KR20110126008A - A method for leukocyte segmentation using texture and non-parametric probability model - Google Patents

Abstract

PURPOSE: A leukocyte cleavage method by non-parametric probability and texture information is provided to shorten cleavage time and to improve cleavage accuracy. CONSTITUTION: A leukocyte cleavage method by non-parametric probability and texture information comprises: a step of extracting chroma and brightness of leukocyte region(S100); a step of generating probability model for the leukocyte nuclear region using the extracted chroma and brightness(S200); a step of generating probability map and separating a sub-image from the probability map(S300); a step of performing domain clustering using mean-shift for the separated sub-image(S400); a step of performing combination of each cluster domain of the sub-image(S600); and a step of determining the cluster domain into finally cleaved leukocyte nuclear region(S700).

Description

White blood cell segmentation using nonparametric probability model and texture information {A METHOD FOR LEUKOCYTE SEGMENTATION USING TEXTURE AND NON-PARAMETRIC PROBABILITY MODEL}

The present invention relates to a method for segmenting white blood cells using a nonparametric probability model and texture information. In particular, a nonparametric method for automatically partitioning the nucleus and cytoplasm of leukocytes from blood cell images using probability maps and texture components created using nonparametric probability models. Leukocyte segmentation using a probability model and texture information.

The analysis of the number of leukocytes by type in the blood cell image plays an important role in identifying the patient's health and predicting various blood-related diseases early. However, the conventional white blood cell differential count method consumes a lot of time because the analyst must manually count the number of leukocytes by type, and there is a problem in that a very subjective result can be obtained according to the analyst. In order to solve such a problem, an automated research that can objectively and efficiently perform leukocyte differentiation coefficient using an image is being conducted.

Yampri performed the division of white blood cells using the RGB color space. In G channel, the automatic binarization algorithm was used to divide the nucleus and the adaptive active contour algorithm was used to segment the cytoplasm. However, this method has a problem that the accuracy can vary depending on the degree of staining of the nucleus.

Dorini used a Scale-Space Toggle Operator to simplify imaging of leukocyte cells converted to gray levels, and then applied a Watershed algorithm to detect nuclei. In the cytoplasm, after binarization of the gray level image, the cytoplasm was divided by setting a structural element in the opening operation to a disk shape larger than a normal red blood cell. However, this method is difficult to divide the exact shape of the cytoplasm because only the approximate shape of the cytoplasm containing the nucleus can be obtained.

Sadeghian proposed a technique to split the nucleus and cytoplasm of leukocytes using a GVF snake and an automated binarization algorithm. When the user directly separates the sub-image from the input blood cell image, the sub-image is transformed into a gray level image, and the Canny Edge algorithm is applied, and the nuclear region is segmented by applying a GVF snake to the edge result image. . In addition, the cytoplasmic region was divided by applying a binary segmentation algorithm to the regions other than the nuclear region. However, this method requires the user to separate the sub-images manually, and if the nuclei of leukocytes are multinucleated, they cannot find all the nuclei, and simply binarizing gray level images excludes red blood cells and other blood cells around the leukocytes. There is a problem that can not be.

As such, existing studies do not use only the binarization algorithm or consider the multinuclear case, resulting in poor accuracy. In addition, the algorithm is not only applied to the whole blood cell image as well as the region in which leukocytes exist, or the separation of sub-images, which are regions in which leukocytes exist, is not effective, resulting in poor efficiency.

The present invention has been proposed to solve the above problems of the proposed methods, to create a probabilistic model of the saturation, brightness values of the white blood cell nuclei, to divide the white blood cell nuclei from blood cell images, and to divide the white blood cell nuclei and blood cell images. The purpose of the present invention is to provide a non-parametric probability model and a leukocyte segmentation method using texture information that can shorten the segmentation time and improve segmentation accuracy by dividing the cytoplasm by using the wavelet energy value in.

In addition, the present invention, non-parametric probability model and texture information that can increase efficiency in terms of time and accuracy by automatically selecting a sub-image that is a candidate leukocyte candidate region predicted to exist in the blood cell image and performing a segmentation process only for this Another object of the present invention is to provide a leukocyte division method.

According to a feature of the present invention for achieving the above object, a white blood cell division method using a non-parametric probability model and texture information,

(1) extracting saturation and brightness values from the leukocyte nucleus region divided in blood;

(2) generating a probabilistic model for the leukocyte nucleus region using the extracted chroma and brightness values;

(3) generating a probability map representing a probability of leukocyte nuclei for each pixel in the blood cell image by using the generated probability model;

(4) separating a sub-image which is a leukocyte candidate region from the generated probability map;

(5) performing region clustering on the separated sub-images using mean-shift;

(6) performing merging between respective cluster regions of the separated sub-images using a combined map combining a distance map and the generated probability map; And

(7) determining the cluster region in which the merging is completed as the final divided leukocyte nucleus region.

Preferably,

(8) detecting the leukocyte cytoplasmic region using the wavelet energy map and the finally divided leukocyte nuclear region.

Preferably, in step (2),

The probability is constructed by constructing a histogram, a non-oarametric method, using the extracted saturation and brightness values, and estimating probability density by applying a Gaussian Smooth-Kernel of the Parzen-window. You can create a model.

