CN107730499A - A kind of leucocyte classification method based on nu SVMs - Google Patents

Abstract

The invention discloses a leukocyte classification method based on nu-support vector machine. First, the median filter method is used to preprocess the color blood microscopic image, and then the initial color space is mapped to the HLS color space to obtain a transformed image. Tone image, then use the grayscale image segmentation method based on nu-support vector machine to roughly segment the tone image, use layer-by-layer screening strategy and mathematical morphology method to detect all white blood cells, and then finely segment each white blood cell image , that is to complete the separation of nucleus, cytoplasm and background, extract the most representative 47 features for each white blood cell, nucleus and cytoplasm, and finally complete the classification of white blood cells with the help of nu-support vector machine. The invention can significantly improve the performance of the entire white blood cell automatic identification and counting system.

Description

A Leukocyte Classification Method Based on nu-Support Vector Machine

technical field

The invention relates to a leukocyte classification method based on a nu-support vector machine, belonging to the technical field of medical image processing.

Background technique

By examining the changes in the number and shape of various types of white blood cells in the blood, it can often provide valuable information for doctors to diagnose and help diagnose some diseases. The emergence of new medical branches such as quantitative cytology, molecular biology, and cellular immunology has made the requirement for rapid and accurate quantitative analysis of cells more urgent. However, it is time-consuming and labor-intensive for experts to inspect with the naked eye through a microscope, and the workload is very heavy, and the recognition error is greatly affected by subjective factors such as the experience and fatigue of the experts. With the rapid development of computer image processing technology, pattern recognition and neural network, using these advanced technologies to assist in the recognition and counting of blood cell morphology has become an inevitable trend in the development of hematology testing technology. Research at home and abroad has shown that white blood cell image segmentation, that is, the separation of nucleus, cytoplasm and background, is the most basic and critical link in the entire white blood cell automatic identification system, and its accuracy and stability directly affect the recognition accuracy and operation of the system. speed. The reason is that objective factors such as illumination and staining will cause the imaging quality of cell microscopic images to decline, and it is difficult to control. Therefore, the same white blood cell may appear different in color, background, and even particles under different external conditions. Sometimes leukocytes in microscopic images may be contaminated with stains due to poor handling. Coupled with the inconsistency of factors such as lighting and staining, it becomes more difficult to distinguish each other.

Contents of the invention

Purpose of the invention: Aiming at the problems existing in the prior art, the present invention provides a leukocyte classification method based on nu-support vector machine. The method significantly improves the performance of the entire white blood cell automatic identification and counting system, greatly reduces the labor intensity of doctors reading images, improves diagnostic accuracy, and is convenient for rapid and accurate quantitative analysis and research on cells.

Technical solution: In order to solve the above technical problems, the technical solution adopted in the present invention is:

A kind of leukocyte classification method based on nu-support vector machine, comprises the steps:

Step A, collecting color blood microscopic image data;

Step B, performing median filtering on the microscopic image data obtained in step A to obtain a median filtering image.

In step C, the median filter image obtained in step B is mapped to the HLS color space to obtain a hue image; the HLS (Hue, Lightness, Saturation) model is a commonly used visual color model.

In step D, the tone image obtained in step C is segmented using a grayscale image segmentation method based on nu-support vector machine to obtain a rough segmented image;

Step E, using the fuzzy cellular neural network (Fuzzy CellularNeural Network—FCNN) to detect the image of the white blood cell region in the coarsely segmented image obtained in step D;

Step F, for each white blood cell region image obtained in step E, use cluster analysis to determine the threshold value, combine threshold segmentation and binary morphology method to perform fine segmentation, and obtain a partial image of the nucleus, a partial image of the cytoplasm and a background image;

Step G, extracting the most representative 47 features from the partial image of the nucleus and the partial image of the cytoplasm obtained in the step F;

Step H, using the 47 features obtained in step G as an input vector, using nu-support vector machine to complete the identification and classification of white blood cells;

In step I, after the processing of all the white blood cell area images obtained in step E is completed, the final classification result of the image data obtained in step A is counted and output.

