CN105023023A - Mammary gland type-B ultrasonic image feature self-learning extraction method used for computer-aided diagnosis - Google Patents

Abstract

The invention relates to a mammary gland type-B ultrasonic image feature self-learning extraction method used for computer-aided diagnosis. First a Convolutional Restricted Boltzmann Machine (CRBM) is obtained through unsupervised training based on mammary gland type-B ultrasonic focus region image set of medium or larger scale, and any one given type-B ultrasonic focus region image is input in the trained CRBM first and initial features of a mammary gland type-B ultrasonic image is extracted through the CRBM; and afterwards, principal component analysis (PCA) is used to perform dimensionality reduction on the initial features, thereby obtaining low-dimensional mammary gland type-B ultrasonic image features capable of being used for computer-aided diagnosis, and completing self-learning extraction of mammary gland type-B ultrasonic image superficial layer features. The mammary gland type-B ultrasonic image feature self-learning extraction method used for computer-aided diagnosis adopts a completely unsupervised form, performs self-learning of features from existing mammary gland type-B ultrasonic image data, thereby reducing the workload, avoiding artificial interference, and the method is flexible to realize and has relatively strong practicability.

Description

A kind of breast sonography characteristics of image self study extracting method for computer-aided diagnosis

Technical field

The present invention relates to technical field of medical image processing, particularly a kind of breast sonography characteristics of image self study extracting method for computer-aided diagnosis.

Background technology

Breast cancer occurs in one of malignant tumour the most general in women colony.In recent years, the investigation display of Fan Ai association of China, the occurrence rate of breast cancer is in cumulative year after year.Therefore the early diagnosis precision improving breast cancer becomes more and more meaningful.

At present, the main method that breast cancer diagnosis adopts is that diagnosis person is analyzed the state of an illness by the image feature such as calcification or lump by image checks such as mammary gland x-ray, B ultrasonic images.But due to the density of the soft tissues such as the body of gland in breast tissue, blood vessel, fat and the density of focal zone all very close, add the factors such as diagnosis person's visual fatigue, make the mistaken diagnosis of early-stage breast cancer and fail to pinpoint a disease in diagnosis still often to occur.Along with the development of Medical Imaging Technology and computer aided technique, utilizing computing machine to carry out auxiliary diagnosis becomes possibility; Such as utilize digital image processing techniques, extract the feature that in breast sonography image, pathology is relevant, use the machine learning methods such as SVM to carry out Classification and Identification etc. according to these features to Diagnosis of Breast tumors.

From the application present situation of computer-aided diagnosis breast cancer, the accuracy of computer-aided diagnosis largely depends on that whether extract B ultrasonic image pathology correlated characteristic effective.At present, medical image features for computer-aided diagnosis extracts and substantially adopts manual location focus area-of-interest, and the general characteristics on some bases of being extracted by the method for primary image process, as: grey level histogram feature, shape facility, gray level co-occurrence matrixes feature, wavelet character etc.But said method has the deficiency of the following aspects: the first, the time and effort consuming of extraction one by one of above-mentioned basic general characteristics; The second, above-mentioned single basic general characteristics itself is not correlated with in field, and the application-specific degree of association of breast cancer is little; Three, design the basic general characteristics combination that effectively can be used for computer-aided diagnosis breast cancer and there is serious uncertainty.Therefore, best settlement mechanism be to provide a kind of can according to breast cancer B ultrasonic image automatic learning in the past go out with pathology about and can be used for the method for the characteristics of image of auxiliary diagnosis.

In " Convolutional Deep BeliefNetworks for Scalable Unsupervised Learning of HierarchicalRepresentations " article that the people such as Honglak Lee deliver, author utilize convolution to deeply convince net carries out feature learning to facial image, by learning the feature that obtains, face is identified.This experimental result shows that convolution deeply convinces that net can be used for the self study of feature.

