CN101727666B

CN101727666B - Image segmentation method and device, and method for judging image inversion

Info

Publication number: CN101727666B
Application number: CN 200810176865
Authority: CN
Inventors: 刘炎; 孙文武
Original assignee: Shenzhen Mindray Bio Medical Electronics Co Ltd
Current assignee: Shenzhen Mindray Bio Medical Electronics Co Ltd
Priority date: 2008-11-03
Filing date: 2008-11-21
Publication date: 2013-07-10
Anticipated expiration: 2028-11-21
Also published as: CN101727666A

Abstract

The invention discloses image segmentation method and device, and a method for judging image inversion and distinguishing the front side and the back side of a sternum. The image segmentation method comprises the following steps of: a. preprocessing the input image; b. searching the outline of the target area, and checking the search result of the outline, wherein the process for checking the search result of the outline comprises the following steps of: judging whether the shape abnormity index of the target area in the outline search process is less than or equal to a preset threshold value; if so, accepting the result of the search process; and if not, rejecting the result of the search process; and c. extracting the target area according to the check result. By increasing the check procedure, the invention can enable a user to reasonably distinguish the input image, thereby solving the problem that the image abnormity affects the segmentation.

Description

Image partition method and device, image inversion determination methods

Technical field

The present invention relates to technical field of image processing, the dividing processing that is specifically related to image with and application in the image processing of digital X line rabat area of computer aided.

Background technology

Image is cut apart, and particularly be the necessary process of the rabat computer-aided diagnosis of digital X line and image aftertreatment cutting apart of rabat lung images.On the one hand, it has extracted lung's area-of-interest, can reduce the false-positive quantity of disease detection in the lung; On the other hand, it is convenient to the balanced realization of rabat tissue; Simultaneously, it also is the basis of calculating the lung parameter automatically.

The realization technology of lung segmentation is a lot, and technology relatively more commonly used at present mainly contains three kinds:

The rule-based dividing method of the first mainly is to utilize feature of image, handles image so that classical Boundary Detection and Region Segmentation can correctly be identified lung.Though this method realizes simple, adaptivity a little less than.At present, the rule-based lung Boundary Extraction that Ji Neiken (Bram van Ginneken) proposes is that the result is best in the existing report, this method can compatible normal picture and the image of disease is arranged, but shortcoming is not high to image detail part segmentation precision, as when stomach gray scale in the image is dark, mistake can appear in left transeptate location.

It two is pixel classification, extract the feature of image, and the design category device is so that lung and other tissue division are more obvious.This procedure is simple, and rule is few, but implements more consuming timely, because often need to carry out a large amount of recycle ratios, and the result normally presents with the probability form, causes the edge of segmentation result unsmooth.

Its three method that is based on model, the noise resisting ability of this method is more intense, and the result is also smoother complete, but is absorbed in local extremum easily, and lung's structure had unusual or have the image of serious disease to extract also can slip up.Classical ASM (Active Shape Model, ASM) in, the component variation of proper vector should be at a limited field, if limited field is very little, because true shape is innumerable, iteration will be difficult to approach true profile.Under a rational limited field, the situation of shape anomaly may appear in the iteration of ASM, thus make extract the result can not be satisfactory, classical ASM can not identify this mistake automatically, in case above-mentioned situation occurred, then existing dividing method can't be avoided the generation of the problems referred to above.

In addition, for the normotopia rabat, present existing dividing method can't compatible image up and down or left and right occurs to be put upside down, in case such image occurs, gross error can appear in the result.

As seen, there is certain problem in the prior art, needs further to improve.

Summary of the invention

The object of the present invention is to provide a kind of image partition method and device, image inversion to judge and the distinguishing front side and back side of sternum method.

To achieve these goals, the present invention adopts following technical scheme:

A kind of profile searching method that mates most provided by the invention, described method may further comprise the steps successively:

The profile in ferret out zone;

Whether the shape abnormity index of judging the target area outline search process is less than or equal to a predetermined threshold value; If then accept the result of this search procedure; If not, then refuse the result of this search procedure.。This method makes it possible to the image of input is rationally differentiated by increasing checking procedure, has improved the degree of accuracy of algorithm.

The dividing method of a kind of image provided by the invention said method comprising the steps of: A, input picture is carried out pre-service; The profile in B, ferret out zone, and the Search Results of the described profile of check; The process of described checking contour Search Results comprises: whether the shape abnormity index of judging the target area outline search process is less than or equal to a predetermined threshold value; If then accept the result of this search procedure; If not, then refuse the result of this search procedure; C, according to assay, extract the target area.The present invention makes it possible to the image of input is rationally differentiated by increasing checking procedure, has solved image and has existed unusual and problem that influence is cut apart.

The present invention also provides a kind of method of utilizing directional derivative filtering to ask extreme value detected image characteristic boundary, and described method may further comprise the steps successively:

Vertical direction along target signature border precognition trend is calculated the directional derivative of Gauss's masterplate and the convolution of pending image;

Extract after the convolution in the image extreme point position and extreme value value with the derivative equidirectional;

Described extreme point is carried out the border in a preset range connect;

Line by line scan, search is satisfied the curve of default Boundary Detection condition as the target signature border.

The present invention utilizes directional derivative filtering to ask the method for extreme value can detect target area each characteristic boundary in image accurately, and connect by advanced row bound in a preset range, avoided adopting prior art detection method the discrete problem of characteristic boundary that may occur.

A kind of image segmenting device provided by the invention, described device comprises: pretreatment module is used for input picture is carried out pre-service; Optimum matching profile search module is used for the profile to pretreated picture search target area, and the Search Results of the described profile of check; Cut apart module, be used for according to assay, extract the target area; The smoothing processing module is used for the image after the local segmentation is carried out smoothing processing.The present invention makes it possible to the image of input is rationally differentiated by increasing optimum matching profile search module, solved image and existed unusual and problem that influence is cut apart, and this system has superior robustness and careful property.

The present invention also provides a kind of inversion determination methods of asymmetric image, and its process may further comprise the steps successively:

T1, carry out iterative process with two initial profiles of putting upside down mutually up and down respectively, and output and export the shape abnormity index shape abnormity index of two iterative process;

T2, judge that to be inverted profile be whether initial profile iteration result's shape abnormity index is less than being initial profile iteration result's shape abnormity index with the normal profile, if judge that then the image of input is for being inverted image.

The present invention has realized the inversion of image is judged, in terms of existing technologies, has improved computational accuracy, has avoided image inversion and the error that causes, has also improved the adaptivity of system's computing.

The present invention also provides a kind of postero-anterior position, front and back automatic distinguishing method of chest film picture picture, and its process may further comprise the steps:

W1, utilize directional derivative filtering to ask the method for extreme value to detect chest line, the vertical diaphragm in a left side and right vertical diaphragm characteristic boundary respectively;

W2, calculate the vertical diaphragm in a described left side and right vertical diaphragm characteristic boundary to the maximum horizontal range of described chest line respectively;

W3, judge the dirty position in rabat of picture centre according to described maximum horizontal range, if described maximum horizontal range is more than or equal to a predeterminable range value, and the vertical diaphragm characteristic boundary in the described right side is indulged diaphragm to the maximum horizontal range of described chest line to the maximum horizontal range of described chest line greater than a described left side, and then heart is positioned at the right side of rabat.

In the method, ask the method for extreme value to realize postero-anterior position, rabat front and back automatic distinguishing by utilizing directional derivative filtering, improved computational accuracy, avoid the error of calculation that causes because of the initial input image error.

