CN106204600A - Cerebral tumor image partition method based on multisequencing MR image related information - Google Patents

Abstract

The present invention proposes a kind of cerebral tumor image partition method based on multisequencing MR image related information, said method comprising the steps of: S1, reads the most homotactic MR image in same focus；S2, chooses the cerebral tumor region of one of them sequence MR image as tumor seed points；S3, with selected cerebral tumor seed points, carries out similarity mode to the MR image of other sequences, finds the potential cerebral tumor seed points in other sequence MR images and background seed points；S4, using the potential cerebral tumor seed points found as priori, cuts algorithm with figure and realizes the Accurate Segmentation in multisequencing MR image midbrain tumors region.

Description

Cerebral tumor image partition method based on multisequencing MR image related information

Technical field

The present invention relates to cerebral tumor MR image partition method based on figure hugger opinion, belong to Biologic Medical Image and process neck Territory.

Background technology

NMR (Nuclear Magnetic Resonance)-imaging art (Magnetic Resonance Imaging, MRI) is the one of fault imaging, and it utilizes Magnetic resonance phenomenon obtains electromagnetic signal from human body, and reconstructs human body information.MR image have the contrast to soft tissue high, Any direction direct layering imaging, Noninvasive fanout free region and have the outstanding features such as higher spatial resolution, become brain and swell First-selected computer-aided diagnosis means in tumor diagnosis, are widely used in medical diagnosis, treatment, preoperative plan, postoperative monitoring etc. Important step.

In brain tumor diagnosis based on MR image, it usually needs with reference to T1 (longitudinal relaxation time), T2 (laterally pine The relaxation time), T2WI+FLAIR (T2 weighted imaging), four kinds of imaging sequence (the most different pulse trains of T1WI (T1 weighted imaging) The image obtained) information, comprehensively as the diagnosis basis of the cerebral tumor, only according to any one MR image be difficult to differentiate tumor class Not and form.The Accurate Segmentation of the cerebral tumor is significant to operation and radiotherapy, and the therapeutic scheme of the cerebral tumor is often based upon brain The diagnosis of tumor and Accurate Segmentation result, less divided can cause therapeutic scheme invalid, and over-segmentation can damage normal brain cells, impact The physiological behavior of patient.Therefore, the automatic Accurate Segmentation of the cerebral tumor is possible not only to the tentative diagnosis assisting doctor to carry out early stage, subtracts The impact of few artificial subjective factor, and the treatment for the cerebral tumor has great meaning.

At present conventional tumor MR image automatic segmentation method substantially can be divided into following several technological means:

(1) tumor automatic division method based on rim detection: by detecting the borderline tumor information in MR image, obtain Take tumor region thus realize segmentation automatically.But the method is difficult to tackle noise jamming, the situation such as image blurring, therefore Its practical application effect is poor.

(2) tumor automatic division method based on cluster: according to the similarity between MR image slices vegetarian refreshments, the picture of image Vegetarian refreshments set is divided into the process of some subsets.Divide criterion seek to result make certain represent clustering result quality function or Criterion is optimum, and the subset of division is i.e. tumor region and background area.It is initial poly-that its shortcoming is that the accuracy of segmentation depends on The selection of class central point.

(3) tumor automatic division method based on region growing and regional split: the basic thought of region growing is to have The set of pixels having similar quality collectively forms region.Domain decomposition method is then contrary, is that seed region is constantly split into son Region, iteration performs, until till each intra-zone is similar.The inherent defect of region growing is that segmentation effect depends on Relying the selection in seed points and succession, the shortcoming of domain decomposition method is to make border be destroyed.

Patent " the lung tumor dividing method of a kind of PET and CT image cut based on figure, application number CN201410140351.2 " mainly solve the automatic segmentation problem of lung tumor of existing PET and CT image.The method first passes through right PET image up-samples, and PET and CT image carries out relative displacement location, then to the tumor locus of image with non- Tumor locus carries out the demarcation of seed points, utilizes figure to cut algorithm and splits lung tumor.

Patent " a kind of liver neoplasm dividing method based on CT image and device, application number CN201510925624.9 ", The method extract from the standard picture of the CT image of liver pathological changes section and normal tissue sections, be divided into positive sample and Negative sample；Build multi-level degree of depth convolutional neural networks, obtained coarse segmentation bianry image and the pixel of tumor by grader The probabilistic image of classification；The coarse segmentation bianry image of tumor is carried out morphological erosion operation, it is thus achieved that figure cuts required prospect Image, then the coarse segmentation bianry image of the bianry image of liver with tumor is made phase reducing and carries out morphological erosion operation, Obtain the background image corresponding to liver normal structure；Building non-directed graph, use figure cuts optimized algorithm and obtains final point of tumor Cut region.

