CN107154047A - Multi-mode brain tumor image blend dividing method and device - Google Patents

Abstract

The invention relates to medical equipment and medical imaging. In order to propose an improved multi-mode brain tumor image hybrid segmentation algorithm, FFCM is used to extract the brain tumor area, and a mixed level set algorithm is used to correct the boundary problem existing in the tumor area. Therefore, the FFCM algorithm and the level set algorithm can be more effectively applied to MRI brain tumor images. The technical solution adopted by the present invention is, the multi-mode brain tumor image hybrid segmentation method, first input the MRI images of three modes including T1C, T2 and FLAIR, and use median filtering to filter the image and perform initial segmentation to obtain the pre-processed image, and then Linear fusion is adopted; then FFCM clustering is performed on the fused image, and the regions with larger gray values are automatically extracted, and the obtained under-segmented tumor regions are subjected to mixed level set segmentation. The invention is mainly applied to the acquisition and processing of medical images.

Description

Method and device for hybrid segmentation of multi-mode brain tumor images

technical field

The invention relates to medical equipment, which is an important aspect in the field of medical imaging. It plays an important role in the fields of brain tumor cutting, brain tumor classification, and brain tumor identification. Specifically, it relates to an improved multi-mode brain tumor image hybrid segmentation method and device.

Background technique

In recent years, the incidence of brain tumors has been on the rise, accounting for about 5% of systemic tumors and 70% of childhood tumors. In 2015, approximately 23,000 new cases of brain tumors were diagnosed in the United States alone. The uncontrolled and unlimited growth of cells leads to the development of brain tumors. Brain tumors, if not diagnosed and treated early, can cause permanent brain damage and even death. Magnetic Resonance Imaging (MRI) can be used to detect abnormal changes in body tissue and is necessary to determine treatment options for brain tumors. In all treatments, any information about the location and size of the tumor is very important, but Due to the complex shape of brain tumors, randomness in size and location, and large differences in types, there is currently no segmentation algorithm that can meet clinical needs, and the real-time performance cannot meet the requirements. Different experts manually segment brain tumor images. There is a big difference, and the labor cost is high. Therefore, it is very important to study accurate automatic brain tumor segmentation algorithms.

Brain tumor automatic segmentation technology has always been a hot research direction. Brain tumor image segmentation methods are divided into manual segmentation, semi-automatic segmentation and fully automatic segmentation. Specific segmentation algorithms are divided into threshold algorithm, clustering algorithm and deformation model algorithm. The threshold algorithm was first used in image segmentation. For the problem of brain tumor images, the OTSU algorithm is an automatic adaptive threshold algorithm that can effectively avoid errors caused by fixed thresholds; a local fuzzy threshold (Fuzzy threshold) for multi-region image segmentation , FTH) algorithm also has a certain effect on complex images such as brain tumors. Due to the complexity of brain tumor images and the insufficient consideration of pixel spatial information by threshold algorithms, threshold algorithm segmentation cannot effectively solve the problem of brain tumor segmentation.

Fuzzy clustering is a kind of algorithm suitable for brain tumor image segmentation, especially the fuzzy C-mean (FCM) algorithm, which has the advantage of simple method implementation, but due to the complexity of medical image information and unclear edges, therefore, the seed Point selection has a great influence on the clustering results, and the FCM algorithm method is difficult to use the spatial information of the image, and the calculation itself is complicated. Therefore, the fast FCM (Fast FCM, FFCM) algorithm is proposed to improve the calculation speed; for the problem of insufficient spatial information, the spatial FCM (Spatial FCM, SFCM) algorithm is used to segment images, and the correlation between spatial information is effectively used, but the SFCM algorithm The calculation speed cannot meet the real-time requirements of medical images; the level set algorithm can effectively deal with various contour problems, and the combination of fuzzy clustering and level set methods (Fuzzy Clustering with Level Set Methods, FCLSM) algorithm can effectively solve the level set edge problem, but the FCLSM algorithm has the problem of real-time and easy to fall into local optimum.

The level set algorithm is a kind of deformation model algorithm. The segmentation algorithm based on the level set is also widely used in the segmentation of brain tumors. However, due to the uneven gray level of brain tumor tissues and the lack of obvious boundaries between brain tumor tissues, this method is used to The algorithm is prone to the problem of edge leakage. The distance regularized level set algorithm (Distance Regularized Level Set Evolution, DRLSE) is an effective algorithm. The distance regularization effect in this algorithm eliminates the need for reinitialization, thereby avoiding the local error caused by it; other methods have mixed levels set algorithm. The method uses object boundaries and region information to achieve robust and accurate segmentation. Boundary information can help detect the precise location of the target object, and region information can prevent boundary leakage, but the level set algorithm cannot solve the problem of being easily trapped in local optimum and strongly dependent on the initial value.

