A brain tumor image segmentation method based on multi-level clustering

Abstract

The present invention is a brain tumor image segmentation method based on multi-level clustering, which comprises the following contents: collecting brain tumor nuclear magnetic resonance image data with tumor as the original brain tumor image, and obtaining the tumor area mask of the image at the same time; segmenting the image by multi-level clustering, including a preprocessing module, a feature extraction module, a hierarchical clustering segmentation module, and a tumor area selection module; the feature extraction module extracts features from the filtered brain tumor image, and constructs a feature vector for each pixel point of the filtered brain tumor image based on seven indicators, namely, the pixel value of the filtered brain tumor image itself, the mean value of the 3×3 neighborhood pixels, the mean value of the 5×5 neighborhood pixels, the energy of the grayscale co-occurrence matrix, the mean value of the grayscale co-occurrence matrix, and the contrast and roughness in Tamura, and the feature vectors of all pixels constitute a feature matrix. The present invention combines the advantages of the SAP clustering method and the DP clustering method, and exceeds the AP algorithm in terms of computing speed.

Description

Technical Field

[0001] The invention belongs to the field of medical image segmentation, and in particular relates to a brain tumor image segmentation method based on multi-level clustering. Background Art

[0002] The incidence of brain tumors has been increasing in recent years, and the requirements for corresponding medical technologies have also become increasingly higher. During the treatment process, accurate segmentation of brain tumor images is a prerequisite for practical applications such as disease screening, surgical planning, and tumor model construction, and has great research significance. Researchers have conducted a lot of research on brain medical image segmentation and have achieved many research results.

[0003] In clinical applications, the segmentation of brain tumor images mainly relies on experienced experts to manually outline the location of the tumor lesion in the original MRI image based on their own experience and pathological knowledge with the help of auxiliary drawing software on the computer. This is time-consuming, inefficient, lacks real-time performance, and the segmentation results are also very different, which cannot meet the needs of batch processing of brain tumor image segmentation. Clustering methods do not require high-precision annotation information, are simple and efficient, and are easy to implement. Therefore, many researchers have conducted research and exploration in this area.

[0004] At present, there are many improved and combined clustering algorithms based on K-Means, FCM, AP, DP and other clustering algorithms. However, the accuracy of brain tumor image segmentation is high, and each clustering algorithm has its own defects. The segmentation effect has a lot of room for improvement. As relatively new clustering algorithms, AP and DP algorithms have advantages such as good segmentation effect and no need to manually specify too many parameters. However, they are limited by the fact that both algorithms need to build a large-scale similarity matrix for the amount of data, which makes it difficult to directly apply them to brain tumor image processing with a large number of data points. Summary of the invention

[0005] In view of the difficulty of brain tumor image segmentation and the difficulty in obtaining large-scale annotation information, the present invention proposes a brain tumor image segmentation method based on multi-level clustering. This segmentation method introduces Tamura texture features combined with GLCM features and grayscale information to improve the segmentation results. It uses standardized Euclidean distance to optimize the classical Euclidean distance commonly used in clustering algorithms, reducing the excessive impact of extreme features on segmentation results. This is a complete brain tumor image segmentation solution, including preprocessing, cluster segmentation and threshold method to select tumor areas. The segmentation part is an optimized and improved SAPDP clustering, which is a multi-level clustering algorithm that combines sparse neighbor propagation clustering (SAP) and density peak clustering (DP).

[0006] The technical solution of the present invention is:

[0007] In a first aspect, the present invention provides a brain tumor image segmentation method based on multi-level clustering, characterized in that the image segmentation method comprises the following contents:

[0008] Collect brain tumor MRI image data with tumor as the original brain tumor image, and simultaneously obtain the tumor area mask of the image to form the algorithm test data set;

[0009] The image is segmented using multi-level clustering, including preprocessing module, feature extraction module, hierarchical clustering segmentation module, and tumor area selection module.

