CN103914845B - The method obtaining initial profile in Ultrasound Image Segmentation based on active contour model - Google Patents

Abstract

The invention discloses a kind of method obtaining initial profile in Ultrasound Image Segmentation based on active contour model, comprise the steps of the textural characteristics of training ultrasonoscopy tumor region, and constitute a standard vector together with tumor priori size in standard ellipse empirical value, ultrasonoscopy to be measured；Ultrasonoscopy to be measured is carried out pretreatment；Image is carried out dynamic threshold segmentation；Extract all closed contours produced in dynamic threshold segmentation result, form corresponding subimage；Calculate the textural characteristics 24 of each subimage, ellipse fitting result parameter 1 and subimage closed contour interior pixels number, and constitute a vector by these 26 data；Calculate the distance of gained vector and the standard vector of tumor region training gained, determine that the closed contour in the corresponding subimage of vector that distance is minimum is the profile of tumor in split ultrasonoscopy.The present invention overcomes the many noises of ultrasonoscopy, the problem of obscurity boundary, initial profile accuracy rate is high.

Description

The method obtaining initial profile in Ultrasound Image Segmentation based on active contour model

Technical field

A kind of method that the present invention relates to obtain initial profile in Ultrasound Image Segmentation based on active profile, belongs to ultrasonic Image processing field.

Background technology

Along with the appearance of high intensity focused ultrasound, non-invasively treating oncologic application more and more extensive, this treatment system The biggest advantage is had if do not operated in clinic, preventing from scar, noinvasive or Micro trauma, can detect treatment, in real time not by swollen Tumor size limits, and total cost is low.These advantages all determine that high-strength focus supersonic therapeutic system has the biggest development and application Prospect.And in high-strength focus supersonic therapeutic system most critical exactly patient's tumor is carried out real-time navigation.Navigation procedure needs Tumor in real-time ultrasonic image is positioned.Traditional manual positioning mode is changed into and is automatically positioned, this will be controlled The therapeutic efficiency for the treatment of system and treatment accuracy bring the biggest help.

Ultrasound Image Segmentation be lesion detection, diagnose and treat in be widely used, its segmentation accuracy directly close It is tied to the location of tumor.Through long-term further investigation, it has been suggested that ultrasonic image division method have based on region, based on Border and based on statistical information, different types of dividing method has different segmentation features.Active contour model (Active Contour Model), be otherwise known as Snake, is that a kind of objective contour proposed by Andrew professor Blake is retouched State method, be mainly used in Target Segmentation based on shape.This model will be appreciated that it for far-ranging one is Row visual problem gives unified solution, and in nearest more than ten years, it is by the success of increasing researcher Be applied to many fields of computer vision, such as edge extracting, image segmentation and classification, motion tracking, three-dimensional reconstruction, three-dimensional Vision matching etc..

But this model needs initial profile, restraining again if manually obtaining initial profile, i.e. wasting time and energy, the most not Enough accurate.Therefore the method for design acquisition initial profile automatically is very important.

Summary of the invention

The technical problem to be solved is to provide in a kind of Ultrasound Image Segmentation based on active contour model The method obtaining initial profile.Operation is simple and reliable for the method, speed fast, the method the ultrasonoscopy tumor obtained initially is taken turns Wide accuracy is the highest.

The method automatically extracting tumor initial profile in a kind of Ultrasound Image Segmentation of the present invention, comprises the steps of

(1) choosing more than 60 width existing ultrasonoscopy composition training set, described existing ultrasonoscopy contains and treats point Cut the tumor that ultrasonoscopy is identical, from training set, every width ultrasonoscopy extracts the 24 of tumor region with gray level co-occurrence matrix Individual textural characteristics, in utilizing mathematical linear to return minimum distance method obtain the standard value of 24 textural characteristics, and and standard In oval empirical value, ultrasonoscopy to be measured, tumor priori size constitutes a standard vector A together₀；

(2) ultrasonoscopy to be split is carried out the pretreatment of anisotropy parameter and histogram equalization；

(3) (2) gained image is carried out dynamic threshold segmentation, and extract all closing of producing in dynamic threshold segmentation result Closing profile, each closed contour forms a corresponding subimage；

(4) utilize gray level co-occurrence matrix to extract 24 textural characteristics of each subimage, calculate ellipse fitting result ginseng Number and subimage closed contour interior pixels number, and constitute a vectorial A by these 26 data₁；

(5) each subimage gained vector A is calculated₁Standard vector A with tumor region training gained₀Distance D, make Obtain the minimum vectorial A of D₁Closed contour in corresponding subimage is the initial profile of tumor in ultrasonoscopy to be split.

In step (1), standard ellipse empirical value is defined as 0.7.

Above-mentioned 24 textural characteristics include image 0 °, 45 °, 90 °, contrast on 135 ° of four directions, energy, relevant Property, unfavourable balance square, entropy and each six features of non-similarity.

