CN102663762A - Segmentation method of symmetrical organs in medical image - Google Patents

Abstract

The invention relates to the technical fields of medical image processing and computer vision and provides a segmentation method of symmetrical organs in a medical image. Symmetrical properties of organs and problems of low speeds and low accuracies are considered in the segmentation method. The technical scheme includes that the segmentation method of the symmetrical organs in the medical image includes the following steps of firstly, confirming a bounding box of the image; secondly, detecting a symmetry axis and calculating symmetrical mapping matrixes; thirdly, obtaining symmetrical similar matrixes; fourthly, introducing symmetrical constraint models to the image; and fifthly, solving an energy equation by means of an image segmentation method to obtain a final segmentation result. The segmentation method of the symmetrical organs in the medical image is mainly applied to medical image processing.

Description

The dividing method of symmetrical organ in the medical image

Technical field

The present invention relates to Medical Image Processing, technical field of computer vision, relate in particular to the dividing method of symmetrical organ in a kind of medical image.

Background technology

Digital Image Processing is meant picture signal is converted to digital signal and utilizes computing machine to its process of handling.Along with the deep development of image processing techniques, the great achievement that Digital Image Processing obtains is the application on medical image.In a width of cloth medical image, people are usually only interested in part organ wherein, and these organs occupy certain zone usually, and image last in some characteristic (like gray scale, profile, color and texture etc.) and that close on has difference.These characteristic differences maybe be very obvious, also maybe be very trickle, so that naked eyes can not be discovered.Medical image segmentation is one of prerequisite of its identification and understanding, and the quality of image segmentation quality directly influences the effect that successive image is handled, even determines its success or failure.Cut apart the still basis of three-dimensional visualization, virtual operation.So medical image cutting method has very important effect in field of medical image processing.

Traditional medical image cutting method [1] comprises that Threshold Segmentation, region growing cut apart that [2], Interactive Segmentation, Live Wire are cut apart, Normalized Cut is cut apart [3], deformation model cuts apart, the cutting apart of fuzzy connection degree, cutting apart based on the proxy machine model.On the one hand; Because the inhomogeneous and wriggling of medical image histoorgan causes the inhomogeneous and fuzzy of pixel grey scale; And local bulk effect and pathological tissues cause tissue and pathology edge fog, contrast low, and these problems have proposed very big challenge to existing dividing method.On the other hand, the tissue of human body and organ much all have symmetry, and the low level characteristic that traditional medical image cutting method utilizes like zone and edge, does not have the symmetric properties of good use image.

Patented claim US2010081931A1 has proposed a kind of method of image segmentation, in a dimeric at least picture, at first around object, manually demarcates several points, goes to the edge of match object to be split with curve.According to detected edge; Picture tentatively is divided into several sections, obtains the statistical distribution of each part gray-scale value then, obtain the shared proportion of each part gray-scale value according to statistical distribution; Because statistical is furnished with a plurality of parameters, uses expectation value maximization (EM) algorithm and carry out parameter estimation.Obtain estimated parameter and weight at last, use the Exploration/Selection algorithm and cut apart.Patent US200913002490 proposes a kind of method of medical image segmentation; At first calculate the volume shape index (SI) of each voxel in the picture; Connect SI characteristic, intensity and locus; Form the set of eigenvectors of one five dimension, introduce expectation value maximization algorithm at last, carry out cutting apart of pathology or tumour picture in conjunction with the information that spatial-intensity provides.This patent has obtained better effects in the medical image of clear-cut, but this patent is not considered the symmetric properties of organ.Patent US7336809B2 proposes the convolution dividing method, is solution speed and precision problem, cuts apart and is divided into two stages: presegmentation and fine segmentation.In the presegmentation stage,, reduce the resolution of picture with the convolution sign picture background of image.Fine segmentation is refined and further is segmented in the anatomical structure sign that the merogenesis stage obtains then, finally obtains the tissue regions of being concerned about.This patent is having performance preferably aspect splitting speed and the precision, but this patent is not considered the symmetric properties of organ.

