Registration method and system for non-rigid multi-modal medical image

Abstract

The present invention discloses a registration method and system for a non-rigid multi-modal medical image. The registration method comprises: obtaining local descriptors of a reference image according to Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of the reference image; obtaining local descriptors of a floating image according to Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of the floating image; and finally obtaining a registration image according to the local descriptors of the reference image and the floating image. In the present, by using self-similarity of the multi-modal medical image and adopting the Zernike moment based local descriptor, the non-rigid multi-modal medical image registration is thus converted into the non-rigid mono-modal medical image registration, thereby greatly improving its accuracy and robustness.

Claims

1. A registration method for a non-rigid multi-modal medical image, comprising: according to Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of a reference image and a floating image, respectively obtaining a local descriptor of the reference image and a local descriptor of the floating image, and obtaining a registration image, wherein the method comprises the following steps: step 1: obtaining the local descriptor ZMLD.sup.A(i) of the reference image I.sup.A(i) according to a Zernike moment Z.sup.A.sub.00(i) of order 0 and repetition 0 and a Zernike moment Z.sup.A.sub.11(i) of order 1 and repetition 1 of a pixel point i in the reference image I.sup.A(i), wherein i is an integer from 1 to M, and M represents the size of the reference image and the floating image; step 2: obtaining the local descriptor ZMLD.sup.B(i) of the floating image I.sup.B(i) according to a Zernike moment Z.sup.8.sub.00(i) of order 0 and repetition 0 and a Zernike moment Z.sup.B.sub.11(i) of order 1 and repetition 1 of a pixel point i at the same position in the floating image I.sup.B(i); and step 3: establishing an objective function g(T) according to the local descriptor ZMLD.sup.A(i) of the reference image and the local descriptor ZMLD.sup.B(i) of the floating image; obtaining a transformation parameter according to the objective function, transforming the floating image according to the transformation parameter, and performing an interpolation process on a transformed floating image, thereby obtaining the registration image, wherein the step 2 further comprises the following sub-steps: step 2-1: obtaining a similarity distance D.sup.A.sub.00(i) corresponding to the Zernike moment of order 0 and repetition 0 and a similarity distance D.sup.A.sub.11(i) corresponding to the Zernike moment of order 1 and repetition 1 of the pixel point i to other pixel points in an image patch centered at the pixel point i in the reference image; simultaneously, obtaining a similarity distance D.sup.B.sub.00(i) corresponding to the Zernike moment of order 0 and repetition 0 and a similarity distance D.sup.B.sub.11(i) corresponding to the Zernike moment of order 1 and repetition 1 of the pixel point i to other pixel points in an image patch centered at the pixel point i in the floating image; step 2-2: obtaining the local descriptor ZMLD.sup.A(i) of the reference image and the local descriptor ZMLD.sup.B (i) of the floating image, ${\mathrm{ZMLD}}^A\left(i\right)=\exp \left(-\frac{D_{00}^A\left(i\right)}{{}_{00}^A\left(i\right)}-\frac{D_{11}^A\left(i\right)}{{}_{11}^A\left(i\right)}\right),{\mathrm{ZMLD}}^B\left(i\right)=\exp \left(-\frac{D_{00}^B\left(i\right)}{{}_{00}^B\left(i\right)}-\frac{D_{11}^B\left(i\right)}{{}_{11}^B\left(i\right)}\right),$ wherein h.sup.A.sub.00(i), h.sup.A.sub.11(i), h.sup.B.sub.00(i) and h.sup.B.sub.11(i) are decay parameters.

2. The registration method of claim 1, wherein in the step 2-1, the image patches in the step 2-2 all have a side length of 3, and in the step 2-2, $h_{00}^A\left(i\right)={\left\{{}_{00}^A\left(i\right)+c_1.Math.1.4826\mathrm{MED}\left[.Math.{}_{00}^A\left(i\right)-\mathrm{MED}\left(.Math.{}_{00}^A\left(i\right).Math.\right).Math.\right]\right\}}^2,{}_{00}^A\left(i\right)=c_2\frac{.Math.8.Math.Z_{00}^A\left(i\right).Math.-.Math..Math.Z_{00}^A\left(j\right).Math..Math.}{\sqrt{72}},h_{11}^A\left(i\right)={\left\{{}_{11}^A\left(i\right)+c_1.Math.1.4826\mathrm{MED}\left[.Math.{}_{11}^A\left(i\right)-\mathrm{MED}\left(.Math.{}_{11}^A\left(i\right).Math.\right).Math.\right]\right\}}^2,{}_{11}^A\left(i\right)=c_2\frac{.Math.8.Math.Z_{11}^A\left(i\right).Math.-.Math..Math.Z_{11}^A\left(j\right).Math..Math.}{\sqrt{72}},h_{00}^B\left(i\right)={\left\{{}_{00}^B\left(i\right)+c_1.Math.1.4826\mathrm{MED}\left[.Math.{}_{00}^B\left(i\right)-\mathrm{MED}\left(.Math.{}_{00}^B\left(i\right).Math.\right).Math.\right]\right\}}^2,{}_{00}^B\left(i\right)=c_2\frac{.Math.8.Math.Z_{00}^B\left(i\right).Math.-.Math..Math.Z_{00}^B\left(j\right).Math..Math.}{\sqrt{72}},h_{11}^B\left(i\right)={\left\{{}_{11}^B\left(i\right)+c_1.Math.1.4826\mathrm{MED}\left[.Math.{}_{11}^B\left(i\right)-\mathrm{MED}\left(.Math.{}_{11}^B\left(i\right).Math.\right).Math.\right]\right\}}^2,{}_{11}^B\left(i\right)=c_2\frac{.Math.8.Math.Z_{11}^B\left(i\right).Math.-.Math..Math.Z_{11}^B\left(j\right).Math..Math.}{\sqrt{72}},$ wherein MED() represents a median operator, c.sub.1 and c.sub.2 are constants of 0.5-1, and j is any pixel point in the image patch centered at the pixel point i.

