CN113570628A - Leukocyte segmentation method based on active contour model - Google Patents

Abstract

The invention discloses a white blood cell segmentation method based on an active contour model, which consists of the following steps: Step 1: Obtain the number of all pixel values in each small cube, and assign all pixel values in each small cube as predetermined pixel values , and obtain the preprocessing image, step 2: Use the density peak clustering segmentation method to perform density clustering on the preprocessing image to obtain the initial white blood cell image, and then use the morphological processing method to smooth the edge contour of the white blood cell nucleus in the initial white blood cell image. Obtain the final leukocyte nucleus image, step 3: use the level set method to process the final leukocyte nucleus image to obtain the leukocyte nucleus contour curve, and use the leukocyte nucleus contour curve as the initial contour curve of the active contour model to obtain the final leukocyte image; Grayscale feature, segment leukocytes with the evolution of the nucleus contour line to the cytoplasmic edge to obtain more accurate results.

Description

A method of leukocyte segmentation based on active contour model

technical field

The invention belongs to the field of medical image processing, in particular to a white blood cell segmentation method based on an active contour model.

Background technique

Medical image segmentation has become one of the main research directions in the field of image processing due to its great help to the fields of auxiliary medicine and clinical medicine. For certain medical images, such as images of white blood cells in peripheral blood cell images. Changes in the number of white blood cells can be used to determine whether the human body is infected. However, usually due to the characteristics of leukocytes or the influence of equipment, the nucleus and cytoplasm of leukocytes are inevitably affected by uneven grayscale and noise during the stage of staining and image acquisition. Algorithm accuracy and speed are still not ideal.

In addition to various types of white blood cells, peripheral blood images also contain a large number of red blood cells, platelets and other substances. Usually, the standard staining method for leukocytes can mark leukocytes well, but due to the characteristics of leukocytes, the nuclei of leukocytes are deeply stained, dark purple, and the cytoplasm is lavender. At the same time, due to the interference of red blood cells in the background, staining leakage and other problems, most segmentation algorithms have excellent segmentation performance for leukocyte nuclei, but the segmentation effect for leukocyte cytoplasm is unsatisfactory.

Usually, the identification and classification system of leukocytes includes the following aspects: image preprocessing, leukocyte segmentation, feature extraction and classification. Segmentation, as a link before the classification task, is crucial to the accuracy of classification. There are two main segmentation methods at present: one is the segmentation method based on the graph theory, mainly including the threshold method, clustering method, etc.; the other is the segmentation method based on the variation theory, mainly including the active contour method. At present, there is no segmentation method based on the atlas theory that can effectively overcome the problem of uneven grayscale of leukocyte images in various situations. However, the segmentation method based solely on variational theory, because of the huge grayscale difference between the leukocyte nucleus and the cytoplasm, leads to activity The evolution curve of the contour is difficult to fit the edge of the cytoplasm or converges to the edge of the nucleus, which makes it difficult to segment the leukocytes.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a leukocyte segmentation method based on an active contour model, which can accurately segment leukocyte nucleus and cytoplasm.

The present invention adopts the following technical scheme: a leukocyte segmentation method based on an active contour model, which consists of the following steps:

Step 1: Divide the RGB pixel space of the original stained white blood cell image into small cubes of equal length, calculate the pixel value at the center point in all the small cubes, that is, the predetermined pixel value, obtain the number of all pixel values in each small cube, and set the All pixel values in each small cube are assigned to predetermined pixel values, and a preprocessed image is obtained,

Step 2: Perform density clustering on the preprocessed image using the density peak clustering segmentation method to obtain an initial leukocyte image, and then use a morphological processing method to smooth the edge contours of leukocyte nuclei in the initial leukocyte image to obtain a final leukocyte nucleus image ,

Step 3: The final leukocyte nucleus image is processed by the level set method to obtain the leukocyte nucleus contour curve, the leukocyte nucleus contour curve is used as the initial contour curve of the active contour model, and the local JS divergence is used to construct the local gray information expression as the edge stop function. The initial contour curve is gradually calculated to obtain the leukocyte edge, thereby obtaining the final leukocyte image.