[0010] The preprocessing module selects mean filtering, inputs the original brain tumor image, outputs the filtered brain tumor image, replaces each pixel point of the original image in the algorithm test data set with the mean of the 3×3 neighborhood of the pixel point, performs average smoothing filtering on each pixel point, and obtains the filtered brain tumor image;

[0011] The feature extraction module extracts features from the filtered brain tumor image, and constructs a feature vector for each pixel point of the filtered brain tumor image using seven indicators: pixel value of the filtered brain tumor image itself, mean value of 3×3 neighborhood pixels, mean value of 5×5 neighborhood pixels, energy of grayscale co-occurrence matrix, mean value of grayscale co-occurrence matrix, contrast and roughness in Tamura. The feature vectors of all pixels constitute a feature matrix:

[0012] The input of the hierarchical clustering segmentation module is the feature matrix obtained by the feature extraction module, and the output is the final cluster center and the category to which each point belongs. The specific process is: the data points belonging to each representative point in the rough classification stage are assigned to the category to which the representative point belongs in the division stage, that is, the newly generated cluster center in the FDP stage obtains all the belonging points of the cluster center in the SAP stage, and the segmentation is completed;

[0013] The tumor region selection module is implemented by a threshold segmentation method. By performing threshold selection on the cluster segmentation results, clusters that are not tumor clusters are filtered out, leaving a tumor region mask binary image, and completing high-precision segmentation of brain tumor images.

[0014] In a second aspect, the present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program is suitable for executing the above-mentioned image segmentation method when loaded by a computer.

[0015] In a third aspect, the present invention provides a brain tumor image segmentation device based on multi-level clustering, characterized in that it includes a memory and a processor, the memory is used to store a computer program, and the computer program includes a preprocessing module, a feature extraction module, a hierarchical clustering segmentation module, and a tumor region selection module.

[0016] The preprocessing module selects mean filtering, inputs the original brain tumor image, outputs the filtered brain tumor image, replaces each pixel point of the original image with the mean value of the 3×3 neighborhood of the pixel point, performs average smoothing filtering on each pixel point, and obtains the filtered brain tumor image;

[0017] The feature extraction module extracts features from the filtered brain tumor image, and constructs a feature vector for each pixel point of the filtered brain tumor image using seven indicators: pixel value of the filtered brain tumor image itself, mean value of 3×3 neighborhood pixels, mean value of 5×5 neighborhood pixels, energy of grayscale co-occurrence matrix, mean value of grayscale co-occurrence matrix, contrast and roughness in Tamura. The feature vectors of all pixels constitute a feature matrix:

[0018] The hierarchical clustering segmentation module inputs the feature matrix obtained by the feature extraction module and outputs the final cluster center and the category to which each point belongs. The specific process is: the data points belonging to each representative point in the rough classification stage are assigned to the category to which the representative point belongs in the division stage, that is, the newly generated cluster center in the FDP stage obtains all the belonging points of the cluster center in the SAP stage, and the segmentation is completed;

[0019] The tumor region selection module is implemented by threshold segmentation method. By selecting the threshold value of the cluster segmentation results, the clusters that are not tumor clusters are filtered out, leaving the tumor region mask binary map, and completing the high-precision segmentation of the brain tumor image. The mask binary map shows the location, size, shape, boundary and other information of the tumor.

[0020] Compared with the prior art, the present invention has the following advantages:

[0021] The present invention can achieve the segmentation of brain tumor images efficiently and accurately, combining the advantages of SAP clustering method and DP clustering, and surpassing AP algorithm in operation speed. The sparse AP algorithm is introduced, and the optimal number t of neighborhood pixels is determined through a large number of experiments. In order to balance the algorithm operation speed and segmentation accuracy, the t value is set to the number of all pixels in the 32×32 neighborhood, so that the scale of the similarity matrix is ​​256 2 ×32 2 =67108864, more than the original 256 4 =4294967296, which reduces the complexity by 64 times.

[0022] The present invention can realize image segmentation without manually specifying cluster centers. The exact number of cluster centers is often unknown in the brain tumor image segmentation task, so the interference of randomly specifying the number of cluster centers on the cluster segmentation result is avoided. Few parameters need to be manually determined, including the bias parameter P in the coarsening stage, the damping factor λ, and the truncation distance d in the division stage. cThese parameters will affect the results of cluster segmentation. In order to adapt to the characteristics of brain tumor images with large grayscale values, unclear cluster boundaries, and many cluster centers, the parameter combination suitable for brain tumor image segmentation can be obtained through experimental adjustment. For this type of data, P is set to the median of all elements in the similarity matrix S; λ is selected as 0.8; d c The average number of neighbors per point is chosen to be around 2% and does not need to be adjusted for each image.

[0023] The present invention introduces the Tamura texture features of the pixel neighborhood, combines the neighborhood grayscale value of the pixel and the neighborhood grayscale co-occurrence matrix features to construct a feature matrix for each pixel, and clusters the multi-dimensional feature vectors of each pixel. Compared with clustering using only grayscale information, the image segmentation effect has been improved by introducing multi-dimensional features for clustering. During the experiment, an average Dice value improvement of 0.03 was obtained.