At the beginning of patent " the lung 4D-CT tumor automatic division method cut based on figure, application number CN201510518535.2 " obtains The geometric center point of tumor on beginning phase place 3D-CT image, obtains initial phase target seed, then calculates each adjacent phase target Moving displacement between seed, is added the target seed of initial phase with moving displacement, it is thus achieved that the target seed of other phase places Position, on each phase object seed obtained, utilizes figure segmentation method to obtain lesion segmentation result.

Patent " dividing method of multi-modal magnetic resonance image (MRI) based on rarefaction representation and device, application number CN201310695295.4 ", by setting up Image Segmentation Model based on MAP-MRF framework, image is carried out Accurate Segmentation, profit With markov random file, take into full account the impact of the neighbor of pixel surrounding space, add the accuracy of image segmentation, The optimization method simultaneously using online dictionary learning method to cut with figure, improves the speed of service.

Different from the selection that the use relative displacement of above-mentioned part patent of invention carries out tumor region seed, MR image becomes with CT As principle is different, it is impossible to use relative displacement carries out the selection of the tumor region seed of different sequence MR image, this patent scheme Use the related information between the most homotactic MR image, tumor region is carried out similarity detection, to realize different sequence The MR image tumor region Primary Location of row, it is not necessary to the repeatedly manual tumor region seed selecting segmentation；Additionally, above-mentioned patent carries The figure segmentation method gone out is not carried out carrying out lung tumor region probabilistic forecasting, lacks expertise, the single sequence of simple dependence The information of row MR image is split, and the priori that this patent scheme uses the probabilistic information of tumor region prediction to cut as figure is known Know, improve the accuracy of segmentation.

Summary of the invention

Present invention aim to overcome that the deficiency of the existing cerebral tumor MR image partition method cut based on figure, especially tumor In the presence of blur margin is clear, fuzzy, the poor effect of segmentation automatically.A kind of brain based on multisequencing MR image related information is provided Tumor image dividing method, the method uses the related information between the most homotactic MR image, carries out tumor region Similarity detects, and to realize the most homotactic MR image tumor region seed Primary Location, greatly reduces manual intervention；This Outward, the priori using the probabilistic information of tumor region prediction to cut as figure, improve the accuracy of segmentation.

For solving above-mentioned technical problem, the present invention adopts the following technical scheme that: a kind of based on the association of multisequencing MR image The cerebral tumor image partition method of information, comprises the following steps:

S1, reads the most homotactic MR image in same case；

S2, chooses the cerebral tumor region of one of them sequence MR image as tumor seed points；

S3, with the cerebral tumor seed points selected by current sequence MR image, carries out similarity to the MR image of other sequences Coupling, finds the potential cerebral tumor seed points in other sequence MR images；

S4, using the potential cerebral tumor region found and background seed points as priori, cuts algorithm with figure and realizes many The segmentation in sequence MR image midbrain tumors region.

Further, in step S1 of the present invention, the most homotactic described MR image includes T1 (longitudinal relaxation time), T2 (transverse relaxation time), T2WI+FLAIR (T2 weighted imaging), four kinds of imaging sequence (the most different arteries and veins of T1WI (T1 weighted imaging) Rush the image that sequence obtains).

Further, the seed points system of selection that described step S2 uses only needs to enter in the MR image of a sequence wherein Row selects, it is not necessary to select in all of sequence MR image, and selection mode includes that click is rule, draws a circle, taken a little etc. Man-machine interaction mode.

Further, in step S3 of the present invention, the Similarity Match Method of employing specifically includes:

(1) cerebral tumor seed points region selected by described current sequence MR image is as template；

(2) the MR image to other sequences carries out pointwise Regional Similarity calculating from top to bottom, from left to right；

(3) the maximum region of similarity is searched out as the potential cerebral tumor region in this sequence MR image, remainder As potential background parts；

Further, in step S4 of the present invention, specifically include:

(1) cerebral tumor region potential in MR image step S3 obtained is as target seed；

(2) MR image being mapped as network, pixel corresponds to the node of figure, and adds extra two nodes: source point And meeting point, there are line, each pixel to have line with source point and meeting point between each two neighbor pixel, these lines are made Limit for network；

(3) Regional Similarity probability step S3 obtained is used for the weights on the limit of setting network figure, and calculates network Energy function；

(4) cut algorithm according to the energy function use figure of figure to solve, using the solution of minima as the segmentation result of figure, Weighing criteria is used to be compared with goldstandard by segmentation result.

