CN102682449B - Automatic segmentation method based on self-adaptive external force level set for magnetic resonance images (MRIs) of soft tissue and realization method thereof - Google Patents

Abstract

The invention provides an automatic segmentation method based on self-adaptive external force level set for magnetic resonance images (MRIs) of soft tissue and a realization method thereof. The automatic segmentation method comprises the steps that object region tissue satisfying a gray threshold is selected under the circumstance of interactive operation; gradient information of a smoothened image is calculated; a curve evolution guiding function is set; a boundary stopping function is set; and if an evolutional curve exceeds the boundary of an object region, parameters of the boundary stopping function are adjusted, then curve evolution is re-conducted and the iteration is conducted again until the evolutional curve is converged within the boundary of the object region. The automatic segmentation method has the characteristics of high real-time performance, high operation efficiency, capability of segmenting multiple discrete regions at the same time, capability of accurately recognizing fuzzy boundaries of the soft tissue of a human body, high segmentation precision, clear image detail characteristics, high intelligent degree, no need of manual intervention, stable and reliable operation, and the like.

Description

Soft tissue nuclear-magnetism image adaptive external force level set auto Segmentation and implementation method

Technical field

The present invention relates to a kind of medical image cutting method.Particularly relate to and a kind ofly have without the need to manual intervention, soft tissue nuclear-magnetism image adaptive external force level set auto Segmentation and the implementation method of the characteristic of destination organization can be gone out by auto Segmentation rapidly and accurately.

Background technology

Iamge Segmentation is one of research topic of the field classics such as image procossing, graphical analysis and computer vision, also be one of difficult point, its objective is the region that image is divided into each tool characteristic and target is separated from background and noise, thus providing infrastructural support for follow-up quantitative, qualitative analysis.Due to various image, especially the mode of medical image is different, inevitably introduce certain random noise again in process, form and the intensity difference of adding interesting target in image are very large, also do not have a kind of general method to be suitable for the segmentation of all images so far.MRI (nuclear magnetic resonance, Magnetic Resonance Imaging) is the clinical diagnostic imaging method of current very important one.MRI utilizes nuclear magnetic resonance principle, according to the decay that the energy discharged is different in the inner different structure environment of material, the electromagnetic wave launched is detected by additional gradient magnetic, can learn and form the nuclear position of this object and kind, the structural images of interior of articles can be depicted as accordingly.Since MRI technology is used for body structures, creates a series of revolutionary diagnostic method in medical circle, greatly promoted developing rapidly of medical science, neuro-physiology and Cognitive Neuroscience.MRI technology has good resolving power to soft tissue, become clinical in indispensable a kind of means.But because MRI is long for sweep time, spatial resolution is not ideal enough, and doctor is difficult to by two-dimentional original image the three-D space structure sketching out focus; Simultaneously due to human body soft tissue have that Nonlinear Large Deformation, contrast are low, the characteristic such as obscurity boundary etc., out-of-shape, make with the naked eye to differentiate clinically the lesion tissue such as soft tissue neoplasm and become difficulty, often can occur owing to identifying unclear and sing misdiagnosis and mistreatment that is that cause; And also very complicated for the medical image segmentation of soft tissue, need the dissection doctor of specialty to carry out manual described point drafting, time and effort consuming, clinical actual requirement can not be adapted to, limit the application development of MRI technology.Therefore, the research carrying out the fast automatic dividing method of medical image to human body soft tissue MRI image is very necessary and urgently to be resolved hurrily.

Traditional medical image cutting method manually puts according to anatomical knowledge and experience the marginal point getting area-of-interest by clinician, i.e. manual segmentation, then simulates smooth contour of object by interpolation etc.The method precision is higher, but takes time and effort, and has very large relation with the experience of doctor.Along with the development of computing machine and image processing techniques, there is a lot of computing machine partitioning algorithm, as the image partition method of classics.At present, the image segmentation algorithm of main flow mainly contains threshold method, watershed method, region growing algorithm and active contour model algorithm etc., and these methods are the special partitioning algorithm for concrete a certain or a certain class sclerous tissues image mostly.Due to human body soft tissue MRI image there is out-of-shape, Nonlinear Large Deformation, the feature such as contrast is low, noise is large, classic method is difficult to therefrom split extract destination organization.