3. A registration system based on the registration method of claim 2, wherein the registration system comprises a first Zernike moment module, a first descriptor module, a second Zernike moment module, a second descriptor module and a registration module; wherein the first Zernike moment module is configured to obtain Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of a reference image and output them to the first descriptor module, and the first descriptor module is configured to obtain local descriptors of the reference image and output them to the registration module, wherein the second Zernike moment module is configured to obtain Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of a floating image and output them to the second descriptor module, and the second descriptor module is configured to obtain local descriptors of the floating image and output them to the registration module, wherein the registration module is configured to obtain a registration image.

4. The registration method of claim 1, wherein the step 3 specifically comprises the following sub-steps: step 3-1: establishing the objective function g(T.sub.)=SSD+R(T.sub.), wherein SSD represents a sum of squared differences between the local descriptors ZMLD.sup.A(i) and ZMLD.sup.B(i), (0<<1) is a constant, R(T.sub.) represents a regularization term, and the number of iterations =1; iteratively solving the objective function g(T.sub.) to obtain the transformation parameter T.sub.; step 3-2: transforming the local descriptor ZMLD.sup.B(i) of the floating image according to the transformation parameter T.sub., performing the interpolation process on the transformed local descriptor, and updating the original local descriptor ZMLD.sup.B(i) with the local descriptor subjected to the interpolation process, =+1; iteratively solving the objective function g(T.sub.) to obtain a transform parameter T.sub.; step 3-3: if the number of iterations is greater than or equal to a threshold of the number of iterations and g(T.sub.)g(T.sub.-1), transforming the floating image according to the transformation parameter T.sub., and performing the interpolation process on the transformed floating image to obtain the registration image; otherwise, returning to the step 3-2.

5. The registration method of claim 4, wherein the similarity metric SSD in the step 3-1 is expressed as: $\mathrm{SSD}=\frac{\underset{i=1}{\overset{M}{.Math.}}{\left[{\mathrm{ZMLD}}^A\left(i\right)-{\mathrm{ZMLD}}^B\left(i\right)\right]}^2}{M}.$

6. A registration system based on the registration method of claim 5, wherein the registration system comprises a first Zernike moment module, a first descriptor module, a second Zernike moment module, a second descriptor module and a registration module; wherein the first Zernike moment module is configured to obtain Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of a reference image and output them to the first descriptor module, and the first descriptor module is configured to obtain local descriptors of the reference image and output them to the registration module, wherein the second Zernike moment module is configured to obtain Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of a floating image and output them to the second descriptor module, and the second descriptor module is configured to obtain local descriptors of the floating image and output them to the registration module, wherein the registration module is configured to obtain a registration image.

7. The registration method of claim 4, wherein an iteration solution is performed by a limited memory Broyden-Fletcher-Goldfarb-Shanno method or a gradient descent method, a transformation is performed by using a B-spline Free-form Deformation model, and the interpolation process is performed by a bilinear interpolation method or a B-spline interpolation method.

8. A registration system based on the registration method of claim 7, wherein the registration system comprises a first Zernike moment module, a first descriptor module, a second Zernike moment module, a second descriptor module and a registration module; wherein the first Zernike moment module is configured to obtain Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of a reference image and output them to the first descriptor module, and the first descriptor module is configured to obtain local descriptors of the reference image and output them to the registration module, wherein the second Zernike moment module is configured to obtain Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of a floating image and output them to the second descriptor module, and the second descriptor module is configured to obtain local descriptors of the floating image and output them to the registration module, wherein the registration module is configured to obtain a registration image.

