CN110210578B - Cervical cancer histopathology microscopic image clustering system based on graph theory - Google Patents

Abstract

The invention discloses a cervical cancer histopathology microscopic image clustering system based on graph theory, which comprises: acquiring microscopic image data of a cervical cancer tissue, and carrying out first-stage clustering; step two, the distribution of cell nucleuses is approximately represented through skeletonized nodes; step three, taking each communicated skeleton as a region, and constructing a minimum spanning tree through skeleton nodes in each region; step four: calculating different statistical values according to the minimum spanning tree structure chart to be used as graph characteristics; step five: and performing second-stage clustering based on the extracted graphic features and the global features. The invention provides a cervical cancer histopathology microscopic image clustering system based on graph theory, which makes full use of nodes generated in skeletonization processing to approximately represent the distribution of cell nuclei, and improves the clustering effect; different types of tissues are expressed in a secondary clustering mode, and special structures and complexity in the tissues are reflected.

Description

Cervical cancer histopathology microscopic image clustering system based on graph theory

Technical Field

The invention relates to an image analysis technology in the field of image medical assistance. More particularly, the invention relates to a cervical cancer histopathology microscopic image clustering system based on graph theory.

Background

In the prior art, when clustering a cervical cancer tissue microscopic image, a flow processing method as shown in fig. 2 is generally used to segment the cervical cancer tissue microscopic image, and the specific steps are as follows:

(a) acquiring an original microscopic section image of a cervical cancer tissue;

(b) converting the original image into a gray image;

(c) reconstructing the image through operator operation to highlight the difference between the cell nucleus and the cell cytoplasm;

(d) filtering out unnecessary parts by using morphological processing, and positioning cell nucleuses in the tissue image by using an automatic threshold value;

(e) map structures are generated with the location of the nuclei.

Further, the segmented image is analyzed based on the region, and a corresponding clustering result as shown in fig. 3 is obtained. As can be seen from fig. 3, (a) is a graph structure of normal tissue, (B) (c) (d) represents three stages of CIN1, CIN2, CIN3, respectively, in which yellow-labeled clusters of Y1, Y2, Y3, Y4 represent basal layers, green-labeled clusters of G1, G2, G3, G4 represent intermediate layers, and blue-labeled clusters of B1, B2, B3, B4 represent surface layers, for analyzing the relationship of grading of CIN and changes in the graph.

For clustering by adopting such a flow method, the following problems generally exist:

(1) in the clustering process, the problem that tissue cells are difficult to divide due to large-area adhesion and overlapping often occurs, the accurate identification of the positions of cell nuclei cannot be achieved by adopting the existing dividing technology, the condition of excessive division or insufficient division can be generated, certain difference is generated between the actual form of the tissue cells and the actual form of the tissue cells, and the judgment of the structure is influenced.

(2) In the prior art, the space structures among cells are represented by distinguishing the triangular areas according to the graph structures and marking the triangular areas with different colors, the method has specificity, the distinguishing degree of different areas and different forms of the cells of the tissues is not high, and because the tissue images have the characteristics of complexity and specificity, the existing method can only represent the density of the cell spaces and cannot explain the difference in the structure, which is very unfavorable for distinguishing cancers.

Disclosure of Invention

An object of the present invention is to solve at least the above problems and/or disadvantages and to provide at least the advantages described hereinafter.

The invention also aims to provide a cervical cancer histopathology microscopic image clustering system based on graph theory, which makes full use of nodes generated in skeletonization processing to approximately represent the distribution of cell nuclei, so that the identification of the spatial structure of the cell nuclei is more accurate, and the clustering effect is improved; and further, different types of tissues are expressed in a secondary clustering mode, and the clustering result can obviously show the special structure and the complexity degree in the tissues.

To achieve these objects and other advantages in accordance with the present invention, there is provided a histopathological microscopic image clustering system for cervical cancer based on graph theory, comprising:

acquiring microscopic image data of a cervical cancer tissue, and carrying out first-stage clustering;

performing skeletonization processing on the result of the first-stage clustering to approximately represent the distribution of cell nuclei through skeletonized nodes;

separating the nodes, taking each communicated skeleton as a region, and constructing a minimum spanning tree through skeleton nodes in each region;

step four: calculating different statistical values according to the generated minimum spanning tree structure diagram, and using the statistical values as graphic features to represent different organizations;

step five: and performing second-stage clustering based on the extracted graphic features and the global features.