Summary of the invention

The present invention is intended to overcome the deficiency of prior art; Provide a kind of symmetric properties of considering organ, can solution speed and the precision problem medical image in the dividing method of symmetrical organ; For achieving the above object; The technical scheme that the present invention takes is that the dividing method of symmetrical organ comprises the steps: in the medical image

Step 1: the qualification frame Bounding Box that confirms image

For the initial pictures that provides, need the user to provide one and limit frame, as initial input;

Step 2 detects axis of symmetry, calculates the symmetrical mapping matrix

Use the MIFT algorithm and obtain discrete match point, MIFT is the prefix abbreviation of Mirror Reflection Invariant Feature Descriptor, by the match point that obtains, uses the RANSAC method to calculate the mapping matrix of 2 dimensions, uses H _TExpression;

Step 3: obtain symmetrical similar matrix

Utilize the symmetrical mapping matrix H _T, obtain pixel u initial matched pixel u ', through the optimizing process of part, obtain the symmetrical more accurately coupling of pixel u, the similar discrimination m (k) of the neighbor point k of definition u ' is:

$m\left(k\right)={\left|\right|z_u-z_k\left|\right|}_2\·\exp \left(\frac{\mathrm{dis}(x_{u^{\′}},x_k)}{10}\right),---\left(1\right)$

Here z _uThe characteristic that comprises brightness, color of remarked pixel u, || z _u-z _k|| ₂Represent detected pixel k and pixel u on feature space, dis (x _{U '}, x _k) remarked pixel x _{U '}And x _kEuclidean distance, the best symmetrical match point v of pixel u with can obtain through minimizing similar discrimination m:

$u\→v=\arg \underset{v\∈N\left(u^{\′}\right)}{\min }\left\{m\left(v\right)\right\},---\left(2\right)$

Here N (u ') is the set of neighborhood pixels point of the symmetric points u ' of u, definition mapping m={m (u) } the similar discrimination set of representative picture, through entropy the symmetrical structure of the overall situation is described, set balance parameters λ and be:

$\mathrm{\λ}=\exp \left\{\frac{\mathrm{Entrop}{y}^3\left(m\right)}{1000}\right\},---\left(3\right)$

The entropy of the similar discrimination set of Entropy (m) expression, 3 top expression cube, 1000 is normalized parameter; Be used for adjustment factor,, obtain different λ values according to the symmetric case of picture; Last by axis of symmetry that obtains and mapping matrix, picture is represented H with symmetrical similar matrix _S, symmetrical similar matrix H _S=[H _S(u)] use the exponential form definition:

$H_S\left(u\right)=2\mathrm{\λexp}(-\frac{m\left(u\right)}{\mathrm{mean}\left(m\right)})+1-\mathrm{\λ},---\left(4\right)$

Wherein mean (m) is the mean value of m, H _S(u) represent symmetrical similarity, H _S(u) value is limited to by parameter lambda between [1-λ, 1+ λ];

Step 4: introduce the symmetry constraint model in the drawings

Here need introduce markov random file to symmetrical similar matrix, obtain the symmetric properties of foreground object:

Traditional Markov stochastic equation:

$E\left(L\right)=\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{D}_u\left(L_u\right)+\underset{(u,v)\∈N}{\mathrm{\Σ}}{V}_{u,v}(L_u,L_v),---\left(5\right)$

Here L={L _u| u ∈ Λ } the two-value label of expression picture Λ, for generally cutting apart L _u∈ 0, and 1}, 0 expression background, 1 expression organ is for pixel v, L _v∈ 0,1}; Data item D _uBe called the data penalty, it obtain prospect and background based on image; V _{U, v}Be level and smooth punishment, it is used for encouraging similar pixel to distribute same label; N comprises the right set of adjacent pixels, uses symmetrical similar matrix to strengthen the extraction of symmetrical prospect as conversion parameter, and revising the data penalty is D ' _u:

$D_u^{\′}\left(u\right)=\left\{\begin{array}{ccc}v(u;L_u)\mathrm{Hs}\left(u\right)& \mathrm{if}& L_u=s\\ v(u;L_u)(2-\mathrm{Hs}\left(u\right))& \mathrm{if}& L_u=t\end{array}\right.,---\left(6\right)$

