CN106919950B - Brain MR Image Segmentation Method Based on Probability Density Weighted Geodesic Distance - Google Patents

Abstract

The invention discloses a brain MR image segmentation method based on probability density weighted geodesic distance, which belongs to the technical field of image processing. Including: read in several simulated brain database images, perform histogram statistics, and obtain the sample value of the most concentrated distribution interval; use the obtained sample value as prior knowledge to calculate the probability density of each pixel on the selected image to be processed Estimation; based on the probability density function, perform superpixel segmentation on the image to be processed; scan the segmented superpixels, screen out the superpixels that do not meet the standards for splitting, and use the FCM algorithm to cluster all the pixels in the superpixels into two categories again, according to The classification results look for connected regions, and take the pixels in each connected region as a new class, update the superpixel classification result matrix; use the FCM algorithm to cluster on the basis of all updated superpixels, and obtain the brain of the image to be processed Tissue segmentation results. The invention improves the accuracy of superpixel segmentation and brain tissue segmentation.

Description

Brain MR Image Segmentation Method Based on Probability Density Weighted Geodesic Distance

technical field

The invention relates to the technical field of image processing, in particular to a brain MR image segmentation method based on probability density weighted geodesic distance.

Background technique

Image segmentation is one of the most classic research topics in the fields of image processing, image analysis, and computer vision, and it is also one of the biggest difficulties. Image segmentation technology plays a key role in many medical image applications, and it is also an important part of various tissues and organs in images. The basis for further analysis of the pathology of the human body, through the use of image segmentation, the region of interest in the image is extracted to provide a basis for clinical diagnosis and treatment, and the brain is an important organ of the human body, so the study of brain region segmentation technology is very important for the brain It is of great significance for the three-dimensional reconstruction of the brain, the study of neural circuits, and the diagnosis of clinical brain diseases.

Superpixel segmentation is an image over-segmentation algorithm that can be used as preprocessing in some image applications, such as segmentation, saliency detection, face recognition, etc. Superpixels can capture redundancies in images, greatly reducing the complexity of subsequent image processing tasks. An existing effective superpixel segmentation method is a superpixel based on geodesic distance, which uses geodesic distance instead of Euclidean distance to measure the similarity between pixels, and the segmentation effect is good for natural images. However, for brain MR (Magnetic Resonance, MR for short) images, this method cannot accurately separate every small superpixel block of brain tissue area.

Fuzzy C-Means clustering algorithm (FCM for short) is the most widely used fuzzy clustering image segmentation algorithm. Compared with other segmentation methods, FCM can retain more information of the original image. However, the traditional FCM algorithm fails to consider the grayscale features of each point and the correlation degree of its neighboring pixels in image segmentation, which leads to the algorithm being sensitive to noise and grayscale inhomogeneity. Aiming at the above problems, the proposed Many improved FCM algorithms, although the improved method has a certain degree of improvement in anti-noise or efficiency, but due to the high complexity of brain images, satisfactory segmentation results cannot be achieved, so the traditional single method segmentation cannot Meet actual requirements.

Contents of the invention

The invention provides a brain MR image segmentation method based on probability density weighted geodesic distance, which improves the accuracy of superpixel segmentation and brain tissue segmentation.

In order to solve the problems of the technologies described above, the present invention provides technical solutions as follows:

A method for brain MR image segmentation based on probability density weighted geodesic distance, comprising:

Step 1: Read in several simulated brain database images, perform histogram statistics on them, and obtain the sample values of the most concentrated distribution intervals of white matter, gray matter or cerebrospinal fluid;

Step 2: Randomly select an image from the simulated brain database image as the image to be processed, and use the obtained sample value of the most concentrated distribution interval as prior knowledge to estimate the probability density of each pixel on the image to be processed , get the probability density function;

Step 3: Perform superpixel segmentation on the image to be processed based on the obtained probability density function, and record the superpixel classification result matrix;

Step 4: Scan the segmented superpixels, screen out superpixels that do not meet the standard according to the superpixel color standard deviation for splitting, use the FCM algorithm to group all pixels in the superpixels into two categories again, and then classify them according to As a result, the connected regions are searched, and the pixels in each connected region are used as a new class, and the superpixel classification result matrix is updated;

Step 5: According to the updated superpixel classification result matrix, the FCM algorithm is used to perform clustering on the basis of all updated superpixels to obtain the brain tissue segmentation result of the image to be processed.