Further, the edge stop function described in step 3 is:

where I(x) represents the local image area, C ₁ and C ₂ are the average gray values of the local area divided into inner and outer parts by the contour curve, respectively, p and q respectively represent the inner and outer parts divided by the contour curve Gray distribution, JS(p, q) represents the JS divergence information between the gray distribution of the inner and outer parts, α is the local threshold, which is taken as 0.02 to measure the uniformity of the local gray, α-JS ( p, q)>0, it means that the grayscale is uniform, otherwise, it means that the grayscale is not uniform, and the local window is selected as 5*5.

Further, step 2 consists of the following steps:

Step 21: First, calculate the local density ρ and the relative distance δ of the preprocessed image pixel by pixel, and determine the truncation distance d _c ;

Among them, the local density ρ of density clustering is:

In the above formula, d _{i, j} represent the distance between pixel i and pixel j, calculated using Euclidean distance, d _cutoff represents the cutoff distance, that is, the effective density radius, which is the only parameter in the density peak clustering , take 0.5% of the maximum relative distance, then ρ _i can represent the distribution of pixel i around the pixel, that is, the local density;

Among them, the relative distance δ of density clustering is:

Step 22: Arrange the product of the local density ρ of all pixel points and the relative distance δ in descending order, take the first N points as the cluster center point, and the cluster center N value is 2, indicating that the cluster structure is divided into two categories, one is the background part, one is the foreground leukocyte nucleus;

Step 23: Classify and divide the remaining points, that is, pixels other than the N cluster center points, according to the cluster centers, and obtain the clustered image, that is, the white blood cell nucleus image, and the white blood cell nucleus image obtained by clustering is obtained. Binarization, the white blood cell nucleus is used as the foreground, and the rest is used as the background. The morphological processing method is used, and the small area is eliminated by morphological expansion and erosion operations; the foreground part is processed, and the edge contour of the white blood cell nucleus is obtained. That is, the image is smoothed by Gaussian filtering, and the Sigma is taken as 1.8 to obtain the final white blood cell nucleus image.

The beneficial effects of the present invention are as follows: the present invention divides the leukocyte image by combining the clustering method of the atlas theory and the active contour method of the difference theory, obtains the rough outline of the leukocyte nucleus through density clustering, and considers the characteristics of the leukocyte itself. The leukocytes are segmented by the evolution of the contour of the nucleus to the edge of the cytoplasm, and more accurate results are obtained.

Description of drawings

Fig. 1 shows the results before and after the processing of step 1 of the stained leukocyte image in Example 1 of the present invention; Fig. 1a is an image before preprocessing, Fig. 1b is an image after preprocessing, and Fig. 1c is an image after preprocessing;

Figure 2 shows the results before and after the processing of leukocyte nuclei by density peak clustering on leukocyte images in Example 1 of the present invention, Figure 2a shows the leukocyte nuclei images obtained after density clustering, Figure 2b shows binarization, morphological processing, Smoothed leukocyte nuclei images;

Figure 3 is the result of fine segmentation of the white blood cell image by the active contour method in Example 1 of the present invention, Figure 3a is the initial contour image of the cell nucleus, Figure 3b is the curve evolution result, and Figure 3c is the final segmentation result.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

The invention discloses a leukocyte segmentation method based on an active contour model, which consists of the following steps:

Step 1: Divide the RGB pixel space of the original stained white blood cell image into small cubes of equal length, calculate the pixel value at the center point in all the small cubes, that is, the predetermined pixel value, obtain the number of all pixel values in each small cube, and set the All pixel values in each small cube are assigned to predetermined pixel values, and a preprocessed image is obtained.

Since the density clustering of the original leukocyte image will face huge data processing problems, the mini-RGBCUBE method is used to preprocess the leukocyte image first to reduce the number of pixels.

Step 2: Use the density peak clustering segmentation method to perform density clustering on the preprocessed image to obtain the initial white blood cell image, and then use the morphological processing method to smooth the edge contour of the white blood cell nucleus in the initial white blood cell image to obtain the final white blood cell nucleus image.

In addition to the density peak clustering segmentation method in step 2, other segmentation methods can of course be used, such as K-Means (K-means), C-means and other segmentation methods, but the color consistency of the segmentation small area and the time complexity of the algorithm are comprehensively considered. The performance of the density peak clustering algorithm is relatively optimal.

