CN101666865B - Method for registering diffusion tensor nuclear magnetic resonance image in local quick traveling mode - Google Patents

Abstract

The invention relates to a method for registrating a diffusion tensor nuclear magnetic resonance image in a local quick traveling mode. The method is technically characterized by comprising the following steps: firstly, extracting the morphological character of certain pixel periphery in a diffusion tensor image by a local FM algorithm to obtain a local FM time pattern of each interested pixel; secondly, registrating a model by triple B sample strips to establish a deformation field and correcting the local FM time characteristic pattern; and considering the deformation and the reorientation of a template pattern while registrating, correcting the deformation field by using the corrected template time characteristic pattern and target time characteristic pattern which are obtained by the local FM algorithm and registrating a target pattern and a template pattern. Compared with other methods, the registration method based on the FM mode can extract richer and more effective tensor characteristics, thereby better defining the corresponding relationships of points among images and generating more stable and accurate registration result.

Description

A kind of method for registrating diffusion tensor nuclear magnetic resonance image of local quick traveling mode

Technical field:

The present invention relates to a kind of method for registrating diffusion tensor nuclear magnetic resonance image of local quick traveling mode, belong to brain science, fields such as computer vision, image understanding and pattern recognition.Be suitable for based on the brain structure of DTI image and the research of function the Clinics and Practices of cerebral nerve property relevant disease, and the fields such as art preplanning of clinical neurosurgery.

Background technology:

The dispersion tensor nuclear magnetic resonance, NMR image (DTI) that shakes makes the visual and quantitative analysis of three-dimensional brain white matter nerve fiber become possibility.It is in multiple sclerosis, apoplexy, and senile dementia, schizophrenia, epilepsy, the successful Application in the research of nerve relevant diseases such as the cerebral tumor makes the DTI technology become one of most popular nmr imaging technique in brain research aspect over nearly 15 years.

In theory research and practical application; For the DTI image that relatively obtains from different objects or same object different time points; Need at first these image registrations to same with reference in the space, so be an important task with the spatial registration of external DTI image in the body.Because the base pixel unit of DTI image is a tensor, so the registration of the registration of DTI image and common scalogram picture is very different.In the past; People's some yardstick information that propose from the tensor image of utilizing more, as fractional anisotropy (Fractional Anisotropy, FA), apparent diffusion coefficient (Apparent Diffusion Coefficient; ADC) and average diffusion coefficient (Mean diffusivity; MD) etc., perhaps only utilize the tensor information of current pixel, come the similarity between the compared pixels.The shortcoming of these class methods is that the former has only utilized part tensor information, does not make full use of structure, the directional information of the cerebral white matter that the tensor image provided; Though the latter has utilized tensor information, but be to calculate, do not consider the dependency of neighborhood, also lacked the of paramount importance information that the DTI image is provided to a great extent about the white matter nerve fiber trend based on current pixel.Deficiency to above method; In this invention; We adopt and a kind ofly advance fast based on the part that (Fast Marching, algorithm FM) extract the morphological characteristic of pixel periphery, promptly in whole brain; Each interested pixel is calculated its local FM time diagram, as the characteristic of tensor image registration.With respect to above-mentioned characteristic; This disperse time diagram based on local FM algorithm is not only considered the tensor direction; Also utilize the attribute of the adjacent tensor information of each pixel as uniqueness; Thereby reflected the comprehensive characteristics of image of complete sum more, can better come the registration of guide image as the similarity measurement characteristic.

Summary of the invention:

For fear of the weak point of prior art, the present invention proposes a kind of method for registrating diffusion tensor nuclear magnetic resonance image based on local quick traveling mode, with reference to (the Fast Marc that advances fast ^hIng, the method for registering of the DTI image that FM) proposes on the basis of algorithm.

