CN101609553A - Non-rigid image registration algorithm based on implicit shape representation and edge information fusion - Google Patents

Abstract

The invention discloses a kind of non-rigid image registration algorithm based on implicit shape representation and edge information fusion, it comprises following steps: at first, the exterior contour alignment with floating image subject to registration and reference picture realizes the global registration based on implicit shape; Then, selection is carried out local registration based on the Free Transform model of the multi-resolution grid of the B batten of edge information fusion.The invention has the beneficial effects as follows: by in the registration framework, introducing tolerance the edge registration, made full use of the significant contribution of marginal information in image registration, improve accuracy of registration and robustness, obtained transform domain level and smooth, continuous and that assurance is shone upon one to one.

Description

Non-rigid image registration algorithm based on implicit shape representation and edge information fusion

Technical field

This algorithm is the medical image registration method that the hierarchical transformation model of a kind of use from the overall situation to the part covers whole transform domain, particularly solves the image registration problem that big local deformation is arranged.

Background technology

Image registration is the prerequisite of image co-registration, is to generally acknowledge the bigger image processing techniques of difficulty, also is the gordian technique of decision Medical image fusion technical development.Since the later stage eighties, the research of medical figure registration is subjected to the attention of medical circle and engineering circle day by day, the expansion in extensive range in the world.All kinds of bibliographical informations are vast as the open sea, and new method emerges in an endless stream.The method for registering images of existing comparative maturity still is mainly used on rigid body or the accurate rigid body, and for the research of non-rigid body still a little less than the relative thin.Because the complicacy of medical structure itself, diversity, clinical in there not being the rigid boundary, even the registration at the position of obscurity boundary has urgent requirement, impelled the research centre of medical figure registration progressively to shift to non-rigid body field.

At present medical figure registration mainly contains two big classifications: based on the image registration (framework is arranged) of surface with based on the image registration (frameless) of image internal feature.Using mutual information to do to measure the registration medical image is present widely used reasonable a kind of method.Mutual information is a key concept in the information theory, is estimating of two stochastic variable statistic correlations.When two width of cloth reached optimal registration based on the image of common anatomical structure, it is maximum that the gray scale mutual information of their respective pixel should reach.Because this is estimated the relation that does not need between different imaging pattern hypograph gray scales and makes any hypothesis, do not need image is cut apart or any pre-service yet, so be widely used in CT/MR, in the Image Registration such as PET/MR, particularly when the data division of one of them image is damaged, also can obtain good registration effect.Based on the problem that the image registration of mutual information exists, mainly show as the robustness problem of registration, promptly in the process of search optimal transformation, exist the interference of a large amount of local maximums, make registration process be trapped in local maximum easily and cause image to mismatch.

Summary of the invention

The present invention is directed to the deficiency of prior art, a kind of non-rigid image registration algorithm based on implicit shape representation and edge information fusion has been proposed, the steps include: that at first the exterior contour alignment with floating image subject to registration and reference picture realizes the global registration based on implicit shape; Then, selection is carried out local registration based on the Free Transform model of the multi-resolution grid of the B batten of edge information fusion.

Global registration concrete grammar based on implicit shape representation is: the exterior contour of at first using the implicit shape representation image, with the exterior contour of two images subject to registration zero level collection as distance function, the range conversion space that impliedly is embedded into high one dimension, calculate its distance map figure respectively, realize global registration by the mutual information of maximization distance map figure then, the alignment exterior contour obtains the global registration image.

Carrying out local registration concrete grammar based on the Free Transform model of the multi-resolution grid of the B batten of edge information fusion is: at first, merge marginal information; Then, select Free Transform model,, and improve image similarity tolerance by the improvement of level method for registering speed, degree of accuracy and the robustness of the multiresolution B batten control mesh from coarse to meticulous based on the B batten.

Edge and view data integration technology concrete grammar are: at first, reference picture and global registration image are carried out operator filtering respectively, obtain edge-detected image separately, pixel value in the edge-detected image is suppressed greater than the strong edge of certain threshold level by the method that multiply by an inhibiting factor; Then, former figure and edge-detected image are carried out image co-registration by certain weight, form new image; At last, the image after merging is carried out registration by multi-resolution grid Free Transform method.

Local method for registering based on the Free Transform of the multi-resolution grid of B batten is: at first, definition is 2 dimension tensor product forms of 1 dimension cubic B-spline based on the Free Transform of B batten, on reference picture, place sparse, regular lattice point, it is the reference mark, obtain the deformation territory by the motion that defines each reference mark then, use spline interpolation to assess the deformation values of calculating between the reference mark, thereby produce a local control, the level and smooth conversion of the overall situation; Then, progressively reduce the spacing at reference mark, promptly progressively increase the resolution of grid, repeat same step and carry out registration; The registration transformation of representing local deformation by the resolution that improves the reference mark grid.