The level set active contour model method that development in recent years is got up has fully utilized region and boundary information, is widely used in the field such as Iamge Segmentation and computer vision with the continuity of its accuracy, automatism and final segmentation result.These class methods are easy to the Prior Knowledge Constraints such as the shape of combination segmentation object, become the study hotspot in the segmentation field of medical image gradually, for the segmentation problem solving human body soft tissue MRI image provides an approach.Level Set Method is a kind of method that Euler method solves implicit partial differential equations, is proposed at first by the people such as Osher, for the tracking at fluid motion and flame propagation interface, is introduced in again computer vision and image processing field subsequently.Its main thought is the level set function φ (x, the y that passive curve C (p, t) are embedded into the high one-dimensional space, t) in, by calculated level set function zero level φ, (C (p, t), the evolution of curve of pursuit is carried out in the position of t)=0.Wherein, p is arbitrary parameter variable, and t represents the time.In practical medical image, the edge of object shows as the most significant part of local strength's change, occurs with the discontinuous form of local feature.Ask first differential can obtain a unimodal function to image, peak value is corresponding with the edge of object.Carry out differential to unimodal function, the differential value at peak value place is zero, and peak value both sides symbol is contrary, and original extreme point corresponds to the zero cross point in second-order differential, can extract the edge of image by detecting zero cross point.At present,

Based in the image partition method of level set, in order to ensure the stability of algorithm numerical evaluation, need in curve evolvement process to repeat symbolic measurement initial work consuming time, greatly limit the real-time application of algorithm, the auto Segmentation of human body soft tissue MRI cannot be realized, clinical practice requirement can not be met.The quick variation level set model of the people such as Li, by interpolation symbolic measurement punishment energy functional, fundamentally solves the periodicity initialization matter of symbolic measurement, drastically increases efficiency and the robustness of algorithm.But, because this method only utilizes simple gradient information controlling curve to develop, often there is boundary leakage phenomenon during the fuzzy or discontinuous image of segmenting edge, the requirement of human body soft tissue MRI Iamge Segmentation cannot be met equally.Therefore, the present invention is based on Level Set Theory, for human body soft tissue MRI picture characteristics, exploitation establishes a kind of self-adaptation external force level set algorithm for the fast automatic segmentation of human body soft tissue MRI image and its implementation.

Summary of the invention

Technical matters to be solved by this invention is, provide a kind of have real-time, operation efficiency is high, can split multiple zone of dispersion simultaneously, accurately automatically can identify soft tissue nuclear-magnetism image adaptive external force level set auto Segmentation and the implementation method of the smeared out boundary of destination organization in human body soft tissue MRI image in prosthetic intervention situation.

The technical solution adopted in the present invention is: a kind of soft tissue nuclear-magnetism image adaptive external force level set auto Segmentation and implementation method.The self-adaptation external force level set automatic division method of soft tissue nuclear-magnetism image, comprises the steps:

1) in interactive operation situation, according to the gray threshold of image sequence and input, the choosing of initial profile frame is carried out to human body soft tissue magnetic resonance sequence two-dimensional medical images, chooses the target area tissue meeting gray threshold;

2) gaussian filtering carried out in two-dimensional space two-dimentional soft tissue nuclear magnetic resonance image is level and smooth, and calculate the gradient information of level and smooth rear image, first adopt gaussian kernel function to the smoothing filtering of original image, remove the high frequency noise in image background, then calculate the gradient information of image after smothing filtering;

3) curve evolvement guidance function is set, makes evolution curve select the form adaptive adjustment evolution direction of profile according to initial block;

4) border is set and stops function, start curve evolvement computing, if evolution curve convergence is to object boundary, then split end; Otherwise return step 3) curve evolvement, proceed iteration, if evolution curve directed overshoot zone boundary enters next step;

5) if evolution curve directed overshoot zone boundary, then stop function parameter adjusting to border, then return step 3) curve evolvement, again iteration, until evolution curve convergence is to target area boundaries.

Step 1) described in initial profile only need to comprise arbitrary target area pixel point, the shape of initial profile can arbitrarily be chosen.