Perform density clustering on the preprocessed color image in step 1, and use Euclidean distance metric to binarize the clustering results, with leukocyte nuclei as the foreground and the rest as the background; The edge contours of the nuclei were smoothed to obtain the final leukocyte nuclear image.

Among them, step 2 consists of the following steps:

Step 21: First, calculate the local density ρ and the relative distance δ of the preprocessed image pixel by pixel, and determine the cutoff distance d _c .

Among them, the local density ρ of density clustering is:

In the above formula, d _i,j represents the distance between pixel i and pixel j, which is calculated using Euclidean distance. d _cutoff represents the cutoff distance, that is, the effective density radius, which is the only parameter in the density peak clustering, and takes 0.5% of the maximum relative distance. Then ρ _i can represent the distribution of pixel i around the pixel, that is, the local density.

Among them, the relative distance δ of density clustering is:

Step 22: Arrange the product of the local density ρ of all pixel points and the relative distance δ in descending order, take the first N points as the cluster center point, and the cluster center N value is 2, indicating that the cluster structure is divided into two categories, one is the background In part, one type is foreground leukocyte nuclei.

Step 23: Classify and divide the remaining points, that is, pixels other than the N cluster center points, according to the cluster centers, and obtain the clustered image, that is, the white blood cell nucleus image, and the white blood cell nucleus image obtained by clustering is obtained. Binarization, the white blood cell nucleus is used as the foreground, and the rest is used as the background. The morphological processing method is used, and the small area is eliminated by morphological expansion and erosion operations; the foreground part is processed, and the edge contour of the white blood cell nucleus is obtained. That is, the image is smoothed by Gaussian filtering, and the Sigma is taken as 1.8 to obtain the final white blood cell nucleus image.

Step 3: The final leukocyte nuclear image is processed by the level set method to obtain the nucleus contour curve, the nucleus contour curve is used as the initial contour curve of the active contour model, and the local JS divergence is used to construct the local gray information expression as the edge stop function, so that the initial The contour curve is calculated step by step to obtain the leukocyte edge, thereby obtaining the final leukocyte image.

Among them, step 3 consists of the following steps:

Step 31: Use the level set method to process the final leukocyte nucleus image to obtain the nucleus contour curve, take the nucleus contour curve as the initial contour line, process the edge contour of the leukocyte nucleus obtained in step 24 through the level set method, set the leukocyte nucleus part to 1, and set the rest part to 1. Set to 0.

Step 32: Build a local JS divergence-driven edge stop function;

The edge stop function is:

Among them, I(x) represents the local image area, and C1 and _C2 are the average gray value of the local area divided into inner and outer parts by the contour curve, respectively. JS(p||q) represents the JS divergence information between the gray distributions of the inner and outer parts, and α is the local threshold, which is taken as 0.02 to measure the uniformity of the local gray.

Step 33: Using the edge stopping function to evolve the active contour curve according to the local gray level change of the white blood cells to obtain the final white blood cell image.

The invention uses the evolution of the active contour curve to perform fine segmentation of leukocytes, uses the nucleus contour curve as the initial contour curve of the active contour model, and uses the local JS divergence to construct a local gray information expression as the edge stop function to guide the evolution of the initial contour curve to white blood cells. edge, so as to obtain an accurate segmentation image of white blood cells.

The invention first preprocesses the original white blood cell image, then roughly divides the cell nucleus, and uses the rough segmentation contour to accurately segment the active contour model of the curve evolution. On the one hand, a two-step method is used to segment leukocytes according to the uneven staining of cytoplasm and nuclei in leukocyte images; The edge stopping function of the active contour model is obtained, and the white blood cells with uneven grayscale are more accurately segmented. Considering the staining characteristics of the leukocyte nuclei in the peripheral blood image, the leukocyte nuclei are obtained by clustering. In view of the uneven gray level of the leukocyte, the edge stop function is used to evolve the contour curve, and a more accurate leukocyte segmentation method is realized by combining the two aspects.

Example 1

Step 1: Preprocess the input original stained white blood cell color image, as shown in Figure 1a, using the RGB-CUBE method to reduce the number of local pixels, and divide the RGB pixel space (256*256*256) into small pixels with a side length of 16. Cube, define side length lenminicube=16.

Calculate the pixel value at the center point in all the small cubes, that is, the predetermined pixel value, obtain the number npixel of all pixel values in each small cube, assign all the pixel values in each small cube to the predetermined pixel value, and obtain the preprocessing image, thereby reducing the number of pixels in the original image.