Preferably, in the first-stage clustering process in the first step, the primary clustering operation is performed on the cervical histopathology images by applying a k-means algorithm by using RGB pixel values as color features;

the K value in the K-means algorithm is set to be 2, and the clustering result is used for representing the cell nucleus area by adopting white color and representing the area except the nucleus by adopting black color, so that the cell nucleus is distinguished from cytoplasm and intercellular substance.

Preferably, in the step one, the method further comprises the steps of fusing the results of Sobel edge detection, Canny edge detection, Otsu thresholding and watershed transformation to improve the clustering effect, and increasing the clustering accuracy by using morphological operations.

Preferably, during the skeletonization process in the second step, redundant nodes generated during the skeletonization process need to be deleted, so as to approximate the distribution of the cell nucleus.

Preferably, in step four, the extracted graphic features are configured to include the average, variance, skewness, kurtosis of the edge length and the angle in each region;

in step five, the extracted global features are configured to include the tissue perimeter, and the independently fitted parameter values for the nodes within each tissue.

Preferably, in the fifth step, the second-stage clustering is configured to implement quadratic clustering by adopting a k-means algorithm;

the k value in the quadratic clustering k-means algorithm is configured to be 3, so that the organization structure is divided into three levels through a more detailed clustering result, the complexity and the specificity of the structure are effectively distinguished, and the cancer risk of the organization is predicted.

Preferably, the method further comprises a sixth step of analyzing and evaluating the clustering result through a silhouette value;

the range of the silhouette value outline is set to be negative 1 to 1, the higher the value of the silhouette value outline is, the more appropriate the adopted clustering method is, and the lower the value of the silhouette value outline is or the negative value of the silhouette value outline is, the more or less clustering is indicated.

The invention at least comprises the following beneficial effects: firstly, the invention improves the prior organization clustering technology, fully utilizes nodes generated in skeletonization processing to solve the problem caused by the deficiency of the segmentation technology in the prior art, approximately represents the distribution of cell nucleuses through the skeletonized nodes, has more accurate identification on the space structure of the cell nucleuses, obviously reduces the difference with the actual shape of the cells, can be beneficial to reducing errors, is beneficial to judging the structure of the cells in the later period and improves the clustering effect;

secondly, the invention expresses different types of tissues in a secondary clustering mode, compared with a triangular expression mode in the prior art, the clustering result can obviously show the special structure and the complexity in the tissues, can effectively explain the difference in structure, and is very beneficial to the discrimination of cancers.

Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention.

Drawings

FIG. 1 is a flow chart of a scheme of a cervical cancer histopathological microscopic image clustering system based on graph theory in one embodiment of the invention;

FIG. 2 is a flow chart of a prior art segmentation of cervical cancer tissue;

FIG. 3 is a graph of the clustering results of cervical cancer tissues in the prior art;

FIG. 4 is a graph illustrating the clustering effect of the second clustering performed by the cervical cancer histopathology microscopic image clustering system based on graph theory according to an embodiment of the present invention;

FIG. 5 is a graph diagram of the effect of the cervical cancer histopathology microscopic image clustering system based on graph theory in the silhouette value evaluation in one embodiment of the invention;

FIG. 6 is a schematic diagram showing comparison between the original image and the clustering results of the primary clustering stage and the secondary clustering stage.

Detailed Description

The present invention is further described in detail below with reference to the attached drawings so that those skilled in the art can implement the invention by referring to the description text.

It will be understood that terms such as "having," "including," and "comprising," as used herein, do not preclude the presence or addition of one or more other elements or groups thereof.

