Summary of the invention
Present invention aims to the problem that medical image is fuzzy, uneven, individual variation is big, it is proposed to a kind of phase locked medical image processing method Network Based, to solve the over-segmentation problem that existing method exists, improve the diagnosis discrimination of medical image.
The technical thought of the present invention is: with dividing ridge method, the image after feature extraction is split, and the zonule after segmentation is corresponded to the node in network, sets up network model;By agitator formula, node is carried out phase place iteration, complete the merging of overdivided region, improve the effect of image segmentation.Implementation step includes as follows:
(1) extract the pixel feature of primitive medicine image, and the pixel feature extracted is carried out watershed segmentation, obtain N number of region, N >=1000;
(2) similarity between calculating each two region, and each region and the similarity of himself, obtain N × (N+1)/2 Similarity value;
(3) with N number of region as each node in network, with N × (N+1)/2 Similarity value as the weights in network;
(4) in [0,2 π], randomly generate N number of number, as the initial phase of each node in network, adopt agitator formula that the N number of node in network is carried out phase place iteration, when all node N meet formula θi(k)-θi(k-3) during < 0.01, finishing iteration, θiK () represents the phase place of node i kth time iteration, iteration total degree is designated as m;
(5) [0,2 π] is divided into n1Individual subinterval, statistics phase place falls the node number in each subinterval, with n1Individual subinterval sequence number is as abscissa, to drop on subinterval interior nodes number as vertical coordinate, draws PHASE DISTRIBUTION curve, and statistics is between adjacent trough all nodes on curve, is classified as a class, and N number of node is divided into L altogether1Class, L1Numerical value identical with trough number, the classification sequence number of N number of node is stored in the vectorial F that length is N, if the arbitrary node i in N number of node belongs to jth class, then the i-th value f in Fi=j, j=1 ..., L1;
(6) with m generation of node and m-3 for the slope of all nodes of phase calculation, N number of slope value is obtained, to L1Each class of apoplexy due to endogenous wind carries out following operation: [-0.01,0.01] is divided into n2Individual subinterval, statistics slope value falls the node number in each subinterval, with n2Individual subinterval sequence number, as abscissa, to drop on subinterval node number as vertical coordinate, draws L1Bar slope distribution curve, adds up being between adjacent trough all nodes on each curve, is classified as a class, and all nodes are divided into L altogether2Class, L2With L1The trough sum of bar slope distribution curve is identical, updates the classification sequence number of each node in vector F;
(7) detection L2The nodes of each class of individual apoplexy due to endogenous wind, if the nodes of certain class is less than threshold value T=N/200, then redistributes such all nodes, the similarity of the arbitrary node i of this apoplexy due to endogenous wind Yu class j is designated as vij, calculate the L of i and all classes2Individual similarity, to L2Individual similarity is ranked up, maximizing, and node i is given class that maximum is corresponding again, updates the classification sequence number of vector F interior joint i;
(8) by the classification sequence number of vector F storage, namely the classification sequence number belonging to the region of pixel, is tagged on each pixel;
(9) compose different gray values, the image after being split to inhomogeneous pixel, and export.
The present invention compared with prior art has the advantage that
1, the present invention adopts agitator formula that the region after watershed algorithm segmentation is carried out phase place iteration, and operand is little, calculates speed fast;
2, the region after segmentation is merged by the present invention by extracting phase place and slope, compensate for the over-segmentation defect of watershed algorithm, improves the diagnosis discrimination of medical image;
3, the present invention fully takes into account the complexity of medical image, adopts the method for sequencing of similarity that the region sorting out mistake is reclassified, it is possible to focal area is accurately positioned.
Simulation result shows, mammogram figure adopts the medical image processing method of the present invention, improves the discrimination of image, contribute to the Accurate Diagnosis of medical image.
Detailed description of the invention
With reference to Fig. 1, the present invention phase locked medical image processing method Network Based, comprise the steps:
Step 1, extracts the feature of primitive medicine image pixel point and pixel feature is carried out watershed segmentation.
(1a) inputting primitive medicine image, this example selects original mammogram figure as shown in Figure 2, is sized to 1024 × 1024;
(1b) feature based on gray level co-occurrence matrixes is extracted, for any pixel i, extract contrast, concordance, energy [0,45,90,135] 12 dimensional features on four direction, then extract 10 dimensional features based on un-decimated wavelet transform conversion, above-mentioned 12 dimensional features and this two feature of 10 dimensional features are merged composition 22 dimensional vectors, as the feature of ith pixel point;
(1c) all pixels in image are repeated step (1b), obtain the feature of all pixels of primitive medicine image;
(1d) the pixel feature extracted is carried out watershed segmentation, obtain N number of region, N >=1000, the mammogram figure after segmentation, as shown in Figure 3.
Step 2, the similarity between zoning.
(2a) calculate the feature in all regions, obtain the feature x in N number of region1,...xN, wherein, the feature x of ith zoneiThe meansigma methods of all pixel features equal in ith zone;
(2b) with the feature x in N number of region1,...xN, calculate the similarity between each two region, and each region and himself similarity, the computing formula of similarity is:
Wherein, wijRepresent the similarity between ith zone and jth region, xiIt is the feature of ith zone, xjBeing the feature in jth region, σ is scale parameter, and span is (0,1).
Step 3, with N number of region as each node in network, with N × (N+1)/2 Similarity value as the weights in network, sets up network.
Step 4, adopts oscillator phase synchronization formula to carry out phase place iteration.
