CN105809703A - Adhesion hemocyte image segmentation method based on improved fractional differential and graph theory - Google Patents

Abstract

The invention relates to an image segmentation method of cohesive blood cells based on improved fractional differential and graph theory. Firstly, aiming at the phenomenon of blurred blood cell images and low contrast, the score of morphological denoising and improved quasi-circle mask operator is The improved fractional differential algorithm preprocesses the blood cell image, and the improved fractional differential algorithm preserves the cell edge details while filtering out the staining pollution and particle noise of the blood cell image; then uses the watershed algorithm to preprocess the image Initially segment the image, and map the over-segmented region into nodes; finally, use the improved graph theory minimum spanning tree (MST) algorithm to re-segment the cell image obtained in step S2. The invention can improve the segmentation accuracy of the cohesive cells in the cell image.

Description

Image Segmentation Method of Adhesive Blood Cells Based on Improved Fractional Differentiation and Graph Theory

technical field

The invention relates to the technical field of medical image segmentation, in particular to an image segmentation method of cohesive blood cells based on improved fractional differential and graph theory.

Background technique

Cells are the basic building blocks of all living organisms. An important direction of biomedicine, which has developed rapidly in recent years, is to diagnose diseases by identifying and counting cells and whether the texture image is distorted. Whether the cell image data can be analyzed accurately depends on whether the cell image can be segmented accurately.

The segmentation method based on graph theory has been a research hotspot at home and abroad in recent years. In 2006, Sharon et al. proposed a top-down hierarchical segmentation method based on graph method in "Nature". The segmentation results are accurate and efficient. Vanhamel et al. proposed a nonlinear multi-scale color image segmentation algorithm based on graph theory; Hu Xuegang et al. proposed a segmentation criterion based on graph theory and normalization, and applied the digital image processing model (LIP model) to image processing; Ye Wei Combined with Mumford-Shah theory, proposed an optimization method, by considering the degree of combination between regions in the image and the geometric properties of each region, calculate the weight based on the degree of combination between regions and add it to the image segmentation of the minimum spanning tree Methods; Zhang et al. proposed an image segmentation method combining watershed and graph theory. AnnaFabijanska et al. proposed an improved algorithm of minimum spanning tree based on graph theory, which improves the speed of image segmentation by reducing the number of vertices in the graph. Biplab Banerjee et al. proposed a minimum spanning tree algorithm based on graph theory, combined with a clustering algorithm and an improved Mean-Shift algorithm to segment multispectral satellite images. In addition, Felzenszwalb and Huttenlocher (referred to as FH) put forward the "small and combined" merging criterion to improve the minimum spanning tree segmentation algorithm. This method has high efficiency and can partially perform different segmentations according to different image characteristics. However, it also has its own disadvantages. The preset k value in the algorithm is difficult to be effectively controlled artificially. If the value is too large, over-merging will occur; if it is too small, it will not be able to effectively suppress the generation of small areas, and more redundant area. This paper improves on the basis of FH algorithm.

Contents of the invention

In view of this, the object of the present invention is to propose a method for segmenting cohesive blood cell images based on improved fractional differential and graph theory, which improves the precision of cohesive cell segmentation in cell images.

The present invention is realized by the following scheme: a method for segmenting cohesive blood cell images based on improved fractional differential and graph theory, comprising the following steps:

Step S1: Aiming at the blurred and low-contrast phenomenon of the blood cell image, the blood cell image is preprocessed by combining the morphological denoising and the improved fractional order differential algorithm of the circle-like mask operator. The improved fractional order differential algorithm The algorithm better preserves the details of the cell edge while filtering out the staining pollution and particle noise of the blood cell image;

Step S2: Use the watershed algorithm to initially segment the preprocessed image, and map the over-segmented regions into nodes;

Step S3: The cell image obtained in step S2 is re-segmented by the improved graph theory minimum spanning tree (MST) algorithm.

