CN105335742A - Robust feature statistics based three-dimensional region adaptive segmentation method - Google Patents

Abstract

The present invention relates to the field of adaptive segmentation technology, in particular to a three-dimensional area adaptive segmentation method based on robust feature statistics, comprising the following steps: A. Adopting the LARFS segmentation method, respectively referring to the principles of the RG segmentation algorithm and the RSS segmentation algorithm and combining them Application; B. Select seed points in the 3D data space to perform self-adaptive modeling of 3D image geometric features; C. According to the LARFS segmentation method, the RSS fast segmentation algorithm based on the level set is used to model the image growth characteristics; D. Using the level The set method evolves the outline of the feature area, combines the traditional area growth algorithm with the statistical method, makes full use of the area information of the image, expresses the local segmentation feature and drives the outline evolution, and adopts the LARFS method, which can effectively prevent noise interference and reduce the use of The operator's intervention operation can get a better segmentation effect.

Description

A 3D Region Adaptive Segmentation Method Based on Robust Feature Statistics

technical field

The invention relates to the technical field of adaptive segmentation, in particular to a method for adaptive segmentation of three-dimensional regions based on robust feature statistics.

Background technique

RG segmentation can better solve the problem of conformal determination of region of interest (ROI) in 3D image processing software, and significantly improves the efficiency of traditional manual segmentation methods, but still requires a lot of manual operations. The ideal ROI segmentation method does not require Too many subjective interventions, fully automatic or semi-automatic adaptive algorithms, only require the operator to correct and confirm the output results.

At present, there are many researches on semi-automatic segmentation algorithms based on 3D images, such as methods based on gray levels, level set methods, graph cutting algorithms, and statistical segmentation methods. The 3D ROI segmentation scenarios targeted by the above algorithms are different. Since the 3D image will be affected by noise, offset field effect, local volume effect or tissue movement during imaging or reconstruction, it often has local inhomogeneity or blur. Quality defects, so it is easy to produce insufficient segmentation or leakage when performing ROI segmentation.

Contents of the invention

The object of the present invention is to aim at the defects and deficiencies of the prior art, and provide a three-dimensional region adaptive segmentation method based on robust feature statistics with good segmentation effect, which is based on a semi-automatic segmentation method of voxel information, effectively prevents noise interference, reduces The doctor's intervention operation can get better segmentation results.

In order to achieve the above object, the technical scheme adopted in the present invention is:

A kind of three-dimensional area self-adaptive segmentation method based on robust feature system of the present invention, comprises the following steps:

A. Using the LARFS segmentation method, respectively refer to the principles of the RG segmentation algorithm and the RSS segmentation algorithm and apply them in combination;

B. Select seed points in the 3D data space to perform adaptive modeling of geometric features of 3D images;

C. According to the LARFS segmentation method, the RSS fast segmentation algorithm based on the level set is used to model the image growth characteristics;

D. Use the level set method to evolve the contour of the feature area.

Further, in the step A, during the evolution process of the image contour, more local robust feature information of the image is extracted based on a statistical method.

Further, the step A is to first manually select a single seed point as input information, use the RG segmentation algorithm to perform image feature modeling according to the seed point and neighborhood attributes, and perform regional adaptive growth in the three-dimensional data field; Afterwards, the adaptation result point set is used as the input feature label for image area feature statistics and modeling; then the level set algorithm is used to evolve the outline of the feature area; finally the segmentation result is obtained.

Further, the RG algorithm is a semi-automatic segmentation method. Before performing the RG algorithm, the operator needs to mark the seed information on the target ROI area in the image. When the algorithm starts, the system will automatically check the adjacent voxels of the seed point. feature value, and decide whether to add the adjacent voxel to the segmentation result according to its similarity with the seed.

Further, the RSS algorithm comprehensively applies image position information and pixel information to model the feature image, and then performs contour evolution through a geometric active contour model.