The acquisition of initial profile to be made is more accurate, is determining standard vector A₀During, need to use reliably Mathematical measure.Such as during training, supersonic tumor image has 60 width, every piece image to be desirable that to obtain a texture feature vector A_i, then how to process these 60 vectors to obtain a standard vector A₀Need to think better of.A is utilized after in conjunction with₀Come Method used when determining tumor region: calculate characteristic vector A extracted from each subimage₁With A₀Distance, then The most also this reference of available range obtains standard vector A₀.Concrete execution method is the linear regression in mathematics, will certain 60 eigenvalues that individual textural characteristics is had correspond in coordinate system, then obtain data in coordinate system so that these are 60 years old Individual value is the shortest to the distance sum of this data place straight line, then these data are exactly the standard value of this textural characteristics, successively may be used In the standard value the most corresponding to obtain these 24 textural characteristics, these standard values and standard ellipse data, ultrasonoscopy to be measured Tumor priori size constitutes a standard vector A together₀。

When in ultrasonoscopy the most to be measured, tumor priori size is by ultrasound detection, doctor or expert's Direct Mark go out tumor Region major and minor axis, unit is generally millimeter, then according to physical length in ultrasonoscopy and pixel ratio, such as every millimeter 4 Pixel, can be calculated the priori size of tumor region.

Above-mentioned each subimage can only comprise a closed contour.

Method used by above-mentioned computed range D is χ²Range formula is tried to achieve:

$D_{x^2}\left(\mathrm{1,0}\right)=\frac{1}{2}{\mathrm{\Σ}}_{k=0}^{25}\frac{{[A_1\left(k\right)-A_0\left(k\right)]}^2}{A_1\left(k\right)+A_0\left(k\right)}$

A₀(k) and A₁K () represents vector A respectively₀And A₁Middle kth parameter, k takes the integer between 0～25.

The invention have benefit that:

(1) in place of changing the clumsiness that traditional manual obtains initial profile.

(2) initial profile acquisition methods is simple, quickly, it is adaptable to movable contour model

(3) initial profile of gained is clearly accurate, has the most excellent as active contour afterwards for convergence Point, substantially increases efficiency and the accuracy of Ultrasound Image Segmentation.

Accompanying drawing explanation

Fig. 1 is the flow chart of automatic profile acquisition methods of the present invention.

Fig. 2 is a width liver tumor image initial profile results figure of present example 1.Fig. 2 (a) is the ultrasonic figure of liver neoplasm As original graph；Fig. 2 (b) is original image pre-processed results after anisotropy parameter and histogram equalization；Fig. 2 (c) is Pretreatment image dynamic threshold segmentation result；Fig. 2 (d) is the initial profile result obtained.

Fig. 3 is two width hysteromyoma ultrasonoscopy initial profile result figures of present example 2, and result is as actively The initial profile of skeleton pattern, obtains the result figure of final tumor boundaries after the convergence under the GGVF field of force drives.

Detailed description of the invention

Carry out the more detailed description present invention below by way of some detailed description of the invention, but the present invention is not limited to these in fact Execute example.

Embodiment 1:

Through liver neoplasm ultrasonoscopy existing to 60 width be trained the standard feature obtaining liver tumor ultrasonoscopy to Amount A₀.The most each width imagery exploitation gray level co-occurrence matrixes is calculated its 24 textural characteristics values, then 60 can be obtained These all characteristic vectors are carried out mathematical linear homing method process, obtain a mark by the group characteristic vector containing 24 elements Quasi-characteristic vector, i.e. by the vectorial A that searching one to the average distance of this stack features vector is the shortest, correspondence in this vector A Each textural characteristics value is exactly the standard value that we need to obtain.Then this feature vector A is added standard ellipse fitting data 0.7 and ultrasonoscopy to be measured in the calculated liver of tumor region major and minor axis that marked by doctor during ultrasound detection swell Two elements of tumor priori size, then just obtained a standard feature vector A containing 26 elements₀。

Fig. 2 (a) is liver tumor ultrasonoscopy I to be split, and Fig. 2 (b) is for carry out anisotropy parameter and Nogata by original image The pre-processed results of gained after figure equilibrium treatment；Fig. 2 (c) is for carry out the knot of dynamic threshold segmentation to gained image in Fig. 2 (b) Really, and extract all closed contours produced in dynamic threshold segmentation result, form corresponding subimage；Utilize gray level symbiosis square Battle array calculates the textural characteristics 24 of each subimage, calculates each subimage ellipse fitting result parameter and subimage Guan Bi wheel Wide interior pixels number, and constitute characteristic vector A by these 26 data₁；Utilize χ²Range formula calculates each subimage Gained vector A₁Standard vector A with tumor region training gained₀Distance D, determine the vectorial A minimum so that D₁Corresponding subgraph Closed contour in Xiang is the initial profile of tumor in ultrasonoscopy to be split, is the initial profile of acquisition shown in Fig. 2 (d) Result.