Contents of the invention

In order to overcome the deficiencies of the prior art, the present invention aims to propose an improved mixed segmentation algorithm for multi-mode brain tumor images, which uses FFCM to extract brain tumor regions, and uses a mixed level set algorithm to correct the boundary problems existing in the tumor regions. Therefore, the FFCM algorithm and the level set algorithm can be more effectively applied to MRI brain tumor images. The technical solution adopted by the present invention is, the multi-mode brain tumor image hybrid segmentation method, first input the MRI images of three modes including T1C, T2 and FLAIR, and use median filtering to filter the image and perform initial segmentation to obtain the pre-processed image, and then Linear fusion is adopted; then FFCM clustering is performed on the fused image, and the regions with larger gray values are automatically extracted, and the obtained under-segmented tumor regions are subjected to mixed level set segmentation.

The specific steps of FFCM clustering and segmentation are to divide the data into c categories through the fuzzy C-means theory. For an M×N image, set {h _i ,i=1,2,…,n}, n=M×N, h _i is a set of pixel intensity values in the image histogram, where M and N are the length and width of the image, {v _j ,j=1,2,…,c} is a set of cluster centers, and μ _j (h _i ) is the membership function that h _i belongs to class j, ||·|| represents the 2 norm, b is a constant greater than 1, then:

If iterative formula (3) (4) satisfies the iteration termination condition, t>T or Then stop, where t represents the number of iterations, ε is the stop condition, and T represents the maximum number of iterations. After the algorithm ends, the pixels are classified according to the maximum degree of membership. If μ _j (h _i )＞μ _j (h _k ), then h _i is classified as the jth type of area, k=1,2,...,c; i≠k.

The specific steps of hybrid level set segmentation are that the zero set of embedding function φ is used to represent the active contour C={X|φ(X)=0}, the points inside/outside the contour have positive/negative φ values, and the proposed minimum The simplified function definition:

In formula (5), I is the image to be segmented, is the boundary feature map associated with the image gradient, is the gradient operator, H(φ) is the Heaviside function, Ω is the image domain, α and β are predefined weights to balance the two items, μ is a predefined parameter indicating the lower limit of the gray level of the target object;

in, is the normal vector pointing to the outside of the curve, so the dominant curve evolution partial differential equation of the active contour is

In the formula and curvature <·,·> is the inner product; since only the geometric change of the curve is of interest in the segmentation, it can be noticed from equation (6) that all points on the curve move in the normal direction, equation (6) the first One term is a propagation term describing the expanding motion of the curved part inside the target object and the contracting motion of the outer part, and the second term is an advective term describing the movement of the curve in the vector field caused by the gradient of g to attract the curve to the target The boundary of the object, the third item describes the curvature flow weighted by the gradient feature map g, which is used to smooth the curve of the weak part of the boundary support;

In horizontal concentration, with Describe the same curve change, if φ is a signed distance function, that is The time-varying derivative of the level set embedding function is

g is a decreasing function, where c is the control slope.

The multi-mode brain tumor image hybrid segmentation device is equipped with a computer for processing T1C, T2 and FLAIR MRI images. The computer includes the following modules: using median filtering to perform filtering processing on the images and an initial segmentation module to obtain pre-processed images; The preprocessed image is input into the linear fusion module; the fused image is input into the FFCM clustering and segmentation module, and the area with a large gray value is automatically extracted to obtain the tumor under-segmented area; then input to the mixed level set segmentation module to obtain the final result.

Features and beneficial effects of the present invention are:

The present invention uses an improved multi-mode brain tumor image hybrid segmentation algorithm to segment MRI images with brain tumors. Compared with some classic methods, its advantages are mainly reflected in:

1) Novelty: For the first time, the FFCM algorithm and mixed level set algorithm are used to segment MRI images with brain tumors. According to the characteristics of MRI brain tumor images, combined with the advantages of FFCM algorithm and mixed level set, the rapid segmentation of brain tumor images is achieved. Purpose.

2) Effectiveness: Using FFCM can quickly and effectively obtain under-segmented regions, and inputting under-segmented regions into the mixed level set can speed up the convergence boundary, thereby effectively overcoming the defects of the algorithm and improving accuracy.

3) Practicability: the current segmentation algorithm is difficult to meet the requirements of practicability and real-time performance. The present invention combines the reasonable part between the mixed level set algorithm and the FFCM algorithm, thereby overcoming the defects of some algorithms and increasing the amount of algorithm to a certain extent. practicality. And further discussion is made for the automatic segmentation of brain tumors.