Thus, the feature self study that this patent proposes to utilize the limited Boltzmann machine of convolution (CRBM) to complete breast sonography image is extracted.

Summary of the invention

In view of this, the object of the invention is to propose a kind of breast sonography characteristics of image self study extracting method for computer-aided diagnosis, reduce workload, avoid artificial interference, and contribute to the Correlation with Pathology of feature.

The following scheme of employing of the present invention realizes: a kind of breast sonography characteristics of image self study extracting method for computer-aided diagnosis, specifically comprises the following steps:

Step S1: a given medium-scale above breast sonography focal area image set, described this image set of medium-scale expression is at least containing breast sonography diagnostic images more than 200 width;

Step S2: in manual extraction step S1, image set is from the breast sonography focal area image ROI of each breast sonography diagnostic image, and using the sample of whole breast sonography focal area image ROI as training set; The size of wherein said breast sonography focal area image ROI is 150 × 150;

Step S3: training training set being used for the limited Boltzmann machine CRBM of convolution extracted towards initial characteristics;

Step S4: the limited Boltzmann machine CRBM of convolution adopting CD Algorithm for Training to extract towards initial characteristics, obtain the concrete setting of the convolution limited Boltzmann machine CRBM extracted towards initial characteristics;

Step S5: a given width breast sonography focal area image, it can be used as the input of the limited Boltzmann machine CRBM of convolution extracted towards initial characteristics, and obtain from the output layer output extraction of the limited Boltzmann machine CRBM of described convolution the initial characteristics that dimension is 117600 dimensions;

Step S6: utilize principal component analysis (PCA) PCA to carry out dimension-reduction treatment to the initial characteristics that the dimension that step S5 obtains is 117600 dimensions, obtain the breast sonography characteristics of image of low-dimensional, wherein low-dimensional represents that dimension is 100 dimensions, and the breast sonography characteristics of image of this 100 dimension can be applicable in subsequent calculations machine auxiliary diagnosis.

Further, the limited Boltzmann machine CRBM of convolution in described step S3 comprises input layer, hidden layer and pooling layer, wherein input layer is breast sonography focal area image, output layer is pooling layer, and the probable value of each unit of pooling layer represents the initial shallow feature of breast sonography focal area image.

Further, as follows towards specifically arranging of the convolution limited Boltzmann machine CRBM of initial characteristics extraction described in step S4: input layer size is set to 150 × 150, i.e. N _vget 150, input layer is biased c and is set to 0; Wave filter size is set to 10 × 10, i.e. N _wget 10, bank of filters number gets 24, and namely K is set to 24, convolutional layer characteristic pattern be split into multiple non-overlapping 2 × 2 wait converge B _αfritter, converges ratio C and gets 2;

Given input layer v, then convolutional layer, namely the conditional probability of Hidden unit is:

$P({h^k}_{i,j}=1|v)=\frac{\exp \left(I\right({h^k}_{i,j}))}{1+{\Σ}_{(i^{\′},j^{\′})\∈B_{\mathrm{\α}}}\exp \left(I\right({h^k}_{i^{\′}{j}^{\′}}\left)\right)};$

Given input layer v, then the conditional probability of pooling layer unit is:

$P(p_{\mathrm{\α}}^k=0|v)=\frac{1}{1+{\mathrm{\Σ}}_{(i^{\′},j^{\′})\∈B_{\mathrm{\α}}}\exp \left(I\right({h^k}_{i^{\′},j^{\′}}\left)\right)};$

Given convolutional layer, i.e. hidden layer h, then the conditional probability of input layer unit is:

$P(v_{i,j}=1|h)=\mathrm{\σ}({(\underset{k}{\Σ}{W}^k*h^k)}_{i,j}+c),$ Input layer is biased c and gets 0;

Wherein represent the signal being transferred to convolutional layer by visible layer, b ^krepresent that the convolutional layer characteristic pattern that a kth wave filter is corresponding is biased, on this characteristic pattern, all unit share same biased b ^k, W ^krepresent the matrix of coefficients (convolution kernel) of a kth wave filter, represent W ^kmatrix portraitlandscape upset simultaneously, B _αrepresent the fritter to be converged of 2 × 2, i, j are at B _αmiddle value, represents row and column index, 1≤i, j≤2, h ^k _i,jany B on a kth characteristic pattern in expression convolutional layer _αthe unit that in fritter, the i-th row j arranges, 1≤i, j≤2, * represent convolution operation, and σ represents sigmoid function.