Description of drawings

Fig. 1 is method flow synoptic diagram of the present invention;

Fig. 2 is apparatus structure synoptic diagram of the present invention;

Fig. 3 is when carrying out lung segmentation, the original DR image of input;

Fig. 4 is when carrying out lung segmentation, pretreated result;

Fig. 5 is the abnormal results of classical ASM iteration, and wherein fine rule is initial profile, and thick line is the iteration result.；

Fig. 6 is for when carrying out lung segmentation, and through mating most the result after outline search process is handled, wherein fine rule is initial profile, and thick line is the iteration result;

Fig. 7 is when carrying out lung segmentation, left transeptate erroneous matching result;

Fig. 8 is for when carrying out lung segmentation, and left diaphragm is corrected result afterwards;

Fig. 9 is when carrying out lung segmentation, the result after local segmentation processing and the smoothing processing;

Figure 10 is for when carrying out lung segmentation, and the extraction result on extremal features border is got in directional derivative filtering;

Figure 11 is most preferred embodiment process flow diagram of the present invention.

Embodiment

As shown in Figure 1, the present invention mainly discloses a kind of dividing method of image, and it at first carries out pre-service to input picture; Then, the profile in ferret out zone, and the Search Results of the described profile of check, the profile here refers to the profile after the initial segmentation; At last, according to assay, extract the target area.This method not only can be applied to the rabat lung areas to be cut apart, and cuts apart but also can be applied to the comparatively fixing target of other gray scales and shape.Below describe main technical schemes of the present invention in all its bearings in detail, wherein, the concept of unique point refers to the turnover of border, target area or excessive bigger point.

First aspect, preprocessing process.

Above-mentioned preprocessing process mainly comprises: the grey scale mapping of image is arrived between same gray area, and image degree of comparing is strengthened the process of handling.Wherein, with the grey scale mapping of image to the process between same gray area, mainly be: with the tonal range linear mapping to 0 of image to 255 intervals, carrying out histogram equalization and gray scale negate then handles, make target area and background area be low gray-scale value, corresponding with reference to characteristic boundary (such as with lung areas as in the object area segmentation, with reference to referring to heart and vertical diaphragm with characteristic boundary) be high gray-scale value, obtain image A.Here, also can adopt contrast enhancement process such as adaptive equalization, direct greyscale transformation to replace histogram equalization.

To the process that image degree of comparing strengthen to be handled, mainly be: two the homalographic rectangular images up and down to image A adopt the iteration threshold algorithm to carry out rough background segment respectively, obtain image B after the combination, have only 0 and 255 two kind of pixel value in the image B.Here, also can replace the iteration threshold algorithm with other adaptive thresholding algorithms such as optimal thresholds.The purpose that is divided into two is in order to resist the inhomogeneous influence of gray scale, gray scale is inhomogeneous may cause in the target area local part (such as with lung areas as the apex pulmonis portion in the object area segmentation) because the dark background that is split into of gray scale.

Be example with lung areas as the target area below, above-mentioned preprocessing process implementation method be described:

Common DR(Digital Radiograph, the digital X-ray photography) size of images is 3000 * 3000 pixels (as shown in Figure 2), in order to reduce the workload of follow-up digital processing, can dwindle size of images, such as when carrying out pre-service, after adopting elder generation that image is reduced into 256 * 256 pixels, carry out above-mentioned grey scale mapping with image again and arrive between same gray area, and image degree of comparing is strengthened the process of handling.In addition, interference for fear of the image head neck, in the segmentation process of lung areas, also need to increase the process of removing most of incidence, such as: the image B after will strengthening above-mentioned contrast is carried out bianry image and is opened, and remove zone less than certain pixel (such as 40 pixels, can obtain by test in advance), topmost search for line by line from image B then, if the value of searching is greater than or equal to a gray-scale value (such as 128) for the contiguous pixels number of non-background value in n is capable, image A after then grey scale mapping being handled is since the 0th row, all pixels capable up to n-20 all are set to 0, namely remove most of incidence.The purpose of removing incidence is to get the interference structure of extreme value testing process in order to reduce directional derivative filtering.In like manner, the lower end of image A is also handled with the same manner, because image may turn upside down.So far, preprocessing part is finished, and original image and pre-service result are respectively as Fig. 3 and Fig. 4.As seen, when carrying out the pre-service that lung areas cuts apart, carry out mainly that size is unified, gray scale is unified and incidence is removed and handled.

Second aspect is mated outline search process most.

As shown in Figure 1, the process of the Search Results of the profile in above-mentioned ferret out zone and checking contour mainly comprises following two aspects:

It at first is the process of ferret out region contour, it can be based on active shape model (ASM) iterative algorithm, image is carried out repeatedly the outline search process of target area, by mean profile being carried out the mode that position or size conversion make it to cover the entire image scope, change the initial profile of each iterative process in each outline search process.The mean profile here is to calculate in the ASM training process, classical ASM search procedure is carried out iteration with it as the input initial profile of ASM search procedure, and will be re-used as the ASM initial profile behind its translation convergent-divergent here, so be to have set up a plurality of search procedures.In this process, only change initial profile, just each search procedure is by carrying out mean profile the mode that conversion such as upper and lower, left and right stepping translation and convergent-divergent make it to cover the entire image scope, change the initial profile of active shape model iteration, carry out repeatedly the initial profile search procedure of target area then.The position here or size conversion can be to carry out conversion such as upper and lower, left and right stepping translation and convergent-divergent.

Next is the process of checking contour Search Results, and it mainly is whether the shape abnormity index of judging the target area outline search process is less than or equal to a predetermined threshold value; If then accept the result of this search procedure; If not, then refuse the result of this search procedure.If the result of all search procedures of refusal then also can accept described shape abnormity index and be the result of minimum search procedure, and then carry out subsequent treatment.

The computing method of the described shape abnormity index here may further comprise the steps: in the repeatedly iterative process of target area outline search process, record the component number that surpasses institute's limited field in the feature value vector of each iterative process, the maximal value of described component number is designated as the shape abnormity index of this outline search process.A target area outline search process is to there being repeatedly profile iterative computation processing procedure.

From said process as can be known, the present invention is by increasing the judgement of a shape abnormity index.Improved classical ASM(active shape model ASM, Active Shape Model) algorithm, make it more effective when the search profile.The present invention is intended to provide a Reasonable Parameters and judges this unusual, its principle is: preestablish a constant C HANGE_LIMIT, than under search for profile a bigger eigenwert hop limit, the component number n that in each iterative process, exceeds institute's limited field in the recording feature value vector, with the maximal value N of n as result of determination shape abnormity index whether reasonably.So when implementing, mean profile with training process carries out conversion such as upper and lower, left and right Pan and Zoom as new initial profile respectively, carry out a plurality of search procedures, if there is the N of a process output to be less than or equal to CHANGE_LIMIT, then accept the iteration result of this process as last result, if the shape abnormity index of all processes is all greater than CHANGE_LIMIT, then the key diagram picture may have big unusual, at this moment refuse the result of all processes, perhaps with the process of shape abnormity index minimum as optimum process and accept its result.And for this process, the present invention is called mates outline search process most.

The present invention can be based on classical ASM(active shape module iterative algorithm) come the profile of target area is searched for, then its theoretical foundation is: the deformation degree of profile in iterative process has been expressed in the variation of feature value vector, classical ASM has illustrated that the component value variation of feature value vector should have a limited field, in order to keep the normal of shape, it replaces the value of off-limits component with the maximum magnitude place, shape can not occur unusually in the time of can only guaranteeing like this that limited field is very little, otherwise shape still bigger variation may occur, as Fig. 5, the profile when 28 eigenwerts have 13 to be limited in the iterative process.And if limited field is very little, because true shape is innumerable, iteration will be difficult to approach true profile.Under a rational limited field, the situation of impelling the iteration of ASM shape anomaly to occur mainly contains two kinds: the one, and the lung profile in the image itself has unusually; The 2nd, the initial profile distance objective is far away excessively, and Fig. 5 is exactly this situation.First kind of situation is the problem of image itself, and ASM will lose effectiveness, and the method that the present invention can adopt based on directional derivative filtering to get the extremal features Boundary Extraction solves, and will mention below.Second kind of situation just can solve with the coupling outline search process that the present invention proposes, by with the translation of mean profile stepping ground or each corner of zooming to image as initial profile, can reduce the distance of initial profile and target, and when this distance is optimum, eigenwert change amount also can be very little after the ASM iteration, therefore be not easy to exceed limited field, can obtain optimum profiles.Make and mate profile search most effectively, must be noted that: eigenwert hop limit ratio can not be too small, not so is easy to go beyond the scope, and determines according to experience, and the eigenwert hop limit is [0,3] than general range when carrying out classical ASM algorithm.