Accompanying drawing explanation

Fig. 1 is the method flow diagram of a kind of embodiment of the present invention.

Fig. 2 is the method schematic diagram of selected seed point

Fig. 3 is the method flow diagram of cerebral tumor Regional Similarity coupling

Fig. 4 is tumor region similarity search process schematic

Fig. 5 is the method flow diagram of figure segmentation method

Detailed description of the invention

Below in conjunction with the accompanying drawings and specific embodiment the present invention is carried out in further detail with complete explanation.May be appreciated It is that specific embodiment described herein is only used for explaining the present invention, rather than limitation of the invention.

As it is shown in figure 1, a kind of based on multisequencing MR image related information the cerebral tumor image partition method of the present invention, bag Include following steps:

S1: obtain same case the most homotactic MR image, specifically include T1 (longitudinal relaxation time), T2 (transverse relaxation Time), T2WI+FLAIR (T2 weighted imaging), four kinds of imaging sequences of T1WI (T1 weighted imaging).

S2: choose the seed points selecting party that the cerebral tumor region of one of them sequence MR image uses as tumor seed points Method only needs to select in the MR image of a sequence wherein, it is not necessary to selects in all of sequence MR image, selects Mode includes that click is rule, draws a circle, taken the man-machine interaction mode such as a little.As in figure 2 it is shown, one of which preferably selects mode it is Click line interactive mode, human brain region 1 comprises tumor region 2, and mouse 3 clicks on left mouse button, draws in tumor region 2 Straight line 4, then the pixel at straight line 4 place is regarded as tumor seed points.

S3: as shown in Figure 3 and Figure 4, with the cerebral tumor seed points selected by current sequence MR image, the MR to other sequences Image carries out similarity mode, finds the potential cerebral tumor seed points in other sequence MR images, is described in detail below:

1), by described current sequence MR image select cerebral tumor seed points region A (i, j) as template, wherein i =1,2 ..., m；J=1,2 ..., n, (i, j) represent current pixel point position, m and n represent respectively region length and Wide；

2) the pointwise region that, as shown in Figure 4, the MR image to other sequences is carried out from top to bottom, from left to right is searched Rope, take out one of them region B (i, j) carries out Similarity Measure, wherein i=1,2 ..., m；J=1,2 ..., n, calculate public affairs Formula is as shown in Equation 1:

$\mathrm{\γ}=\frac{\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{m}{\Σ}}\[A(i,j)-\stackrel{\‾}{A}\]\[B(i,j)-\stackrel{\‾}{B}\]}{\sqrt{\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{m}{\Σ}}{\[A(i,j)-\stackrel{\‾}{A}\]}^2}\·\sqrt{\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{m}{\Σ}}{\[B(i,j)-\stackrel{\‾}{B}\]}^2}}---\left(1\right)$

Wherein γ represent image-region A (i, j) and region B (i, similarity j) ,-1≤γ≤1；

3), travel through this sequence MR image, search out the maximum region of similarity as the potential brain in this sequence MR image Tumor region, remainder is as potential background parts；

S4: as it is shown in figure 5, using the potential cerebral tumor region found as priori, cut algorithm with figure and realize many sequences The segmentation in row MR image midbrain tumors region, is described in detail below:

1) cerebral tumor region potential in the MR image, step S3 obtained is as target seed；

2), MR image being mapped as network, pixel corresponds to the node of figure, and adds extra two nodes: source point And meeting point, there are line, each pixel to have line with source point and meeting point between each two neighbor pixel, these lines are made Limit for network；

3) the Regional Similarity probability, step S3 obtained is used for the weights on the limit of setting network figure and calculates network Energy function, specific as follows:

A) label of all of pixel in given network, is expressed as L={l₁, l₂..., l_s, wherein l_i=1 represents Tumor, l_i=0 represents background, and the weight table on the limit of any two neighbor pixel p and q is shown as B_{＜ p, q ＞}, B_{＜ p, q ＞}With adjacent picture The gray value I of vegetarian refreshments p and q_pAnd I_qSimilarity is directly proportional, as shown in Equation 2.