Preferably, described CD Algorithm for Training process is specific as follows:

S41: for any given width B ultrasonic focal area image (ROI), size is 150 × 150, first image is converted into gray level image, then image array is divided by 255, be converted into [0,1] by image matrix data scope, suppose that the image after processing is v;

S42: according to conditional probability distribution $P({h^k}_{i,j}=1|v)=\frac{\exp \left(I\right({h^k}_{i,j}))}{1+{\mathrm{\Σ}}_{(i^{\′},j^{\′})\∈B_{\mathrm{\α}}}\exp \left(I\right({h^k}_{i^{\′},j^{\′}}\left)\right)}$ Hidden layer (convolutional layer) kth characteristic pattern h is obtained by v ^k, 1≤k≤24;

S43: each probability unit of a kth characteristic pattern is converted into binary mode, obtains state_h ^k, specific as follows: for any B _αfritter, is preferentially stored as the vector v ec of (Isosorbide-5-Nitrae), an any given random number rnd=rand (1) by row, if then in vec, first element is 1, and its excess-three element is 0; If and (wherein 2≤s≤4), then in vec, s element is 1, and all the other elements are 0; Then remap back vec vector 2 × 2 matrixes, to B _αassignment;

S44: according to conditional probability distribution $P(neg\_v_{i,j}=1|h)=\mathrm{\σ}({(\underset{k}{\Σ}{W}^k*state\_h^k)}_{i,j}+c)$ (c gets 0) obtains the visible layer image neg_v after sampling by scale-of-two convolutional layer;

S45: the operation in similar step S42, can obtain a kth characteristic pattern neg_h in the convolutional layer after sampling ^k, 1≤k≤24;

S46: compute gradient:

${\mathrm{dW}}^k=v*\widetilde{h^k}-neg\_v*\widetilde{neg\_h^k;}$

${\mathrm{db}}^k=\frac{\underset{s=1}{\overset{\mathrm{si}ze(h^k,2)}{\Σ}}\underset{t=1}{\overset{size(h^k,1)}{\Σ}}{h^k}_{ts}-\underset{s=1}{\overset{size(h^k,2)}{\Σ}}\underset{t=1}{\overset{\mathrm{si}ze(h^k,1)}{\Σ}}neg\_{h^k}_{ts}}{size(h^k,1)\×size(h^k,2)};$

S47: upgrade filter coefficient and be biased:

W ^k＝W ^k+α×dW ^k；

b ^k＝b ^k+α×db ^k；

Wherein α is learning rate, is set to 0.01, dW in the present invention ^krepresent the gradient matrix of a kth wave filter, size is 10 × 10, db ^krepresenting the biased of a kth wave filter, is a scalar; Size (h ^k, 1) and representation feature figure h ^kline number, size (h ^k, 2) and representation feature figure h ^kcolumns, * represents convolution operation, × represent common scalar multiplication.H ^k _tsrepresentation feature figure h ^kin the element of t capable jth row, neg_h ^k _tssimilar.

Compared with prior art, the present invention adopts complete unsupervised form, from existing breast sonography view data, go self study feature, reduces workload, avoids artificial interference, and the method realizes flexibly, having stronger practicality.

Accompanying drawing explanation

Fig. 1 is the limited Boltzmann machine of convolution (CRBM) structural representation towards the feature extraction of breast sonography image initial in the present invention.

Embodiment

Below in conjunction with drawings and Examples, the present invention will be further described.