Further, in the step S1, a Tiansi operator is proposed based on vector merging of a circle-like improved mask template, which specifically includes the following steps:

Step S11: Construct a 5x5 square Tiansi fractional differential mask template. The Tiansi mask operator is a square structure, and a ₂ at the four corners of the outermost layer is regarded as four vectors whose starting point is at the origin of the template and take half

Step S12: The upper left corner and the lower left corner of the mask operator template obtained after the processing in step S11 Merge into a new horizontally left vector

Step S13: the merged vector obtained after processing the step S12 Superimposed to the vector at the middle position of the original third layer, a circular-like mask template is obtained.

Further, the step S2 specifically includes the following steps:

Step S21: Suppose the image is segmented by the watershed algorithm to obtain 6 over-segmented regions: {1, 2, 3, 4, 5, 6}, and the edge set of the mapped nodes is:

E={e ₁ ,e ₂ ,e ₃ ,e ₄ ,e ₅ ,e ₆ ,e ₇ ,e ₈ ,e ₉ }={(1,2),(1,5),(2,3), (2,5),(3,4),(3,5),(4,5),(4,6),(5,6)};

The weight values of the edge sets are 5, 9, 6, 8, 4, 1, 7, 4, 2 respectively;

Step S22: Map the weight values of over-segmented regions and edges one by one to obtain an initial weighted undirected graph.

Further, the step S3 specifically includes the following steps:

Step S31: Mapping: map the over-segmented region of the input image into nodes, and calculate the weighted undirected graph G={V,E,ψ} of the 8-field system through the improved weight value definition, where |V|=n, |E |=m, the weight value of the edge connecting node v _i and v _j is ψ(v _i , v _j ), and the minimum area of the set area is p;

Step S32: State initialization: Let S ^q-1 be the cut set after the q-1th region merging, set q=1 to make the initial segmentation S ^q-1 =S0, S0=(v ₁ ,v ₂ ,…, v _n ), that is, each element in V is a region, and the internal difference of the initial region area where i=1,2,3...n;

Step S33: Sorting: Arrange all the weight values in the weighted undirected graph G={V,E,ψ} in ascending order to obtain the queue π=(O ₁ ,O ₂ ,...,O _m );

Step S34: If the queue in step S33 is not empty, make the side with the smallest weight value out of the queue, otherwise go to step S37;

Step S35: Construct S ^q from S ^q-1 : Let and They are the areas where the nodes v _i and v _j at both ends of the edge out of the queue in S ^q-1 are located, if:

and

then the region and Merge to get S ^q , and calculate the weight value of the new area and its neighborhood system; otherwise, do not merge, let S ^q =S ^q-1 ;

Step S36: According to the weight value of the new area, update the internal difference, inter-regional difference, and weight value queue π of the new area, set q=q+1, and return to step S34;

Step S37: traverse each area of S ^q , (v _i , v _j )∈E, v _i ∈C1, v _j ∈C2, if |C1|<p or |C2|<p, then replace C1 and C2 merge;

Step S38: assigning the pixels belonging to the same area to the same color, which is convenient for observation and analysis by human eyes;

Step S39: output the segmentation result of the improved minimum spanning tree algorithm in this paper.

Compared with the prior art, the invention has the beneficial effect of effectively improving the accuracy of over-segmentation and under-segmentation in cell segmentation, and has very wide application prospects.

Description of drawings

Fig. 1 is a schematic flow chart of the method of the present invention.

FIG. 2 is a schematic diagram of a Tiansi mask operator according to an embodiment of the present invention.

FIG. 3 is a schematic diagram of a circular-like mask operator according to an embodiment of the present invention.

Fig. 4 is a schematic diagram of the principle of vector merging according to an embodiment of the present invention.

FIG. 5 is a schematic diagram of an over-segmented region of a watershed algorithm according to an embodiment of the present invention.