Advantage of the present invention is: proposed a kind of non-rigid image registration algorithm based on implicit shape representation and edge information fusion that is applicable to complicated image, this algorithm is to improving based on the deformable registration of multi-resolution grid FFD model, in the registration framework, introduce tolerance to the edge registration, made full use of the significant contribution of marginal information in image registration, improve accuracy of registration and robustness, obtained transform domain level and smooth, continuous and that assurance is shone upon one to one.

Description of drawings

Fig. 1 is an image registration algorithm global registration framework of the present invention;

Fig. 2 is the local registration process of image registration algorithm of the present invention.

Embodiment

Below, the present invention is further illustrated with embodiment in conjunction with the accompanying drawings.

The non-rigid image registration based on implicit shape representation and edge information fusion that this algorithm proposes, it comprises the global registration based on implicit shape representation, based on the local registration of the multi-resolution grid Free Transform of the B batten of edge information fusion.Wherein the fundamental purpose of global registration is the registration image exterior contour.The shape center (hereinafter to be referred as " centre of form ") by translation transformation alignment floating image g and reference picture f exterior contour at first, make the translation parameter of overall similarity transformation for very little or be 0, calculate the distance map figure of two width of cloth images after the centre of form is alignd then, utilize similarity transformation, mutual information tolerance, gradient descending method registration distance map figure, obtain the optimal transformation parameter, utilize the parameter that obtains that reference picture is resampled again, finish the global registration process.The global registration framework as shown in Figure 1.

At first extract the exterior contour (closed curve) of image, utilize level set function, promptly use range conversion that exterior contour is embedded into high 1 dimension space as the zero level collection of distance function the exterior contour modeling.For the ease of expression, consider 2 dimension situations.Make Φ: Ω → R ⁺Be the Lipschitz function, the range conversion of expression shape S.Shape S has defined the division of image area Ω: be enclosed in S with interior zone [R _S] and background area [Ω-R _S].Given above-mentioned definition, implicit shape representation is:

${\mathrm{\&Phi;}}_S(x,y)=\left\{\begin{array}{c}0,(x,y)\&Element;S\\ +D((x,y),S)>0,(x,y)\&Element;R_S\\ -D((x,y),S)<0,(x,y)\&Element;[\mathrm{\&Omega;}-R_S]\end{array}\right.$

Here D ((x, y), S) the presentation video location of pixels (x, y) and the minor increment between the shape S.

We use, and near the arrowband the profile has solved the efficiency of implicit shape representation in registration process as sample area in the embedded space, when improving speed, produced like this can with use the comparable result in entire image zone, and prevented the problem of out-of-bounds point generation error in the process of image registration.

Based on the global registration of maximization mutual information in the global registration of implicit shape representation, the mutual information of two width of cloth image A and B can be defined as follows (the distance map figure that A, B represent the floating image g exterior contour after reference picture f and the centre of form alignment respectively):

I(A，B)＝H(A)+H(B)-H(A，B)

Wherein H (A), H (B) and H (A is individual entropy and the combination entropy of stochastic variable A and B B), and it is defined as:

$\left\{\begin{array}{c}H\left(A\right)=\underset{a}{\mathrm{\&Sigma;}}-P_A\left(a\right)\log P_A\left(a\right)\\ H\left(B\right)=\underset{b}{\mathrm{\&Sigma;}}-P_B\left(b\right)\log P_B\left(b\right)\\ H(A,B)=\underset{a,b}{\mathrm{\&Sigma;}}-P_{A,B}(a,b)\log P_{A,B}(a,b)\end{array}\right.$

Here P _A(a) and P _B(b) be marginal probability density function, P _{A, B}(a b) is joint probability density function.

Based on the local registration of the multi-resolution grid Free Transform of B batten, local deformation can assist the global registration model to obtain accurate deformable registration territory.We select to describe local deformation based on the multi-resolution grid Free Transform model of B batten.Make Λ represent reference mark λ _{I, j}A size be n _x* n _yGrid, reference mark λ _{I, j}Proportional spacing δ is arranged.So, Free Transform can be write as 2 dimension tensor product forms of 1 dimension cubic B-spline:

$T_{\mathrm{local}}(x,y)=\underset{m=0}{\overset{3}{\mathrm{\&Sigma;}}}\underset{n=0}{\overset{3}{\mathrm{\&Sigma;}}}{B}_m\left(u\right)B_n\left(v\right){\mathrm{\&lambda;}}_{i+m,j+n}$

Here

B _lRepresent l rank B spline base function:

$\left\{\begin{array}{c}{B}_0\left(u\right)={(1-u)}^3/6\\ B_1\left(u\right)=({3u}^3-{6u}^2+4)/6\\ B_2\left(u\right)=(-{3u}^3+{3u}^2+3u+1)/6\\ B_3\left(u\right)=u^3/6\end{array}\right.$

The B batten is local control, even this makes it also very high for the reference mark counting yield of big figure.Especially, the basis function of cubic B-spline is limited support, changes reference mark λ in other words _{I, j}Only in the local neighborhood of reference mark, influence conversion.