Wherein Hs (u) representes symmetrical similarity, and s representes source point, and t represents meeting point, at data penalty D ' _uIn, encouraging the pixel of symmetry is prospect, asymmetric part is labeled as background;

Identical with the data penalty, the linear symmetry that connects coupling to the level and smooth punishment of Markov model, be expressed as:

$\underset{\{u,v\}\∈N}{\mathrm{\Σ}}{V}_{u,v}(L_u,L_v)+\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{V}_{u\→v}^S\left(L_u\right),---\left(7\right)$

The u here → v remarked pixel v is the symmetric points by the pixel u of equation (2) acquisition, other variable such as level and smooth punishment V _{U, v}With the same in the equation (5), the S of a top, back representes the limit of u to v;

Step 5: find the solution energy equation through the figure segmentation method, obtain final segmentation result

At last, energy equation is defined as:

$E_s\left(L\right)=\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{D}_u^{\′}\left(L_u\right)+\underset{\{u,v\}\∈N}{\mathrm{\Σ}}{V}_{u,v}(L_u,L_v)+\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{V}_{u\→v}^S\left(L_u\right),---\left(8\right)$

Can regard global minimization as this moment cutting apart, i.e. L ^*=argminE _s(L), because the energy equation E that in equation (8), defines _s(L) have submodularity, solve so the minimized problem of energy equation can be cut algorithm through figure.

The qualification frame is a rectangular shaped rim.

The MIFT algorithmic procedure is following:

1) for the input picture, representes with multiscale space;

2) detect yardstick spatial extrema point, from extreme point, screen unique point: resulting extreme point has been formed the alternative set of unique point, screens in the alternative point during unique point will be gathered thus, weeds out the point of low contrast and skirt response in this set;

3) accurate location feature point position: through the unique point that above operation is confirmed, the mode of the three-dimensional quafric curve of employing match is carried out the match of information and is approached, to obtain more precise coordinates and graphical rule information;

4) confirm direction parameter: have rotational invariance in order to satisfy unique point, the information according to gradient direction on the auxiliary neighborhood territory pixel of unique point and size is each unique point assigned direction parameter;

5) unique point descriptor coding.Dynamically adjust original SIFT coded sequence according to direction parameter, with the feature descriptor MIFT that obtains overturning constant;

Coupling by obtaining is right, uses the RANSAC method to calculate the mapping matrix H of 2 dimensions _T, make that Λ is the dot matrix of all pixels of input picture, for each pixel u ∈ Λ, the symmetric coordinates on the geometry are: x _{U '}=H _Tx _u, x here _u=(x _u, y _u, 1) ^TProvide the homogeneous coordinates of each pixel u, subscript T representes transposition, x _{U '}Be the homogeneous coordinates of the geometry symmetric position that obtains according to conversion.

Technical characterstic of the present invention and effect:

In the dividing method of a kind of medical image symmetry organ that the present invention proposes; Through providing the qualification frame, detect the axis of symmetry of picture, utilize axis of symmetry to calculate the mapping matrix of organ in the picture; Calculate the symmetric matrix of organ then; Foundation splits organ according to the restricted model of organ symmetric properties, thereby reaches efficiently and accurately segmentation result.Method of the present invention is considered the symmetry of human organ, tissue, has significant improvement cutting apart on the accuracy rate.The present invention also can be adapted to the cutting apart of picture object of complex background, and method is simple, and directly perceived, time complexity is low, and accuracy is high.

Description of drawings

Fig. 1: general frame of the present invention.

Fig. 2: source point optimal match point synoptic diagram.

Fig. 3: data penalty, the level and smooth punishment of symmetry synoptic diagram.(1) data penalty; (2) symmetry is smoothly punished.

Fig. 4: input picture synoptic diagram.Among the figure, (a) input picture.(b) user provides and limits the picture that has detected axis of symmetry behind the frame.

Fig. 5: with the form ecbatic figure of pictorial information with symmetric matrix.(a) through the pictorial information after the symmetry coupling.(b) the symmetrical similar matrix information that calculates.