Further, the step B is to first calculate a threshold interval based on the gray value of the seed point and its adjacent 26 spatial points; The points in the interval are divided and used as new seed points; after that, the above process is repeated until the growth range is no longer expanded or the number of iterations reaches the preset value position; finally, a set of seed points is obtained, which is the result of geometric feature adaptation.

Further, the step C is to first extract the local position information and pixel information of the image; and then perform contour evolution through the geometric active contour model according to the initial range of the feature region at each moment.

Further, in the step D, the splitting and merging of the closed segmented contour curves is realized by embedding the two-dimensional curves into the three-dimensional plane.

After adopting the above structure, the beneficial effect of the present invention is: a three-dimensional region self-adaptive segmentation method based on robust feature statistics described in the present invention combines traditional region growth algorithms with statistical methods to make full use of image region information , represents the local segmentation feature and drives the contour evolution. Using the LARFS method can effectively prevent noise interference, reduce user intervention, and obtain better segmentation results.

Description of drawings

Fig. 1 is the flow chart of LARFS segmentation method among the present invention;

Fig. 2 is the formula for calculating the threshold interval in the present invention.

Fig. 3 is a flow chart of geometric active contour evolution of the level set method in the present invention.

Fig. 4 is the formula for calculating the change amount of a single voxel in the present invention.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings.

As shown in Figures 1 to 4, a method for adaptive segmentation of three-dimensional regions based on robust feature statistics according to the present invention includes the following steps:

A. Using the LARFS segmentation method, respectively refer to the principles of the RG segmentation algorithm and the RSS segmentation algorithm and apply them in combination;

B. Select seed points in the 3D data space to perform adaptive modeling of geometric features of 3D images;

C. According to the LARFS segmentation method, the RFS fast segmentation algorithm based on the level set is used to model the image growth characteristics;

D. Use the level set method to evolve the contour of the feature area. .

In the step A, during the evolution process of the image contour, more local robust feature information of the image is extracted based on a statistical method.

Described step A is to first manually select a single seed point as input information, use the RG segmentation algorithm to carry out image feature modeling according to the seed point and neighborhood attributes, and perform region-adaptive growth in the three-dimensional data field; The resulting point set is used as the input feature label for image area feature statistics and modeling; then the level set algorithm is used to evolve the outline of the feature area; finally the segmentation result is obtained.

The RG algorithm is a semi-automatic segmentation method. Before performing the RG algorithm, the user needs to mark the seed information in the ROI area in the image. When the algorithm starts, the system will automatically check the feature values of all adjacent voxels of the seed point, and based on The degree of similarity between it and the seed determines whether to add the adjacent voxel to the segmentation result.

The RSS algorithm comprehensively applies image position information and pixel information to model a feature image, and then performs contour evolution through a geometric active contour model.

The step B is to first calculate a threshold value interval based on the gray value of the seed point and its adjacent 26 spatial points; Points are segmented and used as new seed points; after that, the above process is repeated until the growth range is no longer expanded or the number of iterations reaches the preset value position; finally, a set of seed points is obtained, which is the result of geometric feature adaptation.

The step C is to first extract the local position information and pixel information of the image; then perform contour evolution through the geometric active contour model according to the initial range of the feature region at each moment.

In the step D, the splitting and merging of the closed segmented contour curves is realized by embedding the two-dimensional curves into the three-dimensional plane.

As shown in Figure 1, the system flow of LARFS segmentation method among the present invention is as follows:

(1) Three-dimensional image sequence acquisition;

(2) Visual rendering of volume rendering;

(3) Seed point mark;

(4) Three-dimensional region growth segmentation;

(5) output area data information;

(6) Regional feature vector statistics;

(7) Numerical profile evolution;

(8) output three-dimensional data field;

(9) Segmentation result surface reconstruction;

After the step (9), perform the step (2) volume rendering visualization rendering again to generate superimposed display.

