CN101599174A - A Level Set Method for Contour Extraction of Medical Ultrasound Image Regions Based on Edge and Statistical Features - Google Patents

Abstract

The invention provides a level set medical ultrasound image area contour automatic extraction method based on edges and statistical features. The invention designs a level set based on edges and statistical features for the characteristics of low contrast and small signal-to-noise ratio of medical ultrasound images. Contour extraction method of set area. The method is divided into three steps: the first step, the rough extraction of the area outline, the completion of the image inversion, the background suppression of the adaptive Gaussian function, the use of the otsu image automatic threshold algorithm to convert the image into a binary image, the image defat interference operation, and the closed area Contour extraction; the second step is to use a selective anisotropic medical ultrasound image smoothing algorithm to preprocess the original image; the third step is to accurately extract the contour of the image area based on the edge and statistical features. Experimental results show that the method of the present invention can obtain more accurate segmentation results compared with existing methods.

Description

Method for outline extraction of level set medical ultrasonic image area based on edge and statistical nature

(1) technical field

The present invention relates to the Medical Ultrasound Image Processing field, relate in particular to a kind of region contour extracting method of medical ultrasonic image.

(2) background technology

Medical ultrasonic image is being brought into play important effect as one of important medical image form in clinical diagnosis and medical research.In clinical, the morphological feature of area-of-interest and area thereof or volume all are important diagnostic information in the ultrasonoscopy diagnosis, and the calculating of these information depends on the profile in zone.Therefore, extract the key link that region contour is the medical ultrasonic image diagnosis.Yet, because ultrasonic signal decay, speckle, shade, leakage signal and because the defectives such as edge loss that the directivity of Image Acquisition causes cause in the image obscurity boundary between each tissue unclear, thereby make the doctor relatively more difficult to area-of-interest profile identification in the ultrasonoscopy, simultaneously, doctor's manual drawing area-of-interest profile is not only loaded down with trivial details but also have subjectivity, and it is significant that therefore research has high-quality profile extraction method.

The region contour of image extracts inseparable with image segmentation, image segmentation is exactly that image is divided into some treatment technologies that the zone of certain sense is arranged, promptly some characteristic according to image is divided into the different zone of SOME PROPERTIES to image, at each intra-zone identical or close characteristic is arranged, and the characteristic of adjacent area is inequality.According to pixel characteristic in the target area inner character consistance and interregional borderline uncontinuity, medical ultrasonic image commonly used is cut apart employing based on the edge-detected image dividing method with based on image partition method (the J.Alison Noble of provincial characteristics, Djamal Boukerroui.Ultrasound ImageSegmentation:A Survey.IEEE Trans.on Medical Imaging.2006,25:987-1010).Dividing method based on rim detection mainly is the closed boundary of seeking area-of-interest, generally all adopt various differentiating operators to realize, as various gradient operators such as Roberts operator, Sobel operator, Laplace operator and wavelet transformation etc., affected by noise bigger owing to differentiate, so quality is not very high when these class methods are cut apart ultrasonoscopy; Image partition method based on provincial characteristics is that general image is divided into the overlapping zonule of some complementations, certain similarity that makes each regional interior pixel is greater than the similarity between the zone, this type of dividing method comprises threshold method, region-growing method, Watershed algorithm, EM algorithm, Mathematical Morphology Method and statistical nature method etc., these class methods are in carrying out processing procedure, all pixels are all carried out computing, often, therefore be unfavorable for effective extraction at edge with edge fog.

In recent years on the basis that traditional pixel characteristic is cut apart, dividing method based on deformation model has also been proposed, comprise the Snakes model, Optimization Model, how much and statistical nature model, physical model, (M.Alem á n such as random field models, P.Alem á n, L.Alvarez.Semiautomatic Snake-based Segmentation of Solid Breast Nodules onUltrasonography.Lectures Notes in Computer Science (EuroCast 2005), 2005,3634:467-472), and utilize the effective numerical method of level set on solution curve evolvement problem to realize method (the Yu-Len Huang that extracts with profile of cutting apart of image, Yu-Ru Jiang, Dar-Ren Chen.Level Set Contouring forBreast Tumor in Sonography.Journal of Digital Imaging.2007,20 (3): 238-247) etc., but segmentation effect is still good inadequately.

In a word, for the automatic extraction of ultrasonic image area profile, the existing methods quality is not high, can't satisfy the actual needs of present medical diagnosis.

(3) summary of the invention

The object of the present invention is to provide a kind of on the basis of traditional images dividing method, the advantage of performance level set aspect the curve evolvement numerical solution can access the method for outline extraction of level set medical ultrasonic image area based on edge and statistical nature that high-quality ultrasonic image area profile extracts the result.