Select the length of the small cube to be 2 ⁿ = 16 (n=4), the actual n can be taken as (0-9), but as the value of n does not become larger, the pixels in the small cube are attributed to each After the center of a small cube, the overall pixels of the image will decrease sharply, so that the original image is indistinguishable. When n=0, the length of the small cube is 1, and its center pixel is itself, which is equivalent to no change to the image; When n=9, the length of the small cube is 256, and its center pixel is (127, 127, 127), which is equivalent to turning the entire image into a grayscale value, as shown in Figure 1b. Therefore, after experiments, when n=4 is selected, both efficiency and accuracy can be taken into account. At the same time, the calculation method of the center point pixel of each small cube is:

表示向下取整符号，pixel表示原始图像的像素值。in,

Indicates the round-down symbol, and pixel represents the pixel value of the original image.

The RGBCUBE method is used to reduce the number of original pixels to facilitate subsequent clustering processing, thus obtaining a preprocessed image, as shown in Figure 1c.

Step 2: Perform density clustering on the preprocessed images in Step 1, and use the Euclidean distance metric to binarize the clustering results. The white blood cell nuclei are used as the foreground (white), and the rest are used as the background (black). Then, the morphological processing method is used to process the foreground part, and the edge contours of leukocyte nuclei in the initial leukocyte image are smoothed by means of connected domain marking to obtain the final leukocyte nucleus image.

For the preprocessed image in step 1, the number of pixels in each small cube npixel and the pixel value centerpoints of the center point of the small cube are obtained, and then the local density ρ is calculated for each center pixel value centerpoints, where the local density of the density clustering ρ is:

In the above formula, d _{i, j} represents the distance between pixel i and pixel j, which is calculated using Euclidean distance. d _cutoff represents the cutoff distance, that is, the effective density radius, which is the only parameter in the density peak clustering, and takes 0.5% of the maximum relative distance. The experimental test proves that: d _c takes 0.5% of the maximum relative distance to obtain better clustering results; then ρ _i can represent the distribution of pixel i around the pixel, that is, the local density.

Among them, the relative distance δ of density clustering is:

After calculating the local density and relative distance of each pixel point centerpoints, arrange the product of the two in descending order, and take the points whose first N is 2 as the cluster center. The nucleus part, so the number of clusters is selected as 2.

After the cluster center is found, the remaining pixel points are expanded to cluster, and they are assigned to the cluster of data points that are closest to it and whose local density is larger than it, and finally get the density peak clustering result, as shown in Figure 2a. For the results obtained after clustering, morphological dilation and erosion operations are used to eliminate small areas, and the kernel size is 5X5. Gaussian smoothing filtering, smoothing edges, and smoothing operations can also be performed on the morphologically processed image by using median filtering, mean filtering, etc., but using Gaussian filtering to smooth the image can preserve the overall grayscale distribution of the image. feature to obtain the final leukocyte nuclear image, as shown in Figure 2b.

Step 3: Active Contour Algorithm Precisely Segments WBCs

Taking the white blood cell nucleus contour curve as the initial level set, as shown in Figure 3a, the final white blood cell nucleus image obtained in step 2, the nucleus part is set to 1, the rest is set to 0, and then used as the initial level set function.

Calculate the gray mean values C ₁ and C ₂ of the inner and outer parts divided by the contour curve in the local area:

Among them, I is the local image, and H(φ) is the smooth form of the unit step function, which is used to calculate the inner and outer parts of the local area, specifically:

Among them, phi is the contour curve, eps is a minimum constant, and eps=1e-3 is taken.

Using the local grayscale image mean, an edge stopping function driven by the local Jensen-Shannon divergence is constructed:

In the formula, I(x) represents the local image area, C ₁ and C ₂ represent the average gray value of the local area divided into inner and outer parts by the contour curve, respectively, p and q represent the inner and outer parts divided by the contour curve, respectively. Partial grayscale distribution, JS(p, q) represents the JS divergence information between the grayscale distribution of the inner and outer parts, α is the local threshold, which is 0.02 to measure the uniformity of the local grayscale, α- If JS(p, q)>0, it means that the grayscale is uniform, otherwise, it means that the grayscale is not uniform, and the local window is selected as 5*5.