In the method for calculating the threshold interval in the present invention, the field gray mean value MEAN and the standard variance VARIANCE of the seed point set are selected to calculate the confidence interval Ω _N of the bottom N iterations, and the calculation process formula is shown in formula (3-24) and ( 3-25) and (3-26);

Among them, Seeds represents the number of seed points, VoxelSize represents the number of sample voxels in the field of seed point set, and 0 is the growth factor;

After calculating the above-mentioned field gray mean MEAN, standard deviation VARIANCE and confidence interval Ω _N , then calculate the field gray mean mean and square sum seedsqr of a single seed point, the calculation process formula is shown in formula (3-27) in Figure 2 and (3-28);

Among them, voxel ₁ represents the gray value of the seed point i. The above method uses the estimation of the confidence interval in probability theory. The value of the growth factor will directly affect the segmentation effect of the overall method. The range of the segmented area is proportional to the value of θ. For pixel samples that satisfy the normal distribution, θ is taken as 2.5 is sufficient for a 99% confidence level.

In the present invention, the center algorithm needs to input the seed point set of the target area before running, and introduces the robust statistical theory to define three eigenvalues that are not sensitive to noise and can be quickly calculated for each seed, which are respectively the center points of the sample voxels in the field. The number of digits MED(x), the standard interquartile interval IQR(x) and the middle is the standard deviation MAD(x), where MED(x) represents the middle position of the sequence after the variable values in the sample are arranged in order of size Variable value; IQR(x) is the difference between the variable value at the 75% position and the variable value at the 25% position in the series; MAD(x) refers to the new series obtained by subtracting the median from the individuals in the sample one by one The median after the rearrangement;

As shown in Figure 3, the geometric active contour evolution process of the level set method in the present invention is as follows:

(1) Image input machine preprocessing;

(2) Establish an energy model;

(3) change to get the Euler-lagrange equation;

(4) Using the level set method to carry out the curve evolution PDE;

(5) Initial level set function setting;

(6) Iteration (PDE numerical algorithm);

(7) Judging whether the condition for stopping evolution is satisfied;

(8) If the stop evolution condition is met, output the level set function to obtain the segmentation contour; if not, repeat step (6).

In the present invention, the input parameters of the LARFS segmentation method only include the seed point F and the growth factor θ, and the numerical process is as follows:

1) Initialize the memory space

(a) Open up 10 stack spaces in the computer memory, which are marked as active layer L ₁ and state layer S ₁ , (i∈[-2, 2]), where the point set contour value corresponding to Li space is U∈[ i-0.5, i+0.5], L ₀ represents the contour evolution boundary, if the space label is negative and the value is smaller, it means that the layer is more inside the segmented image, otherwise it means that the layer is more outside the segmented image.

2) Geometry adaptation process

(a) The seed point and its adjacent 26 voxels in the three-dimensional space are used as the initial region D ₀ , and the regional voxel mean MEAN ₀ and variance VARIANCE are calculated according to the formulas (3-24) and (3-25) in Figure 2 ₀ ;

(b) According to the regional mean MEAN _N-1 and variance VARIANCE _N-1 of the bottom N-1 iterations, calculate the confidence interval Ω _N of the bottom N iterations according to the formula (3-26) in Figure 2;

(c) Update the region D _N based on Ω _N and D _N-1 , the voxels it contains are connected to D _N-1 and within the confidence interval Ω _N ;

(d) update the mean value MEAN _N and the variance VARIANCE _N of area D _N according to formula (3-24) and (3-25) among Fig. 2;

(e) Repeat the process (b)-(d) until the region no longer expands, or the number of iterations reaches a preset value.

3) Image growth and contour evolution process

(a) place the point set contained in the set adaptive region result _DN in the L ₀ space;

(b) For each voxel point x in L ₀ space, calculate the corresponding change amount Δu _λ according to the formula (3-49) in Fig. 4;

(c) Update the u _λ value of each point in the active set, if the new value Then put x into the response state layer S ₁ according to the following rules; if u _λ +Δu _λ >0.5, then put it into the S ₁ layer; if u _λ +Δu _λ <-0.5, then put it into the S ₁ layer;

(d) Visit each voxel point in layer L ₁ sequentially in the order of i=±1,±2, and update the corresponding μ value (+1 or -1) according to the voxel value in the next layer L _1±1 ;

(e) Each point in the state layer S _±1. S _±2 is processed as follows:

i. For each voxel point x in S ₁ , delete it from the original layer L _1±1 , and put it into L ₁ , when i=±3, delete it from all layers;

ii. Put the points adjacent to L _1±1 into the S _1±1 layer.