Fig. 6 is a weighted undirected graph according to an embodiment of the present invention.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

As shown in Figure 1, this embodiment provides a method for segmenting an image of cohesive blood cells based on improved fractional differential and graph theory, including the following steps:

Step S1: Aiming at the blurred and low-contrast phenomenon of the blood cell image, the blood cell image is preprocessed by combining the morphological denoising and the improved fractional order differential algorithm of the circle-like mask operator. The improved fractional order differential algorithm The algorithm better preserves the details of the cell edge while filtering out the staining pollution and particle noise of the blood cell image;

Step S2: Use the watershed algorithm to initially segment the preprocessed image, and map the over-segmented regions into nodes;

Step S3: The cell image obtained in step S2 is re-segmented by the improved graph theory minimum spanning tree (MST) algorithm.

In this embodiment, in the step S1, a template of an appropriate size is selected for the size of the cells in the cell image, and a Tiansi operator-based vector-merging improved mask template is proposed, which specifically includes the following steps :

Step S11: Construct a 5x5 square Tiansi fractional differential mask template, as shown in Figure 2, the Tiansi mask operator is a square structure, so a ₂ at the four corners of the outermost layer can be regarded as the starting point 4 vectors at template origin and take half

Step S12: The upper left corner and the lower left corner of the mask operator template obtained after the processing in step S11 According to the principle of vector merging in Figure 4, merge into a new horizontal left vector

Step S13: the merged vector obtained after processing the step S12 The vector superimposed on the middle position of the original third layer obtains a circular-like mask template, as shown in Fig. 3 .

In this embodiment, the step S2 specifically includes the following steps:

Step S21: Assume that the image is segmented by the watershed algorithm to obtain 6 over-segmented regions as shown in Figure 5: {1, 2, 3, 4, 5, 6}, and the edge sets of the mapped nodes are:

E={e ₁ ,e ₂ ,e ₃ ,e ₄ ,e ₅ ,e ₆ ,e ₇ ,e ₈ ,e ₉ }={(1,2),(1,5),(2,3), (2,5),(3,4),(3,5),(4,5),(4,6),(5,6)};

The weight values of the edge sets are 5, 9, 6, 8, 4, 1, 7, 4, 2 respectively;

Step S22: Map the weight values of over-segmented regions and edges one by one to obtain an initial weighted undirected graph, as shown in FIG. 6 .

In this embodiment, the step S3 specifically includes the following steps:

Step S31: Mapping: map the over-segmented region of the input image into nodes, and calculate the weighted undirected graph G={V,E,ψ} of the 8-field system through the improved weight value definition, where |V|=n, |E |=m, the weight value of the edge connecting node v _i and v _j is ψ(v _i , v _j ), and the minimum area of the set area is p;

Step S32: State initialization: Let S ^q-1 be the cut set after the q-1th region merging, set q=1 to make the initial segmentation S ^q-1 =S0, S0=(v ₁ ,v ₂ ,…, v _n ), that is, each element (node) in V is a region, and the internal difference of the initial region area where i=1,2,3...n;

Step S33: Sorting: Arrange all the weight values in the weighted undirected graph G={V,E,ψ} in ascending order to obtain the queue π=(O ₁ ,O ₂ ,...,O _m );

Step S34: If the queue in step S33 is not empty, make the side with the smallest weight value out of the queue, otherwise go to step S37;

Step S35: Construct S ^q from S ^q-1 : Let and They are the areas where the nodes v _i and v _j at both ends of the edge out of the queue in S ^q-1 are located, if:

and

then the region and Merge to get S ^q , and calculate the weight value of the new area and its neighborhood system; otherwise, do not merge, let S ^q =S ^q-1 ;

Step S36: According to the weight value of the new area, update the internal difference, inter-regional difference, and weight value queue π of the new area, set q=q+1, and return to step S34;

Step S37: traverse each area of S ^q , (v _i , v _j )∈E, v _i ∈C1, v _j ∈C2, if |C1|<p or |C2|<p, then replace C1 and C2 merge;

Step S38: assigning the pixels belonging to the same area to the same color, which is convenient for observation and analysis by human eyes;

Step S39: output the segmentation result of the improved minimum spanning tree algorithm in this paper.

The above descriptions are only preferred embodiments of the present invention, and all equivalent changes and modifications made according to the scope of the patent application of the present invention shall fall within the scope of the present invention.