A kind of dispersion tensor method for registering images based on local quick traveling mode is characterized in that step is following:

Step 1: with resolution R the DTI image is carried out down-sampling, described resolution is R=5;

Step 2: as the part anisotropy value FA of sampled pixel during greater than threshold value, utilize the communication mode of method Simulated Water molecule in its neighborhood of quick traveling mode FastMarching-FM, calculate the local FM time diagram of this sampled pixel, concrete steps are:

Step a: with gait of march $v_y\left(r\right)=v_y^{\mathrm{Tensor}}\left(r\right)+{\mathrm{\η\; v}}_y^{\mathrm{Inertia}}\left(r\right)$ Be advanced to other candidate points of neighborhood y, wherein v on every side from a y ' _y ^Tensor(r) be main gait of march, v _y ^Inertia(r) be that main gait of march peace slides the into compromise coefficient between the speed for level and smooth gait of march, η;

Described $v_y^{\mathrm{Tensor}}\left(r\right)=w{r}^T{D}_{y^{\′}}r,$ W eliminates the coefficient of advancing fast in cerebrospinal fluid inside, r:y ' → y, r ^TBe the transposition of r, D _{Y '}Dispersion tensor for a y ';

Described w=1/ (1+exp (α (FA-β))), wherein α and β regulate the constant of FA to advancing fast and controlling;

Described $v_y^{\mathrm{Inertia}}\left(r\right)=v_{y^{\′}}(r_{y^{\′}}\·r),$ R wherein _{Y '}And v _{Y '}Be respectively direct of travel and the speed of current point y ', two dot products of ". " expression;

Step b: the t time of advent that calculates each consecutive points y of y ' outside _y=t _{Y '}+ d/v _y, wherein, t _{Y '}Be the time that arrives y ', v _yIt is main gait of march;

Step c: as new forward face point, begin to repeat to obtain the local FM time diagram that develops forward from pixel x from step a then with those consecutive points y of the minimum time of advent in y ' outside; Multiple scope is for being the centre of sphere with pixel x point, and radius is in the ball of 16mm;

Step 3: utilize the local FM time diagram and the part anisotropy value FA figure that obtain to make characteristic, set up DTI template image D _TWith target image D _SBetween the energy equation of deformation field f:

$E\left(f\right)=\underset{x\∈\mathrm{\Ω}}{\mathrm{\Σ}}\{\underset{y\∈N\left(x\right)}{\mathrm{\Σ}}{(t_S(f\left(x\right)+x,f\left(y\right)+y)-t_T(x,y)(1+\frac{{|f\left(y\right)+y-f\left(x\right)-x|}_{x\→y}-{|y-x|}_{x\→y}}{{|y-x|}_{x\→y}}))}^2$

$+\mathrm{\μ}{({\mathrm{FA}}_S(f\left(x\right)+x)-{\mathrm{FA}}_T\left(x\right))}^2\}.$

Wherein: t _T(x y) is the traveling time of template image T the inside from pixel x to pixel y, t _S(f (x)+x, f (y)+y) are the traveling times that target figure S the inside pixel f (x)+x arrives pixel f (y)+y behind the consideration deformation field f, || _{X → y}Be the distance from x to y along travel track, the neighborhood of N (x) represent pixel x, FA _T(x) be the part anisotropic value of template image T the inside pixel x, FA _S(f (x)+x) is the part anisotropic value of target figure S the inside pixel f (x)+x behind the consideration deformation field f, and μ is the compromise coefficient;

Step 4: deformation field f adopts the cubic B-spline registration model to come modeling, and to the control point c of each cubic B-spline, the estimated energy equation is for the gradient at control point

Use $v\left(c\right)=v\left(c\right)-\mathrm{\ϵ}\∂E/\∂c$ Upgrade the deformation values at control point, and upgrade the deformation field that influenced by this control point; Iteration is carried out at all control point, to upgrade deformation field;

Step 5: repeating step 4 is up to convergence; Accomplish under the resolution r after the registration, if r=1 then forwarded for the 6th step to; Otherwise carry out the cubic B-spline up-sampling; A related variation field of more calculating it on the high-resolution, after obtaining deformation field, adopt the reorientation method of Xu that the tensor image is carried out redirect operation; And r=r-1 is set, forwarding for the 2nd step then to carries out next resolution processes;

Step 6: with the final deformation field of cubic B-spline Model Calculation that produces; And the reorientation method redirection target image of employing Xu is to the template space.