Step 2) described in calculating level and smooth after the gradient information of image be:

Original image function is I (x, y), and gaussian kernel function is G _σ, wherein σ is standard variance; Image after smothing filtering is I ', I '=G _σ* I (x, y), its gradient information is

Step 3) described in curve evolvement guidance function be contrary according to both sides, object edge place second-order differential symbol in real image, the characteristic that the extreme point of first differential corresponds to the zero cross point in second-order differential proposes, according to this characteristic, definition level set curve evolvement guidance function:

f _indicator(I)＝kgsign(I″)g(const+|I′|)

Wherein, (I ") represents the second-order differential symbol of original image I to sign, and Section 2 is for accelerating the Evolution Rates of curve near object boundary, even if constant term const >=1 is still greater than zero for ensureing in definitely level and smooth region evolves speed.

Step 4) described in border stop function formula is set:

Wherein, G _σ* I represents and uses gaussian kernel function G _σto the smoothing filtering of image I, represent and ask gradient to the image after level and smooth, threshold parameter ξ can adjust according to parameters such as the resolution of pending image, gray scales, controls evolution curve and finally stops at correct position; When after selected threshold xi, outside between appointment gradient zones (namely ) boundary function value perseverance be zero.

Step 5) described in border stop function parameter adjusting, under quick Variation Model framework, curve evolvement guidance function is weighted to evolution external force term, and introduce new border stopping function, obtain a kind of new self-adaptation external force level set auto Segmentation model, algorithm realizes by minimizing following energy functional:

ε(φ)＝αggξgLength(C)+βgf _indicatorgg _ξgS(C)+vgP _SDF(φ)

Wherein φ is the level set function of high one dimension, under quick Variation Model framework, curve evolvement guidance function is weighted to evolution external force term, and introduce new border stopping function, obtain a kind of new self-adaptation external force level set auto Segmentation model, algorithm realizes by minimizing described energy functional;

Utilize variational calculation energy functional ε relative to the first variation of φ, then have:

$\frac{\∂\mathrm{\ϵ}}{\∂\mathrm{\φ}}=-\mathrm{\δ}\left(\mathrm{\φ}\right)(\mathrm{\α}g\mathrm{div}\left(g_{\mathrm{\ξ}}\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right)+{\mathrm{\βgg}}_{\mathrm{\ξ}}{\mathrm{gf}}_{\mathrm{indicator}})-\mathrm{\γg}(\mathrm{\Δ\φ}-\mathrm{div}\left(\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right))$

If functional ε gets extreme value, then function phi meets Euler-Lagrange equation the gradient descent flow adding time variable t, minimization of energy functional ε is:

$\frac{\∂\mathrm{\ϵ}}{\∂t}=\mathrm{\αg\δ}\left(\mathrm{\φ}\right)g\mathrm{div}\left(g_{\mathrm{\ξ}}\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right)+{\mathrm{\βgg}}_{\mathrm{\ξ}}{\mathrm{gf}}_{\mathrm{indicator}}\mathrm{g\δ}\left(\mathrm{\φ}\right)+\mathrm{\γg}(\mathrm{\Δ\φ}-\mathrm{div}\left(\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right))$

Wherein, Section 1 is weighting arc length item, keeps the flatness of curve in iterative process; Section 2 is self-adaptation external force term, automatically adjusts the direction of propulsion of curve according to the mutual alignment relation of evolution curve and object boundary, ensures that evolution curve approaches towards the border of target all the time; Section 3 is for keeping the symbolic measurement characteristic of evolution function; α, beta, gamma is every weighting coefficient, and δ (φ) is De Lake function, in order to the zero level collection of detection level set function φ.