The present invention has the following beneficial effects:

In the brain MR image segmentation method of probability density weighted geodesic distance of the present invention, first, the sample value obtained by histogram statistics is used as a priori value to estimate the probability density of each pixel on the image, and then weighted based on the probability density The superpixel method of geodesic distance performs superpixel segmentation on the image, and then scans the superpixels to filter out the superpixels that do not meet the standards for splitting, and uses the internal characteristics of the superpixels to perform local splitting optimization. After the splitting is completed, the superpixel classification result matrix is updated, and finally The FCM algorithm is used to perform clustering based on the segmented superpixels to complete the segmentation of the brain image. This method has at least the following advantages: (1) A new weight factor is designed to define the geodesic distance, which incorporates the probability density function, making the contrast between different brain tissues more obvious and the gradient calculation more reasonable; (2) adding local segmentation In the post-processing process, the superpixels are locally split, and the pixels are classified more accurately, which further improves the accuracy of the superpixel segmentation. Even if the simplest and most traditional FCM is used for the final clustering, the result is still very good; (3) Combine superpixel segmentation technology based on probability density weighted geodesic distance, FCM and other methods, and obtain brain tissue segmentation results on the basis of superpixels.

Description of drawings

Fig. 1 is the schematic flow chart of the brain MR image segmentation method of the probability density weighted geodesic distance of the present invention;

Fig. 2 is the schematic diagram of the principle of the brain MR image segmentation method of the probability density weighted geodesic distance of the present invention;

Fig. 3 is the schematic flow chart of step 3 in the brain MR image segmentation method of probability density weighted geodesic distance of the present invention;

Fig. 4 is the specific flow diagram of the brain MR image segmentation method of the probability density weighted geodesic distance of the present invention;

Fig. 5 (a): is the image to be processed in the present invention: synthetic brain MR example image;

Fig. 5 (b): is the superpixel segmentation map of the present invention corresponding to Fig. 5 (a) synthetic brain MR example image;

Fig. 5(c): is the segmentation result diagram of the final white matter, gray matter, cerebrospinal fluid and other regions of the brain MR example image realized according to the present invention;

Fig. 6 (a): is the image to be processed of the present invention: real brain MR example image;

Fig. 6 (b): is the superpixel segmentation diagram of the present invention corresponding to the real brain MR example image in Fig. 6 (a);

Fig. 6(c): It is a diagram of the segmentation results of the final white matter, gray matter, cerebrospinal fluid and other regions of the brain MR example image realized according to the present invention.

Detailed ways

In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will describe in detail with reference to the drawings and specific embodiments.

The present invention provides a brain MR image segmentation method with probability density weighted geodesic distance, as shown in Figure 1-6, comprising:

Step 1: Read in several simulated brain database images, perform histogram statistics on them, and obtain the sample values of the most concentrated distribution intervals of white matter, gray matter or cerebrospinal fluid;

In this step, the read-in simulated brain database image is a synthetic brain MR image in bmp format, and its gray value is read in.

Step 2: Randomly select an image from the simulated brain database image as the image to be processed, use the obtained sample value of the most concentrated distribution interval as prior knowledge to estimate the probability density of each pixel on the image to be processed, and obtain the probability density function;

Step 3: Perform superpixel segmentation on the image to be processed based on the obtained probability density function, and record the superpixel classification result matrix;

Step 4: Scan the segmented superpixels, screen out superpixels that do not meet the standard according to the superpixel color standard deviation for splitting, use the FCM algorithm to group all pixels in the superpixels into two categories again, and then classify them according to Find connected regions as a result, and use the pixels in each connected region as a new class to update the superpixel classification result matrix;

Step 5: According to the updated superpixel classification result matrix, the FCM algorithm is used to perform clustering on the basis of all updated superpixels to obtain the brain tissue segmentation result of the image to be processed.