Described part anisotropy value FA is 0.25 greater than threshold value.

Described η selects 0.9.

Described α selects 50.

Described β selects 0.3.

Described μ selects 1.0.

The method for registrating diffusion tensor nuclear magnetic resonance image based on local quick traveling mode that the present invention proposes is not only considered the tensor direction in registration process, also utilize the attribute of the adjacent tensor information of each pixel as uniqueness; Adopt method for registering to calculate deformation field simultaneously based on cubic spline.In concrete implementation procedure,, need recomputate time diagram to after deformation field changes; The characteristics that amount of calculation is too big; The invention allows for the fast algorithm that a kind of time diagram upgrades, can under the situation that guarantees registration accuracy, improve computational efficiency greatly.A large amount of experiments prove; With respect to other method; The method for registering based on the FM pattern that the present invention proposes can extract abundanter, more effective tensor property, thereby defines the corresponding relation of putting between the image better, can produce more stable, more accurate registration result.

Description of drawings:

Figure one: based on tensor property with based on the comparison of the similarity measurement of FM time diagram: (a) template figure (FA) and a reference point; (b) target figure (FA); (c) distance map that utilizes tensor property to calculate; (d) distance map that utilizes the FM feature calculation to obtain;

Figure two: part (FM) time diagram of advancing fast: (a) local FM time diagram; (b) (c) with (d) explanation in the differentiation course of different time;

Figure three: the example description of shortcut calculation:

Figure four: the analog image instance: (a) template figure; (b) (c) (d) analog image;

Figure five: the registration error scattergram

Figure six: white matter nerve fiber bundles labelling instance: (a) white matter nerve fiber in the template image; (b) in the registration target image, the white matter nerve fiber of extraction; (c) white matter nerve fiber of manual extraction in the original object image.

The specific embodiment:

Combine accompanying drawing that the present invention is done further describes at present:

(1) the DTI image registration algorithm that proposes of the description of problem: this paper and common yardstick image registration algorithm are very similar, and target is to obtain a deformation field f, to reach registration DTI template image D _TWith target image D _SPurpose.Energy equation is defined as follows:

E(f)＝E _sim(D _T，D _S，f)+λE _con(f)，(1)

Wherein, E _Sim(D _T, D _S, be f) according to the similarity measurement between two width of cloth images of current deformation field f calculating, E _Con(f) be the smooth constraint of deformation field.In the present invention, deformation field f adopts the cubic B-spline registration model to come modeling, can guarantee the smooth of deformation field, therefore can ignore second.Thereby, detailed energy function E=E _SimAgain be defined as:

$E=\underset{x\∈\mathrm{\Ω}}{\mathrm{\Σ}}{\left|\right|A\left[D_S(f\left(x\right)+x)\right]-A\left[Q\left(f\left(x\right)\right)D_T\left(x\right)Q^T\left(f\left(x\right)\right)\right]\left|\right|}^2---\left(2\right)$

Wherein x is a pixel in the Ω of template image zone, D _T(x) be the tensor at pixel x place in template image, D _S(f (x)+x) is the tensor of corresponding pixel in the target image, and Q (f (x)) is the spin matrix that calculates according to deformation field f at pixel x place.Different with common yardstick method for registering images is, in the registration process of tensor image, in order to compare with the corresponding tensor of target image, the tensor of template image need be redirected, and Q (f (x)) is that tensor is redirected the spin matrix that calculates just.The feature extraction operation that it is the dispersion tensor at center that A [] has expressed with a pixel, and ‖ A [D ₁]-A [D ₂] ‖ ²Provided the distance of an eigenvectors.If A []=I, the characteristics of image of a pixel are exactly tensor itself, and so ‖ A [D ₁]-A [D ₂] ‖ ²=‖ D ₁-D ₂‖ ²Become a kind of simple tensor distance.In such cases, the expressed implication of formula 2 is just just the same with the common registration Algorithm based on the tensor similarity.