Owing to adopting level set function, need to use space partial derivative when carrying out gradient calculation in curve evolvement process with with time partial derivative all space partial derivatives all adopt intermediate differential method to approach, and time partial derivative adopts forward difference to approach;

In concrete computation process, adopt the second-order differential of Laplace operator computed image, that is:

$I^{\′\′}={\▿}^{2}I(x,y)=\frac{{\∂}^{2}I(x,y)}{{\∂x}^2}+\frac{{\∂}^{2}I(x,y)}{{\∂y}^2}$

A computer implemented method for the self-adaptation external force level set automatic division method of soft tissue nuclear-magnetism image, be realize self-adaptation external force level set algorithm by Narrow bands, specific implementation step is as follows:

1) with initial profile φ ⁰zero cross point z ⁰centered by construct initial narrow band B ⁰, put iterations k=0, turn to step 4);

2) level set function φ is identified ^kin zero cross point z ^k, construct new arrowband B ^k;

3) narrowband region increases element assignment newly, if arrowband B ^kin newly-increased element be P, Q be arrowband B ^k-1the point that middle distance P is nearest, if put φ k (P)=d;

4) in arrowband, level set upgrades, according to formula $\frac{\∂\mathrm{\ϵ}}{\∂t}=\mathrm{\αg\δ}\left(\mathrm{\φ}\right)g\mathrm{div}\left(g_{\mathrm{\ξ}}\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right)+{\mathrm{\βgg}}_{\mathrm{\ξ}}{\mathrm{gf}}_{\mathrm{indicator}}\mathrm{g\δ}\left(\mathrm{\φ}\right)+\mathrm{\γg}(\mathrm{\Δ\φ}-\mathrm{div}\left(\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right))$ Upgrade arrowband B ^kinterior level set function, when meeting the condition of convergence, stopping iteration, otherwise turning to step 2).

Soft tissue nuclear-magnetism image adaptive external force level set auto Segmentation of the present invention and implementation method, have real-time, operation efficiency is high, can split multiple zone of dispersion, can accurately identify the features such as the smeared out boundary of human body soft tissue simultaneously, and segmentation precision is high, image detail feature is clear, intelligence degree is high, without the need to manual intervention, stable and reliable operation, concrete manifestation is as follows:

(1) two-dimentional soft tissue MRI Iamge Segmentation efficiency is high.

(2) choosing of initial profile frame has nothing to do with shape, and cutting procedure is without the need to manual intervention, and the method for reducing realizes difficulty.

(3) two dimensional image segmentation can be carried out to multiple dispersive target region simultaneously.

(4) curve evolvement efficiency is high, and evolution precision is accurate.

(5) the segmentation quality of target area meets medical image quality requirement completely.

(6) process such as filtering, stress release treatment can be carried out to original image, also can carry out enhancing process to target area.

(7) good to the details segmentation effect of the area-of-interest of skeletal muscle soft tissue MRI image, to the form fractionation in interesting target region and simultaneous adaptation strong.

(8) all computings are carried out in unique two-dimensional image data structure, and data space demand is less.

The present invention is mainly used in the fields such as Medical Image Processing, clinical image engineering, Digital Medicine Technology, radiation therapy, surgery planning, surgical navigational.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of soft tissue nuclear-magnetism image adaptive external force level set automatic division method of the present invention;

Fig. 2 is computer implemented method of the present invention and process flow diagram;

Fig. 3 is the split-run test design sketch that the present invention is applied to different initial profile image,

Wherein, (a) is 140 iteration, and (b) is 90 iteration, and (c) is 160 iteration, and (d) is 160 iteration;

Fig. 4 is the split-run test design sketch that self-adaptation external force Level Set Models of the present invention is applied to different soft tissue MRI image,

Wherein, (a) is 480 iteration, and (b) is 300 iteration, and (c) is 650 iteration, and (d) is 450 iteration.

Embodiment

Below in conjunction with embodiment and accompanying drawing, soft tissue nuclear-magnetism image adaptive external force level set auto Segmentation of the present invention and implementation method are described in detail.

Soft tissue nuclear-magnetism image adaptive external force level set automatic division method of the present invention, the attribute contrary according to the second-order differential symbol of zero cross point both sides in image area, with the symbol of second-order differential value for boot entry, driving is positioned at target internal, outside or the evolution curve crossing with target and automatically, exactly approaches to the edge of target.Meanwhile, new boundary function introduced by model, can constantly adjust its parameter according to the feature of image to be split, ensures that evolution curve finally rests on the edge of target exactly.