[0024] The present invention introduces the standardized Euclidean distance to replace the Euclidean distance commonly used in clustering algorithms, which can effectively weaken the adverse effect of excessively large components in the feature metric on the segmentation result. ij If the value of a feature metric is relatively large, then this feature will make other features almost lose their influence on the clustering result of the pixel. The introduction of weighted thinking can weaken this adverse effect and make each group of feature components have an influence on the clustering result.

[0025] The algorithm of the present invention performs image segmentation and compares it with the real segmentation result, calculates the Dice coefficient, and the result shows that the average Dice value can reach 0.89, which is higher than 0.80 of K-Means and 0.82 of SiFCM. The present invention proposes an improved multi-level clustering algorithm that is insensitive to key parameters on the basis of being able to segment large images, and has better segmentation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a data example. The left side of the figure is the original brain tumor MRI image, and the right side is the tumor area mask binary image of this image.

[0027] Figure 2 This is the flow chart of SAPDP algorithm.

[0028] Figure 3 Experimental procedure.

[0029] Figure 4 The segmentation process and result comparison chart shows the input image, gold standard mask image, spatial fuzzy C-means clustering segmentation result, K-means segmentation result, classic AP segmentation result, and the segmentation result of the algorithm in this paper.

[0030] Figure 5 This is a diagram that compares the impact of different numbers of clustering features on the clustering results for the same image. DETAILED DESCRIPTION

[0031] The specific embodiments of the present invention are described more completely and clearly below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without any creative work belong to the scope of the present invention.

[0032] The present invention is a brain tumor image segmentation method based on multi-level clustering. The method is used in the brain tumor image segmentation task and specifically comprises the following steps:

[0033] ⑴Dataset preparation

[0034] The original data of the present invention includes brain tumor MRI images and manual segmentation results (i.e., the tumor area mask corresponding to the medical image) of 110 patients with low-grade gliomas located on the Cancer Genome Atlas (TCGA), wherein the original data includes two types of images: a mask image representing the tumor area and the original medical image that has not been segmented. During the experiment, 36 groups of cases with brain tumor MRI images and manual segmentation results were selected to construct the algorithm test data set. Each group selected five to ten images, totaling 303 images to constitute the algorithm test data set. The original data was collected from the Internet and combined with some images collected from real hospitals to verify the effect of the algorithm proposed in the present invention. Figure 1 As shown in the figure, the left side is the original brain tumor MRI image, and the right side is the tumor area mask binary image of this image. In the actual processing process, only the original image on the left is processed, and the annotation information on the right is used to compare with the segmentation result and obtain the Dice value.

[0035] ⑵Multi-level clustering segmentation image, including preprocessing module, feature extraction module, hierarchical clustering segmentation module, and tumor area selection module.

[0036] The preprocessing module uses mean filtering, inputs the original brain tumor image, outputs the filtered brain tumor image, replaces each pixel point of the original image in the algorithm test data set with the mean of the 3×3 neighborhood of the pixel point, performs average smoothing filtering on each pixel point, and obtains the filtered brain tumor image. The smoothing effect is related to the filter window. The larger the window, the stronger the smoothing effect, but the details and edges are also more blurred. In order to adapt to the grayscale characteristics of the brain tumor image, after experimental comparison, the present invention uses a 3×3 filter window.

[0037] The feature extraction module implementation process is:

[0038] The feature extraction of the filtered brain tumor image is performed, and the seven indicators of the pixel value of the filtered brain tumor image itself, the mean value of the 3×3 neighborhood pixels, the mean value of the 5×5 neighborhood pixels, the energy of the grayscale co-occurrence matrix, the mean value of the grayscale co-occurrence matrix, the contrast and roughness in Tamura are used to construct a feature vector for each pixel point of the filtered brain tumor image. It is a multi-dimensional pixel information vector, and the feature vectors of all pixels constitute a feature matrix:

[0039] The calculation process of each index is as follows: first, the 3×3 neighborhood pixel mean and 5×5 neighborhood pixel mean of each pixel of the filtered brain tumor image are calculated;

[0040] At the same time, in the stage of calculating the GLCM features of the pixel surrounding area of ​​the filtered brain tumor image, the edge pixels are first expanded in reverse order to expand the image boundary by 4 pixel values ​​in each direction. Through experiments, it is determined that the use of a 9×9 slider to calculate the grayscale co-occurrence matrix information around each pixel can best ensure that the grayscale co-occurrence matrix features around each pixel can represent the features of the point, and the texture features of a pixel are distinguished from the texture features of the surrounding pixels. Then, the features including the energy of the grayscale co-occurrence matrix and the mean of the grayscale co-occurrence matrix are obtained.