9. A registration system based on the registration method of claim 4, wherein the registration system comprises a first Zernike moment module, a first descriptor module, a second Zernike moment module, a second descriptor module and a registration module; wherein the first Zernike moment module is configured to obtain Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of a reference image and output them to the first descriptor module, and the first descriptor module is configured to obtain local descriptors of the reference image and output them to the registration module, wherein the second Zernike moment module is configured to obtain Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of a floating image and output them to the second descriptor module, and the second descriptor module is configured to obtain local descriptors of the floating image and output them to the registration module, wherein the registration module is configured to obtain a registration image.

10. A registration system based on the registration method of claim 1, wherein the registration system comprises a first Zernike moment module, a first descriptor module, a second Zernike moment module, a second descriptor module and a registration module; wherein the first Zernike moment module is configured to obtain Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of a reference image and output them to the first descriptor module, and the first descriptor module is configured to obtain local descriptors of the reference image and output them to the registration module, wherein the second Zernike moment module is configured to obtain Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of a floating image and output them to the second descriptor module, and the second descriptor module is configured to obtain local descriptors of the floating image and output them to the registration module, wherein the registration module is configured to obtain a registration image.

11. The registration system of claim 10, wherein the registration module includes a solving unit and a determining unit, in which the solving unit is configured to construct an objective function according to the local descriptors of the reference image and the floating image, obtain a transformation parameter and output the transformation parameter to the determining unit, and the determining unit is configured to determine whether the objective function meets an iteration stopping criterion, if not, the local descriptor of the floating image is transformed according to the transformation parameter, interpolation process is performed on the transformed local descriptor, the original local descriptor of the floating image is updated with the local descriptor of the floating image subjected to the interpolation process; if yes, the floating image is transformed according to the transformation parameter and interpolation process is performed on the floating image to obtain the registration image.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 is a schematic structural diagram of a registration system for a non-rigid multi-modal medical image according to the present invention.

(2) FIG. 2 is a flow chat of a registration method for a non-rigid multi-modal medical image according to Embodiment 1 of the present invention.

(3) FIG. 3a is a reference T2 image used in Embodiment 1 and Comparative Examples 2-4 of the present invention.

(4) FIG. 3b is a floating T1 image used in Embodiment 1 and Comparative Examples 2-4 of the present invention.

(5) FIG. 3c is a floating Gad image used in Embodiment 1 and Comparative Examples 2-4 of the present invention.

(6) FIG. 3d is a registration T1-T2 image obtained by the method according to Embodiment 1 of the present invention.

(7) FIG. 3e is a registration T1-T2 image obtained by the method according to Comparative Example 4 of the present invention.

(8) FIG. 3f is a registration T1-T2 image obtained by the method according to Comparative Example 3 of the present invention.

(9) FIG. 3g is a registration T1-T2 image obtained by the method according to Comparative Example 2 of the present invention.

(10) FIG. 3h is a registration Gad-T2 image obtained by the method according to Embodiment 1 of the present invention.

(11) FIG. 3i is a registration Gad-T2 image obtained by the method according to Comparative Example 4 of the present invention.

(12) FIG. 3j is a registration Gad-T2 image obtained by the method according to Comparative Example 3 of the present invention.

(13) FIG. 3k is a registration Gad-T2 image obtained by the method according to Comparative Example 2 of the present invention.

(14) FIG. 4a is a reference T1-weighted MR image used in Embodiment 1 and Comparative Examples 2-4 of the present invention.

(15) FIG. 4b is a floating image CT used in Embodiment 1 and Comparative Examples 2-4 of the present invention.

(16) FIG. 4c is a registration image obtained by the method according to Embodiment 1 of the present invention.

(17) FIG. 4d is a registration image obtained by the method according to Comparative Example 4 of the present invention.

(18) FIG. 4e is a registration image obtained by the method according to Comparative Example 3 of the present invention.

(19) FIG. 4f is a registration image obtained by the method according to Comparative Example 2 of the present invention.

DETAILED DESCRIPTION OF THE EMBODIMENTS

(20) For clear understanding of the objectives, features and advantages of the present invention, detailed description of the present invention will be given below in conjunction with accompanying drawings and specific embodiments. It should be noted that the embodiments described herein are only meant to explain the present invention, and not to limit the scope of the present invention. Furthermore, the technical features related to the embodiments of the invention described below can be mutually combined if they are not found to be mutually exclusive.

(21) The present invention provides a registration system for a non-rigid multi-modal medical image, which comprises a first Zernike moment module, a first descriptor module, a second Zernike moment module, a second descriptor module and a registration module, in which the registration module further includes a solving unit and a determining unit as shown in FIG. 1.

(22) An output end of the first Zernike moment module is connected to an input end of the first descriptor module, an output end of the first descriptor module is connected to a first input end of the solving unit, an output end of the second Zernike moment module is connected to an input end of the second descriptor module, and an output end of the second descriptor module is connected to a second input end of the solving unit; an output end of the solving unit is connected to an input end of the determining unit, and an output end of the determining unit is connected to an input end of the second Zernike moment module.