Be example with lung areas as the target area below, above-mentioned implementation method of mating outline search process most be described:

ASM needed through off-line training before using, and the parameter that training process of the present invention is used is: training image 44 width of cloth, and the some distributed model does not carry out registration process, the eigenwert interception rate: 99%, surface model gradient vector length: 6, resolution progression: 3.Eigenwert quantity after the present invention's training is 12.Count for configuration sampling, a left side lung and right lung sample respectively 45 and 40 points, totally 85 points, and be labeled as left apex pulmonis summit, left rib diaphragm angular vertex, left cardio-diaphragmatic angle summit, right lung sharp apex, right cardio-diaphragmatic angle summit, right rib diaphragm angular vertex respectively with the 1st, 20,25,46,60,65.The parameter of ASM search procedure is set to: eigenwert hop limit ratio: 2.5, and gradient vector length: 18, maximum iteration time 80, the convergence coefficient is 0.98, adopts the alignment thereof search.The shape abnormity index predetermined threshold value CHANGE_LIMIT that mates most the profile search is set to 0.When the shape abnormity index of mating outline search process most during greater than CHANGE_LIMIT, the present invention refuses the result of this process.Fig. 6 is the result that native system is accepted.

The third aspect, postero-anterior position automatic distinguishing method before and after the rabat, mainly tell about with lung areas as the application in the image dividing processing of target area.

Above-mentioned first aspect has been told about with the image preprocessing process of lung areas as the target area, another innovative point so of the present invention is, when the lung areas of chest film picture picture is carried out dividing processing, increased the process of postero-anterior position, rabat front and back automatic distinguishing, this process can judge whether image is reversed left to right, improves the dividing processing precision of image.

In order to reduce calculated amount, postero-anterior position automatic distinguishing method can be reduced into 200 * 200 pixels with pretreated image A before and after the rabat of the present invention, obtains handling after the image C.Wherein, the process of postero-anterior position automatic distinguishing may further comprise the steps before and after the rabat:

V1, utilize directional derivative filtering to ask the method for extreme value to detect chest line, the vertical diaphragm in a left side and right vertical diaphragm characteristic boundary respectively.Wherein, the process of detection chest line is as follows:

The first step, every row pixel addition of image C is obtained a vector, this vector has reflected the horizontal intensity profile of image C, obtaining this vector from first maximum value position between 150 components of the 50th component to the, will be that the vertical line of horizontal coordinate is as the reference axis of search center line with this position.

In second step, to the filtering result of image C calculating with two-dimentional Gauss's template, the standard deviation of Gauss's template is set at 8 here, selects big standard deviation to be because centre line zone is wideer.If filtered image is D, the local maximum of along continuous straight runs detected image D and minimal value position generate extreme point image E, and the correspondence position of image E is set to 255 and 0 respectively, and other position is set to 127.What image E stored is the extreme point position of image D.At the maximum point place of image D, the value on the relevant position of image E is 255; At the minimal value place of image D, the value on the relevant position of image E is 0.Producing image E is that extreme point connects for convenience.

The 3rd step, image E is carried out vertical extreme point connect, join domain is every some lower-left and 3 * 3 zones, lower right.

In the 4th step, E lines by line scan to image, and the curve of such condition is satisfied in search, and this curve is in maximum position, and vertically height is greater than 30 pixels, and with the minor increment of reference axis be minimum, with the center line of this curve as the thoracic cavity; If do not search center line, then with reference axis as center line.

In the 5th step, this curve is vertically extended up and down, up to top and the lowermost end of image.

In addition, the process of the vertical diaphragm of detection is as follows:

The first step is carried out the gray scale adjustment to image C, and purpose is in order to strengthen the contrast at vertical diaphragm position.Adjustment process is:

Utilize adaptive algorithm split image background, such as other adaptive thresholding algorithms such as iteration threshold algorithm or optimal thresholds; Calculate the average pixel value of background and foreground area respectively, then image C is carried out the gray scale windowing with these two values, obtain image F, the tonal range linearity of image F is adjusted into 0 to 255, utilize the effective tonal range of average gray conduct in background and the body surface, to use the lower boundary gray scale of scope to replace less than the gray scale of this scope, and use the coboundary gray scale of scope to replace greater than the gray scale of this scope, the gray scale adjustment finishes.

In second step, to the first order derivative of image F calculated level direction and with the filtering of two-dimentional Gauss's template, according to the character of convolution, this calculating is equivalent to the derivative that calculates two-dimentional Gauss's template earlier, does convolution with image again, and the standard deviation of Gauss's template is set at 6 here.According to handling filtered image with the same mode of inspection center line, obtain the image G after the border connects, from the 30th row of G, line by line scan, in center line left side, search maximum position, height greater than 30 and the curve of extreme value value maximum as the vertical diaphragm in left side; On the center line right side, search minimal value position, height greater than 30 and the curve of extreme value value minimum as the vertical diaphragm in right side.If do not detect, then the vertical diaphragm of left and right sides is set at the most left and the rightest two vertical line respectively.

V2, calculate the vertical diaphragm in a described left side and right vertical diaphragm characteristic boundary to the maximum horizontal range of described chest line respectively;

V3, judge the dirty position in rabat of picture centre according to described maximum horizontal range, if described maximum horizontal range is more than or equal to a predeterminable range value (such as 10), and the vertical diaphragm characteristic boundary in the described right side is indulged diaphragm to the maximum horizontal range of described chest line to the maximum horizontal range of described chest line greater than a described left side, and then heart is positioned at the right side of rabat;

V4, with image A, C, E, G, chest line and vertical diaphragm border right and left mutually changing.Here, subsequent treatment acquiescence of the present invention is heart deflection image left side, so far image is carried out reversed left to right disposing.The present invention differentiates, filtering, gets extreme value, connects extreme point and utilize the process of rule search to be called derivative filtering at the extreme point image and get the extremal features boundary extraction method image above-mentioned.

Further comprising the steps of before above-mentioned steps V1: gradation of image strengthens step, and its process may further comprise the steps: utilize adaptive algorithm split image background; And utilize background and the interior effective tonal range of average gray conduct of body surface, and will use the lower boundary gray scale of scope to replace less than the gray scale of this scope, use the coboundary gray scale replacement of scope greater than the gray scale of this scope.

Fourth aspect, image turned upside down deterministic process.

As shown in Figure 1, mating most in the outline search process, if accept the result of search procedure, namely shape abnormity index has been less than or equal to CHANGE_LIMIT, and then the key diagram picture has mated finely, is not the inversion image; If the result of refusal search procedure then needs to be inverted judgement, as following process:

At first, carry out iterative process with two initial profiles of putting upside down mutually up and down respectively, and export the shape abnormity index of two iterative process;

Then, judge that to be inverted profile be that initial profile iteration result's shape abnormity index is whether less than being initial profile iteration result's shape abnormity index with the normal profile; If judge that then described input picture is for being inverted image; Otherwise, there is not inversion, keep the result of being inverted before judging.