[0041] Through experiments, the grayscale co-occurrence matrix information around each pixel can best ensure that the grayscale co-occurrence matrix characteristics around each pixel can represent the characteristics of the point and be distinguished from the grayscale co-occurrence matrix characteristics of nearby pixels. A 7×7 calculation slider is used to obtain the Tamura eigenvalue within the 7×7 range around each pixel in the filtered brain tumor image, and the contrast and roughness indicators in the Tamura image texture analysis are calculated.

[0042] The input of the hierarchical clustering segmentation module is the feature matrix obtained by the feature extraction module, and the output is the final cluster center and the category to which each point belongs. The implementation process is as follows:

[0043] Coarsening stage: implemented by sparse AP algorithm. The traditional AP algorithm constructs a similarity matrix for all data, calculates the distance between every two points as the similarity, and defines the distance between each point and itself as the bias parameter P. If the total amount of data is N, the size of the similarity matrix is ​​N×N. The sparse AP algorithm of this embodiment only considers a specified number of the nearest points, which greatly reduces the amount of calculation and memory usage. The sparse AP algorithm inputs the AP data set X={x1,x2,...,x N}, X is the feature matrix obtained by the feature extraction module, x i is the feature vector corresponding to the i-th pixel, referred to as data point x i, N is the total number of pixels, N = 256 × 256, output the cluster center as the cluster data for the next division stage, and the corresponding relationship between each data point and which cluster center it belongs to. The operation is as follows:

[0044] 1.1) Using formula (1) to calculate the AP data set X = {x1, x2, ..., x N}, the size of S is N×t, t is the number of selected neighbor data points (t=32 2 ). In formula (1), the bias parameter P is set to the median of all elements in the sparse similarity matrix S, the distance formula uses the standardized Euclidean distance, and the attraction matrix R and the attribution matrix A are both initialized to N×t zero matrices.

[0045]

[0046] 1.2) Calculate R using formula (2) T (i,j) and A T (i,j):

[0047]

[0048]

[0049] R T (i, j) refers to the attraction value of the attraction matrix at position (i, j) after the Tth iteration, indicating that the data point x j For data point x i The attractiveness of , that is, the degree to which data point j is suitable for representing data point i;

[0050] A T (i, j) refers to the degree value of the degree matrix at position (i, j) after the Tth iteration, indicating that the data point x i Select data point x j As the fitness of the cluster center, when i=j, calculate A T (i,i) represents A T (i,j).

[0051] 1.3) Use formula (3) to update R in the iteration process T , A T In formula (2) and formula (3), (i, j) represents the value of a certain position in the matrix, and no (i, j) represents the matrix itself.

[0052] R T =R T (1-λ)+R T-1 λ

[0053] A T=A T (1-λ)+A T-1 λ (3)

[0054] Wherein λ represents the damping factor, which can prevent the values ​​of R and A from oscillating during the iteration process. The iteration is more stable and generally λ∈[0.5,1]. In the present application, λ is taken as 0.8. When the value is small, the number of iterations will be reduced and the algorithm convergence will be accelerated, but it will increase the instability factors in the clustering process and affect the clustering effect.

[0055] 1.4) If the attraction matrix R and the attribution matrix A no longer update their values ​​with iteration, or the number of iterations T has reached the maximum number of iterations, then execute the judgment operation of step 1.5). If both conditions are not met, then continue to loop through steps 1.2)-1.3) to update the values ​​of R and A.

[0056] 1.5) If the point x K R TK +A TK >0, (R TK For point x K The corresponding attraction matrix of the Tth iteration, A TK For point x K The corresponding attribute matrix of the Tth iteration) then point x K It is a cluster center, i.e., the representative point selected in the coarsening stage. All cluster centers constitute the representative point set. Each non-cluster center point (i.e., non-representative point) is assigned to the cluster cluster where the cluster center (i.e., representative point) with the greatest similarity is located. In the sparse AP algorithm, the bias parameter P is set to the median of all elements in the sparse similarity matrix S. It can also be adjusted according to needs. For example, it can also be set to the average or minimum value of all values ​​in the similarity matrix. The larger the P value, the more cluster centers are generated.