In the brain MR image segmentation method of probability density weighted geodesic distance of the present invention, first, the sample value obtained by histogram statistics is used as a priori value to estimate the probability density of each pixel on the image, and then weighted based on the probability density The superpixel method of geodesic distance performs superpixel segmentation on the image, and then scans the superpixels to filter out the superpixels that do not meet the standards for splitting, and uses the internal characteristics of the superpixels to perform local splitting optimization. After the splitting is completed, the superpixel classification result matrix is updated, and finally The FCM algorithm is used to perform clustering based on the segmented superpixels to complete the segmentation of the brain image. This method has at least the following advantages: (1) A new weight factor is designed to define the geodesic distance, which incorporates the probability density function, making the contrast between different brain tissues more obvious and the gradient calculation more reasonable; (2) adding local segmentation In the post-processing process, the superpixels are locally split, and the pixels are classified more accurately, which further improves the accuracy of the superpixel segmentation. Even if the simplest and most traditional FCM is used for the final clustering, the result is still very good; (3) Combine superpixel segmentation technology based on probability density weighted geodesic distance, FCM and other methods, and obtain brain tissue segmentation results on the basis of superpixels.

Preferably, step 1 is further:

Read in several simulated brain MR images (obtained from the BrainWeb database), first use the K-means algorithm for initial classification, then divide the gray value into N intervals after normalization, and compare the white matter, gray matter or cerebrospinal fluid The gray value range of the gray value is counted, and the K intervals with the most concentrated distribution of white matter, gray matter or cerebrospinal fluid are counted. Among these K intervals, m gray values are selected as samples for each interval, and the value of the sample is the most concentrated. Sample values for the distribution interval. In this step, N, K, and m are all integers greater than 0.

In the present invention, since the gray matter, white matter, and cerebrospinal fluid pixel gray values of different tissues of the brain image overlap, there is no exact boundary between them, and the category to which each pixel belongs is fuzzy. It is difficult to divide it. Therefore, the present invention first performs histogram statistics, and uses the obtained sample values as prior knowledge to estimate the probability density of each pixel on the image in the next step.

Further, step 2 includes:

Step 21: Calculate K probability estimation models for each pixel in the image to be processed according to formula (1):

Among them, x is the gray value of each pixel in the image, x _i is the m gray value samples of the kth (k=1,2,...,K) interval as the prior value, h is the bandwidth (control parameter);

Step 22: Calculate the mixed probability density function according to the obtained K probability estimation models according to the formula (2), and obtain the estimated value of the probability density of each pixel, that is, the probability that each pixel belongs to white matter, gray matter or cerebrospinal fluid:

Among them, p(k) is the influence factor of each probability estimation model on the data points, which is between 0 and 1;

Step 23: Normalize the obtained mixed probability density function to obtain the probability density function:

Among them, wmax and wmin are the maximum and minimum values of P(x).

In the present invention, the probability density function is incorporated, which can make the contrast between different brain tissues such as gray matter, white matter, and cerebrospinal fluid more obvious, and make our subsequent gradient calculation more accurate.

As an improvement of the present invention, as shown in Figure 3, step 3 includes:

Step 31: Initialize the seed points, first evenly distribute n/2 seed points on the image, then use the Adaptive Hexagonal (Adaptive Hexagonal) method to insert other seed points, use formula (4) to calculate the hexagonal shape of each seed point The complexity, find the hexagon with the largest complexity, divide it into six overlapping small hexagons, insert a new seed point in the small hexagon with the largest complexity, and iterate sequentially until n clusters are sampled Class center, the complexity is defined as:

Among them, H _i is the hexagonal area of the seed point s _i , N is the number of pixels in the image I, M is the number of pixels p in curly brackets that meet the conditions, α is the control parameter, is the gradient of pixel p on the image, G _σ is a Gaussian function with standard deviation σ, ω is an adjustment parameter to prevent is zero;

Step 32: disturb the seed points, move each seed point to the lowest gradient position of its local 3*3 area, and record its X, Y coordinate information as a new seed point;

Step 33: Use formula (5) to calculate the seed point sensitivity gradient based on the probability density:

PSSG(s _i ,G(t))＝||S _p (s _i ,G(t))|| (5)

in, S _p is the gradient of P _w (x) calculated with the sobel operator on the geodesic path;

Step 34: Use the FMM (fast marching method) algorithm to calculate the weighted geodesic distance based on the probability density, and perform boundary diffusion to generate a series of pixels with similar attributes to form a super pixel block;

In this step, A: The geodesic distance is calculated by the following method:

A description of the probability density-weighted geodesic distance from a seed point s _i to any pixel point p is as follows: starting from the seed point s _i along a shortest path to the pixel point p, each point on the path is multiplied by a weight function W( s _i , G(t)) minimum arc length integral, which is defined as:

Among them, G(t) is a geodesic path from the seed point _si to the pixel point, t is constantly changing, and takes a value between 0 and 1, and the weight W is set to a pixel belonging to white matter, gray matter or cerebrospinal fluid The gradient of the possibility is used to define the distance increment on the path G from the seed point _si to a certain pixel point:

The geodesic distance is computed by an extended strategy of the Fast Marching Method (FMM) with iterative propagation of the appropriate velocity field, defining a velocity function based on Equation (8), which is calculated as:

B: Boundary diffusion, generating superpixel blocks:

During the diffusion process, the new velocity of each pixel is no longer static, and depends on the nearest seed point (that is, the pixel with the shortest probability density-weighted geodesic distance to it), starting from the given seed point , use the formula (5) to calculate the gradient of its neighborhood pixels, and expand along the pixel with the maximum speed formula (8), the color value of the expanded pixel will be replaced by the color value of the seed point, the probability of the next pixel The sensitive gradient of the seed point will be continuously calculated and updated by the value of the current neighborhood pixel through the diffusion process of FMM with the formula (5), and each time the pixel point with the smallest geodetic distance (maximum velocity function) from the seed point is diffused forward until All pixels are diffused, and the classification result matrix of superpixels is obtained.

Step 35: update the position and color of the seed point with the public notice (9);

Among them, S _l is the lth superpixel, s _l is the seed point of superpixel S _l , x _l ', c _l ' are the position and color value of the updated seed point respectively, A weight function that measures the degree to which a pixel belongs to a seed point;

Step 36: Repeat step 33, step 34, and step 35. The algorithm aims to optimize an energy function, which is defined as formula (10). When the change of the energy function in two consecutive iterations is less than a specific threshold, it stops. The initial superpixel split completed,

In the present invention, aiming at the high complexity of the brain image, a new weight influencing factor is designed to define the geodesic distance. Using the new gradient calculation method, there will be a more obvious boundary at the edge of the blurred area of the brain MR image, which can be compared Accurately segment superpixel blocks of every small brain tissue area.

Further, step 4 includes:

Step 41: Scan the segmented superpixels. If the color standard deviation of the superpixels is greater than a certain threshold T _c , split the superpixels satisfying this condition. The calculation is defined as:

C(s _l )=λStd _l >T _c (11)

where _Stdl is the standard deviation of pixel color within each superpixel, _Tc is the threshold, and λ is the control parameter;

Step 42: Use the FCM algorithm to cluster all the pixels in each superpixel that needs to be split into two categories again, and the objective function of the local FCM is:

Among them, C is the number of clusters, Q _n is the number of pixels in the currently scanned superpixel, μ _ij is the fuzzy membership function of the j-th pixel belonging to the i-th cluster, satisfying the constraint μ _ij ∈ [ 0,1], m is the weight function acting on the fuzzy membership, v _i is the ith cluster center, x _j is the color value of the pixel in the currently scanned superpixel;

Step 43: Find connected regions according to the obtained superpixel classification results, and take the pixels in each connected region as a new class, that is, a new superpixel block, and update the result matrix of superpixel classification;

Step 44: Repeat steps 42 and 43 for the superpixels that need to be split until all superpixels that meet the splitting conditions are split.

In the present invention, after the preliminary segmentation is completed, the local segmentation post-processing is performed, and the internal characteristics of the superpixel are used to perform local segmentation optimization, so that the pixels can be classified more accurately, the accuracy of superpixel segmentation is improved, and the division of superpixels can be more accurate. Accurately separate different brain tissues through further clustering.

Preferably, step 5 is further:

Using superpixels (rather than pixels) as the input of the FCM algorithm, clustering is performed on the basis of the segmented superpixels to complete the segmentation of the brain image. The FCM objective function based on superpixels is:

Among them, C' is the number of clusters, Q _n ' is the number of superpixels in the image, μ′ _ij is the fuzzy membership function of the jth superpixel belonging to the ith cluster, satisfying the constraint m is the weight function acting on the fuzzy membership degree, v′ _i is the ith cluster center, and ξ _i is the color mean of the superpixel S _j .

In the present invention, the superpixel segmentation technology based on the probability density weighted geodesic distance is connected with the FCM algorithm, and the brain tissue segmentation result is obtained on the basis of the superpixel, which improves the segmentation efficiency and accuracy.

As an improvement of the present invention, the read-in image can also be a real brain MR image in dcm format, and the processing method of the real brain MR image is similar to that of the synthetic brain MR image, the difference is that the synthetic brain MR image The image is read in grayscale values, while the real brain MR image is read in density values limited by the window width and level.