For being judged as inverted image, according to the result of image inversion determining step, remedial frames obtains the 3rd image; Serve as to handle the basis with the 3rd image, carry out the process of above-mentioned ferret out region contour and the process of described checking contour Search Results, if assay is to accept the result of search procedure, then carry out characteristic boundary correction process process with this Search Results; If assay is the result of this search procedure of refusal, then keep and mate outline search process result before most, utilize directional derivative filtering to ask extreme value to detect the characteristic boundary of target area.

From above-mentioned narration as can be seen, said process also can be realized based on the ASM algorithm, when being inverted judgement, must set up two search procedures then, mean profile after first search procedure is trained ASM is searched in pretreated image as initial profile, second search procedure turns upside down mean profile and searches in pretreated image as initial profile, perhaps pretreated image is put upside down and mean profile is constant, set up two search procedures.Note, only export the shape abnormity index of each search procedure here, but need not mate the search of profile way of search most, be to use the ASM search here, note shape abnormity index, but only carry out a search procedure, do not carry out the conversion of initial profile.Be example with lung areas as the target area below, the implementation method of above-mentioned inversion deterministic process be described:

1, utilize ASM to carry out iteration with two initial profiles of putting upside down mutually up and down respectively, and export the shape abnormity index of two iterative process, the shape abnormity index of establishing two processes is respectively EXCEPTION1 and EXCEPTION2.

2, calculate that pretreated image distinguishes effectively that the capable average gray of top and the default row of lowermost end or several rows add up and, such as the row of the top 10 row row of gray scale and the lowermost end 10 row gray scale that on average adds up that on average adds up, be made as TOP_CUM and BTM_CUM respectively.Effective district here refers to remove the zone of head.

3, judge whether to satisfy one of following two conditions,

Condition one: topmost the row of default row or several rows on average adds up gray scale greater than the gray scale that on average adds up of the row of default row or several rows bottom, and difference is greater than a predefined constant, be TOP_CUM greater than BTM_CUM, and the difference of the two is greater than 10000;

Condition two: topmost on average add up gray scale and the row of default row or several rows bottom of the row of default row or several rows on average adds up the absolute value of difference of gray scale less than described constant, and be that initial profile iteration result's shape abnormity index is less than being initial profile iteration result's shape abnormity index with the normal profile to be inverted profile, be the absolute value of difference of TOP_CUM and BTM_CUM less than 10000, and EXCEPTION2 is less than EXCEPTION1.

If satisfy one of above-mentioned two conditions, then the present invention judges that input picture is to be inverted image, otherwise, there is not inversion, keep the result of being inverted before judging.Above-mentioned concrete reference constant is to come fixed, different target area to adopt different reference constants according to the characteristic of target area, can or rule of thumb know by the training in advance acquisition.So above-mentioned several determination methods also can be applied on other Region Segmentation disposal routes.In the said process, because lung is asymmetric down in shape, so the height of shape abnormity index has reflected possible image orientation, but for the serious image of disease, two matching degrees may be very nearly the same, so will add up in conjunction with the gray scale of top and bottom and judge.

For the normotopia rabat, because the incidence influence makes non-inversion image apex lower than the gray scale accumulated value that the bottom presents.If detect image for being inverted, then image A, C, E, G, chest line, vertical diaphragm border are exchanged up and down.Be less than or equal to CHANGE_LIMIT if detect inversion and EXCEPTION2, then the result of ASM also being exchanged up and down and accepting this result is that initial profile extracts the result, mates profile search search again on the image A after the exchange most otherwise utilize.

Reason is everywhere corrected about characteristic boundary in the 5th aspect.

No matter be the coupling outline search process of above-mentioned second aspect, or the coupling outline search process of the image inversion of above-mentioned fourth aspect after judging, result for the detection search process accepts, and all utilizes shade of gray and directional derivative filtering extreme value to carry out the characteristic boundary correction process.Its processing procedure may further comprise the steps:

Q1, the shade of gray that utilization is vertical and target signature border precognition is moved towards calculate the seed point on initial characteristics border.This process may further comprise the steps:

In the profile of accepting, extract the unique point of at least two characteristic boundaries, with any point coordinate between this unique point as initial point;

Calculate the grey scale difference vector that described initial point place is vertical and target signature border precognition is moved towards;

Calculate maximal value and the position of described grey scale difference vector, with the seed point of this position as the initial characteristics border.In addition, when the dividing processing of some target areas, when correcting left diaphragm in the dividing processing of lung areas, the maximal value that calculates and position thereof also need to satisfy certain standard, could will describe in detail below as the seed point on initial characteristics border.

Q2, utilize directional derivative filtering to ask the method for extreme value, detect the target signature border of satisfying default Boundary Detection condition.The above-mentioned detection method of utilizing directional derivative filtering to ask extreme value of having mentioned describes in detail here again, and this detection method is following realization in order:

Described extreme point is carried out the border in a preset range connect;

Increasing the border step of connecting in said method, is discontinuous for fear of characteristic boundary.Because correlation technique of the prior art the phenomenon that detected characteristic boundary is made of a plurality of line segments may occur, and detection side's rule of utilizing directional derivative filtering to ask extreme value of the present invention can avoid occurring this problem.Its concrete implementation procedure is referring to the testing process of above-mentioned chest line.It is fixed that the preset range that the border connects is come by the size of image.

Below carry out image with lung areas as the target area and be divided into example, describe the processing procedure that left diaphragm is corrected in detail.Because left side lung is influenced by stomach easily, when stomach had big quantity of fluid or bubble, its gray scale was similar to lung, and the two distance is very near, was difficult to sometimes distinguish.Fig. 7 is that matching process is to the transeptate erroneous matching result in a left side.Introduce concrete processing procedure below.

At first, with mate most outline search process as a result the average coordinates of the 20th unique point (left rib diaphragm angular vertex) in the profile and the 25th unique point (left cardio-diaphragmatic angle summit) as initial point, in image, calculate from this more than vertical direction 76 put following vertical grey scale difference vector at 6 to this (note: this difference vector be the gray scale with the every some lower right point gray scale calculating that deducts this point, this is in order to reduce the influence of stomach), and obtain maximal value and the position of this difference vector, if satisfy following three conditions:

1, this maximal value is greater than or equal to first preset value, such as 10;

2, the vertical direction coordinate distance of this position and initial point is greater than second preset value, such as 5;

3, the vertical range difference of this position and right cardio-diaphragmatic angle summit (being profile the 60th unique point) is less than or equal to the 3rd preset value, the purpose that second and three condition is set such as 10(is because the possibility of result before correcting is correct result or does not comprise diaphragm, at this moment need not correct again), then this position as initial left diaphragm seed point coordinate.Can know that in advance left transeptate precognition trend is level, so adopt the mode of vertically asking the grey scale difference vector in the said process here.

Then, utilize directional derivative filtering to ask extreme value to detect left diaphragm border, its process is:

1, move towards to calculate the directional derivative of Gauss's masterplate and the convolution of pending image along the precognition of characteristic boundary, wherein, derivative direction is and horizontal direction angle 60 degree that Gauss's standard deviation is 2, the connection of extreme point level.

2, extract after the convolution in the image extreme point position and extreme value value with the derivative equidirectional;

3, described extreme point being carried out the border in a preset range connects, whether detect the left diaphragm seed point vertical direction some points in front and back (such as 5 points) in the image after the border connects exists and is in maximum position, length greater than the curve of a constant (such as 15 points), if have, this curve then is used as left diaphragm, otherwise withdraw from diaphragm and correct program, keep the testing result of mating profile most.

When detecting left diaphragm curve, in order to prevent from losing the lung edge, then with the some pixels of the downward translation of the curve that searches, such as with 5 pixels of the downward translation of this curve.Then, the position that high order end and the low order end of described curve extended to image boundary and/or target area center line respectively again, be about to this curve high order end and low order end and extend to image boundary and chest line position respectively, bearing of trend is respectively lower left 45 degree and levels.At last, a preset direction pixel grey scale of described curve correspondence position all is set to a gray scale constant, the lower pixel gray scale that is about to curve correspondence position in the image A all is set to 240, obtains image H; Do like this is in order to suppress the interference of the low gray scale of stomach.