In step D, the process of the rough segmentation of the tone image is as follows:

Step D-1, constructing a histogram for the tone image obtained in step C;

Step D-2, using the nu-support vector machine to perform function fitting on the histogram obtained in step D-1, and find the support vector set;

Step D-3, adaptively select the threshold value in the support vector set found in step D-2, that is, select the support vector located near the inflection point of transition from negative value to positive value as the threshold value according to the first-order derivative information of the fitted curve;

Step D-4, performing threshold segmentation on the tone image obtained in step C by using the threshold obtained in step D-3.

In step G, the feature extraction process of the partial image of the nucleus and the partial image of the cytoplasm is as follows:

Step G-1, extracting 7 morphological characteristic parameters from the partial image of the nucleus and the partial image of the cytoplasm obtained in step F, to quantitatively describe the number of leaves, shape, size, and regularity of the white blood cell and nucleus;

Step G-2, extracting 24 color feature parameters from the partial nucleus image and partial cytoplasm image obtained in step F, to quantitatively describe the brightness, hue and saturation of white blood cells, nucleus and cytoplasm;

In step G-3, 16 statistical texture parameters are extracted from the partial nucleus image obtained in step F, so as to quantitatively describe the texture features of the nucleus.

Beneficial effect: compared with the prior art, the white blood cell automatic recognition method based on nu-support vector machine provided by the present invention utilizes the tone information of blood microscopic image features, and is completed by the grayscale image segmentation method based on nu-support vector machine Coarse segmentation of the tone image; Detect all white blood cells with the help of FCNN; use cluster analysis to determine the threshold, combine threshold segmentation and binary morphology methods to fine-tune the local images containing a single white blood cell; the local images obtained in the previous step On the basis, the most representative white blood cell features are extracted, including 47 features in three categories such as shape, color and texture; the recognition and classification of white blood cells are completed by using nu-support vector machine. The classification and recognition effect of this method is ideal, with high stability and good robustness. Provide valuable information for doctors to diagnose, and help to conduct rapid and accurate quantitative analysis of cells.

Description of drawings

Fig. 1 is the flowchart of the white blood cell classification method based on nu-support vector machine of the present invention;

Fig. 2 is a schematic diagram of calculation of nuclei concavity parameters.

detailed description

Below in conjunction with specific embodiment, further illustrate the present invention, should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention, after having read the present invention, those skilled in the art will understand various equivalent forms of the present invention All modifications fall within the scope defined by the appended claims of the present application.

As shown in Figure 1, the white blood cell classification method based on nu-support vector machine of the present invention, its steps are as follows:

Step 101, collecting color blood microscopic image data;

Step 102, preprocessing the microscopic input image data obtained in step 101;

Step 103, performing HLS color space conversion on the preprocessed image obtained in step 102, to obtain a tone image;

Step 104, for the tone image obtained in step 103, use a grayscale image segmentation method based on nu-support vector machine to segment white blood cells to obtain a rough segmented image;

Step 105, using FCNN to detect all white blood cell region images for the coarsely segmented image obtained in step 104;

Step 106, for each white blood cell region image obtained in step 105, use cluster analysis to determine the threshold value, and combine threshold segmentation and binary morphology method to perform fine segmentation to obtain a partial image of the nucleus, a partial image of the cytoplasm, and a background image;

Step 107, extracting the most representative 47 features from the partial image of the nucleus and partial image of the cytoplasm obtained in step 106;

Step 108, using the 47 features obtained in step 107 as an input vector, using nu-support vector machine to complete the identification and classification of white blood cells;

Step 109, output statistical results after all the partial images of white blood cells have been processed.

The leukocyte classification method based on nu-support vector machine of the present invention will be described in detail below.

1. Input microscopic image

Input a color microscopic image of blood n is the number of pixels in the image, and I( _xi ) is the pixel vector of pixel point _xi .

2. Pretreatment

Here median filtering is used.

3. Color space conversion

Through a large number of comparative experiments, we found that the hue component in HLS is not sensitive to changes in illumination, and it can maintain good consistency for cell microscopic images obtained with different color stains, which is helpful for subsequent processing. So we map the input image from the RGB color space to the HLS color space. The conversion method of the hue component is as follows:

Let the range of each component of the color value (r, g, b) of the RGB space be [0,1,…,255], let

v'=max(r,g,b), u'=min(r,g,b) and

h=60h'

h is the hue component value.