Claims (4)

1. the adhesion blood cell image dividing method based on the fractional order differential improved and graph theory, it is characterised in that: comprise the following steps:

Step S1: for the phenomenon that blood cell image is fuzzy and contrast is not high, being combined by the fractional order differential algorithm of the similar round mask operator of morphology denoising and improvement and blood cell image carries out pretreatment, the fractional order differential algorithm of described improvement pollutes in the dyeing filtering blood cell image and remains cell edges details preferably while grain noise；

Step S2: with watershed algorithm, pretreated image is carried out just segmentation, overdivided region is mapped as node；

Step S3: cell image step S2 obtained with graph theory minimum spanning tree (MST) algorithm improved is split again.

2. a kind of adhesion blood cell image dividing method based on the fractional order differential improved and graph theory according to claim 1, it is characterized in that: in described step S1, propose a kind of Tiansi operator based on vector merge similar round improve mask template, specifically include following steps:

Step S11: the square Tiansi fractional order differential mask template of one 5x5 of structure, Tiansi mask operator is square structure, a on the position of 4 drift angles of outermost layer₂Treating as is the starting point 4 vectors at template initial pointAnd take its half

Step S12: the upper left corner of the mask operator template obtained after described step S11 is processed is with the lower left cornerMerge into level new vector to the left

Step S13: the combined vector obtained after described step S12 is processedBe added to the vector in original third layer centre position, obtains the mask template of similar round.

3. a kind of adhesion blood cell image dividing method based on the fractional order differential improved and graph theory according to claim 1, it is characterised in that: described step S2 specifically includes following steps:

Step S21: assume that image obtains 6 overdivided regions after watershed algorithm is split: 1,2,3,4,5,6}, the limit collection of its mapped node is:

E={e₁,e₂,e₃,e₄,e₅,e₆,e₇,e₈,e₉}={ (1,2), (1,5), (2,3), (2,5), (3,4), (3,5), (4,5), (4,6), (5,6) }；

The weighted value of described limit collection respectively 5,9,6,8,4,1,7,4,2；

Step S22: the weighted value on overdivided region and limit is mapped one by one and obtains initial weighted undirected graph.

4. a kind of adhesion blood cell image dividing method based on the fractional order differential improved and graph theory according to claim 1, it is characterised in that: described step S3 specifically includes following steps:

Step S31: map: the overdivided region of input picture is mapped as node, calculates weighted undirected graph G={V, E, the ψ of 8 neighborhood systems by the weighted value definition improved }, wherein | V |=n, | E |=m, connect node v_iAnd v_jThe weighted value on limit is ψ (v_i,v_j), setting area minimum area is p；

Step S32: state initialization: set S^q-1It is the cut set after q-1 sub-region merges, makes q=1 make initial segmentation be S^q-1=S0, S0=(v₁,v₂,…,v_n), namely in V, each element is a region, by prime area internal diversityRegion areaWherein i=1,2,3 ... n；

Step S33: sequence: weighted undirected graph G={V, E, ψ } in all of weighted value according to ascending order arrangement, obtain queue π=(O₁,O₂,…,O_m)；

Step S34: if the queue in step S33 is not empty, makes the limit dequeue that weighted value is minimum, otherwise forward step S37 to；

Step S35: from S^q-1Structure S^q: setWithIt is S respectively^q-1The two end node v on the limit of middle dequeue_iAnd v_jThe region at place, if:

And

Then by regionWithMerge and obtain S^q, and calculate the weighted value of new region and its neighborhood system；Otherwise nonjoinder, makes S^q=S^q-1；

Step S36: the weighted value according to new region, updates the internal diversity of new region, region difference and weighted value queue π, makes q=q+1, returns step S34；

Step S37: traversal S^qEach region, (v_i,v_j) ∈ E, v_i∈ C1, v_j∈ C2, if | C1 | < p or | C2 | < p, then remerges C1 and C2；

Step S38: be same color by the pixel assignment belonging to the same area, it is simple to the observation analysis of human eye；

Step S39: the segmentation result of the minimal spanning tree algorithm that output improves herein.