After finishing the adjustment of above-mentioned translation and gray scale, adopt the coupling profile searching method of front in image H, to search for profile again, if at this moment Shu Chu shape abnormity index is less than or equal to CHANGE_LIMIT, then accept the result after this result corrects as left diaphragm, otherwise keep the result before correcting.So far, left diaphragm is corrected and is finished correction result such as Fig. 8.

The 6th aspect asks extreme value to detect the characteristic boundary of target area about utilizing directional derivative filtering.

Coupling outline search process after the image inversion of above-mentioned fourth aspect is judged, result for the detection search process refuses, all utilize directional derivative filtering to ask extreme value to detect the characteristic boundary of target area, its processing procedure may further comprise the steps successively:

1, calculates the directional derivative of Gauss's masterplate and the convolution of pending image along the vertical direction of characteristic boundary precognition trend;

3, described extreme point being carried out the border in a preset range connects;

4, line by line scan, search is satisfied the curve of default Boundary Detection condition as characteristic boundary;

5, each characteristic boundary of determining is predicted the extension of trend, made the adjacent feature border of target area occur intersecting;

6, the target area that the characteristic boundary after combination is extended and filling obtain sealing.

Below carry out image with lung areas as the target area and be divided into example, describe in detail and utilize directional derivative filtering to ask extreme value to detect the processing procedure of target area characteristic boundary.

Directional derivative filtering extremal features Boundary Extraction is not subjected to the restriction of lung's shape, therefore when mating most profile search inefficacy, changes this process over to.The front has utilized this method to detect chest line and vertical diaphragm, and the characteristic boundary that also will detect comprises lung center line, lung outer boundaries, diaphragm, apex pulmonis and left side of heart border here.Because the method flow front is mentioned, be example (supposing that left lung is in the image left side) with left lung only here, enumerate parameter and detection rule, i.e. characteristic boundary testing conditions that each characteristic boundary adopts.

(1) the lung center line directly detects in image E, and the line scanning scope is 40 to 195, and stepping is 10, and the column scan scope is for from chest line to its left side 40, testing conditions be in the minimum point place, significant height is maximum curve.Here, significant height refers to the height of the curve segment of the upper right trend in lower-left in the curve.Detect the lung center line just for the further feature border of deriving, its result is not used in and forms last lung border.

(2) lung outer boundaries testing conditions is: curve is in the minimal value position, is in vertical diaphragm left side, highly is maximum.Such as, the lung outer boundaries directly detects in image G, and the line scanning scope is 40 to 195, and stepping is 10, and the column scan scope is for from lung center line to its left side 40, and testing conditions is to be in the minimum point place, highly to be the curve of maximum.

(3) transeptate testing conditions is: curve is in maximum position, length greater than a constant and extreme value value maximum.Such as, diaphragm detects will be to the differentiate of image C elder generation, and the parameter of using is 4 as single order vertical direction derivative, Gauss's standard deviation, and extreme point position level connects.The line scanning scope is that lung center line summit vertical coordinate adds 10 and puts vertical coordinate on earth and add 40,10 points about the some horizontal coordinate at the bottom of the line centered by the column scan scope, testing conditions are to be in maximum point place, length to be greater than or equal to 16 and the curve of extreme value value maximum.

(4) testing conditions of apex pulmonis is: curve is in the minimal value position, is in diaphragm top, length greater than certain constant and extreme value value minimum.Such as, the derivative that the detection apex pulmonis uses is single order vertical direction derivative, and Gauss's standard deviation is 2, and extreme point position level connects.The line scanning scope be above 40 of lung center line summit vertical coordinate the column scan scope is each 5 point about lung center line apex horizontal coordinate to following 15 points, testing conditions is to be in minimum point place, length to be greater than or equal to 8 and the curve of extreme value value minimum.

(5) left side of heart Boundary Detection condition is: curve is in maximum position, in the span maximum of lung outer boundaries and vertical diaphragm by-level direction.Such as, detecting the left side of heart border is in order to prevent that the part cardiac segmentation from being lung, because heart is mainly in the left side.At first obtain the minimum point vertical coordinate on apex pulmonis border and diaphragm border, these two coordinates will be as the vertical range of search cardiac boundary.The horizontal scanning scope is for indulging linea diaphragmatica 40 point ranges left.Extract the first order derivative of image C, derivative direction is and horizontal direction angle 30 degree that Gauss's standard deviation is 4.Testing conditions is the curve that is in maximum point place, horizontal direction span maximum.

(6) vertical diaphragm Boundary Detection condition is: curve is in maximum position, is in the rabat zone line, height is greater than a constant and extreme value value maximum.

After these several Boundary Detection are intact, every border is extended, can not seal because there being the profile of intersection to prevent them, bearing of trend can oneself determine, but to try one's best with reference to the anatomy trend on border, and note avoiding not due intersection, intersect as extension and the diaphragm of apex pulmonis etc.Above-named is the processing of left lung, and right lung will be handled in the same way, just need not extract cardiac boundary again, and the result that the extremal features Boundary Extraction is got in derivative filtering as shown in figure 10.With all boundary combinations after extending, with the lung zone that fill in the zone, two-value morphology is opened, the maximum region reservation obtains sealing.Attention if certain characteristic boundary does not all detect, shows lung's detection failure so in left lung and right lung when the detected characteristics border.

When the result of mating the profile search most is rejected, can also adopt other rule-based dividing method to extract lung border or zone herein.

The local segmentation processing procedure according to assay, can also be carried out in the 7th aspect before extracting the target area, mate the details that the profile search obtains image segmentation result afterwards most for continuing before improving.Its process may further comprise the steps:

C11, the profile that extracts is carried out interpolation obtain closed outline, as the profile of having cut apart;

C12, the profile foundation order profile index to having cut apart are about to profile according to moving towards record profile coordinate clockwise or counterclockwise.；

C13, with default sampling interval to described configuration sampling, sampled point is set up area-of-interest as the regional area center;

C14, adopt adaptive thresholding algorithm to cut apart to described area-of-interest, the result's merging before segmentation result and the local segmentation;

C15, the result who is combined carry out the bianry image finishing, and this dressing process comprises at least two operations in the reservation of burn into maximum region, expansion, the zone filling;

Zone after C16, extraction are filled.

Below carry out image with lung areas as the target area and be divided into example, describe the local segmentation processing procedure in detail.The profile that extracts is carried out linear interpolation, fill then and obtain lung areas, next step carries out local threshold cuts apart.The purpose of local segmentation is in order to improve the details of segmentation result.Foundation order profile index at first.Configuration sampling and also is included into rib diaphragm angular vertex in the sampled point set (in order to obtain rib diaphragm angle accurately) every 10 then, each sampled point is set up 11 * 7 area-of-interest as the regional area center.Adopt the iteration threshold algorithm to cut apart to these zones then, the result before segmentation result and local threshold are cut apart merges.Here, also can replace the iteration threshold algorithm with other adaptive thresholding algorithms such as optimal thresholds.The result who is combined then carries out the bianry image finishing: comprise 3 * 1 corrosion, maximum region keeps, and 3 * 1 expand, and fill in the zone.So far, local threshold is cut apart end.

Eight aspect after carrying out the local segmentation processing, can also be carried out the smoothing processing process to image, thereby eliminates the interference that invalid data produces image.Its process may further comprise the steps:

C21, at least two described clarification of objective point coordinate of record;

C22, according to described characteristic point coordinates, characteristic boundary is divided into some curve segments, and with the trend of each curve segment of vector record;

C23, described each curve segment is sampled with default sampling interval, and carry out interpolation arithmetic according to the trend of described segment;

C24, connection features border make the profile sealing.