The object of the present invention is achieved like this: it comprises three parts: slightly extract based on the region contour of original image (1); (2) based on the ultrasonoscopy pre-service of original image; (3) the level set image-region profile based on edge and statistical nature of bound fraction (1) and part (2) extracts.The image-region profile slightly extracts, image pre-service, image-region profile are smart extracts, the image-region profile is thick to be extracted according to the characteristics of selection area interested (as tumour etc.) in ultrasonoscopy, adopt image negate, Gaussian function background inhibition, binaryzation, morphology opening operation to go to disturb, need not the initialization Level Set Method and extract the regional coarse contours of extraction such as profile, as the initial value of the smart extraction algorithm of follow-up profile; A kind of selective anisotropic medical ultrasonic image smoothing algorithm is adopted in the image pre-service; The image-region profile is smart to extract the level set region contour extraction method that adopts edge and statistical nature to combine.

(1) first: region contour slightly extracts.

The thick purpose of extracting of region contour is to do initialization for the meticulous extraction of back facial contour, and it is divided into following five steps:

1) image negate

In ultrasonoscopy, mostly present echoless or low echo property characteristic such as area-of-interests such as ventricle, tumor of breast (being selection area).The purpose of image negate is for these zones present high brightness, and tissue on every side becomes low-light level, for next step is prepared.

The negate formula is as follows:

μ ₂(i，j)＝255-μ ₁(i，j) (1)

I in the formula, j are respectively the horizontal ordinate and the ordinates of certain pixel in the image, μ ₁, μ ₂It is respectively before the conversion and the image after the conversion.

2) self-adapting Gaussian function background suppresses

The effect of Adaptive Suppression function is to make the area-of-interest that is positioned at picture centre present high brightness, and deep pixel is in low-light level even is zero.So-called self-adaptation is meant the size automatically adjusting of inhibition function window width according to area-of-interest.The present invention has designed a kind of Adaptive Suppression function based on Gaussian function, function expression as shown in the formula:

$N(p;\widehat{\mathrm{\μ}},\widehat{\mathrm{\σ}})=\frac{\exp (-\frac{1}{2}{(p-\widehat{\mathrm{\μ}})}^T{K}^{-1}(p-\widehat{\mathrm{\μ}}))}{2\mathrm{\π}\sqrt{\det K}}---\left(2\right)$

P is the coordinate of certain some pixel in the image in the formula,

Be the centre coordinate of image, $K=\left(\begin{array}{cc}{{\mathrm{\&sigma;}}_x}^2& \\ & {{\mathrm{\&sigma;}}_{\mathrm{<figure-callout\; id="2"\; label="y"\; filenames="A2009100716410014H1.png,A2009100716410017H1.png"\; state="\{\{state\}\}">y</figure-callout>}}}^2\end{array}\right),$ σ _x ², σ _y ²Be respectively the axial variance of image x, y, and σ _x=w/4, σ _y=h/4, wherein w, h are respectively the width and the height of image.

After the adaptive Gauss inhibition, area-of-interests such as image tumour will be than surrounding tissue brightness height.

3) utilization otsu image automatic threshold algorithm is converted into bianry image to image

The otsu algorithm is the automatic selection algorithm of a kind of image threshold of classics, and this algorithm is asked for the maximal value (being maximum between-cluster variance) of following formula by exhaustive method, thereby determines the threshold value of image:

${\mathrm{\&sigma;}}_b^2\left(t\right)={\mathrm{\&omega;}}_1\left(t\right){\mathrm{\&omega;}}_2\left(t\right){[u_1\left(t\right)-u_2\left(t\right)]}^2---\left(3\right)$

ω in the formula ₁(t), ω ₂(t) being image is divided into the probability of two time-likes, u by threshold value t ₁(t), u ₂(t) be the average of two classes respectively.

Adopt otsu automatic threshold method to calculate the threshold value of image, thereby image is become bianry image.

4) operation such as image degrease, other material interference

From view picture ultrasonoscopy shearing area-of-interest the time, inevitably can around the area-of-interest profile, attach the image that fat or other tissue are arranged.These tissues belong to the dark hypoecho tissue, brightness is lower in medical ultrasonic image, behind the image inversion operation, the image of these tissues will present high brightness, though suppressing operation its brightness meeting of back through adaptive Gauss decreases, yet behind otsu adaptive thresholding Value Operations, wherein a part still might keep.For the interference that prevents that these from organizing bright spot, adopt the method for morphology opening operation to remove these bright spots, promptly consider ultrasonoscopy resolution, to the bright spot of those areas, all remove as interference less than certain pixel count.