The image processing results are analyzed below, such as the pictures given in Figure 5(a) and Figure 6(a), using the above method for processing, the processing results are shown in Figure 5(c), Figure 6(c) , Figure 5(b), and Figure 6(b) are superpixel segmentation diagrams.

In Figure 5(b) and Figure 6(b), it can be seen from the visual effect that the superpixel segmentation of the present invention is accurate, and the boundary fits the edge, and each small area can be accurately separated, which is for the subsequent separation The clustering work of brain tissues such as white matter (WM), gray matter (GM), and spinal fluid (CSF) provides a solid foundation. Even with the simplest and most traditional FCM clustering our results are still very good.

In Fig. 5(c) and Fig. 6(c), it can be seen that the present invention can accurately segment the brain tissue image and preserve the original information of the image to the greatest extent.

The above description is a preferred embodiment of the present invention, it should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made, and these improvements and modifications can also be made. It should be regarded as the protection scope of the present invention.

Claims (3)

1. A method for probability density weighted geodesic distance brain MR image segmentation, comprising:

step 1: reading a plurality of simulated brain database images, and performing histogram statistics on the images to obtain sample values of white matter, gray matter or cerebrospinal fluid most centralized distribution intervals;

step 2: randomly selecting one image from the simulated brain database images as an image to be processed, and performing probability density estimation on each pixel point on the image to be processed by using the obtained sample value of the most concentrated distribution interval as prior knowledge to obtain a probability density function;

and step 3: performing superpixel segmentation on the image to be processed based on the obtained probability density function, and recording a superpixel classification result matrix;

and 4, step 4: scanning the segmented super pixels, screening out the super pixels which do not meet the standard according to the standard deviation of the super pixel colors, splitting, clustering all pixels in the super pixels into two classes again by using an FCM algorithm during splitting, then searching for connected regions according to classification results, taking the pixels in each connected region as a new class, and updating a super pixel classification result matrix;

and 5: clustering on the basis of all updated superpixels by using an FCM algorithm according to the updated superpixel classification result matrix to obtain a brain tissue segmentation result of the image to be processed;

the step 1 is further as follows:

reading a plurality of simulated brain MR images, firstly carrying out initial classification on the images by using a K-means algorithm, then dividing the normalized gray values into N intervals, carrying out statistics on the gray value range of white matter, gray matter or cerebrospinal fluid of the N intervals, carrying out statistics on K intervals with the white matter, gray matter or cerebrospinal fluid distributed most intensively, and selecting m gray values as samples in each of the K intervals;

the step 2 comprises the following steps:

step 21: respectively calculating K probability estimation models of each pixel point in the image to be processed according to a formula (1):

wherein x is the gray value of each pixel point in the image, and x_iM gray value samples with the kth interval as a priori value, wherein K is 1, 2.. and K, h is a control parameter;

step 22: calculating a mixed probability density function according to a formula (2) by using the obtained K probability estimation models to obtain a probability density estimation value of each pixel point, namely the possibility that each pixel point belongs to white matter, grey matter or cerebrospinal fluid:

wherein, p (k) is the influence factor of each probability estimation model on the data points, and the influence factor is between 0 and 1;

step 23: normalizing the obtained mixed probability density function to obtain a probability density function:

wherein wmax and wmin are the maximum and minimum values of P (x);

the step 3 comprises the following steps:

step 31: initializing seed points, uniformly distributing n/2 seed points on an image, then inserting other seed points by using a self-adaptive hexagon method, calculating the complexity of a hexagon of each seed point by using a formula (4), finding the hexagon with the maximum complexity, dividing the hexagon into six overlapped small hexagons, inserting a new seed point in the small hexagon with the maximum complexity, and sequentially iterating until n clustering centers are sampled, wherein the complexity is defined as:

wherein H_iIs the seed point s_iN is the number of pixels of image I, M is the number of pixels p in parenthesis that satisfy the condition, alpha is a control parameter, is the gradient of a pixel point p on the image, G_σIs a Gaussian function with a standard deviation sigma, omega is a tuning parameter, preventingA case of zero;

step 32: disturbing the seed points, moving each seed point to the lowest gradient position of a local 3 x 3 area of the seed point, and recording X, Y coordinate information of the seed point as a new seed point;

step 33: the probability density based seed point sensitivity gradient is calculated using equation (5):