Embodiment 2:

It is trained obtaining the standard feature of hysteromyoma ultrasonoscopy through hysteromyoma ultrasonoscopy existing to 60 width Vector A₀.The most each width imagery exploitation gray level co-occurrence matrixes is calculated its 24 textural characteristics values, then can obtain These all characteristic vectors are carried out mathematical linear homing method process, obtain one by 60 groups of characteristic vectors containing 24 elements Standard feature vector is i.e. by the vectorial A that searching one to the average distance of this stack features vector is the shortest, corresponding in this vector A Each textural characteristics value be exactly we need obtain standard value.Then this feature vector A is added standard ellipse matching number Calculated according to the tumor region major and minor axis marked by doctor during ultrasound detection in 0.7 and ultrasonoscopy to be measured Two elements of hysteromyoma priori size, then just obtained a standard feature vector A containing 26 elements₀。

Fig. 3 (a1) is hysteromyoma ultrasonoscopy I to be split₁, Fig. 3 (b1) for original image carried out anisotropy parameter and Histogram equalization processes the pre-processed results of gained later；Fig. 3 (c1) divides for gained image in Fig. 3 (b1) is carried out dynamic threshold The result cut, and extract all closed contours produced in dynamic threshold segmentation result, form corresponding subimage；Utilize gray level Co-occurrence matrix calculates the textural characteristics 24 of each subimage, calculates each subimage ellipse fitting result parameter and subimage Closed contour interior pixels number, and constitute characteristic vector A by these 26 data₁；Utilize χ²Range formula calculates each Subimage gained vector A₁Standard vector A with tumor region training gained₀Distance D, determine the vectorial A minimum so that D₁Right Answering the closed contour in subimage is the initial profile of tumor in ultrasonoscopy to be split, is acquisition shown in Fig. 3 (d1) Initial profile result.Fig. 3 (e1) is then as the initial profile of active contour model using the initial profile of gained in Fig. 3 (d1), It is then passed through GGVF external force field and carries out restraining the final hysteromyoma border result figure of gained.

Fig. 3 (a2) is hysteromyoma ultrasonoscopy I to be split₂, Fig. 3 (b2) for original image carried out anisotropy parameter and Histogram equalization processes the pre-processed results of gained later；Fig. 3 (c2) divides for gained image in Fig. 3 (b2) is carried out dynamic threshold The result cut, and extract all closed contours produced in dynamic threshold segmentation result, form corresponding subimage；Utilize gray level Co-occurrence matrix calculates the textural characteristics 24 of each subimage, calculates each subimage ellipse fitting result parameter and subimage Closed contour interior pixels number, and constitute characteristic vector A by these 26 data₁；Utilize χ²Range formula calculates each Subimage gained vector A₁Standard vector A with tumor region training gained₀Distance D, determine the vectorial A minimum so that D₁Right Answering the closed contour in subimage is the initial profile of tumor in ultrasonoscopy to be split, is acquisition shown in Fig. 3 (d2) Initial profile result.Fig. 3 (e2) is then as the initial profile of active contour model using the initial profile of gained in Fig. 3 (d2), It is then passed through GGVF external force field and carries out restraining the final hysteromyoma border result figure of gained.

In embodiment, the extraction of initial profile does not need implementer manually to provide initial profile, and this clearly alleviates The workload of implementer, improves the efficiency of Ultrasound Image Segmentation.

Claims (2)

1. the method obtaining initial profile in a Ultrasound Image Segmentation based on active contour model, it is characterised in that comprise Following steps:

(1) choosing more than 60 width existing ultrasonoscopy composition training set, described existing ultrasonoscopy contains super with to be split The tumor that acoustic image is identical, extracts 24 stricture of vaginas of tumor region from training set with gray level co-occurrence matrix in every width ultrasonoscopy Reason feature, in utilizing mathematical linear to return minimum distance method obtain the standard value of 24 textural characteristics, and and standard ellipse In empirical value, ultrasonoscopy to be measured, tumor priori size constitutes a standard vector A together₀；

(2) ultrasonoscopy to be split is carried out the pretreatment of anisotropy parameter and histogram equalization；

(3) (2) gained image is carried out dynamic threshold segmentation, and extract all Guan Bi wheels produced in dynamic threshold segmentation result Exterior feature, each closed contour forms a corresponding subimage；

(4) utilize gray level co-occurrence matrix to extract 24 textural characteristics of each subimage, calculate ellipse fitting result parameter with And subimage closed contour interior pixels number, and constitute a vectorial A by these 26 data₁；

(5) each subimage gained vector A is calculated₁Standard vector A with tumor region training gained₀Distance D so that D is Little vectorial A₁Closed contour in corresponding subimage is the initial profile of tumor in ultrasonoscopy to be split；

Described 24 textural characteristics include image 0 °, 45 °, 90 °, contrast on 135 ° of four directions, energy, dependency, inverse Difference square, entropy and six features of non-similarity.

In method the most according to claim 1, it is characterised in that in step (1), standard ellipse empirical value is defined as 0.7.