Fig. 6: several Zhang Butong have a symmetric medical picture.

Fig. 7: experimental result picture: the left side is an input picture, and the right is output.

Embodiment

The present invention is directed to the problem that the traditional medicine image partition method exists, in based on the image partition method of low level, introduce high-rise symmetry constraint.Fig. 1 has shown the framework of this method: at first, the user provides the qualification frame to the input picture; Obtain the symmetrical mapping matrix through symmetry detection algorithm then; Then introduce symmetrical similar matrix and strengthen symmetrical mapping; Introduce the symmetry constraint model at last in the drawings, and cut apart picture through the figure segmentation method.

It is the symmetric properties of utilizing organ that the present invention mainly contributes, and promotes the segmentation effect of organ, and the symmetry constraint model that the present invention proposes has submodularity [4], can in polynomial time, solve through the figure segmentation method.

Image partition method based on axis of symmetry has five steps, shown in figure (1).Be detailed description below to these five steps:

Step 1: the qualification frame (Bounding Box) of confirming image

The present invention needs the user to provide one and limits frame, as initial input for the initial pictures that provides.Because for the preceding background segment of image, need provide an initial input to eliminate the ambiguity of picture prospect, background (promptly specifying the fore/background initial range) by the user.Provide cut apart [5] that the method that limits frame is widely used in display foreground, background user interactions, application of the present invention be input as a rectangular shaped rim.

Step 2 detects axis of symmetry, calculates the symmetrical mapping matrix

Symmetry is the build-in attribute of a lot of organs, and in two-dimentional geometric figure, the symmetry on plane can be carried out the projection conversion.In the present invention, use MIFT [6] (A Mirror Reflection Invariant Feature Descriptor) algorithm and obtain discrete match point.The MIFT algorithmic procedure is following:

1. for the input picture, represent with multiscale space.

2. detect yardstick spatial extrema point, from extreme point, screen unique point.Resulting extreme point has been formed the alternative set of unique point, screens in the alternative point during unique point will be gathered thus.Because be not the requirement that all extreme points all satisfy unique point.In this set, also exist the point of low contrast and skirt response, outstanding inadequately as their uniqueness of characteristic and the stability of image, so will these two types of points be weeded out.

3. accurate location feature point position.Passed through above operation; Unique point is confirmed; Because conversion and the influence on the pixel unit size on the yardstick possibly cause unique point coordinate and some deviation of yardstick information; For the information that guarantees unique point is accurate, adopt the mode of the three-dimensional quafric curve of match to carry out the match of information and approach, to obtain more precise coordinates and graphical rule information.

4. confirm direction parameter.Have rotational invariance in order to satisfy unique point, the information according to gradient direction on the auxiliary neighborhood territory pixel of unique point and size is each unique point assigned direction parameter.

5. the unique point descriptor is encoded.Dynamically adjust original SIFT coded sequence according to direction parameter, with the feature descriptor MIFT that obtains overturning constant.

Coupling by obtaining is right, uses RANSAC method [7] to calculate the mapping matrix of 2 dimensions, uses H here _TExpression.Make that Λ is the dot matrix of all pixels of input picture, for each pixel u ∈ Λ, the symmetric coordinates on the geometry are: x _{U '}=H _Tx _u, x here _u=(x _u, y _u, 1) ^TProvide the homogeneous coordinates of each pixel u, subscript T representes transposition, x _{U '}Be the homogeneous coordinates of the geometry symmetric position that obtains according to conversion.

Step 3: obtain symmetrical similar matrix

Utilize mapping matrix H _T, pixel u can obtain its initial matched pixel u '.Third step obtains the symmetrical more accurately coupling of pixel u through local optimizing process.Should be adjacent on this matching theory with initial matched pixel u '.The similar discrimination m (k) that we define the neighbor point k of u ' is:

$m\left(k\right)={\left|\right|z_u-z_k\left|\right|}_2\·\exp \left(\frac{\mathrm{dis}(x_{u^{\′}},x_k)}{10}\right),---\left(1\right)$

Here z _uThe characteristic of remarked pixel u (like brightness, color).As shown in Figure 2: || z _u-z _k|| ₂Represent that detected pixel k and source point u are on feature space.Dis (x _{U '}, x _k) remarked pixel x _{U '}And x _kEuclidean distance, be in order to regulate the order of magnitude divided by 10.