5) the closed region profile extracts

Adopt a kind of initialization Level Set Method that need not to extract profile, can obtain the area-of-interest profile of a sealing, its Evolution Equation is:

$\frac{\&PartialD;\mathrm{\&phi;}}{\&PartialD;t}=\mathrm{\&mu;}[\mathrm{\&Delta;\&phi;}-\mathrm{div}\left(\frac{\&dtri;\mathrm{\&phi;}}{|\&dtri;\mathrm{\&phi;}|}\right)]+\mathrm{\&lambda;\&delta;}\left(\mathrm{\&phi;}\right)\mathrm{div}\left(g\frac{\&dtri;\mathrm{\&phi;}}{|\&dtri;\mathrm{\&phi;}|}\right)+\mathrm{vg\&delta;}\left(\mathrm{\&phi;}\right)---\left(4\right)$

Div is a divergence operator in the formula,

Be gradient operator, δ (φ) is a Dirac function, and parameter lambda is positive weight constant, $g=\frac{1}{1+{|{\&dtri;G}_{\mathrm{\&sigma;}}*\mathrm{\&mu;}|}^2},$ * be convolution algorithm, standard deviation is the Gaussian function G of σ _σ(x)=(2 π σ) ^-1/2Exp (| x| ²/ 4 σ).

This Level Set Method has overcome traditional level set needs continuous initialization level set function in evolution process shortcoming, compares with traditional Level Set Method to have that speed is fast, the simple advantage of Numerical Implementation.

(2) second portion: ultrasonoscopy pre-service.

Because ultrasonoscopy structure more complicated, signal to noise ratio (S/N ratio) are low, the meticulous profile that directly extracts area-of-interests such as tumour is difficulty relatively, so earlier ultrasonoscopy is done pre-service.

At the characteristics of medical ultrasonic image, adopt a kind of selective anisotropic medical ultrasonic image smoothing algorithm as image pre-processing method, its expression formula is:

$\left\{\begin{array}{c}\frac{\&PartialD;u(x,y,t)}{\&PartialD;t}=g\left(\right|G_{\mathrm{\&sigma;}}*\&dtri;u\left|\right)|\&dtri;u|\mathrm{div}\left(\frac{\&dtri;u}{|\&dtri;u|}\right)\\ \mathrm{\&mu;}(x,y,0)={\mathrm{\&mu;}}_0(x,y)\end{array}\right.---\left(5\right)$

In the formula

Be gradient operator, * is a convolution algorithm, and div is a divergence operator.Spread function g (| s|) be a monotonic decreasing function, when | s| → ∞, g (| s|) → 0, as the control rate of propagation, to strengthen the image border.Usually spread function is elected as $g\left(s\right)=\frac{1}{1+{(s/k)}^2},$ K wherein is the Grads threshold parameter.G in the formula _σ(x)=(2 π σ) ^-1/2Exp (| x| ²/ 4 σ).Because $G_{\mathrm{\&sigma;}}*\&dtri;u=\&dtri;(G_{\mathrm{\&sigma;}}*u),g\left(\right|G_{\mathrm{\&sigma;}}*\&dtri;u\left|\right)$ Essence be image to be done based on the smoothing processing of Gaussian function to remove noise earlier, the Grad of the image after the computing then is according to Grad decision image smoothing degree.

(3) third part: the level set image-region profile based on edge and statistical nature extracts.

This part is that the essence of image outline is extracted.Elder generation is embedded in the coarse contour image that extracts through first finishing the pretreated image of second portion, as the smart level set initialization of extracting of third part, begins the level set region contour extraction of third part based on edge and statistical nature then.

Be based on the curve differentiation method of geometric deformation model based on the level set region contour extracting method of edge and statistical nature.Suppose to develop curve C (s, t)=[X (s, t), Y (s, t)], wherein s is the arbitrary parameter variable, t is time (R.Malladi, J.A.Sethian, B.C.Vemuri.Shape Modeling with Front Propagation:Level Set Approach.IEEE Trans.on Pattern Analysis and Machine Intelligence, 1995,17 (2): 158-175).If developing the interior of curve is N to unit normal vector, curvature is k, and then curve can be represented with following partial differential equation along the evolution process of its unit normal vector direction:

$\frac{\&PartialD;C}{\&PartialD;t}=V\left(k\right)N---\left(6\right)$

V in the formula (k) is a velocity function, every evolution speed on the decision curve.