Using the constructed edge stop function to guide the evolution of the contour curve, it is assumed that the gray density of the target area in the image is greater than that of the background area (and vice versa), when JS(p, q)>α, that is, α-JS(p, q) <0, g _jspf <0, which means that the contour curve moves to the area with large grayscale changes, and finally is at the edge of the image; when JS(p, q)<α, that is, α-JS(p, q)>0, g _jspf > 0 means that the evolution curve is in the flat gray area of the image, and the gray value of each pixel is not much different. When the contour curve evolves, the symbol of g _jspf is continuously detected, and the contour curve keeps approaching the target edge and finally stops at the target boundary.

Through the constant change of the symbol of g _jspf , the contour curve finally converges at the cell edge. At this time, the symbol inside the contour curve is negative, and the symbol outside the contour curve is positive, so as to achieve the purpose of accurately segmenting white blood cells, as shown in Figure 3b and Figure 3c .

The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection of the present invention. within the range.

Claims (3)

1. A leukocyte segmentation method based on an active contour model is characterized by comprising the following steps:

step 1: dividing an RGB pixel space of an original stained leukocyte image into small cubes with equal length, calculating pixel values positioned at a central point in all the small cubes, namely predetermined pixel values, obtaining the number of all the pixel values in each small cube, assigning all the pixel values in each small cube to be the predetermined pixel values, and obtaining a preprocessed image,

step 2: performing density clustering on the preprocessed image by using a density peak value clustering segmentation method to obtain an initial white blood cell image, then performing smooth processing on the white cell nucleus edge contour in the initial white blood cell image by using a morphological processing method to obtain a final white cell nucleus image,

and step 3: and processing the final white cell nucleus image by using a level set method to obtain a white cell nucleus contour curve, taking the white cell nucleus contour curve as an initial contour curve of the active contour model, constructing a local gray information expression by using local JS divergence as an edge stop function, and gradually calculating the initial contour curve to obtain a white cell edge so as to obtain a final white cell image.

2. The method for segmenting white blood cells based on the active contour model according to claim 1, wherein the edge stop function in step 3 is:

wherein I (x) denotes a local image region, C₁And C₂The gray scale values are respectively the average gray scale value of the inner part and the outer part divided by the contour curve in the local area, p and q respectively represent the gray scale distribution of the inner part and the outer part divided by the contour curve, JS (p, q) represents JS divergence information between the gray scale distribution of the inner part and the outer part, alpha is a local threshold value, 0.02 is taken to measure the uniformity of the local gray scale, alpha-JS (p, q) > 0 represents that the gray scale is uniform, otherwise, the gray scale is non-uniform, and a local window is selected to be 5.

3. The method for segmenting white blood cells based on the active contour model according to claim 1, wherein the step 2 comprises the following steps:

step 21: firstly, calculating the local density rho and the relative distance delta pixel by pixel of a preprocessed image, and determiningFixed cut-off distance d_c；

Wherein, the local density rho of the density cluster is:

in the above formula, d_i，jRepresenting the distance between pixel i and pixel j, calculated using the Euclidean distance, d_cutoffRepresents the truncation distance, i.e. the effective density radius, which is the only parameter in the clustering of density peaks, taking 0.5% of the maximum relative distance, then ρ_iThe distribution condition of the pixel points i around the pixel points, namely the local density, can be represented;

wherein, the relative distance δ of the density cluster is:

step 22: arranging products of local density rho and relative distance delta of all pixel points in a descending order, taking the first N points as clustering center points, taking the N value of the clustering center as 2, and representing that clustering structures are divided into two types, one type is a background part, and the other type is a foreground white cell nucleus;

step 23: classifying and dividing the rest points, namely pixel points except the N clustering center points according to the clustering centers to obtain clustered images, namely white cell nucleus images, binarizing the clustered white cell nucleus images, taking the white cell nuclei as a foreground, taking the rest parts as a background, and eliminating small areas by using a morphological processing method and utilizing morphological expansion and corrosion operation; and processing the foreground part to obtain a white cell nucleus edge contour, and then smoothing the white cell nucleus edge contour, namely smoothing the edge of the image by Gaussian filtering, and taking 1.8 from Sigma to obtain a final white cell nucleus image.

Images