In the present invention, neighborhood tensor property A [] is defined as through carrying out local FM algorithm and obtains time diagram.Though the tensor in the single pixel field is an a kind of higher-dimension system, simply based on the algorithm of FM, we can effectively extract the tensor pattern around the pixel through a kind of.Therefore, main idea of the present invention is utilized the resulting more effective tensor property of local FM algorithm exactly, improves the precision of DTI image registration.

Except the tensor property that is the basis with local FM, the FA value is also stressed current tensor points as a characteristic vector element more with more.Why to be the tensor property on basis with FM have the better property distinguished in order to illustrate, and Fig. 1 has showed respectively with simple tensor distance with based on FM tensor property calculated distance figure.Here, the FM algorithm with each pixel is being the centre of sphere, and radius is to carry out in the ball of 16mm.Can find out, a large amount of with the selected closely similar point of reference pixel is arranged in whole template image if only use the tensor distance; Tensor property based on FM then can better help to distinguish reference pixel and other pixels.Therefore, the characteristic of this uniqueness can be used for registration to obtain more stable result.

(2) based on the dispersion tensor feature extraction of local FM:

In the present invention, to each interested pixel, we use local FM algorithm simulation hydrone communication mode in its neighborhood, are used to instruct image registration as the tensor property of current pixel.At first, the evolution front begins towards periphery from current pixel x that neighborhood carries out evolution, supposes to have arrived some y ' at certain time point that be advanced to other peripheral candidate point y from y ' then, the speed of evolution can be defined as:

$v_y\left(r\right)=v_y^{\mathrm{tensor}}\left(r\right)+{\mathrm{\ηv}}_y^{\mathrm{inertia}}\left(r\right)---\left(3\right)$

Wherein, v _y ^Tensor(r) be main evolution speed, r:y ' → y, and:

$v_y^{\mathrm{tensor}}\left(r\right)=w{r}^T{D}_{y^{\′}}r---\left(4\right)$

Wherein, w is at the inner coefficient of eliminating quick evolution of cerebrospinal fluid.Little then w is little for the FA value, and vice versa.Therefore w can calculate through w=1/ (1+exp (α (FA-β))), and wherein α and β are the constants of control FA effect.Second of formula 3 is used for guaranteeing the level and smooth of evolution:

$v_y^{\mathrm{inertia}}\left(r\right)=v_{y^{\′}}(r_{y^{\′}}\·r)---\left(5\right)$

R wherein _{Y '}And v _{Y '}Be respectively evolution direction and the speed of current point y ', ". " table two dot product.In the FM algorithm, the evolution time of advent of evolution front and the relation of evolution speed can be described with the Eikonal equation:

$|\▿T|V=d---\left(6\right)$

Wherein,

Be from a point to another evolution required time, v is the evolution speed between them, therefore d is the distance between these points, arrives the time t of forward face point y _yBe:

t _y＝t _y′+d/v _y (7)

Wherein, t _{Y '}Be the time that arrives y ', v _yBe formula 3 defined evolution speed.For each consecutive points of y ' outside, we utilize formula 7 to calculate its time of arrival, and the point of selecting to have the minimum time of advent is as new forward face point.Like this, x develops forward from pixel, can form one the time of advent of x neighborhood point and include the figure of tensor information characteristics on every side, is referred to as local FM time diagram.In invention, η selects 0.9, and α selects 50, and β selects 0.3.

Fig. 2 has shown the local FM time diagram of a given pixel, can find out, the distribution of time diagram has reflected the tensor distributed intelligence on every side of this pixel, has also reflected the distributed intelligence of moving towards of the interior white matter nerve fiber of given neighborhood of pixels.In this paper equation 2, it is tensor property A [] that local FM time diagram is used as.Detailed enforcement will be described at next joint.