Description of drawings:

Fig. 1 is a flow chart of the improved multi-mode brain tumor image hybrid segmentation algorithm of the present invention to segment MRI brain tumors.

Fig. 2 is the similarity coefficient (Dice) of the algorithm of the present invention in 10 brain tumor images.

detailed description

1 Fast FCM theory based on histogram

The core idea of FFCM is to find the appropriate membership degree and cluster center for the pixel intensity value, so that the variance and iteration error of the cost function within the cluster are minimized. The value of the cost function is the weighted cumulative sum of the 2-norm measure from the pixel to the cluster center. The FFCM clustering and segmentation algorithm is to divide the data into c categories through the fuzzy C-means theory. For an M×N image, suppose {h _i ,i=1,2,…,n}, n=M×N, h _i is a collection of pixel intensity values in the image histogram. {v _j ,j=1,2,…,c} is a set of cluster centers, and μ _j (h _i ) is the membership function of h _i belonging to class j, so the objective function of FFCM is

and

In the formula, ||·|| represents the 2-norm, and b is a constant greater than 1, which controls the ambiguity of the clustering results. To calculate the minimum value of J _f such that

It can be deduced from formula (1) (2)

If iterative formula (3) (4) satisfies the iteration termination condition, t>T or Then stop, where t represents the number of iterations, ε is the stopping condition, and T represents the maximum number of iterations. After the algorithm ends, classify the pixels according to the maximum degree of membership, if μ _j (h _i )>μ _j (h _k ), then h _i will be classified as the jth type of area, k=1,2,...,c ; i≠k.

2 Mixed model level set principle

Osher and Sethian proposed a level set method to represent low-dimensional curves as zero-level sets of high-dimensional surfaces. At any time, as long as φ is known, its zero level set curve can be obtained, where φ(X) is the level set function. To deal with the variation of the surface, it is only necessary to represent the surface as a zero-level set in a higher one-dimensional space. Level set models have significant advantages over particle and parametric models, and the concept and numerical implementation are adaptable to solve problems of any size; regions inside and outside active contours can be easily determined.

Since the MRI images of brain tumors are extremely complex, the present invention uses a segmentation method based on a level set to integrate boundary problems and region information, and at the same time make up for the boundary problems left by the FFCM algorithm. Before describing the model, several parameters need to be explained. The zero set of the embedding function φ is used to represent the active contour C={X|φ(X)=0}, and points inside/outside the contour have positive/negative φ values. The proposed function definition that needs to be minimized

In formula (5), I is the image to be segmented, is the boundary feature map associated with the image gradient, where g is a decreasing function, where a is the control slope, H(φ) is the Heaviside function, Ω is the image domain, and α and β are predefined weights to balance the two terms. μ is a predefined parameter indicating the lower limit of the gray level of the target object. is the gradient operator.

in, is the normal vector pointing outside the curve. Thus the partial differential equation for the dominant curve evolution of the activity profile is

In the formula and curvature <·,·> is the inner product. Since only the geometric changes of the curve are of interest in segmentation, it can be noticed from Equation (6) that all points on the curve are shifted in the normal direction. The first term of Equation (6) is the propagation term describing the extension motion of the curve part inside the target object and the contraction motion of the outer part, and the second term is the advection term, which describes the curve movement in the vector field caused by the gradient of g, with Attracts the curve to the boundary of the target object, and the third term describes the curvature flow weighted by the gradient feature map g, which acts to smooth the curve in the part where the boundary support is weak.

In horizontal concentration, with Describe the same curve change, if φ is a signed distance function, that is The time-varying derivative of the level set embedding function is

Table 1

Table 1 shows the results of processing 47 brain tumor images by the algorithm of the present invention, among which Jaccard coefficient, similarity coefficient (Dice) and recall are the most commonly used evaluation criteria.

Since the MRI brain tumor image itself is not of high quality and cannot be used for direct segmentation, the present invention first uses the hybrid segmentation algorithm framework shown in Figure 1. Since the three modal images can provide partially irrelevant and complementary information for tumor segmentation, this paper The invention first inputs three modes of MRI images including T1C, T2 and FLAIR. Due to the presence of certain noise in the image itself, the median filter is used to filter the image and the initial segmentation to obtain the preprocessed image, and then linear fusion is used; the fused image is segmented by the FFCM clustering algorithm, and the area with a large gray value is automatically extracted. The obtained tumor under-segmented regions are segmented by mixed level set, and finally the segmentation results are evaluated. The combination of the FFCM clustering algorithm and the mixed model level set algorithm not only speeds up the speed of the level set algorithm itself, but also improves the lack of dependence of the mixed model level set algorithm on the initial value. In order to test a suitable fusion ratio, the present invention compares various ratios, and finally finds that the more suitable ratio is FLAIR:T2:T1C=5:4:1.