Soft tissue nuclear-magnetism image adaptive external force level set automatic division method of the present invention, comprises the steps:

1) in interactive operation situation, according to the gray threshold of image sequence and input, the choosing of initial profile frame is carried out to human body soft tissue magnetic resonance sequence two-dimensional medical images, choose the interesting target regional organization meeting gray threshold, described initial profile only needs to comprise arbitrary target area pixel point, and the shape of initial profile can arbitrarily be chosen;

2) gaussian filtering carried out in two-dimensional space two-dimentional soft tissue nuclear magnetic resonance image is level and smooth, and calculate the gradient information of level and smooth rear image, first adopt gaussian kernel function to the smoothing filtering of original image, remove the high frequency noise in image background, then calculate the gradient information of image after smothing filtering; After described calculating is level and smooth, the gradient information of image is:

Original image function is I (x, y), and gaussian kernel function is G _σ, wherein σ is standard variance.Image after smothing filtering is I ', I '=G _σ* I (x, y), its gradient information is

3) curve evolvement guidance function is set, enables evolution curve select the form adaptive adjustment evolution direction of profile according to initial block.In real image, the edge of object shows as the most significant part of local strength's change, shows the discontinuous characteristic of local feature.By asking first differential can obtain a unimodal function to image, peak value is corresponding with the edge of object.To unimodal function subdifferential again, the differential value at peak value place is zero, and peak value both sides symbol is contrary, and original extreme point corresponds to the zero cross point in second-order differential.Namely the curve evolvement guidance function that the present invention proposes is that the characteristic contrary according to this zero cross point both sides symbol proposes, definition level set curve evolvement guidance function:

f _indicator(I)＝kgsign(I″)g(const+|I′|)(1)

Wherein, (I ") represents the second-order differential symbol of original image I to sign, and Section 2 is for accelerating the Evolution Rates of curve near object boundary, even if constant term const >=1 is still greater than zero for ensureing in definitely level and smooth region evolves speed.

4) border is set and stops function, start curve evolvement computing.Border stops function directly control the position of evolution curve stopping, if evolution curve convergence is to object boundary, then splits end; Otherwise return step 3) curve evolvement, proceed iteration, if evolution curve directed overshoot zone boundary enters next step.

Border stopping function proposed by the invention is:

Wherein, G _σ* I represents and uses gaussian kernel function G _σto the smoothing filtering of image I, represent and ask gradient to the image after level and smooth, threshold parameter ξ can adjust according to parameters such as the resolution of pending image, gray scales, controls evolution curve and finally stops at correct position.Its threshold parameter of human body soft tissue MRI image ξ of different parts is also different, generally chooses destination organization edge gradient value.When after selected threshold xi, outside between appointment gradient zones (namely ) boundary function value perseverance be zero, therefore, even if the iteration curve that goes down to develop also there will not be boundary leakage phenomenon always.

5) if evolution curve directed overshoot zone boundary, then stop function parameter adjusting to border, then return step 3) curve evolvement, again iteration, until evolution curve convergence is to target area boundaries.

Described stops function parameter adjusting to border, under quick Variation Model framework, curve evolvement guidance function is weighted to evolution external force term, and introduce new border stopping function, obtain a kind of new self-adaptation external force level set auto Segmentation model, algorithm realizes by minimizing following energy functional:

$\mathrm{\ϵ}\left(\mathrm{\φ}\right)={\mathrm{\αgg}}_{\mathrm{\ξ}}\mathrm{gLength}\left(C\right)+\mathrm{\β}{\mathrm{gf}}_{\mathrm{indicator}}{\mathrm{gg}}_{\mathrm{\ξ}}\mathrm{gS}\left(C\right)+\mathrm{vg}{P}_{\mathrm{SDF}}\left(\mathrm{\φ}\right)$

$=\mathrm{\α}{\mathrm{gg}}_{\mathrm{\ξ}}g{\∫}_{\mathrm{\Ω}}\mathrm{\δ}\left(\mathrm{\φ}\right)|\▿\mathrm{\φ}|\mathrm{dxdy}+\mathrm{\β}{\mathrm{gf}}_{\mathrm{indicator}}{\mathrm{gg}}_{\mathrm{\ξ}}g{\∫}_{\mathrm{\Ω}}H(-\mathrm{\φ})\mathrm{dxdy}+\mathrm{vg}{\∫}_{\mathrm{\Ω}}\frac{1}{2}\left(\right|\▿\mathrm{\φ}|-1)\mathrm{dxdy}---\left(3\right)$

Wherein φ is the level set function of high one dimension, under quick Variation Model framework, curve evolvement guidance function is weighted to evolution external force term, and introduces new border stopping function, obtains a kind of new self-adaptation external force level set auto Segmentation model.Algorithm realizes by minimizing above-mentioned energy functional.