Reference mark Λ as B batten Free Transform parameter and the degree of freedom of non-rigid shape deformations, its modeling mainly depends on the resolution of Λ grid.Big reference mark spacing allows overall non-rigid shape deformations, and little reference mark spacing can be to the highly non-rigid shape deformations modeling of localization.Simultaneously, the resolution of reference mark grid has defined the number of degree of freedom, thereby has determined the complexity of calculating.Make Λ ¹..., Λ ^LThe reference mark grid of expression different resolution.The spacing of supposing the reference mark is from Λ ^lProgressively be reduced to Λ ^L+1, promptly the resolution of control mesh progressively increases.Each control mesh Λ ^lDefined the local deformation T of this resolution stage with its B batten FFD that is associated _Local ^l, they with form partial transformation T _Local:

$T_{\mathrm{local}}(x,y)=\underset{l=1}{\overset{L}{\mathrm{\&Sigma;}}}{T_{\mathrm{local}}}^l(x,y)$

In this case, partial transformation is the associating of the B batten conversion of reference mark each resolution stage of grid.For prevent to calculate separately several times B batten Free Transform expense, we by the reference mark grid progressively refinement single B batten Free Transform represent partial transformation.Like this, by inserting the reference mark grid that new reference mark grid forms the l+1 level in the l level.Suppose to reduce by half, so reference mark λ in each step reference mark spacing _{2i, 2j} ^L+1With λ _{I, j} ^lThe position be the same, and new reference mark Λ ^L+1Value can use B batten algorithm of subdivision directly from Λ ^lValue obtain.

To reference picture f and global registration image g ₁Carry out operator filtering respectively, obtain edge-detected image f separately _Edge(x, y) and g _Ledge(x, y).Owing to use the similarity measurement standard, the introducing at strong edge can seriously weaken the contribution of other weak edge to similarity measurement, therefore need be suppressed strong edge.Method is if pixel value is greater than certain threshold level in the edge-detected image, then it be multiply by one less than 1 inhibiting factor.Then former figure and edge-detected image are carried out image co-registration by certain weight, form new image f ', g ₁', at last to image f ', g after merging ₁' carry out registration by above-mentioned multi-resolution grid FFD method.Whole local registration process as shown in Figure 2.

Claims (5)

1. non-rigid image registration algorithm based on implicit shape representation and edge information fusion is characterized in that comprising following steps: at first, the exterior contour alignment with floating image subject to registration and reference picture realizes the global registration based on implicit shape; Then, selection is carried out local registration based on the Free Transform model of the multi-resolution grid of the B batten of edge information fusion.

2. the non-rigid image registration algorithm based on implicit shape representation and edge information fusion as claimed in claim 1, it is characterized in that, global registration concrete grammar based on implicit shape representation is: the exterior contour of at first using the implicit shape representation image, with the exterior contour of two images subject to registration zero level collection as distance function, the range conversion space that impliedly is embedded into high one dimension, calculate its distance map figure respectively, realize global registration by the mutual information of maximization distance map figure then, the alignment exterior contour obtains the global registration image.

3. the non-rigid image registration algorithm based on implicit shape representation and edge information fusion as claimed in claim 1, it is characterized in that, local registration concrete grammar based on the Free Transform model of the multi-resolution grid of the B batten of edge information fusion is: at first, merge marginal information; Then, select Free Transform model,, and improve image similarity tolerance by the improvement of level method for registering speed, degree of accuracy and the robustness of the multiresolution B batten control mesh from coarse to meticulous based on the B batten.

4. as claim 1 or 3 described non-rigid image registration algorithms based on implicit shape representation and edge information fusion, it is characterized in that, the fusion method of described marginal information is: at first, reference picture and global registration image are carried out operator filtering respectively, obtain edge-detected image separately, pixel value in the edge-detected image is suppressed greater than the strong edge of certain threshold level by the method that multiply by an inhibiting factor; Then, former figure and edge-detected image are carried out image co-registration by certain weight, form new image; At last, the image after merging is carried out registration by multi-resolution grid Free Transform method.

5. as claim 1 or 3 described non-rigid image registration algorithms based on implicit shape representation and edge information fusion, it is characterized in that, multi-resolution grid Free Transform method for registering based on the B batten is: at first, definition is 2 dimension tensor product forms of 1 dimension cubic B-spline based on the Free Transform of B batten, on reference picture, place sparse, the lattice point of rule, it is the reference mark, obtain the deformation territory by the motion that defines each reference mark then, use spline interpolation to assess the deformation values of calculating between the reference mark, thereby produce a local control, the conversion that the overall situation is level and smooth; Then, progressively reduce the spacing at reference mark, promptly progressively increase the resolution of grid, repeat same step and carry out registration; The registration transformation of representing local deformation by the resolution that improves the reference mark grid.

Images