(23) The first Zernike moment module is configured to obtain and output Zernike moments Z.sup.A.sub.00(i) of order 0 and repetition 0 and Zernike moments Z.sup.A.sub.11(i) of order 1 and repetition 1 of a reference image A(i), and the first descriptor module is configured to obtain and output local descriptors ZMLD.sup.A(i) of the reference image; the second Zernike moment module is configured to obtain and output Zernike moments Z.sup.B.sub.00(i) of order 0 and repetition 0 and Zernike moments Z.sup.B.sub.11(i) of order 1 and repetition 1 of a floating image I.sup.B(i), the second descriptor module is configured to obtain and output local descriptors ZMLD.sup.B(i) of the floating image, the solving unit is configured to construct an objective function g(T.sub.) according to the local descriptors of the reference image and the floating image and obtain a transformation parameter T.sub., and the determining unit is configured to determine whether the objective function meets an iteration stopping criterion. If it is not, the local descriptor of the floating image is transformed according to the transformation parameter, interpolation process is performed on the transformed local descriptor, the original local descriptor of the floating image is updated with the local descriptor of the floating image subjected to interpolation process; Otherwise, the floating image is transformed according to the transformation parameter and interpolation process is performed on the floating image to obtain the registration image.

(24) Specifically, registration of the non-rigid multi-modal medical image by the registration system includes the following steps:

(25) Step 1: obtaining a local descriptor ZMLD.sup.A(i) of the reference image I.sup.A(i) according to the Zernike moment Z.sup.A.sub.00(i) of order 0 and repetition 0 and the Zernike moment Z.sup.A.sub.11(i) of order 1 and repetition 1 of the pixel point i in the reference image I.sup.A(i), wherein i is an integer from 1 to M, and M is the size of the reference image and the floating image (that is, when the images have a length of X and a width of Y, M=XY).

(26) Step 1-1: calculating the Zernike moment by the following formulas:

(27) $\begin{array}{cc}{Z}_{\mathrm{nm}}=\frac{n+1}{{}_N}\underset{x=0}{\overset{N-1}{.Math.}}\underset{y=0}{\overset{N-1}{.Math.}}{R}_{\mathrm{nm}}\left({}_{\mathrm{xy}}\right)e^{-\mathrm{jm}{}_{\mathrm{xy}}}f\left(x,y\right)& \left(1\right)\\ R_{\mathrm{nm}}\left(\right)=\underset{s=0}{\overset{\left(n-.Math.m.Math.\right)/2}{.Math.}}{\left(-1\right)}^s\frac{\left(n-s\right)!}{s!\left(\left(\left(n+.Math.m.Math.\right)/2\right)-s\right)!(\left(\left(n-.Math.m.Math.\right)/2-s\right)!}{}^{n-2s}{}_{\mathrm{xy}}=\frac{\sqrt{{\left(2x-N+1\right)}^2+{\left(2y-N+1\right)}^2}}{N},{}_{\mathrm{xy}}={\tan }^{-1}\left(\frac{N-1-2y}{2x-N+1}\right)& \left(2\right)\end{array}$
wherein N represents the side length of the first image patch centered at the pixel point i (N is usually an odd number between 3 and 11 and can be selected by taking into comprehensive consideration of the complexity, calculation efficiency and image registration accuracy; when the image complexity is high, the registration accuracy requirement is not high or a high calculation efficiency is required, a smaller value such as 3 or 5 may be taken, otherwise, a larger value may be taken); (x,y) represents an image function of the first image patch centered at the pixel point i, and and respectively represent a polar angle and a polar axis of the pixel point in the image function (x,y) (when the pixel point i is located at the edges of the reference image and the floating image, an image function value of a pixel point in the first image patch which exceeds the range of the original reference image and floating image is filled with the image function value of the pixel adjacent thereto); x and y represent coordinates of any pixel in the image function (x,y), relative to the center point i of the first image patch, in the first image patch; j={square root over (1)}, .sub.N is a normalization factor, .sub.N=N.sup.2; m represents the repetition of the Zernike moment, n represents the order of the Zernike moment, and s=0(n|m|)/2 (in the present invention, s=0, R.sub.00()=1 and R.sub.11()= since n=m=1 or n=m=0).

(28) When n=m=0, Zernike moments of order 0 and repetition 0 of the image are obtained, an when n=m=1, Zernike moments of order 1 and repetition 1 of the image are obtained; thus, the Zernike moment Z.sup.A.sub.00(i) of order 0 and repetition 0 and the Zernike moment Z.sup.A.sub.11(i) of order 1 and repetition 1 of the pixel point i in the reference image can be obtained by the above formulas, and meanwhile, the Zernike moment Z.sup.B.sub.00(i) of order 0 and repetition 0 and the Zernike moment Z.sup.B.sub.11(i) of order 1 and repetition 1 of the pixel point i at the same position in the floating image are obtained.