The tone image correction formula is:

Wherein, H represents the corrected tone value, and the corrected value h _t is the corresponding value of the peak in the histogram of the tone image.

Gray-scale linear transformation is performed on the tone image to enhance the contrast effect.

4. Rough segmentation

Firstly, the histogram of the tone image is generated, and the items with a value of zero are deleted, and the remaining non-zero items form the final histogram.

Treat the histogram as a functional relationship, and find a sparse set of support vectors with the help of nu-SVM. For nu-support vector regression, since each component w _i in the weight vector w corresponds to an adjustable regularization parameter λ _i , this allows us to obtain a sparser solution, that is, most components of w are 0. And we call the hue value h _i corresponding to the non-zero component w _i a support vector. Support vector sets have two nice properties. First, it can characterize the original histogram well. The support vectors are often located near the local maximum and minimum points. For image segmentation, it is sufficient to select several of them as thresholds to meet the segmentation requirements. Second, the support vector set is usually very sparse, accounting for only a small fraction of the total number of samples n. The sparseness of the number of support vectors allows us to select the segmentation threshold only from a small number of support vectors to ensure the efficiency of segmentation.

When the number of elements in the support vector set is more than expected, further screening is required. Specifically, according to the first-order derivative information of the fitting curve, the support vector located near the inflection point of transition from negative to positive is selected as the threshold. Use the obtained threshold to threshold the tone image.

5. Detection of white blood cells

In blood microscopic images, in addition to white blood cells, there are also some secondary image areas, such as red blood cells, thrombus cells, and stains. They are distinct from white blood cells in color and shape. The gray value of the white blood cell area is generally smaller than that of the red blood cell area, and because the white blood cell nucleus is embedded in the connected cytoplasmic area, the white blood cells are clustered; after the rough segmentation process, the red blood cells generally only have the edge part, which is ring-shaped. So this step is to eliminate these interferences and extract white blood cells with complete margins.

FCNNs are useful tools for solving image processing problems. FCNN is able to take into account hue information and structural knowledge in the same process. This is one of the reasons we consider adopting FCNN. Another important reason is that FCNN has unique advantages in real-time image processing and is easy to implement in hardware, which will undoubtedly be of great help in improving the processing speed of the system. Here, FCNN is used to achieve morphological grayscale reconstruction, and the parameter template is used as follows:

B = 0, A _fmin = no need to define, A _fmax = no need to define, B _fmin = no need to define, B _fmax - = 0, R _x = 1, I = 0, u = image after rough segmentation, x ₀ = arbitrary, y = tone image;

After such FCNN function, the interference of residual red blood cells, stains and blood coagulation cells in the image is eliminated at one time, and the effect is very good. Even in the case that the gray value of part of the red blood cell area is similar to that of the white blood cell cytoplasm area, the image of the white blood cell area can still be obtained well.

Calculate the first-order moment of equal gray value for each white blood cell area, determine the coordinate position of each cell, and take it as the center, set the window adaptively according to the maximum radial size of the white blood cell, and restore the tone image in the window, so that it can be extracted at one time Multiple images of individual leukocyte areas in the field of view.

6. Regional image segmentation

For the image of a single white blood cell area, cluster analysis method was used to determine the thresholds Tn and Tc to achieve three-valued segmentation. Tn represents the threshold between the nucleus and the cytoplasm, and Tc represents the threshold between the cytoplasm and the background. The advantage of this method is that the total variance within the class is taken as the criterion, and there is always a minimum value in the range of Tn and Tc, that is, the optimal threshold can always be given. Tn and Tc that minimize the total variance within the class are the optimal thresholds. At this time, the inter-class variance of the gray value between the nucleus, cytoplasm and background reaches the maximum.

Perform r erosion on the binary image after removing the background, and then expand d times with the residue as the sprout, so as to obtain a binary image that removes the interference of scattered noise, generally r is less than 5, and d is less than 5; the shape of the region is used Factors (area and circularity) constitute a feature function for discrimination and exclusion to detect the nucleus area. What remains is the cytoplasmic region. Finally, the partial image of the nucleus, the partial image of the cytoplasm and the background image are obtained.