As shown in Figure 1, present embodiments provide a kind of breast sonography characteristics of image self study extracting method for computer-aided diagnosis, specifically comprise the following steps:

Step S1: a given medium-scale above breast sonography focal area image set, described this image set of medium-scale expression is at least containing breast sonography diagnostic images more than 200 width;

Step S2: in manual extraction step S1, image set is from the breast sonography focal area image ROI of each breast sonography diagnostic image, and using the sample of whole breast sonography focal area image ROI as training set; The size of wherein said breast sonography focal area image ROI is 150 × 150;

Step S3: training training set being used for the limited Boltzmann machine CRBM of convolution extracted towards initial characteristics;

Step S4: the limited Boltzmann machine CRBM of convolution adopting CD Algorithm for Training to extract towards initial characteristics, obtain the concrete setting of the convolution limited Boltzmann machine CRBM extracted towards initial characteristics;

Step S5: a given width breast sonography focal area image, it can be used as the input of the limited Boltzmann machine CRBM of convolution extracted towards initial characteristics, and obtain from the output layer output extraction of the limited Boltzmann machine CRBM of described convolution the initial characteristics that dimension is 117600 dimensions;

Step S6: utilize principal component analysis (PCA) PCA to carry out dimension-reduction treatment to the initial characteristics that the dimension that step S5 obtains is 117600 dimensions, obtain the breast sonography characteristics of image of low-dimensional, wherein low-dimensional represents that dimension is 100 dimensions, and the breast sonography characteristics of image of this 100 dimension can be applicable in subsequent calculations machine auxiliary diagnosis.

In the present embodiment, the limited Boltzmann machine CRBM of convolution in described step S3 comprises input layer, hidden layer and pooling layer, wherein input layer is breast sonography focal area image, output layer is pooling layer, and the probable value of each unit of pooling layer represents the initial shallow feature of breast sonography focal area image.

In the present embodiment, as follows towards specifically arranging of the convolution limited Boltzmann machine CRBM of initial characteristics extraction described in step S4: input layer size is set to 150 × 150, i.e. N _vget 150, input layer is biased c and is set to 0; Wave filter size is set to 10 × 10, i.e. N _wget 10, bank of filters number gets 24, and namely K is set to 24, convolutional layer characteristic pattern be split into multiple non-overlapping 2 × 2 wait converge B _αfritter, converges ratio C and gets 2;

Given input layer v, then convolutional layer, namely the conditional probability of Hidden unit is:

$P({h^k}_{i,j}=1|v)=\frac{\exp \left(I\right({h^k}_{i,j}))}{1+{\mathrm{\Σ}}_{(i^{\′},j^{\′})\∈B_{\mathrm{\α}}}\exp \left(I\right({h^k}_{i^{\′},j^{\′}}\left)\right)};$

Given input layer v, then the conditional probability of pooling layer unit is:

$P(p_{\mathrm{\α}}^k=0|v)=\frac{1}{1+{\mathrm{\Σ}}_{(i^{\′},j^{\′})\∈B_{\mathrm{\α}}}epx\left(I\right({h^k}_{i^{\′},j^{\′}}\left)\right)};$

Given convolutional layer, i.e. hidden layer h, then the conditional probability of input layer unit is:

$P(v_{i,j}=1|h)=\mathrm{\σ}({(\underset{k}{\Σ}{W}^k*h^k)}_{i,j}+c),$ Input layer is biased c and gets 0;

Wherein represent the signal being transferred to convolutional layer by visible layer, b ^krepresent that the convolutional layer characteristic pattern that a kth wave filter is corresponding is biased, on this characteristic pattern, all unit share same biased b ^k, W ^krepresent the matrix of coefficients (convolution kernel) of a kth wave filter, represent W ^kmatrix portraitlandscape upset simultaneously, B _αrepresent the fritter to be converged of 2 × 2, i, j are at B _αmiddle value, represents row and column index, 1≤i, j≤2, h ^k _i,jany B on a kth characteristic pattern in expression convolutional layer _αthe unit that in fritter, the i-th row j arranges, 1≤i, j≤2, * represent convolution operation, and σ represents sigmoid function.