(29) Step 1-2: in the reference image, obtaining similarity distances D.sup.A.sub.00(i) and D.sup.A.sub.11(i) of the pixel point i to other pixel points in the second image patch centered at the pixel point i, the calculation formula being as follows:

(30) $D_{\mathrm{nm}}\left(i\right)=\frac{1}{{\left(.Math.r.Math.\right)}^2-1}\underset{jr}{.Math.}{\left(.Math.Z_{\mathrm{nm}}\left(i\right).Math.-.Math.Z_{\mathrm{nm}}\left(j\right).Math.\right)}^2,$
wherein |r| represent the side length of the second image patch, which is usually an odd number larger than or equal to 3,j is any pixel point other than the pixel point i in the second image block, and |Z.sub.nm(i)| and |Z.sub.nm(j)| are modulus of the Zernike moments of order n and repetition m of the pixel point i and the pixel point j, respectively; when the pixel point i is located at the edge of the reference image, the modulus of the Zernike moment of a pixel point in the second image patch that exceeds the range of the reference image is filled with the modulus of the Zernike moment of the pixel point adjacent thereto; thus, D.sup.A.sub.00(i), D.sup.A.sub.11(i), D.sup.B.sub.00(i) and D.sup.B.sub.11(i) can be obtained.

(31) Steps 1-3: obtaining the local descriptors ZMLD, the formula being as follows:

(32) $\mathrm{ZMLD}\left(i\right)=\exp \left(-\frac{D_{00}\left(i\right)}{h_{00}\left(i\right)}-\frac{D_{11}\left(i\right)}{h_{11}\left(i\right)}\right),$
wherein h.sub.00(i,r) and h.sub.11(i,r) are decay parameters, and the calculation formula thereof is as follows:
h.sub.nm(i)=[.sub.nm.sup.l(i)+.sub.nm.sup.g(i)].sup.2, n=m=0 or n=m=1,
for example, when |r|=3, .sub.nm.sup.l(i) and .sub.nm.sup.g(i) are respectively expressed as:

(33) ${}_{\mathrm{nm}}^g\left(i\right)=c_1.Math.1.4826\mathrm{MED}\left[.Math.{}_{\mathrm{nm}}^l\left(i\right)-\mathrm{MED}\left(.Math.{}_{\mathrm{nm}}^l\left(i\right).Math.\right).Math.\right],{}_{\mathrm{nm}}^l\left(i\right)=c_2{\mathrm{.Math.}}_i=c_2\frac{.Math.8.Math.Z_{\mathrm{nm}}\left(i\right).Math.-\underset{jr}{.Math.}.Math.Z_{\mathrm{nm}}\left(j\right).Math..Math.}{\sqrt{72}},$
wherein MED() represents the median operator, and c.sub.1 and c.sub.2 are adjustment coefficients with 0.5<c.sub.11 and 0.5<c.sub.21.

(34) According to the above formulas, local descriptors ZMLD.sup.A(i) of the reference image can be respectively obtained.

(35) Step 2: by the same calculation method as in the step 1, obtaining a local descriptor ZMLD.sup.B(i) of the floating image according to the Zernike moment Z.sup.B.sub.00(i) of order 0 and repetition 0 and the Zernike moment Z.sup.B.sub.11(i) of order 1 and repetition 1 of the pixel point i at the same position in the floating image.

(36) Step 3, establishing an objective function according to the local descriptors ZMLD.sup.A(i) and ZMLD.sup.B(i), and finally achieving registration of the reference image and the floating image; taking an example of establishing an objective function with the B-spline Free-form Deformation (FFD) model as the transformation model, the registration process is described, the registration process specifically comprising the following sub-steps:

(37) Step 3-1: establishing an objective function g(T.sub.)=SSD+R(T.sub.), in which the smaller the value of g(T.sub.), the more similar the floating image is to the reference image, SSD represents the similarity metric between the local descriptors ZMLD.sup.A(i) and ZMLD.sup.B(i),

(38) $\mathrm{SSD}=\frac{\underset{i=1}{\overset{M}{.Math.}}{\left[{\mathrm{ZMLD}}^A\left(i\right)-{\mathrm{ZMLD}}^B\left(i\right)\right]}^2}{M},$
wherein (0<<1) represents a weighting parameter, T.sub. is a third-order B-spline function related to the coordinates (x, y) of the pixel point i, which represents a transformation parameter for transforming the floating image into a registration image; R(T.sub.) is a regularization term, and the calculation formula thereof is as follow:

(39) $R\left(T_{}\right)=\frac{1}{XY}\underset{x=1}{\overset{X}{.Math.}}\underset{y=1}{\overset{Y}{.Math.}}\left[{\left(\frac{{}^2{T}_{}\left(x,y\right)}{x^2}\right)}^2+2{\left(\frac{{}^2{T}_{}\left(x,y\right)}{\mathrm{xy}}\right)}^2+{\left(\frac{{}^2{T}_{}\left(x,y\right)}{y^2}\right)}^2\right],$
wherein, x and y respectively represents the horizontal coordinate and vertical coordinate of the pixel point i in the floating image, and X and Y respectively represent the length and width of the floating image, then XY=M; and the number of iterations is initially set to be 1; the objective function g(T.sub.) is solved iteratively by the limited memory Broyden-Fletcher-Goldfarb-Shanno method or the gradient descent method to obtain a transform parameter T.sub.(x,y).