7. Feature extraction

Feature extraction is a quantitative description of cells, which plays a very important role in the automatic classification of cells and directly affects the recognition rate of the classification system. Generally, the following two types of features can be extracted for identification: mathematical model features and structural features. For the method of extracting image features with mathematical models, the key to classification and recognition is the extraction and selection of features. Whether the feature selection is appropriate will directly affect the effect of classification and recognition. For white blood cell images, there are many features that can be extracted, and the methods are also flexible and diverse. The key is to find the most effective invariant feature parameters based on the separability of categories. That is, those most representative attributes should be selected as features. Under the guidance of clinical cytopathology experts, on the basis of referring to the cell atlas and observing a large number of actual cell images, 47 most representative parameters were selectively extracted from many features, and the mathematical model of the corresponding features was established. for quantitative analysis by computer.

(1) Morphological characteristic parameters

They are quantitative descriptions of the number of lobes, shape, size, and outline of cells and nuclei.

(1a) Cell area G ₁ = total number of pixels in the partial image of the cell.

(1b) Cytoplasm to cell area ratio G ₂ =total number of pixels in the partial image of cytoplasm/total number of pixels in the partial image of cells. _G2 is smaller for lymphocytes and larger for monocytes.

(1c) Cell circularity G ₃ =the square of the number of pixels of the cell outline/(4π×the total number of pixels in the partial image of the cell). It is an important characteristic parameter for distinguishing lymphocytes from other types of cells. For lymphocytes, the value is close to 1; while neutrophils and monocytes are the smallest.

(1d) The number of nuclear lobes of cells G ₄ = the number of nuclear lobes. This is an important characteristic parameter that differentiates the neutrophil segmented granulocytes from other types of cells. For neutral lobulated nucleocytes, G ₄ is between 2 and 5; for neutral rod-shaped granulocytes, monocytes, and lymphocytes, G ₄ is 1; for eosinophils and basophils, G ₄ is less than 3.

(1e) Nucleus circularity G ₅ =square of the number of pixels of the nucleus outline/(4π×the total number of pixels in the partial image of the nucleus), meaning the same as G ₃ . G ₅ of lymphocytes is close to 1; while neutrophils and monocytes are the smallest.

(1f) Elongation of the nucleus. To describe the elongation of neutrophil rod nuclei, a measure of nuclear elongation was defined.

G ₆ =D _max /D _min

Among them, D _max and D _min represent the maximum value and the minimum value of the projection of the local image of the cell nucleus in various directions, respectively. This is an important feature to distinguish neutrophils, lymphocytes, and monocytes, and it is the largest for neutrophils _G6 .

(1g) Concavity of the nucleus. Since the nuclei of monocytes are kidney-shaped, it is necessary to give a measure of concavity. G ₇ =1-ρ _i max(θ ₁ ,θ ₂ ), where ρ _i =1/180°, the algorithm is explained in conjunction with Fig. 2 as follows: first find the symmetry axis AB of the local image of the cell nucleus. If the axis of symmetry does not exist, the axis with the smallest symmetry difference is used as the approximate axis of symmetry. Then find two points C and D, and make their tangent line perpendicular to the tangent line of point A. If C is not unique, take the median value. Then set points G and H so that their tangents are parallel to the tangents of point A. Next, determine F and E, so that Determine I, J, so that Finally, calculate the included angle θ ₁ between the tangent line at point E and the tangent line at point F, and then calculate the included angle θ ₂ between the tangent line at point I and the tangent line at point J.

This type of feature is more intuitive and easy to find and extract. It is best for distinguishing typical white blood cells with large morphological differences, such as lobulated granulocytes, rod-shaped cells, and lymphocytes, but it is powerless for distinguishing granulosa cells, and the effect is poor. So other types of features must also be extracted.

(2) Color characteristic parameters

The brightness of different types of white blood cells is different, which is reflected in the corresponding patterns on the histogram of the cell brightness image, such as gray scale deviation, number of peaks and valleys, peak size, etc. Hue and saturation have similar characteristics. Therefore, its characteristics can be described by color characteristic parameters. We extract the following 8 parameters from the histograms of the cell brightness image, hue image and saturation image respectively, a total of 24 color features: the average value of the cytoplasm; the variance of the cytoplasm; the average value of the nucleus; the variance of the nucleus; the average value of the cell ; cell variance; nucleoplasmic integral ratio; the ratio of the range of variation between cells and nuclei.