Below carry out image with lung areas as the target area and be divided into example, describe the smoothing processing process in detail.Border as a result after the process local threshold is cut apart can produce burr, handles so will carry out edge smoothing.In order to make last segmentation result not produce too significantly sawtooth on original image size, smoothing processing of the present invention is carried out under 512 * 512 sizes.Bianry image after local threshold cut apart is scaled 512 * 512, obtains image I, in order to keep some feature corners of lung profile, as rib diaphragm angle, cardio-diaphragmatic angles etc. are at first noted these characteristic point coordinates, and this can obtain from the coupling profile Search Results of front.Be example with left lung, the present invention extracts rib diaphragm angular vertex: the 20th point, the cardio-diaphragmatic angle summit: the 25th point, apex pulmonis is indulged the diaphragm flex point: the 43rd point, apex pulmonis outside flex point: the 4th point.Notice that these are the coordinates on 256 * 256 images, they be multiply by 2.Then according to these characteristic point coordinates, detected characteristics border in image I, be example with left lung equally, comprise border, lung side, diaphragm, vertical diaphragm, apex pulmonis, and with the trend on these borders of vector record, as the trend on this several characteristic border be respectively vertically, level, vertically, level.The record delimitation trend is very important, because smoothly will utilize interpolation, the interpolation direction must be corresponding with the alignment for the boundary line, otherwise interpolation is with discontinuous even mistake.Utilize cubic spline function that every characteristic boundary is carried out interpolation then, interpolation sampling is 15 points at interval.Carry out the border connection at last and make profile sealing, result such as Fig. 9.Profile filling after level and smooth is zoomed to original size 3000 * 3000 then, obtained last result.

Correct at above-mentioned characteristic boundary and to manage everywhere and utilize directional derivative filtering to ask after extreme value detects the characteristic boundary of target area, will carry out local segmentation respectively handles and smoothing processing, different with said process is, ask after extreme value detects the characteristic boundary of target area utilizing directional derivative filtering, in its smoothing processing process, the extraction of feature corners will utilize the point of crossing on border in the characteristic boundary extension process to specify.

As shown in figure 11, most preferred embodiment of the present invention is, be applied on the lung areas dividing processing of rabat, this embodiment carries out pre-service, the automatic distinguishing processing of postero-anterior position, rabat front and back, mates profile search processing most the image of input successively, when accepting search procedure, carry out left diaphragm correction process, local segmentation processing, smoothing processing successively; When refusal during search procedure, to be inverted judgement, and to re-execute and mate the profile search most and handle, this moment is when accepting search procedure, carries out successively that left diaphragm correction process, local segmentation are handled, smoothing processing; Successively utilize directional derivative filtering ask the characteristic boundary of extreme value detection target area, carry out local segmentation processing and smoothing processing when refusing search procedure this moment.The application of the method for the invention described above on the rabat lung segmentation, by a large amount of PA(postero-anterior position), the AP(anteroposterior position), be inverted the image collection emulation testing that the rabat image blend is formed, the image that various diseases is arranged comprising lung, simulation result is: 90.16% segmentation result and legitimate reading coincide accurately, and the relative legitimate reading of 9.84% segmentation result has small part to lose or unnecessary.

Said method can realize that then the present invention also provides a kind of image segmenting device by computer programming, and as shown in Figure 2, described device comprises:

Pretreatment module is used for input picture is carried out pre-service;

Optimum matching profile search module is used for the profile to pretreated picture search target area, and the Search Results of the described profile of check;

Cut apart module, be used for according to assay, extract the target area.

In addition, in order to improve the details that extract the target area, in described cutting apart in the module local segmentation module can also be set, be used for according to assay, image is carried out the local threshold dividing processing, its process is referring to above-mentioned related description.

For fear of the interference of border burr, in described cutting apart in the module smoothing processing module can also be set, carry out smoothing processing for the image after local threshold is cut apart, its process is referring to above-mentioned related description.

As shown in Figure 2, above-mentioned optimum matching profile search module comprises:

The profile search unit is for the outline search process of image being carried out repeatedly the target area;

Verification unit is used for judging whether the shape abnormity index of each target area outline search process is less than or equal to a predetermined threshold value, and the result of search procedure or the result of refusal search procedure are accepted in output.

As shown in Figure 2, said apparatus also comprises: be used for the result of described verification unit output refusal search procedure is inverted the inversion judge module of judgement, this module comprises:

Iteration unit is used for carrying out iterative process with two initial profiles of putting upside down mutually up and down respectively, and exports the shape abnormity index of two iterative process;

Judging unit be used for to judge that to be inverted profile be that initial profile iteration result's shape abnormity index is whether less than being initial profile iteration result's shape abnormity index with the normal profile; If export described input picture when satisfying above-mentioned condition for being inverted the result of determination of image.

In addition, in order to improve the judgement degree of accuracy,, above-mentioned inversion judge module also comprises: the gray scale unit that adds up, be used for calculating be positioned at image topmost and bottom the gray scale of default row or several rows add up and; Judging unit so, can also be used for judging whether to satisfy one of following two conditions at least: condition one: topmost the row of default row or several rows on average adds up gray scale greater than the gray scale that on average adds up of the row of default row or several rows bottom, and difference is greater than a predefined constant; Condition two: topmost on average add up gray scale and the row of default row or several rows bottom of the row of default row or several rows on average adds up the absolute value of difference of gray scale less than described constant, and is that initial profile iteration result's shape abnormity index is less than being initial profile iteration result's shape abnormity index with the normal profile to be inverted profile; And when satisfying one of above-mentioned two conditions, export described input picture for being inverted the result of determination of image.

As shown in Figure 2, said apparatus also comprises: correcting unit, and this unit connects the output terminal of described judging unit, is used for the remedial frames as a result according to described image inversion determining step, obtain the 3rd image, and the 3rd image is fed through described profile search unit.

As shown in Figure 2, said apparatus also comprises: the image that utilizes shade of gray and directional derivative filtering extreme value that described verification unit is exported carries out the correction process module of characteristic boundary correction process, and this module is accepted between search procedure result's the port for output at input end and the described verification unit of described local segmentation module.

As shown in Figure 2, above-mentioned correction process module comprises: utilize shade of gray vertical and target signature border precognition trend to calculate the first module of initial characteristics border seed point and utilize directional derivative filtering to ask Unit second on extreme value detected characteristics border;

Described first module comprises:

Execution is extracted the unique point of at least two characteristic boundaries in the profile of accepting, and with the unit of any point coordinate between this unique point as initial point;

Be used for calculating described initial point place vertically and the unit of the grey scale difference vector of target signature border precognition trend;

Be used for calculating maximal value and the position of described grey scale difference vector, with the unit of this position as the seed point on initial characteristics border;

Described Unit second comprises:

Be used for calculating along the vertical direction of characteristic boundary precognition trend the unit of the convolution of the directional derivative of Gauss's masterplate and pending image;

For image and the extreme point position of derivative equidirectional and the unit of extreme value value after the extraction convolution;

Be used for described extreme point is carried out the unit that the border connects in a preset range;

Be used for lining by line scan, search is satisfied the curve of default Boundary Detection condition as the unit of characteristic boundary.

As shown in Figure 2, said apparatus also comprises: utilize directional derivative filtering to ask extreme value to detect the characteristic boundary detection module of target area characteristic boundary, this module is used between output refusal search procedure result's the port at input end and the described verification unit of described local segmentation module.