(40) Step 3-2: transforming the local descriptor ZMLD.sup.B(i) of the floating image according to the transformation parameter T.sub.(x,y), performing interpolation process on the transformed local descriptor by the bilinear interpolation method or the B-spline interpolation method, and updating the original local descriptor ZMLD.sup.B(i) with the local descriptor subjected to the interpolation process, =+1; iteratively solving the objective function g(T.sub.) to obtain a transform parameter T.sub..

(41) Step 3-3, if the number of iterations r is greater than or equal to a threshold (generally, 5020) of the number of iterations and g(T.sub.)g(T.sub.-1), transforming the floating image according to the transformation parameter T.sub.(x,y), and performing interpolation process on the transformed floating image by the bilinear interpolation method or the B-spline interpolation method to obtain the registration image I.sup.B(i), thereby completing image registration; otherwise; returning to the step 3-2.

Embodiment 1

(42) The present invention provides a registration method for a non-rigid multi-modal medical image, and as shown in FIG. 1, the method comprises the following steps:

(43) Step 1: calculating Zernike moments of order 0 and repetition 0 and Zernike moments of order 1 and repetition 1 of the first image patch centered at each pixel point i in the reference image and the floating image according to the formulas (1) and (2); finally obtaining moment features Z.sup.A.sub.00(i) and Z.sup.A.sub.11(i) of the reference image I.sub.A and moment features Z.sup.B.sub.00(i) and Z.sup.B.sub.11(i) of the floating image I.sub.B.

(44) Step 2: calculating Zernike moment based local descriptors (referred to as ZMLD) according to the moment features.

(45) Step 2-1: based on the image self-similarity, calculating similarity distances D.sup.A.sub.00(i), D.sup.A.sub.11(i), D.sup.B.sub.00(i) and D.sup.B.sub.11(i) of the pixel point i to other pixel points in the second image patch centered at the pixel point i in the reference image and the floating image, the formula being as follow:

(46) 0$D_{\mathrm{nm}}\left(i,r\right)=\frac{1}{{\left(.Math.r.Math.\right)}^2-1}\underset{jr}{.Math.}{\left(.Math.Z_{\mathrm{nm}}\left(i\right).Math.-.Math.Z_{\mathrm{nm}}\left(j\right).Math.\right)}^2,$
n=m=0 or n=m=1
wherein |Z.sub.nm(i)| and |Z.sub.nm(j)| are modulus of Zernike moments of order n and repetition m of the pixel point i and the pixel point j, respectively; |r| represent the side length of the second image patch, and when |r|=3, a high registration accuracy can be ensured; in the present embodiment, when |r|>3, the registration accuracy is not increased and the computational efficiency decreases, and when |r|<3, the registration accuracy is obviously affected; j is any pixel point other than the pixel point i in the second image patch.

(47) Step 2-2: obtaining local descriptors ZMLD, the calculation formula being as follows:

(48) $\mathrm{ZMLD}\left(i\right)=\exp \left(-\frac{D_{00}\left(i\right)}{h_{00}\left(i\right)}-\frac{D_{11}\left(i\right)}{h_{11}\left(i\right)}\right),$
wherein h.sub.00(i,r) and h.sub.11(i,r) are decay factors, and the calculation formula thereof is as follows:
h.sub.nm(i)=[.sub.nm.sup.l(i)+.sub.nm.sup.g(i)].sup.2, n=m=0 or n=m=1

(49) Since |r|=3 in the present embodiment, .sub.nm.sup.l(i) and .sub.nm.sup.g(i) are respectively expressed as below:

(50) ${}_{\mathrm{nm}}^g\left(i\right)=c_1.Math.1.4826\mathrm{MED}\left[.Math.{}_{\mathrm{nm}}^l\left(i\right)-\mathrm{MED}\left(.Math.{}_{\mathrm{nm}}^l\left(i\right).Math.\right).Math.\right],{}_{\mathrm{nm}}^l\left(i\right)=c_2{\mathrm{.Math.}}_i=c_2\frac{.Math.8.Math.Z_{\mathrm{nm}}\left(i\right).Math.-\underset{jr}{.Math.}.Math.Z_{\mathrm{nm}}\left(j\right).Math..Math.}{\sqrt{72}},$
wherein MED() represents the median operator, and c.sub.1 and c.sub.2 represent adjustment coefficients (in present embodiment, c.sub.1=c.sub.2=0.8).