[0057] Division phase: FDP algorithm is used to implement the FDP algorithm. The FDP algorithm inputs the DP data set Y = {y1, y2, ..., y M}, Y is the set of feature vectors of the pixels in the cluster center obtained in the coarse classification stage, y p is the pth data in the DP data set, and M is the number of cluster center points after the coarse clustering stage. The cluster centers of the partitioning stage and the belonging points of each cluster center are output as follows.

[0058] 2.1) Use the standardized Euclidean distance to calculate the distance between every two points in the DP data set Y. If there are M data in Y, these distance values ​​are stored in an M×M matrix to form the distance matrix of the FDP stage, which is used to calculate the local density and distance in the next step. Some eigenvalues ​​have too much influence on the clustering results because of their large values. Using the standardized Euclidean distance can avoid this situation.

[0059] 2.2) The local density ρ formula uses the Gaussian kernel function. Compared with the truncated kernel function, the Gaussian kernel function increases the amount of calculation, but can better reduce incorrect clustering.

[0060]

[0061] In formula (4), d pq For point y p With point y q The distance between them is the value at the position (p,q) of the distance matrix, where p and q are integers ranging from 1 to M. c is the cutoff distance, and then the average number of neighbors of each point is automatically searched to obtain the d c value, take this d c The value is the cutoff distance used in this method. Arrange ρ in descending order, that is, After sorting, δ S The calculation method of is shown in formula (5):

[0062]

[0063] 2.3) Determine the cluster center. The points with large δ and ρ are selected as the cluster center. The average of the local maximum density and the local minimum density is taken as the density threshold, and the average of the distance maximum and the distance minimum is taken as the distance threshold. All points are traversed, and the points with local density and distance greater than the respective set thresholds are selected as the cluster center points. The remaining points with the closest distance and high local density are in the class where the selected cluster center point is located.

[0064] The data points belonging to each representative point in the coarse classification stage are assigned to the category to which the representative point belongs in the division stage, that is, the newly generated cluster center in the FDP stage obtains all the belonging points of the cluster center in the SAP stage, and the segmentation is completed.

[0065] The feature extraction part of the present invention adds information of seven dimensional indicators such as Tamura feature texture and neighborhood grayscale value, which have good performance in the medical field, to construct a feature matrix. The information dimension is richer, and the standardized Euclidean distance is adopted, and the weighting factor is introduced to solve the problem of large numerical differences of each feature component, and the SAPDP algorithm is successfully applied to the brain tumor image segmentation. Using the standardized Euclidean distance to calculate the pixel distance can improve the problem that the feature metric value is too large, which causes the metric to have too much influence on the segmentation result.

[0066] The tumor region selection module is implemented by threshold segmentation method. By selecting the threshold of the cluster segmentation result, the clusters that are not tumor clusters are filtered out, leaving the tumor region mask binary map, completing the high-precision segmentation of the brain tumor image, and realizing the positioning and shaping of the tumor region. The mask binary map shows the location, size, shape, boundary and other information of the tumor.

[0067] The image segmentation results of the method of the present invention were compared with the real segmentation results, and the Dice coefficient was calculated. The results showed that the average Dice value of the method of the present invention could reach 0.89, which was significantly higher than 0.80 of K-Means and 0.82 of SiFCM.

[0068] The original image is segmented by the above segmentation method to obtain the segmentation result. The tumor area obtained by segmentation is compared with the mask information of the data set, and the Dice parameter is calculated. The traditional K-means, spatial fuzzy C-means, and classic AP clustering are compared. In the comparison process, the traditional K-means and spatial fuzzy C-means algorithms have parameters that need to be specified, such as the cluster number K of K-means, the cluster number K of SIFCM and the fuzzy coefficient m. In order to obtain better segmentation effect and more reasonably compare with the method in this paper, the relevant parameters of the traditional algorithm are adjusted for many experiments, and the segmentation results of the optimal parameter combination are selected for comparison. Due to memory overflow, the classic AP algorithm cannot directly segment the images in the data set, so some brain MRI images are reduced to adapt to the classic AP algorithm. The evaluation index is the Dice value of the segmentation result and the actual result of manual segmentation. The Dice coefficient is a comparison relationship at the pixel level. The real target appears in a certain area A, and the target area of ​​the model prediction result is B. Then the Dice coefficient formula is as shown in formula (6).

[0069]

[0070] The value range of the Dice coefficient is 0-1. The closer it is to 1, the better the segmentation result. Figure 3 shown.