(3) based on the registration Algorithm of local FM time diagram

According to above method, can obtain a local FM time diagram to all interested pixels of whole brain.The as can beappreciated from fig. 2 disperse speed or the time of advent are very different along different directions, and this depends on DTI image itself.In fact, other directions are slow along the direction evolution of white matter nerve fiber is fast.Therefore, the pattern that exists that has reflected pixel nerve fiber on every side from the time diagram that obtains based on tensor FM.Introduce the realization of registration Algorithm below:

Get back to formula 2, come the compute tensor characteristic with local FM time diagram as, so formula 2 can be write:

$E=\underset{x\∈\mathrm{\Ω}}{\mathrm{\Σ}}\{\underset{y\∈N\left(x\right)}{\mathrm{\Σ}}{(t_S(f\left(x\right)+x,f\left(y\right)+y)-t_{Q\left(f\right)D_T{Q}^T\left(f\right),f}(x,y))}^2+\mathrm{\μ}{({\mathrm{FA}}_S(f\left(x\right)+x)-{\mathrm{FA}}_T\left(x\right))}^2\},---\left(8\right)$

Wherein, t _S(x y) is the evolution time of target figure S the inside from pixel x to pixel y,

Be to consider that deformation field f and tensor are redirected the distortion time of rear pattern plate figure from pixel x to pixel y, the neighborhood of N (x) represent pixel x.

Obviously; In case deformation field f has changed; Time diagram around each template pixel also will inevitably change accordingly; So new time diagram just must recomputate on the basis of distortion and redirected rear pattern plate image; This is a job very consuming time, so we simplify the realization of

, specifically is embodied as:

$t_{Q\left(f\right)D_T{Q}^T\left(f\right),f}(x,y)\≈t_T(x,y)(1+\frac{{|f\left(y\right)+y-f\left(x\right)-x|}_{x\→y}-{|y-x|}_{x\→y}}{{|y-x|}_{x\→y}}),---\left(9\right)$

Wherein, || _{X → y}Be the distance from x to y along travel track, final, the energy function of simplification can be defined as:

$E=\underset{x\∈\mathrm{\Ω}}{\mathrm{\Σ}}\{\underset{y\∈N\left(x\right)}{\mathrm{\Σ}}{(t_S(f\left(x\right)+x,f\left(y\right)+y)-t_T(x,y)(1+\frac{{|f\left(y\right)+y-f\left(x\right)-x|}_{x\→y}-{|y-x|}_{x\→y}}{{|y-x|}_{x\→y}}))}^2---\left(10\right)$

$+\mathrm{\μ}{({\mathrm{FA}}_S(f\left(x\right)+x)-{\mathrm{FA}}_T\left(x\right))}^2\}.$

Fig. 3 has provided the example description of this simplification: as far as slick strain field, relatively very little along local angle's variation of FM track.Such as, the FM track of red tensor evolution tensor to the upper right corner is before and after smooth strain, about the same from the lower left corner.In the method for the invention, suppose to change very for a short time that along the angle of these evolution tracks therefore, the change of time diagram can be calculated in proximate variation according to the distortion longitudinal separation.

(4) more than comprehensive, can sum up as follows based on the dispersion tensor method for registering images of the quick evolution pattern in part:

Step 1 is selected resolution R, and the DTI image of importing is carried out down-sampling, is used for Multi-Resolution Registration; And current resolution r=R is set;

Step 2 is under resolution r, to each pixel, if its FA value, is calculated its local FM time diagram greater than given threshold value;

Step 3, to the control point c of each cubic B-spline, estimation objective function is for the gradient at control point, as: Use $v\left(c\right)=v\left(c\right)-\mathrm{\ϵ}\∂E/\∂c$ Upgrade the deformation values at control point, and upgrade the deformation field that influenced by this control point; Iteration is carried out at all control point, to upgrade deformation field.

Step 4, repeating step 3 is up to convergence.Accomplish under the resolution r after the registration, if r=1 then forward step 5 to otherwise carries out the cubic B-spline up-sampling, a related variation field of more calculating it on the high-resolution.After obtaining deformation field, adopt the reorientation method of Xu that the tensor image is carried out redirect operation.R=r-1 is set, forwards step 2 then to and carry out next resolution processes.