(51) According to the above formulas, local descriptors ZMLD.sup.A(i) of the reference image and local descriptors ZMLD.sup.B(i) of the floating image can be respectively obtained.

(52) Step 3-1: establishing an objective function g(T.sub.)=SSD+R(T.sub.) by taking the B-spline Free-form Deformation (FFD) model as the transformation model, in which SSD represents the similarity metric between the local descriptors ZMLD.sup.A(i) and ZMLD.sup.B(i),

(53) $\mathrm{SSD}=\frac{\underset{i=1}{\overset{M}{.Math.}}{\left[{\mathrm{ZMLD}}^A\left(i\right)-{\mathrm{ZMLD}}^B\left(i\right)\right]}^2}{M},$
wherein M represents the image size (in the present embodiment, since both the length X and the width Y of the reference image and the floating image are 256, M=XY=256256); a represents a weighting parameter (=0.015), T represents a transformation parameter for transforming the floating image into a registration image; and R(T) represents a regularization term, the calculation formula thereof being as follow:

(54) $R\left(T_{}\right)=\frac{1}{XY}\underset{x=1}{\overset{X}{.Math.}}\underset{y=1}{\overset{Y}{.Math.}}\left[{\left(\frac{{}^2{T}_{}\left(x,y\right)}{x^2}\right)}^2+2{\left(\frac{{}^2{T}_{}\left(x,y\right)}{\mathrm{xy}}\right)}^2+{\left(\frac{{}^2{T}_{}\left(x,y\right)}{y^2}\right)}^2\right],$

(55) Step 3-2: iteratively solving the objective function g(T) with the L-BFGS method with the aim of minimizing the value of g(T), and when g(T) gets the minimum value and the number of iterations is greater than or equal to a threshold of the number of iterations, stopping the iteration to obtain a transformation parameter T.

(56) Step 3-3: transforming the floating image by using the transformation parameter T obtained in the step 3-2, and performing interpolation process by the bilinear interpolation method to obtain a registration image corresponding to the floating image, thereby completing image registration.

Comparative Example 1

(57) The registration was carried out according to the NMI method in the prior art (Pattern Recognit. 32(1) (1999) 71-86.).

Comparative Example 2

(58) The registration was carried out according to the ESSD method in the prior art (Med. Image Anal. 16(1) (2012) 1-17.), in which the specific parameters are as follows: selecting 77 image patches, and calculating the entropy corresponding to the image patches by using the Gaussian weight, local normalization method and Parzen-window estimation, thereby obtaining the ESSD corresponding to the entire image.

Comparative Example 3

(59) The registration was carried out according to the WLD method in the prior art (Sensors 13(6) (2013) 7599-7613), in which the specific parameters are as follows: the radius for WLD calculation is R=1 and R=2, the patch size for the similarity metric is 77, and the weight term =0.01.

Comparative Example 4

(60) The registration was carried out according to the MIND method in the prior art (Med. Image Anal. 16(7) (2012) 1423-1435), in which the specific parameters are as follow: the patch size is 33.

(61) Analysis of Results

(62) In order to further embody the advantages of the present invention, comparison of the registration accuracy of Embodiment 1 and Comparative Examples 1-4 was made. The registration accuracy is evaluated using the target registration error (TRE), wherein TRE is defined as:

(63) $\mathrm{TRE}=\frac{{.Math.T_s\left(R\right)-T_c\left(R\right).Math.}_2}{.Math.R.Math.}$
wherein T.sub.s represents the random deformation and is also the gold standard for evaluation, T.sub.c represents the deformation obtained by the registration method, and R represents the number of pixels used for image registration performance evaluation.

(64) The simulated MR image was used for the registration accuracy test. The simulated T1, T2 and PD weighted MR images used in Example 1 were taken from the BrainWeb database. Table 1 lists the standard deviation and mean of the TRE obtained by all evaluated methods. It can be seen from Table 1 that when registration is performed on different MR images, the TRE provided in Embodiment 1 has lower mean and standard deviation than other methods, which indicates that the method proposed by the present invention has the highest registration accuracy among all the compared methods.

(65) TABLE-US-00001 TABLE 1 Comparison of TREs (mm) for all evaluated methods performed on T1-T2, PD-T2 and T1-PD images Registration method T1-T2 PD-T2 T1-PD Standard Standard Standard Mean deviation Mean deviation Mean deviation No registration 4.35 3.32 4.50 3.52 4.49 2.27 Comparative 1.81 2.33 1.95 1.98 2.12 1.94 Example 1 Comparative 1.53 1.41 1.92 1.66 1.97 2.19 Example 2 Comparative 1.46 1.53 0.97 1.10 1.63 1.35 Example 3 Comparative 1.02 1.16 0.96 0.99 0.97 1.06 Example 4 Embodiment 1 0.65 0.58 0.73 0.62 0.89 0.95