[0071] ⑶Test the improved SAPDP image segmentation effect

[0072] like Figure 4As shown in the figure, the comparison diagram shows the input image, the gold standard mask image, the spatial fuzzy C-means clustering segmentation result, the K-means segmentation result, the classic AP segmentation result and the segmentation result of the algorithm in this paper. The image of this segmentation task has 65536 pixels and a size of 256×256. Directly using the classic AP algorithm for segmentation will result in insufficient memory. Therefore, the original image is reduced to a new image of 64×64 and segmented. The segmentation result will be blurred and a lot of detail information is lost. However, by calculating the DICE coefficient, it is found that the accuracy is still 0.72. The low accuracy of the AP algorithm is not due to the poor clustering effect of the algorithm itself, but the loss of a lot of information during the image reduction operation. SIFCM and K-means clustering as partitioning methods have good effects on the definition of tumor boundaries, but it is difficult to distinguish the skull part that does not belong to the tumor area but has close grayscale information. The processing results did not perform morphological operations. It can be seen that our algorithm has the best skull removal ability while achieving good tumor boundary segmentation. The proposed method achieves the best segmentation for this image, from which the location, size and shape of the tumor can be determined, and the tumor boundary is also clearly displayed. The DICE coefficients of this group of data processing results are shown in Table 1.

[0073] Table 1 Comparison of DICE values ​​of single group images

[0074]

[0075] This paper also studies the image segmentation effect of multi-level clustering using only grayscale information. Figure 5 For the same image, compare the impact of different numbers of clustering features on the clustering results. Figure 5 The second image is the result of segmentation using only grayscale values, and the calculated DICE coefficient is 0.93. After adding Tamura feature metric and GLCM feature metric for clustering segmentation, the DICE coefficient is increased to 0.97. Through a large number of experimental comparisons, adding the texture feature metric used in this method for clustering can effectively improve the clustering effect. In the experimental process of 36 groups of images, the average DICE value is increased by more than 0.03.

[0076] Table 2 is a partial example of the DICE values ​​of each method in all experimental groups. The average Dice value of multiple images in each experimental group is calculated as the evaluation index of the group of images. As can be seen from the table, some images (such as group 35) are difficult to accurately segment the tumor area due to their own characteristics, but our algorithm is still better than other algorithms in segmentation effect.

[0077] Table 2 DICE segmentation results of some groups

[0078]

[0079]

[0080] During the experiment, 36 groups of brain MRI images were segmented and compared. The average DICE values ​​of 303 images in 36 groups are shown in Table 3. During the calculation process, a large number of parameters of K-means and SIFCM algorithms were adjusted. For the algorithm in this paper, the bias parameter P in the coarsening stage and the cutoff distance d in the refinement stage were adjusted. c After confirmation, no adjustment is required, and no additional parameters need to be specified.

[0081] Table 3 Average DICE values ​​of image processing in all groups

[0082]

[0083] As can be seen from Table 3, the algorithm of the present invention has effectively improved the accuracy of tumor segmentation compared with the classical algorithm, and has the advantage of lower computational complexity over the AP algorithm. It should be noted that the average DICE coefficients of the results of processing the reduced-size images using SIFCM and K-means are 0.62 (K-means) and 0.65 (SIFCM), respectively, and the algorithm of the present invention does not process the reduced-size images.

[0084] The brain tumor image segmentation method based on multi-level clustering of the present invention includes three parts: an image preprocessing module, an image segmentation module, and a tumor area selection module. Before segmentation, the original image is first filtered to weaken the influence of noise in the image. Next, the image after the filtering operation is subjected to Tamura texture feature extraction and GLCM feature extraction to obtain partial texture feature measurements of the image and combine them with the pixel grayscale information after filtering to obtain the feature vector of each pixel point, and the feature vectors of all pixels constitute a feature matrix. The feature matrix is ​​clustered, and sparse AP clustering is adopted in the coarsening stage to obtain a large number of point clusters; in the division stage, the density peak clustering algorithm is used to cluster the point cluster centers of the first stage to obtain new cluster centers; finally, in the refinement stage, all points are divided into the cluster centers of the division stage to obtain the image segmentation results. Targets of the same category are represented by the same color, and a mask binary image is made by a threshold segmentation algorithm to extract the tumor part in the clustering result. Compared with traditional K-Means clustering, improved fuzzy C-means clustering and neighbor propagation clustering algorithms, better segmentation can be achieved. The introduction of Tamura texture features has a significant positive impact on the method. Experiments have shown that the effect is better than that of the group without texture features.