Preferably, in the present embodiment, described CD Algorithm for Training process is specific as follows:

S41: for any given width B ultrasonic focal area image (ROI), size is 150 × 150, first image is converted into gray level image, then image array is divided by 255, be converted into [0,1] by image matrix data scope, suppose that the image after processing is v;

S42: according to conditional probability distribution $P({h^k}_{i,j}=1|v)=\frac{\exp \left(I\right({h^k}_{i,j}))}{1+{\mathrm{\Σ}}_{(i^{\′},j^{\′})\∈B_{\mathrm{\α}}}\exp \left(I\right({h^k}_{i^{\′},j^{\′}}\left)\right)}$ Hidden layer (convolutional layer) kth characteristic pattern h is obtained by v ^k, 1≤k≤24;

S43: each probability unit of a kth characteristic pattern is converted into binary mode, obtains state_h ^k, specific as follows: for any B _αfritter, is preferentially stored as the vector v ec of (Isosorbide-5-Nitrae), an any given random number rnd=rand (1) by row, if then in vec, first element is 1, and its excess-three element is 0; If and (wherein 2≤s≤4), then in vec, s element is 1, and all the other elements are 0; Then remap back vec vector 2 × 2 matrixes, to B _αassignment;

S44: according to conditional probability distribution $P(neg\_v_{i,j}=1|h)=\mathrm{\σ}({(\underset{k}{\Σ}{W}^k*state\_h^k)}_{i,j}+c)$ (c gets 0) obtains the visible layer image neg_v after sampling by scale-of-two convolutional layer;

S45: the operation in similar step S42, can obtain a kth characteristic pattern neg_h in the convolutional layer after sampling ^k, 1≤k≤24;

S46: compute gradient:

${\mathrm{dW}}^k=v*\widetilde{h^k}-neg\_v*\widetilde{neg\_h^k;}$

${\mathrm{db}}^k=\frac{\underset{s=1}{\overset{size(h^k,2)}{\Σ}}\underset{t=1}{\overset{size(h^k,1)}{\Σ}}{h^k}_{ts}-\underset{s=1}{\overset{size(h^k,2)}{\Σ}}\underset{t=1}{\overset{size(h^k,1)}{\Σ}}neg\_{h^k}_{ts}}{size(h^k,1)\×size(h^k,2)};$

S47: upgrade filter coefficient and be biased:

W ^k＝W ^k+α×dW ^k；

b ^k＝b ^k+α×db ^k；

Wherein α is learning rate, is set to 0.01, dW in the present invention ^krepresent the gradient matrix of a kth wave filter, size is 10 × 10, db ^krepresenting the biased of a kth wave filter, is a scalar; Size (h ^k, 1) and representation feature figure h ^kline number, size (h ^k, 2) and representation feature figure h ^kcolumns, * represents convolution operation, × represent common scalar multiplication.H ^k _tsrepresentation feature figure h ^kin the element of t capable jth row, neg_h ^k _tssimilar.

The foregoing is only preferred embodiment of the present invention, the feature self study extracting method that this patent proposes is not limited to breast sonography image itself, also in easily extensible to other medical image, the self study of pathology correlated characteristic is extracted, all equalizations done according to the present patent application the scope of the claims change and modify, and all should belong to covering scope of the present invention.