(f) Repeat processes (b)-(e) until a preset number of iterations, time or output volume is reached.

4) Result output

(a) Output a set of three-dimensional space points with u≤0, which can form the final segmented image;

(b) Use the surface reconstruction method to form the output result into a 3D model and integrate it into the scene display.

A three-dimensional area self-adaptive segmentation method based on robust feature statistics described in the present invention combines traditional area growth algorithms with statistical methods, fully utilizes image area information, represents local segmentation features and drives contour evolution, and adopts The LARFS method can effectively prevent noise interference, reduce user intervention, and obtain better segmentation results.

The above is only a preferred embodiment of the present invention, so all equivalent changes or modifications made according to the structure, features and principles described in the scope of the patent application of the present invention are included in the scope of the patent application of the present invention.

Claims (8)

1. A three-dimensional region adaptive segmentation method (LARFS) based on robust feature statistics, characterized in that: comprise the steps: A. Using the LARFS segmentation method, respectively refer to the principles of the Region Growth Segmentation Algorithm (RG) and the Robust Statistical Segmentation Algorithm (RSS) and apply them in combination; B. Select seed points in the 3D data space to perform adaptive modeling of geometric features of 3D images; C. According to the LARFS segmentation method, the RSS fast segmentation algorithm based on the level set is used to model the image growth characteristics; D. Use the level set method to evolve the contour of the feature area. 2. A kind of three-dimensional region self-adaptive segmentation method based on robust feature statistics according to claim 1, characterized in that: in the step A, in the image contour evolution process, more images are extracted based on statistical methods Local robust feature information. 3. A kind of three-dimensional region self-adaptive segmentation method based on robust feature statistics according to claim 1, characterized in that: described step A is to manually select a single seed point as input information, and utilize the RG segmentation algorithm Image feature modeling is performed according to seed points and neighborhood attributes, and region adaptive growth is performed in the 3D data field; then the adaptation result point set is used as the input feature label for image region feature statistics and modeling; then the level set algorithm is used to The contour of the feature area is evolved; the segmentation result is finally obtained. 4. A method for self-adaptive segmentation of three-dimensional regions based on robust feature statistics according to claim 1, characterized in that: the RG algorithm is a semi-automatic segmentation method, before performing the RG algorithm, the doctor needs to analyze the The target organ marks the seed information. When the algorithm starts, the system will automatically check the feature values of all adjacent voxels of the seed point, and decide whether to add the adjacent voxels to the segmentation result according to their similarity with the seed. 5. A kind of three-dimensional region self-adaptive segmentation method based on robust feature statistics according to claim 1, characterized in that: the RSS algorithm comprehensively applies image position information and pixel information to model the feature image, and then through geometric active The contour model performs contour evolution. 6. A kind of three-dimensional area self-adaptive segmentation method based on robust feature statistics according to claim 1, characterized in that: said step B is first based on the gray level of the seed point and its adjacent 26 spatial points value to calculate a threshold interval; then a seed point grows outward as a reference, and the points whose gray value in the image area is within the threshold interval are divided as new seed points; after that, the above process is repeated until the growth range is no longer expanded or iterated The number of times reaches the preset value position; the final set of seed points is obtained, which is the result of geometric feature adaptation. 7. A method for self-adaptive segmentation of three-dimensional regions based on robust feature statistics according to claim 1, characterized in that: said step C is to first extract local position information and pixel information of the image; and then according to each moment The initial extent of the feature region undergoes contour evolution through a geometric active contour model. 8. A three-dimensional area adaptive segmentation method based on robust feature statistics according to claim 1, characterized in that: in the step D, the closed segmentation contour curve is realized by embedding the two-dimensional curve into the three-dimensional plane split and merge.