Fig. 1 shows an implementation form of a cervical cancer histopathology microscopic image clustering system based on graph theory according to the invention, which comprises:

acquiring microscopic image data of a cervical cancer tissue, and carrying out first-stage clustering;

and step two, skeletonizing the result of the first-stage clustering to approximately express the distribution of cell nuclei through skeletonized nodes, wherein the first-stage clustering result shows that sparse cell nuclei can be well clustered, but the cell nuclei of high-viscosity cells (overlapped cells and adherent cells) are difficult to identify. In addition, the clustering result is greatly influenced by the sharpness and the dyeing method, so the scheme provides a method for generating skeletonized nodes to approximately represent the distribution of the kernels;

separating the nodes, taking each communicated skeleton as an area, wherein the separation can promote the differentiation of different types of tissues, further constructing a minimum spanning tree through the skeleton nodes in each area, and calculating graph characteristics such as side length, triangle number and the like to perform optimal clustering on the images by adopting a triangulation algorithm to replace the minimum spanning tree; or adopting a Thiessen polygon algorithm to replace a minimum spanning tree, and calculating image characteristics such as the number of edges, the area and the like to perform optimal clustering on the images;

step four: calculating different statistical values according to the generated minimum spanning tree structure diagram, representing different tissues as graphic features, wherein the seed points generate the minimum spanning tree by a prim algorithm, the length of the spanning tree, the average distance of line segments and the variance distribution are calculated, and the minimum spanning tree structure diagram is classified according to the length, the distance and the variance distribution;

step five: based on the extracted graphical features (statistics) and global features (geometry), a second stage of clustering is performed, which is used to obtain more detailed results through the second stage of clustering process for predicting the cancer risk of the tissue. The scheme uses a non-supervision learning method based on a graph to solve the cervical histopathology image clustering task, divides the tissue structure into three levels, and effectively distinguishes the complexity and the particularity of the structure.

In another embodiment, in the first-stage clustering process in the first step, the primary clustering operation is carried out on the cervical histopathology images by applying a k-means algorithm by using RGB pixel values as color features;

the K value in the K-means algorithm is set to be 2, and the clustering result is used for representing the cell nucleus area by adopting white color and representing the area except the nucleus by adopting black color, so that the cell nucleus is distinguished from cytoplasm and intercellular substance. In the scheme, the clustering result is represented by black and white, foreground objects (cell nucleus) and background objects (cytoplasm and intercellular substance) are distinguished, and the algorithm can perform first-stage optimized clustering on the images by replacing a fused clustering algorithm with a GrabCT algorithm.

In another embodiment, in the step one, the method further comprises the step of fusing the results of Sobel edge detection, Canny edge detection, Otsu thresholding and watershed transformation to improve the clustering effect, wherein the number of cell nuclei identified by the k-means algorithm is large and is close to a real numerical value, but impurities are more; the outline identified by sobel edge detection is accurate, the algorithm is greatly influenced by noise, and the identification rate of cell nuclei with unobvious boundaries is low; part of the areas obtained by canny edge detection are discontinuous and incomplete edge information, but the detection speed is high, and the noise influence is large; the segmentation graph obtained by the threshold method is clear, the calculation is simple, the segmentation graph is not influenced by the change of the brightness and the contrast of the picture, the application range is wide, but the outline of part of cell nuclei is not obvious, and the calculation efficiency is not high; the watershed method is accurate in positioning, but has the problem of excessive segmentation in recognition. In order to obtain a better segmentation result, the methods are fused, so that the fused algorithm is more accurate, the cell nucleus outline is obvious, excessive segmentation cannot be generated, and the method has the advantages of high precision and wide applicability in the histopathological microscopic image analysis of the cervical cancer.

And increases the accuracy of clustering by employing morphological operations. In this scheme, since the binary image of the cell obtained by threshold segmentation generally has noise and the cell edge is not smooth enough, and a large error is generated when the binary image is directly used for cytoskeletonization, the binary image obtained by threshold segmentation is morphologically processed, so that the precision of the post-cytoskeletonization is higher, and the morphological operation is specifically: and setting the area of the pixels of the eight-connected background area in each cell binary image to be smaller than a first preset value as a foreground area. Because the holes in the cells are all background elements, the filling of the holes in the cells can be realized. And performing an on operation on each cell binary image to remove some cell edge noise and some small noise points and separate some lightly adhered cells, wherein the on operation is the prior art and is not described herein again. And setting the area of the pixels of the eight-connected foreground region in each cell binary image to be smaller than a second preset value as a background region. The cell binary image processed by the method still has some larger impurity noise points which cannot be removed, and the impurity noise points are all foreground elements, so that the impurity noise points are removed.