Utilize variational calculation energy functional ε relative to the first variation of level set function φ, then have:

$\frac{\∂\mathrm{\ϵ}}{\∂\mathrm{\φ}}=-\mathrm{\δ}\left(\mathrm{\φ}\right)(\mathrm{\α}g\mathrm{div}\left(g_{\mathrm{\ξ}}\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right)+{\mathrm{\βgg}}_{\mathrm{\ξ}}{\mathrm{gf}}_{\mathrm{indicator}})-\mathrm{\γg}(\mathrm{\Δ\φ}-\mathrm{div}\left(\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right))---\left(4\right)$

If functional ε gets extreme value, then level set function φ meets Euler-Lagrange equation the gradient descent flow adding time variable t, minimization of energy functional ε is:

$\frac{\∂\mathrm{\ϵ}}{\∂t}=\mathrm{\αg\δ}\left(\mathrm{\φ}\right)g\mathrm{div}\left(g_{\mathrm{\ξ}}\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right)+{\mathrm{\βgg}}_{\mathrm{\ξ}}{\mathrm{gf}}_{\mathrm{indicator}}\mathrm{g\δ}\left(\mathrm{\φ}\right)+\mathrm{\γg}(\mathrm{\Δ\φ}-\mathrm{div}\left(\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right))---\left(5\right)$

Wherein, Section 1 is weighting arc length item, keeps the flatness of curve in iterative process; Section 2 is self-adaptation external force term, automatically adjusts the direction of propulsion of curve according to the mutual alignment relation of evolution curve and object boundary, ensures that evolution curve approaches towards the border of target all the time; Section 3 is for keeping the symbolic measurement characteristic of evolution function.α, beta, gamma is every weighting coefficient, and δ (φ) is De Lake function, in order to the zero level collection of detection level set function φ.

Owing to adopting level set function, need to use space partial derivative when carrying out gradient calculation in curve evolvement process with with time partial derivative all space partial derivatives all adopt simple intermediate differential method to approach, and time partial derivative adopts forward difference to approach.

In concrete computation process, adopt the second-order differential of Laplace operator computed image, that is:

$I^{\′\′}={\▿}^{2}I(x,y)=\frac{{\∂}^{2}I(x,y)}{{\∂x}^2}+\frac{{\∂}^{2}I(x,y)}{{\∂y}^2}---\left(6\right)$

Because Laplace operator is responsive especially to noise, bilateral effect can be produced, the direction of Edge detected can not be used for.Therefore, we first carry out filtering with Gaussian function to image, then ask second derivative by Laplace operator to filtered image.

For improving counting yield further, Dirac function adopts regularization format to describe, and adopts deposited (smeared-out) method of opening of single order precision to approach.

${\mathrm{\δ}}_{{\mathrm{\ϵ}}_3}\left(z\right)=\left\{\begin{array}{cc}0,& \left|z\right|>{\mathrm{\ϵ}}_3\\ (1+\cos \left(\frac{\mathrm{\π}\·z}{{\mathrm{\ϵ}}_3}\right))/2{\mathrm{\ϵ}}_3,& \left|z\right|\≤{\mathrm{\ϵ}}_3\end{array}\right.---\left(7\right)$

Wherein, ε ₃determine the bandwidth that numerical value is level and smooth, usual value ε ₃=1.5 Δ x.