Step 5 is with the final deformation field of cubic B-spline Model Calculation that produces; If necessary, then the redirection target image to the template space.

Notice that in the realization of multiresolution, in order to eliminate the error that produces owing to shortcut calculation, we recomputate local FM time diagram to the deforming template image under each resolution.

(5) registration result and analysis

In order to assess the registration Algorithm that the present invention proposes, two groups of experiments have been carried out.In first group of experiment, come analog data with known deformation field, the deformation field that obtains through the algorithm with deformation field and the present invention's proposition then compares, and comes the accuracy of assessment algorithm.In second experiment, on real dispersion tensor image, we use this algorithm and carry out registration, to detect its performance.

A experiment one: based on the experiment of emulated data

At first, the elasticity simulation algorithm based on statistical model of using Xue is simulated 10 DTI images.Concrete grammar is: through 10 strain fields of method simulation of Xue, the DTI image registration algorithm of using Xu then is registrated to a known template image in the object space through mimic 10 deformation fields in front earlier.Like this, we just can simulate to 10 DTI images, and the strain field from template image to these analog images is known.Calculate fractional anisotropy image (FA) simultaneously, be used for image registration as the one-dimensional characteristic of this algorithm.Figure four has shown the FA figure of the analog image of template image and three picked at random, can see that these are different from template image significantly through the FA image that simulation obtains.Because the true strain field between these analog images and the primary template image is known, invent template image that the algorithm that proposes obtains and the deformation field between the true picture through true strain field relatively with utilizing this, can carry out the evaluation of algorithm.As relatively, use traditional DTI image registration algorithm, promptly utilize the registration Algorithm registration yardstick FA image of higher-dimension free form, obtain as algorithm deformation field relatively.

Accomplish after the registration of template image and emulating image, the distortion inaccuracy that calculates between registration deformation field and the true strain field distributes.Suppose that the true strain field is f ^*, the deformation field of calculating is f, both error image Δ f are f and f ^*Between the Euclidean distance of each pixel.Fig. 5 has shown registration Algorithm and the traditional registration Algorithm resultant average deformation error map after 10 emulating images and template image of using the present invention's proposition.It is thus clear that the algorithm of the present invention of having used neighborhood tensor information has littler distortion inaccuracy and more accurate registration result.

B experiment two: real DTI image experiment

The purpose of second group of experiment is the performance of inspection algorithm that this paper carries on true DTI image.For performance that can visual assessment algorithm; We with the algorithm of carrying on DTI automatic image registration to a template image of 14 normal adults; From template image, manually select certain concrete fibre bundle fibre bundle (being designated as fibre bundle A) as a reference, in the target image of registration, choose the fibre bundle close as the fibre bundle interested in the target image (being designated as fibre bundle B) according to the Hausdorff distance then with fibre bundle A distance.Manual simultaneously chooses in the original object image space and the of a sort fibre bundle of fibre bundle A (being designated as fibre bundle C).Relatively fibre bundle B and fibre bundle C can do the checking based on vision to the registration Algorithm that this invention proposes.

It is example that Fig. 6 shows with the corpus callosum, the result that this algorithm is verified.Fig. 6 (a) shows the corpus callosum fibre bundle of manually selecting from template image; Fig. 6 (b) has shown the fiber interested at the automatic labelling of target image of registration; Fig. 6 (c) has shown the corpus callosum fibre bundle of hand labeled from the original object image.Can find out that manually selection is closely similar with the fibre bundle that calculates automatically, thereby the registration Algorithm of proof the present invention proposition has registration accuracy preferably, and utilize this algorithm can realize the automatic extraction of some nerve fibre bundle.