Fig. 2: u is a source point, and u ' is the detected initial symmetric points of second step.Because u ' possibly not be optimal match point, so at dis (x _{U '}, x _k) distance in look for one with source point optimal match point k.

The L2 normal form is used to the distance in controlling features space, and exponential term dis () is the punishment of space length, and make symmetric points on the space geometry position, satisfy constraint: the point near how much symmetric positions is worth bigger.At last, the best symmetrical match point v of pixel u with can obtain through minimizing similar discrimination m:

$u\→v=\arg \underset{v\∈N\left(u^{\′}\right)}{\min }\left\{m\left(v\right)\right\},---\left(2\right)$

Here N (u ') is the set of neighborhood pixels point of the symmetric points u ' of u.Definition mapping m={m (u) } the similar discrimination set of representative picture.Through observing us find, with respect to the foreground area of symmetry, the histogram of the similar discrimination of asymmetric background area is more level and smooth.Therefore can describe the symmetrical structure of the overall situation through entropy, we set balance parameters λ and are:

$\mathrm{\λ}=\exp \left\{\frac{\mathrm{Entrop}{y}^3\left(m\right)}{1000}\right\},---\left(3\right)$

The entropy of the similar discrimination set of Entropy (m) expression here, 3 top expression cube.1000 is normalized parameter, is used for adjustment factor.According to the symmetric case of picture, obtain different λ values.Last by axis of symmetry that obtains and mapping matrix, picture (is annotated: the mapping matrix H that is different from the front with symmetric matrix _T) expression, symmetric matrix H _S=[H _S(u)] use the exponential form definition:

$H_S\left(u\right)=2\mathrm{\λexp}(-\frac{m\left(u\right)}{\mathrm{mean}\left(m\right)})+1-\mathrm{\λ},---\left(4\right)$

Wherein mean (m) is the mean value of m, symmetrical similar matrix H _S(u) codomain is limited to by parameter lambda between [1-λ, 1+ λ].

Step 4: introduce the symmetry constraint model in the drawings

Here need introduce markov random file to symmetrical similar matrix, obtain the symmetric properties of foreground object.As shown in Figure 3, the symmetry constraint model among the present invention is made up of two parts: (1) data penalty; (2) symmetry is smoothly punished.

Fig. 3: s is a source point, and t is a meeting point.The thickness of line is represented the punishment size of this edge, in the drawings, and some u and some v symmetry.(a) do not consider symmetric properties, its segmenting structure possibly be wrong.(b) the red dotted line in is represented axis of symmetry, and the line of the yellow of in (b), introducing is that u punishes that for the symmetry of v is level and smooth our method can be to u, and v provides correct label.

Traditional Markov stochastic equation:

$E\left(L\right)=\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{D}_u\left(L_u\right)+\underset{(u,v)\∈N}{\mathrm{\Σ}}{V}_{u,v}(L_u,L_v),---\left(5\right)$

Here L={L _u| u ∈ Λ } the two-value label of expression picture Λ.For generally cutting apart L _u∈ 0, and 1}, 0 expression background, 1 expression organ is for pixel v, L _v∈ 0,1}; Data item D _uBe called the data penalty, it obtain prospect and background based on image; V _{U, v}Be level and smooth punishment, it is used for encouraging similar pixel to distribute same label; N comprises the right set of adjacent pixels.In the present invention, we use symmetrical similar matrix to strengthen the extraction of symmetrical prospect as conversion parameter, and revising the data penalty is D ' _u:

$D_u^{\′}\left(u\right)=\left\{\begin{array}{ccc}v(u;L_u)\mathrm{Hs}\left(u\right)& \mathrm{if}& L_u=s\\ v(u;L_u)(2-\mathrm{Hs}\left(u\right))& \mathrm{if}& L_u=t\end{array}\right.,---\left(6\right)$

Wherein Hs (u) representes symmetrical similarity, and s representes source point, and t represents meeting point.At data penalty D ' _uIn, encouraging the pixel of symmetry is prospect, asymmetric part is more prone to be labeled as background.