[0085] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A brain tumor image segmentation method based on multi-level clustering, characterized in that: The image segmentation method includes the following contents: Collect brain tumor MRI image data with tumor as the original brain tumor image, and simultaneously obtain the tumor area mask of the image to form the algorithm test data set; The image is segmented using multi-level clustering, including preprocessing module, feature extraction module, hierarchical clustering segmentation module, and tumor area selection module. The preprocessing module selects mean filtering, inputs the original brain tumor image, outputs the filtered brain tumor image, replaces each pixel point of the original image in the algorithm test data set with the mean of the 3×3 neighborhood of the pixel point, performs average smoothing filtering on each pixel point, and obtains the filtered brain tumor image; The feature extraction module extracts features from the filtered brain tumor image, and constructs a feature vector for each pixel point of the filtered brain tumor image using seven indicators: pixel value of the filtered brain tumor image itself, mean value of 3×3 neighborhood pixels, mean value of 5×5 neighborhood pixels, energy of grayscale co-occurrence matrix, mean value of grayscale co-occurrence matrix, contrast and roughness in Tamura. The feature vectors of all pixels constitute a feature matrix: The hierarchical clustering segmentation module inputs the feature matrix obtained by the feature extraction module and outputs the final cluster center and the category to which each point belongs. The specific process is: the data points belonging to each representative point in the rough classification stage are assigned to the category to which the representative point belongs in the division stage, that is, the newly generated cluster center in the FDP stage obtains all the belonging points of the cluster center in the SAP stage, and the segmentation is completed; The tumor region selection module is implemented by using a threshold segmentation method. By performing threshold selection on the cluster segmentation results, clusters that are not tumor clusters are filtered out, leaving a tumor region mask binary image, and completing high-precision segmentation of the brain tumor image. The multi-level clustering segmentation module is divided into three steps: coarsening stage, partitioning stage, and refinement stage; the coarsening stage uses a sparse AP clustering algorithm to reduce a large-scale data set to a data set of appropriate size and maintain the important features and properties of the original data set; The partitioning stage is implemented by the FDP algorithm, which performs clustering operations on the reduced-scale data set; the refinement process assigns the data points belonging to each representative point in the coarse-divided stage to the category to which the representative point belongs in the partitioning stage, that is, the newly generated cluster center in the FDP stage obtains all the belonging points of the cluster center in the SAP stage, completing the partitioning; The situation in the coarsening stage is: The coarsening stage is implemented by the sparse AP algorithm. The sparse AP algorithm inputs the data set X = {x1, x2, ..., x N }, output the preliminary division results and representative point set, the steps are as follows: 1.1) Initialization. Use formula (1) to calculate the AP data set X = {x1, x2, ..., x N }, the size of S is N×t, where the bias parameter P is set to the median of all elements in the similarity matrix S, the distance formula uses the standardized Euclidean distance, and t is the number of selected neighbors; The attraction matrix R and the attribution matrix A are initialized as N×t zero matrices. 1.2) Calculate R using formula (2) T (i,j) and A T (i,j): R T (i, j) refers to the attraction value of the attraction matrix at position (i, j) after the Tth iteration, indicating that the data point x j For data point x i The degree of attraction of data point j, that is, the degree to which data point j is suitable for representing data point i; A T (i, j) refers to the degree value of the degree matrix at position (i, j) after the Tth iteration, indicating that the data point x i Select data point x j As the fitness of the cluster center, when i=j, calculate A T (i,i) represents A T (i,j); 1.3) Use formula (3) to update R in the iteration process T , A T Value R T =R T (1-λ)+R T-1 l A T =A T (1-λ)+A T-1 λ(3) Where λ is the damping factor, which is used to ensure the stability of algorithm iteration, λ∈[0.5,1]; 1.4) If the attraction matrix R and the attribution matrix A no longer update their values ​​with iteration, or the number of iterations T has reached the maximum number of iterations, then execute the judgment operation of step 1.5). If both conditions are not met, then continue to loop through steps 1.2)-1.3) to update the values ​​of R and A; 1.5) If the point x K R TK +A TK >0, then point x K is a cluster center, i.e., the representative point selected in the coarsening stage. All cluster centers constitute a representative point set. Each non-cluster center point is assigned to the cluster cluster where the cluster center with the greatest similarity is located. The conditions of the division phase are: The partitioning phase of the algorithm is implemented using the FDP algorithm. The FDP algorithm inputs the DP data set Y = {y1, y2, ..., y M }, Y is the set of feature vectors of the pixels in the cluster center obtained in the coarse classification stage, y p is the pth data in the DP data set, outputs the cluster centers of the partitioning stage and the belonging points of each cluster center, the steps are as follows: 2.1) Use the standardized Euclidean distance to calculate the distance between every two points in the DP data set Y. If there are M data in Y, these distance values ​​are stored in an M×M matrix to form the distance matrix of the FDP stage, which is used to calculate the local density ρ and distance δ in the next step; 2.2) The formula for local density ρ is: Formula (4) uses the Gaussian kernel function, In formula (4), d pq For point y p With point y q The distance between them, p and q are integers ranging from 1 to M; d c is the cutoff distance, and ρ is arranged in descending order, that is, After sorting, δ s The calculation method is formula (5): 2.3) Determine the cluster center. Points with large δ and ρ are selected as cluster centers. The average of the local maximum density and the local minimum density is taken as the density threshold, and the average of the distance maximum and the distance minimum is taken as the distance threshold. All points are traversed and the points with local density and distance greater than the respective set thresholds are selected as cluster centers. The remaining points with the closest distance and high local density are in the class where the selected cluster center is located.