Claims (4)

1. A multi-mode brain tumor image hybrid segmentation method is characterized in that, at first three kinds of modes are input and comprise the MRI image of T1C, T2 and FLAIR, adopt median filtering to carry out filter processing and initial segmentation to image and obtain preprocessing image, afterwards Linear fusion is adopted; then FFCM clustering is performed on the fused image, and the regions with larger gray values are automatically extracted, and the obtained under-segmented tumor regions are subjected to mixed level set segmentation. 2. The multi-mode brain tumor image hybrid segmentation method as claimed in claim 1, wherein the specific step of FFCM clustering segmentation is to divide the data into c classes by fuzzy C-means theory, and for an M×N image, Let {h _i ,i=1,2,...,n}, n=M×N, h _i is a set of pixel intensity values in the image histogram, where M and N are the length and width of the image, {v _j ,j=1,2,…,c} is a set of clustering centers, and μ _j (h _i ) is the membership function that _hi belongs to class j, ||·|| represents the 2-norm, and b is A constant greater than 1, then: <mrow> <msub> <mi>v</mi> <mi>j</mi> </msub> <mo>=</mo> <mfrac> <mrow> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msup> <mrow> <mo>&amp;lsqb;</mo> <msub> <mi>&amp;mu;</mi> <mi>j</mi> </msub> <mrow> <mo>(</mo> <msub> <mi>h</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> </mrow> <mi>b</mi> </msup> <msub> <mi>h</mi> <mi>i</mi> </msub> </mrow> <mrow> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msup> <mrow> <mo>&amp;lsqb;</mo> <msub> <mi>&amp;mu;</mi> <mi>j</mi> </msub> <mrow> <mo>(</mo> <msub> <mi>h</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> </mrow> <mi>b</mi> </msup> </mrow> </mfrac> <mo>,</mo> <mi>j</mi> <mo>=</mo> <mn>1</mn> <mo>,</mo> <mn>2</mn> <mo>,</mo> <mo>...</mo> <mo>,</mo> <mi>c</mi> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>3</mn> <mo>)</mo> </mrow> </mrow> <mrow> <msub> <mi>&amp;mu;</mi> <mi>j</mi> </msub> <mrow> <mo>(</mo> <msub> <mi>h</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <msup> <mrow> <mo>(</mo> <mn>1</mn> <mo>/</mo> <mo>|</mo> <mo>|</mo> <msub> <mi>h</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>v</mi> <mi>j</mi> </msub> <mo>|</mo> <msup> <mo>|</mo> <mn>2</mn> </msup> <mo>)</mo> </mrow> <mrow> <mn>1</mn> <mo>/</mo> <mrow> <mo>(</mo> <mi>b</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> </mrow> </msup> <mrow> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>k</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>c</mi> </munderover> <msup> <mrow> <mo>(</mo> <mn>1</mn> <mo>/</mo> <mo>|</mo> <mo>|</mo> <msub> <mi>h</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>v</mi> <mi>k</mi> </msub> <mo>|</mo> <msup> <mo>|</mo> <mn>2</mn> </msup> <mo>)</mo> </mrow> <mrow> <mn>1</mn> <mo>/</mo> <mrow> <mo>(</mo> <mi>b</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> </mrow> </msup> </mrow> </mfrac> <mo>,</mo> <mi>i</mi> <mo>=</mo> <mn>1</mn> <mo>,</mo> <mn>2</mn> <mo>,</mo> <mo>...</mo> <mo>,</mo> <mi>n</mi> <mo>;</mo> <mi>j</mi> <mo>=</mo> <mn>1</mn> <mo>,</mo> <mn>2</mn> <mo>,</mo> <mo>...</mo> <mo>,</mo> <mi>c</mi> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>4</mn> <mo>)</mo> </mrow> </mrow> If iterative formula (3) (4) satisfies the iteration termination condition, t>T or Then stop, where t represents the number of iterations, ε is the stop condition, and T represents the maximum number of iterations. After the algorithm ends, the pixels are classified according to the maximum degree of membership. If μ _j (h _i )>μ _j (h _k ), then h _i is classified as the jth type of area, k=1,2,...,c; i≠k. The specific steps of hybrid level set segmentation are that the zero set of embedding function φ is used to represent the active contour C={X|φ(X)=0}, the points inside/outside the contour have positive/negative φ values, and the proposed minimum The simplified function definition: <mrow> <mi>M</mi> <mi>i</mi> <mi>n</mi> <mi>&amp;epsiv;</mi> <mrow> <mo>(</mo> <mi>&amp;phi;</mi> <mo>)</mo> </mrow> <mo>=</mo> <mo>-</mo> <mi>&amp;alpha;</mi> <munder> <mo>&amp;Integral;</mo> <mi>&amp;Omega;</mi> </munder> <mrow> <mo>(</mo> <mi>I</mi> <mo>-</mo> <mi>&amp;mu;</mi> <mo>)</mo> </mrow> <mi>H</mi> <mrow> <mo>(</mo> <mi>&amp;phi;</mi> <mo>)</mo> </mrow> <mi>d</mi> <mi>&amp;Omega;</mi> <mo>+</mo> <mi>&amp;beta;</mi> <munder> <mo>&amp;Integral;</mo> <mi>&amp;Omega;</mi> </munder> <mi>g</mi> <mo>|</mo> <mo>&amp;dtri;</mo> <mi>H</mi> <mrow> <mo>(</mo> <mi>&amp;phi;</mi> <mo>)</mo> </mrow> <mo>|</mo> <mi>d</mi> <mi>&amp;Omega;</mi> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>5</mn> <mo>)</mo> </mrow> </mrow> In formula (5), I is the image to be segmented, is the boundary feature map associated with the image gradient, is the gradient operator, H(φ) is the Heaviside function, Ω is the image domain, α and β are predefined weights to balance the two items, μ is a predefined parameter indicating the lower limit of the gray level of the target object; in, is the normal vector pointing to the outside of the curve, so the dominant curve evolution partial differential equation of the active contour is <mrow> <msub> <mi>C</mi> <mi>t</mi> </msub> <mo>=</mo> <mi>&amp;alpha;</mi> <mrow> <mo>(</mo> <mi>I</mi> <mo>-</mo> <mi>&amp;mu;</mi> <mo>)</mo> </mrow> <mover> <mi>N</mi> <mo>&amp;RightArrow;</mo> </mover> <mo>-</mo> <mi>&amp;beta;</mi> <mo>&lt;</mo> <mo>&amp;dtri;</mo> <mi>g</mi> <mo>&amp;CenterDot;</mo> <mover> <mi>N</mi> <mo>&amp;RightArrow;</mo> </mover> <mo>&gt;</mo> <mo>+</mo> <mi>&amp;beta;</mi> <mi>g</mi> <mi>&amp;kappa;</mi> <mover> <mi>N</mi> <mo>&amp;RightArrow;</mo> </mover> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>6</mn> <mo>)</mo> </mrow> </mrow> In the formula and curvature <·,·> is the inner product; since only the geometric change of the curve is of interest in the segmentation, it can be noticed from equation (6) that all points on the curve move in the normal direction, equation (6) the first One term is a propagation term describing the expanding motion of the curved part inside the target object and the contracting motion of the outer part, and the second term is an advective term describing the movement of the curve in the vector field caused by the gradient of g to attract the curve to the target The boundary of the object, the third item describes the curvature flow weighted by the gradient feature map g, which is used to smooth the curve of the weak part of the boundary support; In horizontal concentration, with Describe the same curve change, if φ is a signed distance function, that is The time-varying derivative of the level set embedding function is <mrow> <msub> <mi>&amp;phi;</mi> <mi>t</mi> </msub> <mo>=</mo> <mi>&amp;alpha;</mi> <mrow> <mo>(</mo> <mi>I</mi> <mo>-</mo> <mi>&amp;mu;</mi> <mo>)</mo> </mrow> <mo>+</mo> <mi>&amp;beta;</mi> <mi>d</mi> <mi>i</mi> <mi>v</mi> <mrow> <mo>(</mo> <mi>g</mi> <mo>&amp;dtri;</mo> <mi>&amp;phi;</mi> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mo>(</mo> <mn>7</mn> <mo>)</mo> <mo>.</mo> </mrow> 1 3. The multi-mode brain tumor image hybrid segmentation method as claimed in claim 1, wherein g is a decreasing function, where c is the control slope. 4. A multi-mode brain tumor image hybrid segmentation device is characterized in that it is provided with a computer for processing the MRI images of T1C, T2 and FLAIR, and the computer includes the following modules: Median filtering is used to carry out filter processing and initial segmentation of the image module to obtain the preprocessed image; then the preprocessed image is input into the linear fusion module; the fused image is input into the FFCM clustering and segmentation module, and the area with a larger gray value is automatically extracted to obtain the under-segmented tumor area; and then input into the mixed level set Split module processing to get the final result.