Identical with the data penalty, the linear symmetry that connects coupling to the level and smooth punishment of Markov model, be expressed as:

$\underset{\{u,v\}\∈N}{\mathrm{\Σ}}{V}_{u,v}(L_u,L_v)+\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{V}_{u\→v}^S\left(L_u\right),---\left(7\right)$

The u here → v remarked pixel v is the symmetric points by the pixel u of equation (2) acquisition.Other variable such as level and smooth punishment V _{U, v}With the same in the equation (5), the S of a top, back representes the limit of u to v.

Step 5: find the solution energy equation through the figure segmentation method, obtain final segmentation result

At last, energy equation is defined as:

$E_s\left(L\right)=\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{D}_u^{\′}\left(L_u\right)+\underset{\{u,v\}\∈N}{\mathrm{\Σ}}{V}_{u,v}(L_u,L_v)+\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{V}_{u\→v}^S\left(L_u\right),---\left(8\right)$

Can regard global minimization as this moment cutting apart, i.e. L ^*=arg min E _s(L).Because the energy equation E of definition in equation (8) _s(L) have submodularity, solve so the minimized problem of energy equation can be cut algorithm through figure.

In the dividing method of a kind of medical image symmetry organ that the present invention proposes; Through providing the qualification frame, detect the axis of symmetry of picture, utilize axis of symmetry to calculate the mapping matrix of organ in the picture; Calculate the symmetric matrix of organ then; Foundation splits organ according to the restricted model of organ symmetric properties, thereby reaches efficiently and accurately segmentation result.Our method is considered the symmetry of human organ, tissue, has significant improvement cutting apart on the accuracy rate.The present invention also can be adapted to the cutting apart of picture object of complex background, and method is simple, and directly perceived, time complexity is low, and accuracy is high.

Overall flow of the present invention is following:

1, input picture and provide the user and to limit frame

As figure (4) (a) shown in to the input picture, the user provides according to the prospect of oneself wanting to cut apart and limits frame and calculate axis of symmetry, like figure (4) (b), providing picture respectively is example:

2, carry out the symmetry coupling

According to the axis of symmetry of having set up, calculate the symmetrical mapping matrix H _TBy formula (1), obtain the similar discrimination of neighbor point, the image of right ends is carried out the symmetry coupling, (a) like figure (5).

3, calculate symmetrical similar matrix

Here be our method of example explanation with the left-right symmetric, for other symmetric case, through using different mapping matrixes, method of the present invention still is suitable for.The axis of symmetry and the mapping matrix that utilize the front to obtain, application of formula (4) is represented pictorial information with the form of symmetric matrix, the result as figure (5) (b) shown in:

4, picture is cut apart

Symmetric properties according to object is set up restricted model, and application drawing cuts algorithm object segmentation is come out.Through above-mentioned steps, obtain final segmentation result, figure (6) is the symmetric medical picture of having of several Zhang Butong.Be respectively that application GrabCut algorithm [8] and our symmetry division method are cut apart, picture provides the comparative result of GrabCut and method of the present invention.

Below be more experimental result:

The left side is an input picture, and the right is output.

Conclusion: when cutting apart symmetrical prospect, the present invention proposes a kind of symmetry constraint model.Our symmetry constraint model and very natural the combining of MRF model, and can both be suitable under even the situation about not having strong, weak in symmetric case.Another advantage is that the present invention just uses the symmetry constraint reconstructed image, does not influence to cut apart optimization.Can in polynomial time, solve through our restricted model of figure segmentation method.Our method not only is applicable to the image of minute surface symmetry, for other situation (for example: rotation) can introduce this framework equally.