Claims (6)

1. dispersion tensor method for registering images based on local quick traveling mode is characterized in that step is following:

Step 1: with resolution R the DTI image is carried out down-sampling, described resolution is R=5;

Step 2: as the part anisotropy value FA of sampled pixel during greater than threshold value, utilize the communication mode of method Simulated Water molecule in its neighborhood of quick traveling mode Fast Marching-FM, calculate the local FM time diagram of this sampled pixel, concrete steps are:

Step a: put y ' certainly with gait of march

and be advanced to other candidate points of neighborhood y on every side, wherein

is that main gait of march,

are that main gait of march peace slides the into compromise coefficient between the speed for level and smooth gait of march, η;

Described

W eliminates the coefficient of advancing fast in cerebrospinal fluid inside, r:y ' → y, r ^T

Be the transposition of r, D _y' be the dispersion tensor of some y ';

Described w=1/ (1+exp (α (FA-β))), wherein α and β regulate the constant of FA to advancing fast and controlling;

Described

R wherein _y' and v _y' be respectively direct of travel and the speed of current point y ', " "

Represent two dot products;

Step b: the t time of advent that calculates each consecutive points y of y ' outside _y=t _y'+d/v _y, wherein, t _y' be the time that arrives y ', v _yIt is main gait of march;

Step c: as new forward face point, begin to repeat to obtain the local FM time diagram that develops forward from pixel x from step a then with those consecutive points y of the minimum time of advent in y ' outside; Multiple scope is for being the centre of sphere with pixel x point, and radius is in the ball of 16mm;

Step 3: utilize the local FM time diagram and the part anisotropy value FA figure that obtain to make characteristic, set up DTI template image D _TWith target image D _SBetween the energy equation of deformation field f:

$E\left(f\right)=\underset{x\∈\mathrm{\Ω}}{\mathrm{\Σ}}\{\underset{y\∈N\left(x\right)}{\mathrm{\Σ}}{(t_S(f\left(x\right)+x,f\left(y\right)+y)-t_T(x,y)(1+\frac{|f\left(y\right)+y-f\left(x\right)-x{|}_{x\→y}-|y-x{|}_{x\→y}}{|y-x{|}_{x\→y}}))}^2$

$+\mathrm{\μ}{({\mathrm{FA}}_S(f\left(x\right)+x)-{\mathrm{FA}}_T\left(x\right))}^2\}.$

Wherein: t _T(x y) is the traveling time of template image T the inside from pixel x to pixel y, t _S(f (x)+x, f (y)+y) are the traveling times that target figure S the inside pixel f (x)+x arrives pixel f (y)+y behind the consideration deformation field f, || _{X → y}Be the distance from x to y along travel track, the neighborhood of N (x) represent pixel x, FA _T(x) be the part anisotropic value of template image T the inside pixel x, FA _S(f (x)+x) is the part anisotropic value of target figure S the inside pixel f (x)+x behind the consideration deformation field f, and μ is the compromise coefficient;

Step 4: deformation field f adopts the cubic B-spline registration model to come modeling; Control point c to each cubic B-spline; The estimated energy equation upgrades the deformation values at control point for the gradient

at control point with

, and upgrades the deformation field that influenced by this control point; Iteration is carried out at all control point, to upgrade deformation field;

Step 5: repeating step 4 is up to convergence; Accomplish under the resolution r after the registration, if r=1 then forwarded for the 6th step to; Otherwise carry out the cubic B-spline up-sampling; A related variation field of more calculating it on the high-resolution, after obtaining deformation field, adopt the reorientation method of Xu that the tensor image is carried out redirect operation; And r=r-1 is set, forwarding for the 2nd step then to carries out next resolution processes;

Step 6: with the final deformation field of cubic B-spline Model Calculation that produces; And the reorientation method redirection target image of employing Xu is to the template space.

2. the dispersion tensor method for registering images based on local quick traveling mode according to claim 1 is characterized in that: described part anisotropy value FA threshold value is 0.25.

3. the dispersion tensor method for registering images based on local quick traveling mode according to claim 1 is characterized in that: described η selects 0.9.

4. the dispersion tensor method for registering images based on local quick traveling mode according to claim 1 is characterized in that: described α selects 50.

5. the dispersion tensor method for registering images based on local quick traveling mode according to claim 1 is characterized in that: described β selects 0.3.

6. the dispersion tensor method for registering images based on local quick traveling mode according to claim 1 is characterized in that: described μ selects 1.0.

Images