As shown in Figure 2, above-mentioned characteristic boundary detection module comprises:

The unit of the convolution of the directional derivative of Gauss's masterplate and pending image is calculated in execution along the vertical direction of characteristic boundary precognition trend;

Carry out to extract after the convolution in the image and the extreme point position of derivative equidirectional and the unit of extreme value value;

Execution is carried out the unit that the border connects with described extreme point in a preset range;

Execution is lined by line scan, and the unit of the curve of default Boundary Detection condition is satisfied in search;

Be used for each characteristic boundary of determining is predicted the extension of trend, make the adjacent feature border of target area the unit of intersection occur;

The unit that is used for the characteristic boundary after combination is extended and fills the target area that obtains sealing.

As shown in Figure 2, if described correction process module is corrected the left diaphragm in the rabat, then described correction process module also comprises: profile is search unit again, this unit is used for the result of described Unit second output is fed through the described profile search module that mates most, according to the output result of described verification unit, accept Search Results or keep left diaphragm to correct result before.

In addition, when the present invention is used for the lung areas of rabat cut apart, increase postero-anterior position, rabat front and back automatic distinguishing module between described pretreatment module and optimum matching profile search module, this module comprises:

Execution utilizes directional derivative filtering to ask the method for extreme value to detect the unit of chest line, the vertical diaphragm in a left side and right vertical diaphragm characteristic boundary respectively;

Carry out and calculate the vertical diaphragm characteristic boundary of the vertical diaphragm in a described left side and the right side respectively to the unit of the maximum horizontal range of described chest line;

Be used for judging the dirty unit in the rabat position of picture centre according to described maximum horizontal range, if described maximum horizontal range is more than or equal to a predeterminable range value, and the vertical diaphragm characteristic boundary in the described right side is indulged diaphragm to the maximum horizontal range of described chest line to the maximum horizontal range of described chest line greater than a described left side, and then this unit output heart is positioned at the result of determination on rabat right side.

In sum, the present invention not only provides a kind of image partition method, but also postero-anterior position automatic distinguishing method, asymmetric image inversion determination methods, characteristic boundary correction process method before and after the rabat is provided, utilizes directional derivative filtering to ask extreme value to detect method, local segmentation disposal route, the smoothing processing method of target area characteristic boundary, be specially adapted to cutting apart of rabat lung areas.In terms of existing technologies, there is following advantage in the present invention:

1, adopt method of the present invention, not only can identify the image that up and down or left and right is put upside down automatically, and to normal and have the image of disease all can extract correct lung outlines.

2, the difference of ultimate range that vertical diaphragm branch is clipped to the chest line about utilization of the present invention has improved processing accuracy as judging whether foundation reversed left to right of image, and the adaptivity of method.

3, for being inverted image, the present invention utilizes active shape model (Active Shape Model, ASM) search for normal and inverted initial lung profile respectively, which matching degree height the design shape abnormal index differentiates then, the gray scale accumulated value of combining image top and bottom judges whether image is inverted again, improved processing accuracy, and the adaptivity of method.

4, be the image of normal orientation to correcting, the present invention has designed mates outline search process and directional derivative filtering most and gets two kinds of methods of extreme value and cut apart, cut apart for two kinds and act on normal rabat and unusual rabat respectively, these two kinds of rabats also utilize shape abnormity index to judge automatically.Shape abnormity index is an important parameter of the present invention, the present invention by the feature value vector intensity of variation in the statistics ASM iterative process as shape abnormity index, when this index during greater than certain threshold value, the refusal Search Results, re-use conversion foundation such as mean profile translation, convergent-divergent and mate the profile way of search most, can reduce the jump degree of eigenwert, the profile that is mated most.

5, the present invention has also corrected left transeptate location of mistake in conjunction with vertical shade of gray and directional derivative extreme point, makes left diaphragm coupling accurately, guarantees the precision that image is handled.

6, the characteristic boundary that extreme value is got in filtering for directional derivative extracts, the present invention calculates the single order directional derivative and smoothly gets extreme value, again extreme point is carried out assigned direction and connect with the Gaussian filter of various criterion difference, image after connection utilizes the rule detection characteristic boundary then, comprise: vertical diaphragm, lung side profile, diaphragm, apex pulmonis and left side of heart border, at last all borders are carried out the extension of assigned direction.Mate most profile search and derivative filtering and get after extreme value extracts the lung border utilizing, the present invention has carried out local detail to the profile border to be cut apart, and the splines interpolation is carried out on the border, smoothly segmentation result.

7, utilize method of the present invention to carry out lung segmentation because increased extra off-line training, judge upside-down image, normal rabat and unusual rabat different disposal automatically, improved robustness and the careful property of method and device.

Illustrating of above-mentioned each concrete steps is comparatively concrete, can not therefore think the restriction to scope of patent protection of the present invention, and scope of patent protection of the present invention should be as the criterion with claims.In addition, embodiments of the invention are based on mainly that the ASM iterative algorithm is elaborated, but the present invention is not limited to this, same the whole bag of tricks disclosed by the invention and the device of being suitable in the algorithm of similar and ASM iterative algorithm, can also adopt active surface model iterative algorithm to realize the profile search such as the present invention, and in active surface model iterative algorithm, calculate shape abnormity index.

Claims

1. one kind mates the profile searching method most, it is characterized in that, described method may further comprise the steps successively:

The profile in ferret out zone;

Whether the shape abnormity index of judging the target area outline search process is less than or equal to a predetermined threshold value; If then accept the result of this search procedure; If not, then refuse the result of this search procedure;

The process of described ferret out region contour may further comprise the steps:

Image is carried out repeatedly the outline search process of target area, by mean profile being carried out the mode that position or size conversion make it to cover the entire image scope, change the initial profile of each iterative process in each outline search process;

The computing method of described shape abnormity index may further comprise the steps: in the repeatedly iterative process of target area outline search process, record the component number that surpasses institute's limited field in the feature value vector of each iterative process, the maximal value of described component number is designated as the shape abnormity index of this outline search process.

2. method according to claim 1 is characterized in that, if the result of all search procedures of refusal then accepts described shape abnormity index and is the result of minimum search procedure.

3. the dividing method of an image is characterized in that, said method comprising the steps of:

A, input picture is carried out pre-service;

The profile in B, ferret out zone, and the Search Results of the described profile of check;

The process of described checking contour Search Results comprises: whether the shape abnormity index of judging the target area outline search process is less than or equal to a predetermined threshold value; If then accept the result of this search procedure; If not, then refuse the result of this search procedure;

C, according to assay, extract the target area;

4. method according to claim 3 is characterized in that, among the described step B, if the result of all search procedures of refusal then accepts described shape abnormity index and is the result of minimum search procedure.

5. method according to claim 3 is characterized in that, among the described step B, if the result of refusal search procedure then carries out the image inversion determining step, its process comprises the steps:

R1, carry out iterative process with two initial profiles of putting upside down mutually up and down respectively, and export the shape abnormity index of two iterative process;

R2, judge that to be inverted profile be that initial profile iteration result's shape abnormity index is whether less than being initial profile iteration result's shape abnormity index with the normal profile; If judge that then described input picture is for being inverted image.

6. method according to claim 5 is characterized in that, and is further comprising the steps of among the described step R2:

Calculating be positioned at image topmost and bottom the capable average gray of default row or several rows add up with;

Judge whether to satisfy at least one of following two conditions:

Condition one: topmost the row of default row or several rows on average adds up gray scale greater than the gray scale that on average adds up of the row of default row or several rows bottom, and difference is greater than a predefined constant;

Condition two: topmost on average add up gray scale and the row of default row or several rows bottom of the row of default row or several rows on average adds up the absolute value of difference of gray scale less than described constant;

If satisfy one of above-mentioned two conditions, judge that then described input picture is for being inverted image.