(66) To more intuitively show the superiority of the present invention over the rest of the methods, visual comparisons of the registration images corresponding to Embodiment 1 and Comparative Examples 2-4 were provided, as shown in FIG. 3. FIG. 3a shows a reference T2 image, FIG. 3b shows a floating T1 image, FIG. 3c shows a floating Gad image, FIG. 3d shows a registration T1-T2 image obtained by the method in Embodiment 1, FIG. 3e shows a registration T1-T2 image obtained by the method in Comparative Example 4, FIG. 3f shows a registration T1-T2 image obtained by the method in Comparative Example 3, FIG. 3g shows a registration T1-T2 image obtained by the method in Comparative Example 2, FIG. 3h shows a registration Gad-T2 image obtained by the method in Embodiment 1, FIG. 3i shows a registration Gad-T2 image obtained by the method in Comparative Example 4, FIG. 3j shows a registration Gad-T2 image obtained by the method in Comparative Example 3, and FIG. 3k shows a registration Gad-T2 image obtained by the method in Comparative Example 2.

(67) By taking the partial contour of the image shown in the box as an example, for the registration T1-T2 image, the deformation of the lowermost portion of the contour cannot be effectively corrected in Comparative Examples 2-4. For the registration Gad-T2 image, Comparative Examples 3-4 do not provide a good correction for the deformation of the right portion of the box, while the above deformation can be effectively corrected in Embodiment 1. It can be seen that the registration image obtained in Embodiment 1 is more similar to the reference image than those obtained in Comparative Examples 2-4. The above visual comparison demonstrates that the local descriptor of the present invention has rotation invariance, and more accurate representation of the image, so that the present invention is superior in the non-rigid multi-mode medical image registration.

(68) The registration accuracy of the listed methods is evaluated by using the T1, T2 and Grad weighted MR images in the Altas database, and Table 2 gives corresponding TRE results of the respective methods. The results in the Table 2 indicate that the method proposed in the present invention can provide a lower TRE than other methods, and thus has a higher registration accuracy.

(69) TABLE-US-00002 TABLE 2 Comparison of TREs (mm) for all evaluated methods performed on T1-T2, Gad-T2 and Gad-T1 images Registration method T1-T2 Gad-T2 Gad-T1 Standard Standard Standard Mean deviation Mean deviation Mean deviation No registration 5.21 2.70 4.86 2.41 6.50 2.15 Comparative 3.38 2.15 3.17 1.94 3.86 1.98 Example 1 Comparative 2.83 1.89 3.05 1.76 3.18 1.82 Example 2 Comparative 2.61 1.77 2.74 1.62 2.99 1.66 Example 3 Comparative 2.45 1.58 2.34 1.56 2.60 1.53 Example 4 Embodiment 1 2.12 1.40 2.19 1.24 2.44 1.37

(70) For the comparison of the registration accuracy between CT and MR images, five kinds of random deformation processing on the floating CT images were performed. FIG. 4 shows registration results of real CT and MR images by the methods of Embodiment 1 and Comparative Examples 2-4. FIG. 4a shows a reference T1 weighted MR image, FIG. 4b shows a floating CT image, FIG. 4c shows a registration image obtained by the method of Embodiment 1, FIG. 4d shows a registration image obtained by the method of Comparative Example 4, FIG. 4e shows a registration image obtained by the method of Comparative Example 3, and FIG. 4f shows a registration image obtained by the method of Comparative Example 2. It can be seen that it is difficult for the methods in Comparative Examples 3-4 to effectively correct the deformation of the contour in the box, and the outermost contour of the registration image provided in Comparative Example 2 is partially distorted, while the method in Embodiment 1 can provide a more satisfactory registration image. Table 3 shows the average value of the TRE in registration of each group of real CT image and MR image, in which images in Group 5 are the real CT image and MR image used in FIG. 4.

(71) TABLE-US-00003 TABLE 3 Comparison of TREs (mm) for all evaluated methods performed on CT-MR images TRE Registration method Group 1 Group 2 Group 3 Group 4 Group 5 No registration 3.81 5.74 3.94 4.49 5.67 Comparative Example 1 3.19 4.29 3.48 3.99 4.64 Comparative Example 2 2.92 3.87 3.24 3.41 4.06 Comparative Example 3 2.81 3.66 2.92 3.36 3.93 Comparative Example 4 2.80 3.06 2.97 2.63 3.09 Embodiment 1 2.64 2.50 2.89 2.52 2.75

(72) It can be seen from Table 3 that the method of the present invention can achieve a lower TRE in registration of the CT-MR images of Groups 1-5 than other registration methods, which indicates that the method of the present invention can achieve higher registration accuracy in registration of the CT-MR images than the algorithms in the Comparative Examples.

(73) It should be readily understood to those skilled in the art that the above description is only preferred embodiments of the present invention, and does not limit the scope of the present invention. Any change, equivalent substitution and modification made without departing from the spirit and scope of the present invention should be included within the scope of the protection of the present invention.