The basic thought of geometric deformation model is that speed of deformation (with curvature and/or constant) and view data are combined, and evolving speed is stopped at object edge.How according to view data and geometrical property design proper speed function, make contour curve develop as early as possible and stop at the core content that is based on geometry deformation model image edge extracting on the object boundary exactly.The geometry deformation model is generally realized (R.Malladi with Level Set Method, J.A.Sethian, B.C.Vemuri.Shape Modeling withFront Propagation:Level Set Approach.IEEE Trans.on Pattern Analysis and Machine Intelligence, 1995,17 (2): 158-175).

Suppose a level set function φ (x, y, t), its zero level collection be C (s, t), then have φ (C (and s, t), t)=0, ask its differential, and utilize the chain type rule to obtain t:

$\frac{\&PartialD;\mathrm{\&phi;}}{\&PartialD;t}+\&dtri;\mathrm{\&mu;}\&CenterDot;\frac{\&PartialD;C}{\&PartialD;t}=0---\left(7\right)$

Be without loss of generality, can suppose φ in the zero level collection for negative, for just, the interior of level set curve to unit normal vector is so outside the zero level collection $N=-\frac{\&dtri;\mathrm{\&phi;}}{|\&dtri;\mathrm{\&phi;}|}.$ Wushu (6) is brought formula (7) into, so have:

$\frac{\&PartialD;\mathrm{\&phi;}}{\&PartialD;t}=V\left(k\right)|\&dtri;\mathrm{\&phi;}|---\left(8\right)$

Curvature k at zero level Ji Chu is:

$k=\&dtri;\&CenterDot;\frac{\&dtri;\mathrm{\&phi;}}{|\&dtri;\mathrm{\&phi;}|}=\frac{{\mathrm{\&phi;}}_{\mathrm{xx}}{{\mathrm{\&phi;}}_y}^2-{2\mathrm{\&phi;}}_{\mathrm{xy}}{\mathrm{\&phi;}}_x{\mathrm{\&phi;}}_y+{\mathrm{\&phi;}}_{\mathrm{yy}}{{\mathrm{\&phi;}}_x}^2}{{({{\mathrm{\&phi;}}_x}^2+{{\mathrm{\&phi;}}_y}^2)}^{\frac{3}{2}}}---\left(9\right)$

Formula (8) is the fundamental equation of level set, establishes F=V (k), and function F just is based on the curve evolving speed function of curvature so, and it depends on:

1. the forward position local characteristics that develops of curve for example develops the local curvature of curve;

2. with the relevant external parameter of input data, for example gradient;

3. Fu Jia propagation item.

At the characteristics of ultrasonoscopy, synthetic image marginal information and local texture statistics feature, level set equation speed develops function and elects as:

F＝g(x，y)[λ(k(x，y)-k)-probn(x，y)] (10)

Wherein g (x, y)=(1-η (x, y)) ², (x y) is point (x, the edge confidence degree of y) locating to η.Based on the rim detection of edge confidence degree than have based on traditional gradient operator edge detection method calculate simple, efficient is high, to advantages such as insensitive for noise (Meer Peter, Georgescu Bogdan.Edge Detection with Embedded Confidence.IEEE Trans.Pattern Analysis andMachine Intelligence.2001,23 (12): 1351-1365).Because (x, y)＜1, (x is y) just more near 1, so (x y) can be very little, and the speed of differentiation will approach zero for g when curve develops edge near edge η more for 0＜η.K is a mean curvature, and λ is positive parameter.

Probn (x, y) represent current point (x is the possibility size of area-of-interest outside organization y), and it is drawn by following formula:

$\mathrm{probn}\left(p\right)=\mathrm{SLmap}\left(f\left({\stackrel{\&RightArrow;}{t}v}_9\right)\right)---\left(11\right)$

Wherein p be wait to classify a little coordinate (x, y), mapping function $\mathrm{SLmap}\left(s\right)=\frac{1}{1+e^{-\mathrm{\&beta;s}}},$ β is positive adjusting parameter, and f () is the support vector machine decision function (V.Vapnik.Statistical Learning Theory.Wiley, 1998) based on gaussian kernel function, and its expression formula is:

$f\left(x\right)=\underset{i=1}{\overset{N_s}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_i{y}_{i}K(x_i,x)+{\mathrm{\&omega;}}_0---\left(12\right)$

In the formula (11)

Be nine texture feature vectors, they are respectively mean square deviation coefficients (1) and based on the contrast of grey scale difference image, angle second moment, average, entropy (0 °, 90 ° both direction totally 8).Its expression formula is:

1) mean square deviation coefficient $\mathrm{\&eta;}=\frac{u}{\mathrm{\&sigma;}}---\left(13\right)$