(3) Texture feature parameters

Texture features play an important role in recognition because they contain important information about the arrangement of cell tissue surface structures. Compared with other class features, it can better reflect the macroscopic and microstructural properties of cell images. The following are three methods suitable for texture analysis of leukocyte images, and we extracted 16 statistical texture parameters from three transformation matrices. They are all extracted from the partial image of the nucleus. The three image transformation matrices are defined as follows:

(3a) Gray-level variance correlation matrix: The matrix element is defined as the probability that the local variance u of the δ neighborhood of a pixel in the image and the local variance v of the δ neighborhood of a pixel with a distance of d in the θ direction co-occur in the image . The advantage of this array is that it overcomes the disadvantage that the feature is sensitive to gray scale. It is not affected by the depth of cell staining and the image input lighting conditions. It is only related to the local variance of the image and has nothing to do with the absolute value of gray scale. The local variance reflects the rate of change of the local grayscale. If the variance is large, it means that the local grayscale is uneven and the texture is fine; on the contrary, if the variance is small, it means that the texture is coarse. There are few but larger blue-black granules in basophils, which often cover the nuclei and present a coarse texture; for monocytes and lymphocytes, the gray scale of the area is relatively uniform, and the fine texture appears; eosinophils The cytoplasm is full of transparent dense small particles between the two. In order to reflect the differences in these textures, 5 features of angle second moment, contrast moment, entropy, contrast and correlation coefficient were extracted from the normalized matrix. In order to extract the rotation invariant, we take the mean value of the feature values in the four directions of 0°, 45°, 90°, and 135° to represent the five texture features.

(3b) Gray-level variance gradient correlation matrix: Matrix elements are defined as pairs of image points that co-occur with a certain gray-level variance value and a certain gradient value in the normalized gray-level variance image and the normalized gradient image number. The gradient image is obtained by applying the gradient operator to the normalized gray variance image. The characteristic of the gray-level variance gradient correlation matrix is that it reflects the image gray level and image structure information intensively, and has nothing to do with the absolute value of the gray level. For images with coarse textures, such as larger particles in basophil images, the elements in the matrix are distributed close to the gray axis, while for fine textures, such as monocytes and neutrophils images, the elements in the matrix are distributed away from the gray scale. The degree axis spreads out along the direction of the gradient axis. We extracted 7 texture features from the normalized matrix, large (small) gradient dominance, gray level (gradient) distribution non-uniformity, entropy and contrast.

(3c) Neighbor gray-level correlation matrix: The features extracted from this matrix are independent of the spatial rotation of the image and the linear transformation of the gray value, which is very attractive in the actual identification of cells. The elements in the matrix are defined as the probability of occurrence of pixels whose gray values differ by no more than a among all adjacent pixels whose gray value is k and whose distance is less than d in the image. From the normalized matrix, four rotation-invariant features are extracted, namely large and small weights, numerical uniformity and second-order moments.

Texture features reflect the properties of granules in the nucleus, such as granule size, distribution density, and nuclear staining structure. The distinction between eosinophils, basophils, and neutrophils in leukocytes mainly depends on these features.

8. Classification and identification

The nu-support vector machine was used for quantitative analysis, and the 47-dimensional feature vector obtained in the previous step was used as the input vector to make a judgment on the type of white blood cells to be identified.

9. Statistical results and output

Count the percentage of various types of white blood cells in the blood microscopic image, and display or print the analysis data results.

Through the foregoing embodiments, it can be seen that the present invention has the following advantages:

(1) This method uses the grayscale image segmentation method based on nu-support vector machine to complete the rough segmentation of tone image, mainly by introducing nu-support vector machine, and obtains a limited sparse support vector set while fitting, and then directly Filter out the desired segmentation threshold from it. This method is suitable for color microscopic image segmentation, and can effectively overcome the interference of objective factors such as illumination and staining. The recognition accuracy of the system lays a solid foundation.