In another embodiment, in the skeletonization process in the second step, redundant nodes generated in the skeletonization process need to be deleted, so as to approximately represent the distribution of cell nuclei. And then, thinning the skeletonized result. Specifically, a Zhang parallel rapid thinning algorithm is adopted to perform cytoskeleton processing on each morphologically processed cell binary image; the histopathology image to be processed with 46 cell nucleuses in one image is subjected to skeletonization, the skeletonization process can be subdivided into the steps of firstly carrying out grayization process on the image to be processed, then carrying out image segmentation on the image subjected to the grayization process to obtain a cell binary image, then carrying out morphological process on the cell binary image, and further carrying out cytoskeletonization process on the cell binary image, and the skeletonized node distribution can be known, so that the correct rate can reach 93.5% according to the fact that the number of the skeletonized nodes in the image subjected to the cytoskeletonization process is consistent with the number of cells, and under the general condition, the skeleton of each cell can generate two nodes, and the skeleton nodes are similar to the spatial distribution of the cell nucleuses. Finally, redundant nodes are deleted and the remaining nodes are used to approximate the distribution of the nuclei. Experiments show that the overlapping rate of the actual nucleus and the position of the skeleton node reaches 90 percent, so that the nucleus can be approximately expressed in a skeleton node mode.

In another embodiment, in step four, the extracted graph features are configured to include the average, variance, skewness, kurtosis of the edge length and angle within each region;

in the fifth step, the extracted global features are configured to include the circumferences of the tissues and the parameter values of the independent fitting of the nodes in each tissue, and in the scheme, the tissues in the image are described and expressed together through the cooperation of the global features and the graphic features, so that the secondary clustering precision is higher and the secondary clustering effect is better.

In another embodiment, in step five, the second-stage clustering is configured to implement quadratic clustering by adopting a k-means algorithm;

the k value in the quadratic clustering k-means algorithm is configured to be 3, so that the organization structure is divided into three levels through a more detailed clustering result, the complexity and the specificity of the structure are effectively distinguished, and the cancer risk of the organization is predicted. In the scheme, the tissues are grouped into three types according to the characteristics of the graph, because experiments show that the identification effect is best when k is 3. The clustering results of the tissue cells are shown in fig. 4, and the clustering results of the three tissues have obvious difference in morphological structure and complexity. From (a) to (c), the topological structure changes remarkably, showing excellent discriminative power of the graph characteristics.

In another embodiment, the method further comprises a sixth step of analyzing and evaluating the clustering result through a silhouette value, wherein the silhouette value is used for measuring the quality of the classification, and the silhouette value is obtained as a distance value of each data point in the data set, and the value is a dissimilarity degree of each sample with other samples in the same class and a relation value of the dissimilarity degrees with samples in other classes, and the larger the value is, the better the value is;

the range of the silhouette value outline is set to be negative 1 to 1, the higher the value of the silhouette value outline is, the more appropriate the adopted clustering method is, and the lower the value of the silhouette value outline is or the negative value of the silhouette value outline is, the more or less clustering is indicated. In this scheme, in order to further analyze the clustering result, the evaluation of the silhouette value is used to illustrate the relationship between each data point in one cluster and other data points in the neighboring clusters, which is an extremely effective numerical evaluation method. Specifically, as shown in fig. 5, the result of silhouette value contour estimation of the present invention is shown, and it can be seen from the result, because the present invention uses actual medical data, the number of different types of tissues is greatly different, the uniformity of clustering cannot be guaranteed, and the more complex the structure is, the smaller the number is, and the larger the difference is. When k is 3, the average values of the silhouettes of the three clusters are respectively 92%, 87% and 71%. The first level clusters are compared to the second level clusters. As shown in fig. 6, (a) represents the original tissue image, (B) represents the first-stage clustering result, and (c) represents the corresponding second-stage clustering result, wherein blue B5, green G5, and red R1 in the (c) graph represent three types of tissues respectively, so that the special structure and complexity in the tissues can be clearly seen according to the clustering result.

