dimensional lung nodule identification method based on convolutional neural network |
| CN110223743A (en) * | 2019-06-14 | 2019-09-10 | 吉林大学第一医院 | A kind of structuring processing method and system for pulmonary cancer diagnosis record |
| CN111340209A (en) * | 2020-02-18 | 2020-06-26 | 北京推想科技有限公司 | Network model training method, image segmentation method and focus positioning method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107103334A (en) | It is a kind of based on the Lung neoplasm sorting technique of convolutional neural networks and dictionary to study | |
| Wu et al. | Vessel-Net: Retinal vessel segmentation under multi-path supervision | |
| CN106408562B (en) | Eye fundus image Segmentation Method of Retinal Blood Vessels and system based on deep learning | |
| CN107292915A (en) | Method for tracking target based on convolutional neural networks | |
| CN109034210A (en) | Object detection method based on super Fusion Features Yu multi-Scale Pyramid network | |
| CN107506761A (en) | Brain image dividing method and system based on notable inquiry learning convolutional neural networks | |
| CN110210555A (en) | Rail fish scale hurt detection method based on deep learning | |
| CN107169974A (en) | It is a kind of based on the image partition method for supervising full convolutional neural networks more | |
| CN107169956A (en) | Yarn dyed fabric defect detection method based on convolutional neural networks | |
| CN108648191A (en) | Pest image-recognizing method based on Bayes's width residual error neural network | |
| CN107220506A (en) | Breast cancer risk assessment analysis system based on deep convolutional neural network | |
| CN107437092A (en) | The sorting algorithm of retina OCT image based on Three dimensional convolution neutral net | |
| CN107590489A (en) | Object detection method based on concatenated convolutional neutral net | |
| CN106372648A (en) | Multi-feature-fusion-convolutional-neural-network-based plankton image classification method | |
| CN108108699A (en) | Merge deep neural network model and the human motion recognition method of binary system Hash | |
| CN108154519A (en) | Dividing method, device and the storage medium of eye fundus image medium vessels | |
| CN107358262A (en) | The sorting technique and sorter of a kind of high-definition picture | |
| CN107316294A (en) | One kind is based on improved depth Boltzmann machine Lung neoplasm feature extraction and good pernicious sorting technique | |
| CN106445919A (en) | Sentiment classifying method and device | |
| CN110084803A (en) | Eye fundus image method for evaluating quality based on human visual system | |
| CN110276402A (en) | A kind of salt body recognition methods based on the enhancing of deep learning semanteme boundary | |
| CN107944369A (en) | A kind of pedestrian detection method based on tandem zones generation network and enhancing random forest | |
| CN106408001A (en) | Rapid area-of-interest detection method based on depth kernelized hashing | |
| CN106991386A (en) | A kind of gesture identification method based on depth residual error network | |
| CN108038844A (en) | A kind of good pernicious Forecasting Methodology of Lung neoplasm based on legerity type CNN |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170829 |