(2) According to the experience of clinical cytologists, the present invention extracts a total of 47 characteristic parameters of three types of leukocytes that are far more than human eyes can distinguish, and uses nu-support vector machine to realize the automatic classification of six types of leukocytes, and the classification effect is ideal , high stability and good robustness.

Claims (6)

1. a leukocyte classification method based on nu-support vector machine, is characterized in that, comprises the steps: Step A, collecting color blood microscopic image data; Step B, performing median filtering on the microscopic image data obtained in step A to obtain a median filtering image. Step C, mapping the median filter image obtained in step B to the HLS color space to obtain a tone image; In step D, the tone image obtained in step C is segmented using a grayscale image segmentation method based on nu-support vector machine to obtain a rough segmented image; Step E, using the fuzzy cellular neural network (Fuzzy Cellular Neural Network—FCNN) to detect the image of the white blood cell region in the coarsely segmented image obtained in step D; Step F, for each white blood cell region image obtained in step E, use cluster analysis to determine the threshold value, combine threshold segmentation and binary morphology method to perform fine segmentation, and obtain a partial image of the nucleus, a partial image of the cytoplasm and a background image; Step G, extracting the most representative 47 features from the partial image of the nucleus and the partial image of the cytoplasm obtained in the step F; Step H, using the 47 features obtained in step G as an input vector, using nu-support vector machine to complete the identification and classification of white blood cells; In step I, after the processing of all the white blood cell area images obtained in step E is completed, the final classification result of the image data obtained in step A is counted and output. 2. a kind of leukocyte classification method based on nu-support vector machine according to claim 1, is characterized in that, in step D, the process of described tone image rough segmentation is as follows: Step D-1, constructing a histogram for the tone image obtained in step C; Step D-2, using the nu-support vector machine to perform function fitting on the histogram obtained in step D-1, and find the support vector set; Step D-3, adaptively select the threshold value in the support vector set found in step D-2, that is, select the support vector located near the inflection point of transition from negative value to positive value as the threshold value according to the first-order derivative information of the fitted curve; Step D-4, performing threshold segmentation on the tone image obtained in step C by using the threshold obtained in step D-3. 3. a kind of leukocyte classification method based on nu-support vector machine according to claim 1, is characterized in that, in step G, the process of described nucleus local image and cytoplasm local image feature extraction is as follows: Step G-1, extracting 7 morphological characteristic parameters from the partial image of the nucleus and the partial image of the cytoplasm obtained in step F, to quantitatively describe the number of leaves, shape, size, and regularity of the white blood cell and nucleus; Step G-2, extracting 24 color feature parameters from the partial nucleus image and partial cytoplasm image obtained in step F, to quantitatively describe the brightness, hue and saturation of white blood cells, nucleus and cytoplasm; In step G-3, 16 statistical texture parameters are extracted from the partial nucleus image obtained in step F, so as to quantitatively describe the texture features of the nucleus. 4. a kind of leukocyte classification method based on nu-support vector machine according to claim 2, it is characterized in that, generate the histogram of tone image, delete the item wherein value is zero, remaining non-zero item forms final histogram; Consider the histogram as a functional relationship, and use the nu-support vector machine to find a sparse support vector set; when the number of elements in the support vector set is more than expected, further screening is required, according to the first-order derivative information of the fitting curve , select the support vector located near the inflection point of transition from negative value to positive value as the threshold. 5. a kind of leukocyte classification method based on nu-support vector machine according to claim 3, is characterized in that, morphological feature parameter comprises cell area, cytoplasm and cell area ratio, cell circularity, the nuclei leaf number of cell , nucleus circularity, nucleus elongation and nucleus concavity. 6. a kind of leukocyte classification method based on nu-support vector machine according to claim 3, is characterized in that, has extracted 16 statistical texture parameters from three transformation matrices, and three image transformation matrices are defined as follows: (3a) Gray-level variance correlation matrix: The matrix element is defined as the probability that the local variance u of the delta neighborhood of a pixel in the image and the local variance v of the delta neighborhood of a pixel with a distance of d in the θ direction co-occur in the image ; (3b) Gray-level variance gradient correlation matrix: Matrix elements are defined as pairs of image points that co-occur with a certain gray-level variance value and a certain gradient value in the normalized gray-level variance image and the normalized gradient image number; (3c) Neighbor gray correlation matrix.