A kind of computing machine Narrow bands proposed by the invention, carries out computing machine realization to above-mentioned human body soft tissue MRI image adaptive external force level set automatic segmentation algorithm.In segmentation curve evolvement process, interested region is only the very narrow region of centered by zero level collection one, and outline line new after a time step only can from a current location inwards or outwards mobile very little segment distance.Therefore, in each iterative process, without the need to considering the remote point outside narrowband region, the Evolution Rates of arrowband internal net point only need be calculated.After curve evolvement arrives border, arrowband, more again centered by zero level collection curve, construct new arrowband.In iterative process, computational fields is limited in a belt-like zone, and the computation complexity of algorithm is from O (n ³) be reduced to O (kn ²), wherein k is the width of arrowband, greatly can improve the implementation efficiency of algorithm.

As shown in Figure 2, the computer implemented method of the self-adaptation external force level set automatic division method of soft tissue nuclear-magnetism image of the present invention, concrete steps are as follows:

1) with initial profile φ ⁰zero cross point z ⁰centered by construct initial narrow band B ⁰, put primary iteration number of times k=0, turn to step 4);

2) level set function φ is identified ^kin zero cross point z ^k, construct new arrowband B ^k, the same step 1) of method;

3) narrowband region increases element assignment newly, if arrowband B ^kin newly-increased element be P, O be arrowband B ^k-1the point that middle distance P is nearest, if put φ ^k(P)=d;

4) in arrowband, level set upgrades, and upgrades arrowband B according to formula (8) ^kinterior level set function φ, when meeting the condition of convergence, stopping iteration, otherwise turning to step 2).

$\frac{\∂\mathrm{\ϵ}}{\∂t}=\mathrm{\αg\δ}\left(\mathrm{\φ}\right)g\mathrm{div}\left(g_{\mathrm{\ξ}}\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right)+{\mathrm{\βgg}}_{\mathrm{\ξ}}{\mathrm{gf}}_{\mathrm{indicator}}\mathrm{g\δ}\left(\mathrm{\φ}\right)+\mathrm{\γg}(\mathrm{\Δ\φ}-\mathrm{div}\left(\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}\right))---\left(8\right)$

Fig. 3 is the split-run test design sketch that the self-adaptation external force level set automatic segmentation algorithm for human body soft tissue nuclear magnetic resonance image of the present invention and computer implemented method are applied to different initial profile image.Get ξ=20, β=2.0, initial profile lays respectively at the diverse location of target image, all converges to the edge of object after several times iteration.

Fig. 4 is the split-run test design sketch that self-adaptation external force Level Set Models is applied to different soft tissue MRI image.By figure (a), (b), (c) and (d) is known, and in image, each interesting target all exists edge fog phenomenon in various degree, and the edges intersect of initial profile and target is pitched.By the parameters of continuous adjustment model, evolution curve finally converges to the edge of target all exactly.In figure, in the first row, closed curve is respectively initial profile, and the curve in the second row is the final evolution result of initial profile.Wherein, parameter ξ=2.25 in the segmentation of first row image, β=4.0, ξ=7.25 in secondary series, β=-2.5, ξ=6.25 in the 3rd row, β=-2.5, ξ=7.20 in last row, β=2.0.

Claims (6)

1. a self-adaptation external force level set automatic division method for soft tissue nuclear-magnetism image, is characterized in that: comprise the steps:

1) in interactive operation situation, according to the gray threshold of image sequence and input, the choosing of initial profile frame is carried out to human body soft tissue magnetic resonance sequence two-dimensional medical images, chooses the target area tissue meeting gray threshold;

2) gaussian filtering carried out in two-dimensional space two-dimentional soft tissue nuclear magnetic resonance image is level and smooth, and calculate the gradient information of level and smooth rear image, first adopt gaussian kernel function to the smoothing filtering of original image, remove the high frequency noise in image background, then calculate the gradient information of image after smothing filtering;

3) curve evolvement guidance function is set, evolution curve is made to select the form adaptive adjustment evolution direction of profile according to initial block, described curve evolvement guidance function is contrary according to both sides, object edge place second-order differential symbol in real image, the characteristic that the extreme point of first differential corresponds to the zero cross point in second-order differential proposes, according to this characteristic, definition level set curve evolvement guidance function:

f _indicator(I)＝k·sign(I″)·(const+|I′|)