7. method according to claim 6 is characterized in that, described step B is further comprising the steps of:

According to the result of described image inversion determining step, remedial frames obtains the 3rd image;

Serve as to handle the basis with described the 3rd image, carry out the process of described ferret out region contour and the process of described checking contour Search Results, if accept the result of search procedure, then carry out characteristic boundary correction process process with this Search Results; If the result of refusal search procedure then keeps this target area profile search result before, utilize directional derivative filtering to ask extreme value to detect the characteristic boundary of target area.

8. according to claim 3 or 7 described methods, it is characterized in that among the described step B, if the result of the outline search process of accepting then utilizes shade of gray and directional derivative filtering extreme value to carry out characteristic boundary and corrects reason everywhere, this processing procedure may further comprise the steps:

Q1, the shade of gray that utilization is vertical and target signature border precognition is moved towards calculate the seed point on initial characteristics border;

Q2, utilize directional derivative filtering to ask the method for extreme value, detect the target signature border of satisfying default Boundary Detection condition.

9. method according to claim 8 is characterized in that, described step Q2 may further comprise the steps:

Described extreme point is carried out the border in a preset range connect;

10. method according to claim 9 is characterized in that, if the lung areas of chest film picture picture is carried out dividing processing, then when correcting left diaphragm, the unique point of described characteristic boundary is left rib diaphragm angular vertex and left cardio-diaphragmatic angle summit; The process that described utilization shade of gray vertical and target signature border precognition trend calculates the seed point on initial characteristics border may further comprise the steps:

Calculate described initial point place perpendicular to the grey scale difference vector of target signature border precognition trend;

Calculate maximal value and the position of described grey scale difference vector, with the seed point of this position as the initial characteristics border;

Carry out following three determining steps:

Judge whether described maximal value is greater than or equal to first preset value;

Judge that whether the vertical direction coordinate distance of described position and described initial point is greater than second preset value;

Whether the vertical range difference of judging described position and right cardio-diaphragmatic angle summit is less than or equal to the 3rd preset value;

Above-mentioned three judged results are when being, with the seed point of described position as the initial characteristics border;

Default Boundary Detection condition among the described step Q2 is: described curve is made of some points before and after the described seed point vertical direction, and is in maximum position, length greater than a constant.

11. method according to claim 10 is characterized in that, and is further comprising the steps of after the described step Q2:

With the some pixels of the downward translation of the curve that searches;

The position that high order end and the low order end of described curve extended to image boundary and chest line respectively;

Described curve lower pixel gray scale all is set to a gray scale constant;

Carry out the process of described ferret out region contour and the process of described checking contour Search Results, if the shape abnormity index of search procedure is less than or equal to described predetermined threshold value, then accept the result of this search procedure, and with the result of this search procedure result as this left diaphragm correction process; Otherwise, keep left diaphragm and correct result before.

12. method according to claim 7 is characterized in that, among the described step B, if the result of refusal outline search process then utilizes directional derivative filtering to ask extreme value to detect the characteristic boundary of target area, its process comprises the steps:

Vertical direction along characteristic boundary precognition trend is calculated the directional derivative of Gauss's masterplate and the convolution of pending image;

Described extreme point is carried out the border in a preset range connect;

Line by line scan, search is satisfied the curve of default Boundary Detection condition as characteristic boundary;

Each characteristic boundary of determining is predicted the extension of trend, make the adjacent feature border of target area occur intersecting;

The target area that characteristic boundary after combination is extended and filling obtain sealing.

13. method according to claim 3 is characterized in that, also comprises among the described step C according to assay carrying out the process that local segmentation is handled, this process may further comprise the steps:

C12, the profile foundation order profile index to having cut apart;

Zone after C16, extraction are filled.

14. method according to claim 3 is characterized in that, described step C also comprises the process of image being carried out smoothing processing, and this process may further comprise the steps:

The unique point coordinate of C21, at least two described target area profiles of record;

C24, connection features border make the profile sealing.

15. method according to claim 3 is characterized in that, the preprocessing process of described steps A comprises: the grey scale mapping of image is arrived between same gray area, and image degree of comparing is strengthened the process of handling.

16. method according to claim 3, it is characterized in that, if the lung areas of chest film picture picture is carried out dividing processing, then between described preprocessing process and described step B, also comprise the process of postero-anterior position, rabat front and back automatic distinguishing, its process may further comprise the steps:

V1, utilize directional derivative filtering to ask the method for extreme value to detect chest line, the vertical diaphragm in a left side and right vertical diaphragm characteristic boundary respectively;

V3, judge the dirty position in rabat of picture centre according to described maximum horizontal range, if described maximum horizontal range is more than or equal to a predeterminable range value, and the vertical diaphragm characteristic boundary in the described right side is indulged diaphragm to the maximum horizontal range of described chest line to the maximum horizontal range of described chest line greater than a described left side, and then heart is positioned at the right side of rabat.

17. image segmenting device, it is characterized in that, described device comprises: be used for input picture is carried out pretreated pretreatment module, be used for pretreated picture search target area profile and check the optimum matching profile search module of described target area profile Search Results, and be used for extracting the object area segmentation module according to assay;

Described optimum matching profile search module comprises:

Verification unit is used for judging whether the shape abnormity index of each target area outline search process is less than or equal to a predetermined threshold value, and the result of search procedure or the result of refusal search procedure are accepted in output;

18. device according to claim 17 is characterized in that, described device also comprises: be used for the result of described verification unit output refusal search procedure is inverted the inversion judge module of judgement, this module comprises:

19. device according to claim 18, it is characterized in that, described device also comprises: correcting unit, this unit connects the output terminal of described judging unit, be used for the remedial frames as a result according to described image inversion determining step, obtain the 3rd image, and the 3rd image is fed through described profile search unit.

20. device according to claim 17, it is characterized in that, described device also comprises: the image that utilizes shade of gray and directional derivative filtering extreme value that described verification unit is exported carries out the correction process module of characteristic boundary correction process, and this module is accepted between search procedure result's the port for output at described input end and the described verification unit of cutting apart module.

21. device according to claim 20, it is characterized in that described correction process module comprises: utilize shade of gray vertical and target signature border precognition trend to calculate the first module of initial characteristics border seed point and utilize directional derivative filtering to ask Unit second on extreme value detected characteristics border;

Described first module comprises:

Be used for calculating described initial point place perpendicular to the unit of the grey scale difference vector of target signature border precognition trend;

Described Unit second comprises:

22. device according to claim 17, it is characterized in that, described device also comprises: utilize directional derivative filtering to ask extreme value to detect the characteristic boundary detection module of target area characteristic boundary, this module is used between output refusal search procedure result's the port at described input end and the described verification unit of cutting apart module;

Described characteristic boundary detection module comprises:

23. device according to claim 21, it is characterized in that, if described correction process module is corrected the left diaphragm in the rabat, then described correction process module also comprises: profile is search unit again, this unit is used for the result of described Unit second output is fed through described optimum matching profile search module, according to the output result of described verification unit, accept Search Results or keep left diaphragm to correct result before.

24. device according to claim 17, it is characterized in that, when described device is used for the lung areas of rabat cut apart, increase postero-anterior position, rabat front and back automatic distinguishing module between described pretreatment module and optimum matching profile search module, this module comprises:

25. the inversion determination methods of an asymmetric image is characterized in that, its process may further comprise the steps successively:

T1, carry out iterative process with two initial profiles of putting upside down mutually up and down respectively, and export the shape abnormity index of two iterative process;

T2, judge that to be inverted profile be whether initial profile iteration result's shape abnormity index is less than being initial profile iteration result's shape abnormity index with the normal profile, if judge that then the image of input is for being inverted image;

26. method according to claim 25 is characterized in that, and is if the lung areas of chest film picture picture is carried out dividing processing, then further comprising the steps of among the described step T2:

Judge whether to satisfy at least one of following two conditions:

If satisfy one of above-mentioned two conditions, judge that then the image of described input is for being inverted image.