PSSG(s_i,G(t))＝||S_p(s_i,G(t))|| (5)

wherein,S_pis P calculated by sobel operator on geodetic path_w(x) A gradient of (a);

step 34: calculating weighted geodesic distance based on probability density by adopting FMM algorithm, and performing boundary diffusion to generate a series of pixel points with similar attributes to form a superpixel block;

in this step, a: the geodesic distance is calculated by the following method:

from the seed point s_iOne description of the p probability density weighted geodesic distance to any pixel point is: from the seed point s_iStarting to reach the pixel point p along a shortest path, and multiplying each point on the path by a weight function W(s)_iG (t)), defined as:

wherein G (t) is from the seed point s_iA geodesic path from the pixel point is obtained, t is constantly changed, a value between 0 and 1 is taken, the weight W is set as the gradient of the probability that a pixel belongs to white matter, gray matter or cerebrospinal fluid, and the gradient is used for defining the seed point s_iDistance increment to a certain pixel point path G:

geodesic distances are calculated by the extended strategy of the fast marching method FMM with iterative propagation of the appropriate velocity field, defining a velocity function based on equation (7) which is calculated as:

b: boundary diffusion, generating superpixel blocks:

in the diffusion process, the new speed of each pixel point is not static any more, the new speed depends on the closest seed point, namely the pixel point with the shortest probability density weighted geodesic distance to the closest seed point, the gradient of the neighborhood pixel point is calculated by a formula (5) from the given seed point, the pixel point with the maximum speed formula (8) is expanded, the color value of the expanded pixel point is replaced by the color value of the seed point, the probability seed point sensitivity gradient of the next pixel point is updated by the diffusion process of FMM by calculating the value of the current neighborhood pixel point by the formula (5), the pixel point with the minimum geodesic distance to the seed point, namely the maximum speed function, is diffused forwards each time until all pixels are diffused completely, and the superpixel classification result matrix is obtained;

step 35: updating the position and color of the seed point with the public display (9);

wherein S is_lIs the ith super pixel, s_lIs a super pixel S_lSeed point of (2), x_l'，c_l' the location and color value of the seed point after updating respectively,measuring a weight function of the degree of the pixel points belonging to the seed points;

step 36: repeating steps 33, 34, 35, the algorithm aims to optimize an energy function, defined as formula (10), stopping when the energy function changes less than a certain threshold in two successive iterations, completing the preliminary superpixel segmentation,

2. the method for probability density-weighted geodesic distance brain MR image segmentation as claimed in claim 1, characterized in that said step 4 comprises:

step 41: scanning the segmented superpixels, if the color standard deviation of the superpixels is larger than a certain threshold value T_cThen the superpixels that satisfy this condition are split, which is calculated as:

C(s_l)＝λStd_l＞T_c (11)

wherein Std_lIs the standard deviation, T, of the pixel color within each superpixel_cIs a threshold, λ is a control parameter;

step 42: and (3) clustering all pixels in each super pixel needing to be split into two classes again by using an FCM algorithm, wherein the target function of the local FCM is as follows:

wherein C is the number of clusters, Q_nIs the number of pixels, mu, in the currently scanned superpixel_ijIs the jth pixelFuzzy membership function belonging to ith cluster and satisfying constraintm is a weighting function acting on the fuzzy membership, v_iIs the ith cluster center, x_jIs the color value of a pixel point within the currently scanned superpixel;

step 43: searching connected regions according to the obtained super-pixel classification result, taking the pixels in each connected region as a new class, namely a new super-pixel block, and updating the super-pixel classification result matrix;

step 44: and repeating the steps 42 and 43 for the superpixels needing to be split until all the superpixels meeting the splitting condition are split.

3. The method for segmenting the MR image of the brain according to the probability density weighted geodesic distance of the claim 2, characterized in that the step 5 is further as follows:

taking the super-pixels as the input of an FCM algorithm, clustering on the basis of the segmented super-pixels to complete the segmentation of the brain image, wherein the FCM objective function based on the super-pixels is as follows:

wherein C 'is the number of clusters, Q'_nIs the number of super pixels in the image, mu'_ijIs a fuzzy membership function of the jth super pixel belonging to the ith cluster and satisfying the constraintm is a weight function, v ', acting on the fuzzy membership'_iIs the ith cluster center, ξ_iIs a super pixel S_jThe color mean value of (1).