Wherein, sign (I ") represents the second-order differential symbol of original image I; (const+|I ' |) for accelerating the Evolution Rates of curve near object boundary; even if constant term const >=1 is for ensureing the image after definitely level and smooth region evolves speed is still greater than zero, I ' expression smothing filtering;

4) border is set and stops function, start curve evolvement computing, if evolution curve convergence is to object boundary, then split end; Otherwise return step 3) curve evolvement, proceed iteration, if evolution curve directed overshoot zone boundary enters next step;

5) if evolution curve directed overshoot zone boundary, then stop function parameter adjusting to border, then return step 3) curve evolvement, again iteration, until evolution curve convergence is to target area boundaries,

Described stops function parameter adjusting to border, under quick variation level set model framework, curve evolvement guidance function is weighted to evolution external force term, and introduce new border stopping function, obtaining a kind of new self-adaptation external force level set auto Segmentation model, realizing by minimizing following energy functional:

ε(φ)＝α·g _ξ·Length(C)+β·f _indicator·g _ξ·S(C)+γ·P _SDF(φ)

Wherein φ is the level set function of high one dimension, g _ξrepresent that border stops function;

Utilize variational calculation energy functional ε relative to the first variation of φ, then have:

If functional ε gets extreme value, then function phi meets Euler-Lagrange equation the gradient descent flow adding time variable t, minimization of energy functional ε is:

Wherein, Section 1 is weighting arc length item, keeps the flatness of curve in iterative process; Section 2 is self-adaptation external force term, automatically adjusts the direction of propulsion of curve according to the mutual alignment relation of evolution curve and object boundary, ensures that evolution curve approaches towards the border of target all the time; Section 3 is for keeping the symbolic measurement characteristic of evolution function; α, beta, gamma is every weighting coefficient, and δ (φ) is De Lake function, in order to the zero level collection of detection level set function φ.

2. the self-adaptation external force level set automatic division method of a kind of soft tissue nuclear-magnetism image according to claim 1, it is characterized in that, step 1) described in initial profile only need to comprise arbitrary target area pixel point, the shape of initial profile can arbitrarily be chosen.

3. the self-adaptation external force level set automatic division method of a kind of soft tissue nuclear-magnetism image according to claim 1, is characterized in that, step 2) described in calculating smothing filtering after the gradient information of image be:

Original image function is I (x, y), and gaussian kernel function is G _σ, wherein σ is standard variance; Image after smothing filtering is I ', I '=G _σ* I (x, y), its gradient information is

4. the self-adaptation external force level set automatic division method of a kind of soft tissue nuclear-magnetism image according to claim 1, is characterized in that, step 4) described in border stop function formula is set:

Wherein, G _σ* I represents and uses gaussian kernel function G _σto the smoothing filtering of image I, represent and ask gradient to the image after level and smooth, threshold parameter ξ adjusts according to the resolution of pending image, grey parameter, controls evolution curve and finally stops at correct position; When after selected threshold xi, outside between appointment gradient zones, namely boundary function value perseverance be zero.

5. the self-adaptation external force level set automatic division method of a kind of soft tissue nuclear-magnetism image according to claim 1, is characterized in that, owing to adopting level set function, needs to use space partial derivative when carrying out gradient calculation in curve evolvement process with with time partial derivative all space partial derivatives all adopt intermediate differential method to approach, and time partial derivative adopts forward difference to approach;

In concrete computation process, adopt the second-order differential of Laplace operator computed image, that is:

。

6. an implementation method for the self-adaptation external force level set automatic division method of soft tissue nuclear-magnetism image according to claim 1, is characterized in that, is to realize self-adaptation external force level set automatic division method by Narrow bands, and specific implementation step is as follows:

1) with initial profile φ ⁰zero cross point z ⁰centered by construct initial narrow band B ⁰, put iterations k=0, turn to step 4);

2) level set function φ is identified ^kin zero cross point z ^k, construct new arrowband B ^k;

3) narrowband region increases element assignment newly, if arrowband B ^kin newly-increased element be P, Q be arrowband B ^k-1the point that middle distance P is nearest, if put φ ^k(P)=d;

4) in arrowband, level set upgrades, according to formula upgrade arrowband B ^kinterior level set function, when meeting the condition of convergence, stopping iteration, otherwise turning to step 2).