2. The brain tumor image segmentation method based on multi-level clustering according to claim 1, characterized in that: The calculation process of each index in the feature extraction module is: First, the pixel mean of the 3×3 neighborhood and the pixel mean of the 5×5 neighborhood of each pixel in the filtered brain tumor image are calculated; At the same time, in the stage of calculating the GLCM features of the pixel surrounding area of ​​the filtered brain tumor image, the edge pixels are first expanded in reverse order to expand the image boundary by 4 pixel values ​​in each direction, and a 9×9 slider is used to calculate the gray-level co-occurrence matrix information around each pixel, and then the features including the energy of the gray-level co-occurrence matrix and the mean of the gray-level co-occurrence matrix are obtained; A 7×7 calculation slider is used to obtain the Tamura eigenvalue within a 7×7 range around each pixel of the filtered brain tumor image, and the contrast and roughness indicators in the Tamura image texture analysis are calculated.

3. The brain tumor image segmentation method based on multi-level clustering according to claim 1, characterized in that: The segmentation method performs image segmentation and calculates the Dice coefficient, the average Dice value is not less than 0.89, and the t value is set to be the number of all pixels in a 32×32 neighborhood; λ is selected as 0.8, P is set to the median of all elements in the similarity matrix S, d c The average number of neighbors per point is chosen to be 2%.

4. A computer-readable storage medium, wherein a computer program is stored in the storage medium, and the computer program is suitable for executing the method of claim 1 when loaded by a computer.

5. A brain tumor image segmentation device based on multi-level clustering, characterized in that: Executing the method of claim 1, comprising a memory and a processor, wherein the memory is used to store a computer program, wherein the computer program comprises a preprocessing module, a feature extraction module, a hierarchical clustering segmentation module, and a tumor region selection module; The preprocessing module selects mean filtering, inputs the original brain tumor image, outputs the filtered brain tumor image, replaces each pixel point of the original image with the mean value of the 3×3 neighborhood of the pixel point, performs average smoothing filtering on each pixel point, and obtains the filtered brain tumor image; The feature extraction module extracts features from the filtered brain tumor image, and constructs a feature vector for each pixel point of the filtered brain tumor image using seven indicators: pixel value of the filtered brain tumor image itself, mean value of 3×3 neighborhood pixels, mean value of 5×5 neighborhood pixels, energy of grayscale co-occurrence matrix, mean value of grayscale co-occurrence matrix, contrast and roughness in Tamura. The feature vectors of all pixels constitute a feature matrix: The hierarchical clustering segmentation module inputs the feature matrix obtained by the feature extraction module and outputs the final cluster center and the category to which each point belongs. The specific process is: the data points belonging to each representative point in the rough classification stage are assigned to the category to which the representative point belongs in the division stage, that is, the newly generated cluster center in the FDP stage obtains all the belonging points of the cluster center in the SAP stage, and the segmentation is completed; The tumor region selection module is implemented by a threshold segmentation method. By performing threshold selection on the cluster segmentation results, clusters that are not tumor clusters are filtered out, leaving a tumor region mask binary image, and completing high-precision segmentation of brain tumor images.

Images