The main reference document

[1] Tian Jie, Bao Shanglian, Zhou Mingquan. " medical image processing and analysis ", Electronic Industry Press, 2003

[2] Senthilkumar., B.Umamaheswari, G.Karthik, J. " a kind of new region growing partitioning algorithm on breast cancer detects ", computer intelligence and computing research international conference, 2010, P1-4

[3] Jianbo Shi; Malik, J. " Normalized Cut and image segmentation ", pattern analysis and machine intelligence periodical, 2000, P888-905

[4] Iwata, S.; Nagano, K. " modular function that covers under the constraint minimizes ", the meeting of computer science basic science, 2009, P671-680

[5] V.Lempitsky, P.Kohli, C.Rother and T.Sharp. " based on the image partition method that limits the frame priori ", international computer visual conference, 2009, P277-284

[6] X, Guo., X, Cao., J, Zhang. and X, Li. " MIFT: a kind of direct reflection unchangeability feature description ", Asia computer vision meeting, 2009, P536-544

[7] Richard Szeliski. " computer vision: algorithm and application ", Springer publishing house, 2010

[8] C.Rother, V.Kolmogorov, A.Blake. " GrabCut: the mutual foreground extracting method that adopts iteration diagram to cut ", international graphics conference, 2004,23 (3)

Claims (3)

1. the dividing method of symmetrical organ in the medical image is characterized in that, comprises the steps:

Step 1: the qualification frame Bounding Box that confirms image

For the initial pictures that provides, need the user to provide one and limit frame, as initial input;

Step 2: detect axis of symmetry, calculate the symmetrical mapping matrix

Use the MIFT algorithm and obtain discrete match point, MIFT is the prefix abbreviation of Mirror Reflection Invariant Feature Descriptor, by the match point that obtains, uses the RANSAC method to calculate the mapping matrix of 2 dimensions, uses H _TExpression;

Step 3: obtain symmetrical similar matrix

Utilize the symmetrical mapping matrix H _T, obtain pixel u initial matched pixel u ', through the optimizing process of part, obtain the symmetrical more accurately coupling of pixel u, the similar discrimination m (k) of the neighbor point k of definition u ' is:

$m\left(k\right)={\left|\right|z_u-z_k\left|\right|}_2\·\exp \left(\frac{\mathrm{dis}(x_{u^{\′}},x_k)}{10}\right),---\left(1\right)$

Here z _uThe characteristic that comprises brightness, color of remarked pixel u, || z _u-z _k|| ₂Represent detected pixel k and pixel u on feature space, dis (x _{U '}, x _k) remarked pixel x _{U '}And x _kEuclidean distance, the best symmetrical match point v of pixel u with can obtain through minimizing similar discrimination m:

$u\→v=\arg \underset{v\∈N\left(u^{\′}\right)}{\min }\left\{m\left(v\right)\right\},---\left(2\right)$

Here N (u ') is the set of neighborhood pixels point of the symmetric points u ' of u, definition mapping m={m (u) } the similar discrimination set of representative picture, through entropy the symmetrical structure of the overall situation is described, set balance parameters λ and be:

$\mathrm{\λ}=\exp \left\{\frac{\mathrm{Entrop}{y}^3\left(m\right)}{1000}\right\},---\left(3\right)$

The entropy of the similar discrimination set of Entropy (m) expression, 3 top expression cube, 1000 is normalized parameter; Be used for adjustment factor,, obtain different λ values according to the symmetric case of picture; Last by axis of symmetry that obtains and mapping matrix, picture is represented H with symmetrical similar matrix _S, symmetrical similar matrix H _S=[H _S(u)] use the exponential form definition:

$H_S\left(u\right)=2\mathrm{\λexp}(-\frac{m\left(u\right)}{\mathrm{mean}\left(m\right)})+1-\mathrm{\λ},---\left(4\right)$

Wherein mean (m) is the mean value of m, H _S(u) represent symmetrical similarity, H _S(u) value is limited to by parameter lambda between [1-λ, 1+ λ];

Step 4: introduce the symmetry constraint model in the drawings

Here need introduce markov random file to symmetrical similar matrix, obtain the symmetric properties of foreground object:

Traditional Markov stochastic equation:

$E\left(L\right)=\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{D}_u\left(L_u\right)+\underset{(u,v)\∈N}{\mathrm{\Σ}}{V}_{u,v}(L_u,L_v),---\left(5\right)$

Here L={L _u| u ∈ Λ } the two-value label of expression picture Λ, for generally cutting apart L _u∈ 0, and 1}, 0 expression background, 1 expression organ is for pixel v, L _v∈ 0,1}; Data item D _uBe called the data penalty, it obtain prospect and background based on image; V _{U, v}Be level and smooth punishment, it is used for encouraging similar pixel to distribute same label; N comprises the right set of adjacent pixels, uses symmetrical similar matrix to strengthen the extraction of symmetrical prospect as conversion parameter, and revising the data penalty is D ' _u:

$D_u^{\′}\left(u\right)=\left\{\begin{array}{ccc}v(u;L_u)\mathrm{Hs}\left(u\right)& \mathrm{if}& L_u=s\\ v(u;L_u)(2-\mathrm{Hs}\left(u\right))& \mathrm{if}& L_u=t\end{array}\right.,---\left(6\right)$

Wherein Hs (u) representes symmetrical similarity, and s representes source point, and t represents meeting point, at data penalty D ' _uIn, encouraging the pixel of symmetry is prospect, asymmetric part is labeled as background;

Identical with the data penalty, the linear symmetry that connects coupling to the level and smooth punishment of Markov model, be expressed as:

$\underset{\{u,v\}\∈N}{\mathrm{\Σ}}{V}_{u,v}(L_u,L_v)+\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{V}_{u\→v}^S\left(L_u\right),---\left(7\right)$

The u here → v remarked pixel v is the symmetric points by the pixel u of equation (2) acquisition, other variable such as level and smooth punishment V _{U, v}With the same in the equation (5), the S of a top, back representes the limit of u to v;

Step 5: find the solution energy equation through the figure segmentation method, obtain final segmentation result

At last, energy equation is defined as:

$E_s\left(L\right)=\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{D}_u^{\′}\left(L_u\right)+\underset{\{u,v\}\∈N}{\mathrm{\Σ}}{V}_{u,v}(L_u,L_v)+\underset{u\∈\mathrm{\Λ}}{\mathrm{\Σ}}{V}_{u\→v}^S\left(L_u\right),---\left(8\right)$

Can regard global minimization as this moment cutting apart, i.e. L ^*=argminE _s(L), because the energy equation E that in equation (8), defines _s(L) have submodularity, solve so the minimized problem of energy equation can be cut algorithm through figure.

2. the dividing method of symmetrical organ is characterized in that in the medical image as claimed in claim 1, and the qualification frame is a rectangular shaped rim.

3. the dividing method of symmetrical organ is characterized in that in the medical image as claimed in claim 1, and the MIFT algorithmic procedure is following:

1) for the input picture, representes with multiscale space;

2) detect yardstick spatial extrema point, from extreme point, screen unique point: resulting extreme point has been formed the alternative set of unique point, screens in the alternative point during unique point will be gathered thus, weeds out the point of low contrast and skirt response in this set;

3) accurate location feature point position: through the unique point that above operation is confirmed, the mode of the three-dimensional quafric curve of employing match is carried out the match of information and is approached, to obtain more precise coordinates and graphical rule information;

4) confirm direction parameter: have rotational invariance in order to satisfy unique point, the information according to gradient direction on the auxiliary neighborhood territory pixel of unique point and size is each unique point assigned direction parameter;

5) unique point descriptor coding.Dynamically adjust original SIFT coded sequence according to direction parameter, with the feature descriptor MIFT that obtains overturning constant;

Coupling by obtaining is right, uses the RANSAC method to calculate the mapping matrix H of 2 dimensions _T, make that Λ is the dot matrix of all pixels of input picture, for each pixel u ∈ Λ, the symmetric coordinates on the geometry are: x _{U '}=H _Tx _u, x here _u=(x _u, y _u, 1) ^TProvide the homogeneous coordinates of each pixel u, subscript T representes transposition, x _{U '}Be the homogeneous coordinates of the geometry symmetric position that obtains according to conversion.

Images