$B_{<p,q>}\&Proportional;\exp (-\frac{{(I_p-I_q)}^2}{2{\mathrm{\&sigma;}}^2})---\left(2\right)$

B) calculate the energy function of network, as shown in Equation 3, the energy function E (L) of network by area item R (L) and Border item B (L) two parts composition, w₁And w₂Represent area item and the weight coefficient of border item respectively.Wherein area item R (L) represents The energy of this pixel region, as shown in Equation 4；Border item B (L) then represents this pixel and neighbor in network The energy of contact between point, as shown in Equation 5.

E (L)=w₁R(L)+w₂B(L) (3)

$\begin{array}{c}R\left(L\right)=\underset{i=1}{\overset{s}{\&Sigma;}}{R}_i\left(l_i\right)\\ R_i\left(1\right)=-\ln \mathrm{Pr}\left(l_i\right|obj)\\ R_i\left(0\right)=-\ln \mathrm{Pr}\left(l_i\right|bkg)\end{array}---\left(4\right)$

$\begin{array}{c}B\left(L\right)=\underset{\{p,q\}\&Element;N}{\&Sigma;}{B}_{<p,q>}\&CenterDot;\mathrm{\&delta;}(l_p,l_q)\\ \mathrm{\&delta;}(i_p,l_q)=\left\{\begin{array}{cc}0,& l_p=l_q\\ 1,& l_p\&NotEqual;l_q\end{array}\right.\end{array}---\left(5\right)$

4), cut algorithm according to the energy function use figure of figure and solve, using the solution of minima as the segmentation result of figure, Weighing criteria is used to be compared with goldstandard by segmentation result, specific as follows:

Given U₁, U₂Cut segmentation result and goldstandard for figure, then use the segmentation result such as formula 6 of both DSC coefficients comparisons Shown in.

$DSC(U_1,U_2)=2\&CenterDot;\left|\frac{U_1\&cap;U_2}{U_1\&cup;U_2}\right|---\left(6\right)$

DSC coefficient value is the least, represents that segmentation result is the best.

Upper described only the preferred embodiments of the present invention, are not limited to the present invention, for those skilled in the art Speech, the present invention can have various change and change.All made within spirit and principles of the present invention any amendment, equivalent replace Change, improvement etc., should be included within the scope of the present invention.

Claims (5)

1. a cerebral tumor image partition method based on multisequencing MR image related information, it is characterised in that include following step Rapid:

S1, reads the most homotactic MR image in same case；

S2, chooses the cerebral tumor region of one of them sequence MR image as tumor seed points；

S3, with the cerebral tumor seed points selected by current sequence MR image, carries out similarity mode to the MR image of other sequences, Find the potential cerebral tumor seed points in other sequence MR images；

S4, using the potential cerebral tumor seed points found and background seed points as priori, cuts algorithm with figure and realizes many sequences The Accurate Segmentation in row MR image midbrain tumors region.

A kind of cerebral tumor image partition method based on multisequencing MR image related information the most according to claim 1, its Be characterised by, in described step S1 the most homotactic MR image include T1 (longitudinal relaxation time), T2 (transverse relaxation time), T2WI+FLAIR (T2 weighted imaging), four kinds of imaging sequence (figures that the most different pulse train obtains of T1WI (T1 weighted imaging) Picture).

A kind of tumor image dividing method based on multisequencing MR image related information the most according to claim 1, it is special Levying and be, the seed points system of selection that described step S2 uses includes that click is rule, draws a circle, taken the man-machine interaction side such as a little Formula..

A kind of cerebral tumor image partition method based on multisequencing MR image related information the most according to claim 1, its Being characterised by, the Similarity Match Method that described step S3 uses specifically includes:

(1) cerebral tumor seed points region selected by described current sequence MR image is as template；

(2) the MR image to other sequences carries out pointwise Regional Similarity calculating from top to bottom, from left to right；

(3) the maximum region of similarity is searched out as the potential cerebral tumor region in this sequence MR image, remainder conduct Potential background parts.

A kind of cerebral tumor image partition method based on multisequencing MR image related information the most according to claim 1, its Being characterised by, described step S4 specifically includes:

(1) cerebral tumor region potential in MR image step S3 obtained is as target seed；

(2) MR image being mapped as network, pixel corresponds to the node of figure, and adds extra two nodes: source point and remittance Point, has line, each pixel to have line with source point and meeting point between each two neighbor pixel, these lines are as net The limit of network figure；

(3) Regional Similarity probability step S3 obtained is used for the weights on the limit of setting network figure, and calculates the energy letter of figure Number；

(4) cut algorithm according to the energy function use figure of figure to solve, using the solution of minima as the segmentation result of figure, use Segmentation result is compared by weighing criteria with goldstandard.