2) contrast $\mathrm{CON}=\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{i}^2{p}_{\mathrm{\&Delta;}}\left(i\right)---\left(14\right)$

3) the angle second order is apart from (energy) $\mathrm{ASM}=\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{p_{\mathrm{\&Delta;}}}^2\left(i\right)---\left(15\right)$

4) average $u=\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{\mathrm{ip}}_{\mathrm{\&Delta;}}\left(i\right)---\left(16\right)$

5) entropy $\mathrm{RNTR}=-\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{p}_{\mathrm{\&Delta;}}\left(i\right){\log }_2\left(p_{\mathrm{\&Delta;}}\left(i\right)\right)---\left(17\right)$

Formula (11) finally is:

$\mathrm{probn}\left(p\right)=\mathrm{SLmap}(\underset{i=1}{\overset{\mathrm{Ns}}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_i{y}_{i}K({\stackrel{\&RightArrow;}{v}}_i,{\stackrel{\&RightArrow;}{t}v}_9)+{\mathrm{\&omega;}}_0)---\left(18\right)$

Wherein K () is a gaussian kernel function, promptly $K(x_i,x)=\exp (-\mathrm{\&gamma;}{\left|\right|x_i-x\left|\right|}^2),$ , λ _i, ω ₀Be the parameter that obtains in the training stage in order to the structure decision function, y _iBe the v of i support vector _iThe class label, N _sIt is the number of support vector.

U, σ are respectively the average and the variances of image in the window.

p _Δ(i) distribution probability of presentation video μ difference, it finds the solution as follows: establish (x, y) be in the image a bit, this point and the point that faces mutually (x+dx, grey scale difference y+dy) is defined as:

Δμ(x，y)＝μ(x，y)-μ(x+dx，y+dy) (19)

Obtain the difference value of every of entire image, as the brightness of image and do normalized, the grey scale difference image Δ μ that composing images μ is ordered according to the histogram of difference image, obtains p this difference result _Δ(i), 0≤i≤L-1, L are the maximum possible difference values.If i is less, but bigger frequency values is arranged, this situation represents that texture is more coarse; If histogram of difference is more smooth, illustrate that then texture is more careful.Generally when calculating based on the grey scale difference parametric texture, do 0 °, 90 ° both direction grey scale differences, get dx=1, dy=1, the textures windows size elects 15 * 15 as, like this every bit is had 9 proper vectors.

As seen from formula (18), when the probn of certain some value more near 1 the time, just think the tissue of area-of-interest outside.If approach 0, this point belongs to area-of-interest.See (10) formula now again, establish level set initialization curve in area-of-interest the time, curve begins to develop, and at this moment function probn approaches 0, and only ((x y)-k) decides k evolving speed, and speed is bigger by curvature part λ.When function probn approached 1, the curve evolving speed was by curvature that develops curve and the decision of probn difference, and evolving speed reduces.By the method for adding image local textural characteristics, the evolving speed that develops curve is not only determined by the edge at this place, also relevant with the image property that develops wavefront place image-region, the wavefront differentiation in weak edge also can be stopped.

Of low quality at present existing ultrasonic image area profile extraction method, can't satisfy the problem of the actual needs of medical diagnosis, the present invention proposes a kind of region contour extracting method of medical ultrasonic image, this method is considered the characteristics of medical ultrasonic image, organically combining based on rim detection with based on the method for provincial characteristics, on the basis of traditional images dividing method, the advantage of performance level set aspect the curve evolvement numerical solution obtained high-quality ultrasonic image area profile and extracted the result.

The medical ultrasonic image area profile extraction method that the present invention proposes can be divided into three parts: (1) region contour slightly extracts; (2) ultrasonoscopy pre-service; (3) the level set image-region profile based on edge and statistical nature extracts.Wherein first and second portion are based on all that original image carries out, and third part is to realize on first and the synthetic result's of second portion basis.Can carry out first's operation earlier, also can carry out the second portion operation earlier, perhaps two parts carry out simultaneously.

The above concrete steps of the medical ultrasonic image area profile extraction method that proposes for the present invention.The present invention combines rim detection and based on the method for provincial characteristics, and the application level collection carries out the curve evolvement numerical solution, therefore can access higher-quality result images.

(4) description of drawings

Fig. 1 is the method for outline extraction of level set medical ultrasonic image area structured flowchart based on edge and statistical nature;

Fig. 2 is the thick leaching process synoptic diagram of profile;

Fig. 3 is embedded into after the pre-service result schematic diagram in the image for the ultrasonoscopy pre-service and a thick profile that extracts;

Fig. 4 extracts the net result synoptic diagram for the area-of-interest profile;

Fig. 5 is a segmentation result method for measuring similarity synoptic diagram.