Compared with the prior art, the invention forms a complete clustering process and makes evaluation. The pathological images of the cervical cancer tissues are divided into different categories according to the space structure of the nucleus by using the graph theory characteristics, so that the pathological images of the cervical cancer tissues can be applied to the daily practice of a histologist, show great potential in the field of cancer risk prediction, assist a doctor in making judgment, accelerate the diagnosis time and improve the diagnosis accuracy. The structural characteristics of the different clustering results are related to tumors of different grades and different risks, which can be used as a basis for assessing the presence, grade, risk and outcome of cancer.

The above scheme is merely illustrative of a preferred embodiment, and is not intended to be limiting. When the invention is implemented, appropriate replacement and/or modification can be carried out according to the requirements of users.

The number of apparatuses and the scale of the process described herein are intended to simplify the description of the present invention. The application, modifications and variations of the inventive graph-theory based cervical cancer histopathological microscopic image clustering system will be apparent to those skilled in the art.

While embodiments of the invention have been disclosed above, it is not intended to be limited to the uses set forth in the specification and examples. It can be applied to all kinds of fields suitable for the present invention. Additional modifications will readily occur to those skilled in the art. It is therefore intended that the invention not be limited to the exact details and illustrations described and illustrated herein, but fall within the scope of the appended claims and equivalents thereof.

Claims (1)

1. A cervical cancer histopathology microscopic image clustering system based on graph theory is characterized by comprising:

acquiring microscopic image data of a cervical cancer tissue, and carrying out first-stage clustering;

performing skeletonization processing on the result of the first-stage clustering to approximately represent the distribution of cell nuclei through skeletonized nodes;

separating the nodes, taking each communicated skeleton as a region, and constructing a minimum spanning tree through skeleton nodes in each region;

step four: calculating different statistical values according to the generated minimum spanning tree structure diagram, and using the statistical values as graphic features to represent different organizations;

step five: clustering in the second stage is carried out based on the extracted graphic features and the global features;

in the fifth step, the second-stage clustering is configured to realize secondary clustering by adopting a k-means algorithm;

the k value in the quadratic clustering k-means algorithm is configured to be 3, so that the organization structure is divided into three levels through a more detailed clustering result, the complexity and the specificity of the structure are effectively distinguished, and the cancer risk of the organization is predicted;

analyzing and evaluating the clustering result through the silhouette value;

the range of the silhouette value outline is set to be negative 1 to 1, the higher the value is, the more appropriate the adopted clustering method is, and the lower the value is or the negative value is, the more or less clustering is indicated;

in the first-stage clustering process of the first step, the RGB pixel values are used as color features, and primary clustering operation is carried out on the cervical histopathology images by applying a k-means algorithm;

setting the K value in the K-means algorithm as 2, and adopting white to represent a cell nucleus area and black to represent an area except the nucleus in a clustering result so as to distinguish the cell nucleus from cytoplasm and intercellular substances;

in the first step, the method further comprises the steps of fusing results of Sobel edge detection, Canny edge detection, Otsu thresholding and watershed transformation to improve the clustering effect, and increasing the clustering precision by adopting morphological operation;

in the skeletonization process in the step two, redundant nodes generated in the skeletonization process need to be deleted, so as to approximately represent the distribution of cell nuclei, wherein the redundant node deletion process is configured to include:

s1, performing cytoskeleton processing on each morphologically processed cell binary image by adopting a Zhang parallel rapid thinning algorithm;

s2, performing skeletonization on the histopathology image to be processed, which has 46 cell nuclei in one image, namely performing graying processing on the image to be processed, performing image segmentation on the image subjected to graying processing to obtain a cell binary image, and performing morphological processing on the cell binary image to enable the number of the skeletonized nodes in the image subjected to cytoskeletonization processing to be consistent with the number of cells;

s3, deleting redundant nodes to approximately represent the distribution of cell nuclei through the rest nodes;

in step four, the extracted graphic features are configured to include the average, variance, skewness, kurtosis of the edge length and angle within each region;

in step five, the extracted global features are configured to include the tissue perimeter, and the independently fitted parameter values for the nodes within each tissue.

Images