(4a) arithmetic mean of instantaneous value of all weights N × (N+1)/2 in computing networkGenerator matrix E, the value e of the i-th row jth row in this matrix EijFor:
(4b) in [0,2 π], produce N number of random number, as the initial phase of N number of node, be designated as θ (0), θ (0)={ θ1(0),...θi(0),...,θN(0) }, wherein, θi(0) it is the initial phase of i-th node;
(4c) the initial phase θ according to i-th nodei(0), obtaining oscillator phase synchronization formula is:
Wherein, N is the nodes of network, and k is iterations, θiK () is the phase place of node i kth time iteration, Γ is stiffness of coupling, works as eijWhen=1, Γ=3, work as eijWhen=0, Γ=-1,
(4d) from k=0, the initial phase θ (0) of N number of node is substituted in above-mentioned oscillator phase synchronization formula, N number of node is carried out phase place iteration;
(4e) as iterations k > 3, it is judged that whether node reaches stable, if all N number of nodes all meet formula θi(k)-θi(k-3) < 0.01, then node phase reaches stable, and finishing iteration, iteration total degree is designated as m;
(4f) when iterations k≤3, or also not all node all meets formula θi(k)-θi(k-3) < 0.01, then the kth of N number of node for phase theta1(k),...θNK () substitutes in above-mentioned oscillator phase synchronization iterative formula, calculate kth+1 generation phase place of N number of node, and add 1 by iterations k, returns step (4e).
Step 5, classifies to node according to phase place.
[0,2 π] is divided into n1Individual subinterval, statistics phase place falls the node number in each subinterval, with n1Individual subinterval sequence number, as abscissa, to drop on subinterval interior nodes number as vertical coordinate, draws PHASE DISTRIBUTION curve, and statistics is in all nodes between adjacent trough on curve, is classified as a class, and N number of node is divided into L altogether1Class, L1Numerical value identical with trough number, the classification sequence number of N number of node is stored in the vectorial F that length is N, if the arbitrary node i in N number of node belongs to jth class, then the i-th value f in Fi=j, j=1 ... L1。
Step 6, classifies to node according to slope.
(6a) by slope S for all node N of phase calculation of m generation of node and m-3:
S={s1,...si,...sN,
Wherein, the slope s of i-th nodeiFor:
(6b) to L1Each class of apoplexy due to endogenous wind carries out following operation: [-0.01,0.01] is divided into n2Individual subinterval, statistics slope value falls the node number in each subinterval, with n2Individual subinterval sequence number, as abscissa, to drop on subinterval node number as vertical coordinate, draws L1Bar slope distribution curve, adds up being between adjacent trough all nodes on each curve, is classified as a class, and all nodes are divided into L altogether2Class, L2With L1The trough sum of bar slope distribution curve is identical, updates the classification sequence number of each node in vector F.
Step 7, redistributes the node of possible misclassification.
(7a) L is detected2The nodes of individual apoplexy due to endogenous wind any type, if such nodes is less than threshold value T=N/200, then calculates the L of this apoplexy due to endogenous wind i-th node and all classes2Individual similarity vij:
(7b) to L2Individual similarity is ranked up, maximizing, i-th node is given class that maximum is corresponding again, and updates the node classification sequence number in vector F;
(7c) to all L2Individual class repeats step (7a)-(7b), is reclassified by all nodes class interior joint less than threshold value T.
Step 8, by the classification sequence number of vector F storage, namely the classification sequence number belonging to the region of pixel, is tagged on each pixel.
Step 9, composes different gray values, the image after being split to inhomogeneous pixel, and exports.
(9a) the gray value h of the i-th class is calculatedi:
By hiIt is assigned to all pixels of the i-th class;
(9b) make i=i+1, repeat step (9a), until i=L2, now L2All pixels all assignment of individual class are complete, and the gray level image formed by all pixels after assignment is as final segmentation image;
(9c) final segmentation image is exported.
The effect of the present invention can be further illustrated by the following emulation experiment to mammogram figure:
1, simulated conditions
The emulation of the present invention is at windowsXP, SPI, CPUInterCore2Duo, fundamental frequency 2.33Ghz, and software platform is that Matlab2007b runs.The original mammogram that emulation is selected derives from common data sets MIAS, obtains 330 original mammograms altogether, and wherein Fig. 2 is a width canceration mammogram therein.
2, emulation content and result
Emulation 1, carries out watershed segmentation by the present invention to canceration mammogram shown in Fig. 2, obtains the mammogram after Fig. 3 segmentation.
Emulation 2, by the present invention and existing Kmeas method and existing Fcm method, the image after Fig. 3 is split carries out region classification, and result is Fig. 4 such as.Wherein Fig. 4 a is the result adopting the network phase synchronous method of the present invention that Fig. 3 carries out region classification;Fig. 4 b is the result adopting existing Kmeas method that Fig. 3 carries out region classification;Fig. 4 c is the result adopting existing Fcm method that Fig. 3 carries out region classification.
From Fig. 4 a it can be seen that the present invention can detect the lesion region of mammary gland exactly, and compared with Fig. 4 b and Fig. 4 c, the lesion region that the present invention detects is without assorted point, and region consistency is higher, and in the local detail process to tumor region, advantage is comparatively obvious.
Above-mentioned simulation result shows, the present invention can improve the order of accuarcy of image segmentation compared to traditional method, it is thus achieved that better segmentation result, is a kind of effective medical image processing method.
Above example is carried out under premised on technical solution of the present invention, gives detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to above-described embodiment.