(5) embodiment

The present invention is further illustrated with concrete embodiment below in conjunction with accompanying drawing:

Among Fig. 1: 101 slightly extract, 102 for being the meticulous extraction of level set region contour based on edge and statistical nature based on anisotropic ultrasonoscopy pre-service, 103 for the area-of-interest profile; Among Fig. 2: 201 for the tumor of breast original image, 202 for image inversion operation result, 203 for the image result of application self-adapting Gaussian function after handling, 204 for slightly extracting the result for profile for the image, 206 after degrease and the interference of other materials based on the bianry image of OTSU adaptive threshold, 205; Among Fig. 3: 301 are embedded into result in the pretreated image for coarse contour for ultrasonoscopy pre-service result, 302; Among Fig. 4: 401 is that original tumor of breast image, 402 is for the level set medical ultrasonic image area profile based on edge and statistical nature extracts the result, 403 contour extraction methods that propose for the present invention extract result's (redness) and medical expert's manual drawing result (white) contrast; Among Fig. 5: the profile that 501 profiles of cutting apart for the inventive method, 502 are cut apart for the inventive method for the profile of medical expert's manual drawing, 503 and the profile phase of medical expert's manual drawing are relatively, the profile that the area of lost part, 504 is cut apart for the inventive method and the profile phase of medical expert's manual drawing are relatively, the profile that the area of lap, 505 is cut apart for the inventive method and the profile phase of medical expert's manual drawing relatively have more the area of part.

In conjunction with Fig. 1, present embodiment is as follows based on the specific implementation step of the level set medical ultrasonic image area profile extraction method of edge and statistical nature:

(1) the region of interest area image is carried out cutting.The purpose of reducing is to reduce the calculated amount that profile extracts automatically, the method of reducing is to comprise whole image of interest zone, and Edge Distance area-of-interest profile is unsuitable excessive, is generally hundreds of to about several ten thousand pixels, is chosen as 96 * 64 pixels in the present embodiment;

(2) extract the initial profile of area-of-interest, the tumor of breast original image for after reducing shown in 201 among Fig. 2 utilizes formula (1) to carry out the image negate, obtains image 202; Pass through formula (2) Gaussian function background again and suppress, obtain image 203; Adopt otsu automatic threshold method to calculate the threshold value of image, image is become bianry image, shown among Fig. 2 204; Adopt morphology opening operation method to remove the discontinuous bright spot at random of image again, in the present embodiment area is thought that less than the bright spot of 5 pixels bright spot is removed at random, reach the purpose of interference such as degrease, obtain image 205; Utilize formula (4) to realize that the sealing image outline extracts, and can obtain the initial profile of a tumour, shown among Fig. 2 206;

(3) the tumor of breast original image shown in 201 in to Fig. 2 utilizes formula (5) to carry out the anisotropy image smoothing, obtains the pretreated image shown in 301 in the accompanying drawing 3;

(4) initial profile 206 of step (2) gained is embedded in the pretreated image 301, obtains image 302;

(5) utilize formula (19) to calculate the difference image of pretreated image;

(6) utilize formula (13)-(18) to calculate the proper vector of region of interest tract tissue

(7) the improvement iteration cross term verification method [7] that proposes of utilization Carl Staelin is determined the parameter of gaussian kernel function support vector machine (C γ), is got C=512 promptly 2 in the present embodiment ⁹, γ=0.03125 promptly 2 ^-5

(8) using formula (10) is cut apart image, extracts the profile of final area-of-interest.

Describe for beneficial effect of the present invention below in conjunction with embodiment.Picked at random three width of cloth malignant galactophore supersonic tumor images and three width of cloth breast ultrasound benign tumour images.The similarity measurement of the profile that the employing aforementioned embodiments is extracted and the tumour profile of medical expert's manual drawing is estimated the profile extraction effect.The medical expert comprises that radiation expert and clinical expert form, so these manual profiles have certain authority, can estimate the effect of automatic extraction.As shown in Figure 5,501 profiles of cutting apart for the inventive method, 502 is the profile of medical expert's manual drawing.The profile phase of 503 profiles of cutting apart for the inventive method and medical expert's manual drawing relatively, the area of lost part is designated as S2; 504 is the area of lap, is designated as S3; 505 for having more the area of part, is designated as S1, and then similarity measurement is defined as follows:

$O=\frac{S3}{\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="A2009100716410017H1.png"\; state="\{\{state\}\}">S</figure-callout>}1+\mathrm{<figure-callout\; id="2"\; label="S"\; filenames="A2009100716410014H1.png,A2009100716410017H1.png"\; state="\{\{state\}\}">S</figure-callout>}2+S3}---\left(20\right)$

The value of O when O=0, illustrates with two profiles and do not occur simultaneously between 0 and 1, when O=1, illustrates that the profile cut apart and the profile of medical expert's manual drawing overlap fully.The similarity measurement result of aforesaid 6 width of cloth tumor images is respectively: 97.84%, 98.45%, 98.33%, 92.12%, 96.62%, 98.39%, and average similarity tolerance is 96.96%.Adopt aforementioned embodiments that 240 breast ultrasound evaluate image (wherein malignant tumour 128 width of cloth, optimum 112 width of cloth) are carried out the tumour profile and extract research, average similarity tolerance is 91.37%.

Document Yu-Len Huang, Yu-Ru Jiang, Dar-Ren Chen.Level Set Contouring for BreastTumor in Sonography.Journal of Digital Imaging.2007,20 (3): the tumour profile of mammary gland is extracted in the 238-247 employing based on smothing filtering and traditional level set contour extraction method, (3 width of cloth are the benign tumour image to have provided 6 width of cloth images in evaluation of algorithm, 3 width of cloth are the malignant tumour image in addition) profile extraction result, similarity measurement is respectively: 95%, 86%, 95%, 95%, 94%, 91%, this 6 width of cloth image similarity tolerance average is 92.67%.Adopt 118 width of cloth image libraries (pernicious 34 width of cloth, optimum 84 width of cloth) that the algorithm that this article proposes is assessed simultaneously in the document, average similarity tolerance is 87.64%.

From above analysis as can be known, the region contour extracting method of the medical ultrasonic image of the present invention's proposition can obtain higher profile and extract quality.

Claims (5)

1. A level set medical ultrasound image region contour automatic extraction method based on edge and statistical features is characterized in that: the method is divided into three steps: image region contour rough extraction, ultrasonic image preprocessing, image region contour fine extraction, image region Rough extraction of area outline According to the characteristics of the selected area of the original image in the ultrasonic image, the rough extraction of the extracted area is performed by using image inversion, Gaussian function background suppression, binarization, morphological opening operation to remove interference, and level set method without initialization. Contour, as the initial value of the follow-up contour fine extraction algorithm; image preprocessing uses a selective anisotropic medical ultrasound image smoothing algorithm; image region contour fine extraction, using the level set region contour automatic extraction method combining edge and statistical features . 2. the level set medical ultrasonic image region contour automatic extraction method based on edge and statistical feature according to right 1, is characterized in that: the level set regional contour automatic extraction method of described edge and statistical feature is combined, calculates first The texture statistical features of the area where each point on the contour is located are input into the trained support vector machine to obtain the signed distance between these texture statistical features and the classification hyperplane, and at the same time introduce a mapping function to map these signed distances to a In the interval, to characterize the possibility that the point belongs to normal tissue, and finally combine the edge confidence and curvature of the point to determine the curve evolution speed of the point, so as to realize the automatic and accurate extraction of the contour of the image region of interest. 3. the level set medical ultrasonic image region contour automatic extraction method based on edge and statistical features according to right 1, is characterized in that: described image region contour rough extraction step comprises following five steps: 1) Image inversion In ultrasound images, regions of interest such as ventricles and breast tumors are mostly anechoic or hypoechoic; the inverse formula is as follows: μ ₂ (i, j) = 255-μ ₁ (i, j) (1) In the formula, i and j are the abscissa and ordinate of a certain pixel in the image respectively, μ ₁ and μ ₂ are the images before and after conversion respectively; 2) Adaptive Gaussian function background suppression The role of the adaptive suppression function is to make the region of interest located in the center of the image show high brightness, and the pixels far away from the center are at low brightness or even zero. An adaptive suppression function based on Gaussian function is set. The function expression is as follows: $\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ;</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; K</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}\sqrt{\mathrm{<span\; class="notranslate">\; det</span>}\mathrm{<span\; class="notranslate">\; K</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$ where p is the coordinates of a pixel in the image,