Claims (3)

1., for a breast sonography characteristics of image self study extracting method for computer-aided diagnosis, it is characterized in that comprising the following steps:

Step S1: a given medium-scale above breast sonography focal area image set, described this image set of medium-scale expression is at least containing breast sonography diagnostic images more than 200 width;

Step S2: in manual extraction step S1, image set is from the breast sonography focal area image ROI of each breast sonography diagnostic image, and using the sample of whole breast sonography focal area image ROI as training set; The size of wherein said breast sonography focal area image ROI is 150 × 150;

Step S3: training training set being used for the limited Boltzmann machine CRBM of convolution extracted towards initial characteristics;

Step S4: the limited Boltzmann machine CRBM of convolution adopting CD Algorithm for Training to extract towards initial characteristics, obtain the concrete setting of the convolution limited Boltzmann machine CRBM extracted towards initial characteristics;

Step S5: a given width breast sonography focal area image, it can be used as the input of the limited Boltzmann machine CRBM of convolution extracted towards initial characteristics, and obtain from the output layer output extraction of the limited Boltzmann machine CRBM of described convolution the initial characteristics that dimension is 117600 dimensions;

Step S6: utilize principal component analysis (PCA) PCA to carry out dimension-reduction treatment to the initial characteristics that the dimension that step S5 obtains is 117600 dimensions, obtain the breast sonography characteristics of image of low-dimensional, wherein low-dimensional represents that dimension is 100 dimensions, and the breast sonography characteristics of image of this 100 dimension can be applicable in subsequent calculations machine auxiliary diagnosis.

2. a kind of breast sonography characteristics of image self study extracting method for computer-aided diagnosis according to claim 1, it is characterized in that: the limited Boltzmann machine CRBM of the convolution in described step S3 comprises input layer, hidden layer and pooling layer, wherein input layer is breast sonography focal area image, output layer is pooling layer, and the probable value of each unit of pooling layer represents the initial shallow feature of breast sonography focal area image.

3. a kind of breast sonography characteristics of image self study extracting method for computer-aided diagnosis according to claim 1, it is characterized in that: as follows towards specifically arranging of the convolution limited Boltzmann machine CRBM of initial characteristics extraction described in step S4: input layer size is set to 150 × 150, i.e. N _vget 150, input layer is biased c and is set to 0; Wave filter size is set to 10 × 10, i.e. N _wget 10, bank of filters number gets 24, and namely K is set to 24, convolutional layer characteristic pattern be split into multiple non-overlapping 2 × 2 wait converge B _αfritter, converges ratio C and gets 2;

Given input layer v, then convolutional layer, namely the conditional probability of Hidden unit is:

$P({h^k}_{i,j}=1|v)=\frac{\exp \left(I\right({h^k}_{i,j}))}{1+{\mathrm{\Σ}}_{(i^{\′},j^{\′})\∈B_{\mathrm{\α}}}\exp \left(I\right({h^k}_{i^{\′},j^{\′}}))};$

Given input layer v, then the conditional probability of pooling layer unit is:

$P(p_{\mathrm{\α}}^k=0|v)=\frac{1}{1+{\mathrm{\Σ}}_{(i^{\′},j^{\′})\∈B_{\mathrm{\α}}}\exp \left(I\right({h^k}_{i^{\′},j^{\′}}))};$

Given convolutional layer, i.e. hidden layer h, then the conditional probability of input layer unit is:

$P(v_{i,j}=1|h)=\mathrm{\σ}({\left(\underset{k}{\Σ}{W}^k*h^k\right)}_{i,j}+c),$ Input layer is biased c and gets 0;

Wherein represent the signal being transferred to convolutional layer by visible layer, b ^krepresent that the convolutional layer characteristic pattern that a kth wave filter is corresponding is biased, on this characteristic pattern, all unit share same biased b ^k, W ^krepresent the matrix of coefficients of a kth wave filter, i.e. convolution kernel, represent W ^kmatrix portraitlandscape upset simultaneously, B _αrepresent the fritter to be converged of 2 × 2, i, j are at B _αmiddle value, represents row and column index, 1≤i, j≤2, h ^k _i,jany B on a kth characteristic pattern in expression convolutional layer _αthe unit that in fritter, the i-th row j arranges, 1≤i, j≤2, * represent convolution operation, and σ represents sigmoid function.