is the center coordinate of the image, $K=\left(\begin{array}{cc}{{\mathrm{\&sigma;}}_x}^2& \\ & {{\mathrm{\&sigma;}}_{\mathrm{the\; y}}}^2\end{array}\right),$ σ _x ² , σ _y ² are the variances of the x and y axis directions of the image respectively, and σ _x =w, 4, σ _y =h/4, where w and h are the width and height of the image respectively; 3) Use the otsu image automatic threshold algorithm to convert the image into a binary image The Otsu algorithm is a classic image threshold automatic selection algorithm. The algorithm uses the exhaustive method to find the maximum value of the following formula, that is, the maximum inter-class variance, so as to determine the image threshold: ${\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; b</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$ In the formula, ω ₁ (t), ω ₂ (t) are the probability when the image is divided into two categories by the threshold t, and u ₁ (t), u ₂ (t) are the mean values of the two categories respectively; it is calculated by the otsu automatic threshold method The threshold of the image, thus turning the image into a binary image; 4) Image defat, other material interference and other operations Use the method of morphological open operation to remove the bright spots of images with fat or other tissues around the outline of the area, that is, considering the resolution of the ultrasound image, those bright spots with an area smaller than a certain number of pixels are removed as interference; 5) Contour extraction of closed area Using a level set method without initialization to extract the contour, a closed contour of the region of interest can be obtained, and its evolution equation is: $\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; \&phi;</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; \&Delta;\&phi;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; div</span>}<span\; class="notranslate">\; (</span>\frac{<span\; class="notranslate">\; \&dtri;</span>\mathrm{<span\; class="notranslate">\; \&phi;</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; \&dtri;</span>\mathrm{<span\; class="notranslate">\; \&phi;</span>}<span\; class="notranslate">\; |</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&lambda;\&delta;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&phi;</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; div</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; g</span>}\frac{<span\; class="notranslate">\; \&dtri;</span>\mathrm{<span\; class="notranslate">\; \&phi;</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; \&dtri;</span>\mathrm{<span\; class="notranslate">\; \&phi;</span>}<span\; class="notranslate">\; |</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; vg\&delta;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&phi;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$ where div is the divergence operator,

is the gradient operator, δ(φ) is the Dirac function, and the parameter λ is a positive weight constant, $g=\frac{1}{1+{|\&dtri;G_{\mathrm{\&sigma;}}*\mathrm{\&mu;}|}^2},$ * is the convolution operation, the Gaussian function G _σ (x)=(2πσ) ^-1/2 exp(-|x| ² /4σ) with standard deviation σ. 4. The level set medical ultrasonic image region contour automatic extraction method based on edge and statistical features according to claim 1, characterized in that: the ultrasonic image preprocessing step comprises: aiming at the characteristics of medical ultrasonic images, using a The selective anisotropic medical ultrasound image smoothing algorithm is used as an image preprocessing method, and its expression is: $\left\{\begin{array}{c}\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}<span\; class="notranslate">\; *</span><span\; class="notranslate">\; \&dtri;</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; \&dtri;</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; div</span>}<span\; class="notranslate">\; (</span>\frac{<span\; class="notranslate">\; \&dtri;</span>\mathrm{<span\; class="notranslate">\; u</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; \&dtri;</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; |</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$ In the formula

is the gradient operator, * is the convolution operation, and div is the divergence operator. Diffusion function g(|s|) is a monotonically decreasing function. When |s|→∞, g(|s|)→0 is used to control the diffusion speed to enhance the edge of the image. Usually the diffusion function is chosen as $g\left(\mathrm{the\; s}\right)=\frac{1}{1+{(\mathrm{the\; s}/k)}^2},$ where k is the gradient threshold parameter, where G _σ (x)=(2πσ) ^-1/2 exp(-|x| ² /4σ), because $G_{\mathrm{\&sigma;}}*\&dtri;u=\&dtri;(G_{\mathrm{\&sigma;}}*u),$ The essence of the algorithm is to first smooth the image based on the Gaussian function to remove noise, then calculate the gradient value of the processed image, and determine the smoothness of the image according to the gradient value. 5. the method for automatically extracting the contour of the level set medical ultrasound image region based on the edge and the statistical feature according to claim 1, characterized in that: the described extraction step of the contour of the level set image region based on the edge and the statistical feature comprises: first After completing the second part of ultrasound image preprocessing, embed it into the rough contour image extracted in the first part, as the level set initialization of the third part fine extraction, and then start the third part based on the edge and statistical features of the level set area Contour extraction; the level set area contour extraction method based on edge and statistical features is a curve evolution method based on a geometric deformation model, assuming the evolution curve C(s, t) = [X(s, t), Y(s, t)] , where s is any parameterized variable, t is time; suppose the inward unit normal vector of the evolution curve is N, and the curvature is k, then the evolution process of the curve along the direction of its unit normal vector can be expressed by the following partial differential equation: $\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; C</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; V</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$ In the formula, V(k) is a speed function, which determines the evolution speed of each point on the curve.

Images