Preferably, the step (3)

(3a) converting the blood cell image from a red green blue (RGB) color space to a hues saturation lightness (HSL) color space;

(3b) mapping the probability value corresponding to each pixel of the blood cell image to the pixel brightness value based on the generated probability model; And

(3c) generating the probability map from the blood cell image by using the pixel brightness value corresponding to each pixel.

Preferably, step (4),

(4a) binarizing the generated probability map by applying an automatic binarization algorithm;

(4b) removing noise by performing a morphological opening operation on the generated probability map; And

(4c) separating the candidate regions remaining after the morphological opening operation into sub-images.

More preferably, in step (4c),

In order to include the cytoplasm around the nucleus, a region including the candidate region and 1.1 to 1.5 times larger in height and width than the candidate region may be separated into the sub image.

Preferably, in step (5),

Considering that the color of the leukocyte nucleus is different from the surrounding area, the mean-shift is performed by setting the kernel size for the image space to 15, the kernel size for the color space to 10, and the minimum region size per cluster to 20, or In order to ensure that the leukocyte cytoplasm is clearly clustered without being confused with the surrounding red blood cells, the mean-shift is performed by setting the kernel size for the image space to 10, the kernel size to 8 for the color space, and the minimum region size per cluster to 12. can do.

Preferably, step (6) is

(6a) obtaining a distance map from the separated sub-images, and generating a combined map by combining the obtained distance map with the generated probability map;

(6b) detecting a seed region of a leukocyte nucleus in the separated sub-image using the generated binding map; And

(6c) merging between cluster regions according to a predetermined condition based on the detected seed region.

More preferably, the step (6b),

(6ba) calculating an average of the combined map values for each cluster region using the generated combined map and the number of pixels of the cluster region;

(6bb) selecting the cluster region as a seed region if the average of the calculated combined map values and the number of pixels of the cluster region are equal to or greater than a predetermined value;

(6bc) repeating steps 6ba and 6bb until inspection is complete for all cluster regions; And

(6bd) if none of the seed regions is selected, selecting a region having the largest combined map value as a seed region,

The step (6c) is,

(6ca) merging the abutted cluster region with the seed region if the saturation and lightness of the cluster region abutting the detected seed region are greater than or equal to a predetermined value; And

(6cb) it may include repeating step (6ca) until the inspection is complete for all cluster regions.

More preferably, step (8) is

(8a) combining vertical, horizontal and diagonal components obtained through wavelet transform in the separated sub-image into one single wavelet energy map;

(8b) classifying the cluster area into a background area set if an edge length of the cluster area that meets the edge of the separated sub-image is equal to or greater than a predetermined value;

(8c) calculating an average of wavelet energy map values for each cluster region by using the wavelet energy map and the number of pixels of the cluster region;

(8d) selecting the cluster region as a cytoplasmic region if the average of the wavelet energy map values is equal to or greater than a predetermined value;

(8e) repeating steps (8b), (8c), (8d) until the inspection is completed for all cluster regions;

(8f) merging the abutted cluster region into the cytoplasmic region if the mean of the wavelet energy map values for the cluster region abutting the cytoplasmic region is greater than or equal to a predetermined value; And

(8g) repeating step (8f) until the inspection is completed for all cluster regions.

According to the leukocyte segmentation method using the nonparametric probability model and texture information proposed in the present invention, a leukocyte nucleus is divided from blood cell images by creating a probabilistic model of saturation and brightness values of the leukocyte nuclei, and the divided leukocyte nuclei and blood cell images. By dividing the cytoplasm by using the wavelet energy value in, the segmentation time can be shortened and the segmentation accuracy can be improved.

In addition, the leukocyte segmentation method using the non-parametric probability model and texture information according to the present invention, by automatically selecting a sub-image that is a candidate leukocyte candidate region in the blood cell image, and performing the segmentation process only, time and accuracy In terms of efficiency can be increased.

1 is a flow chart of a white blood cell division method using a non-parametric probability model and texture information according to an embodiment of the present invention.
2 is a diagram illustrating a probability density function estimation result obtained by applying a Gaussian smooth kernel in a leukocyte division method using a nonparametric probability model and texture information according to an embodiment of the present invention.
3 is a detailed flowchart of step S300 of the leukocyte division method using a non-parametric probability model and texture information according to an embodiment of the present invention.
4 is a detailed flowchart of step S400 of the leukocyte division method using a non-parametric probability model and texture information according to an embodiment of the present invention.
5 is a diagram illustrating a sub-image selected according to a leukocyte division method using a nonparametric probability model and texture information according to an embodiment of the present invention.
FIG. 6 is a diagram illustrating a region clustering using Mean-Shift for a sub image according to a non-parametric probability model and a leukocyte segmentation method using texture information according to an embodiment of the present invention. FIG.
7 is a detailed flowchart of step S600 of the leukocyte division method using a non-parametric probability model and texture information according to an embodiment of the present invention.
8 is a diagram illustrating a combined map of a leukocyte division method using a nonparametric probability model and texture information according to an embodiment of the present invention.
FIG. 9 is a detailed flowchart of steps S620 and S630 of a white blood cell division method using a non-parametric probability model and texture information according to an embodiment of the present invention. FIG.
10 is a view showing the results of merging leukocyte nuclear regions in a leukocyte segmentation method using a nonparametric probability model and texture information according to an embodiment of the present invention.
11 is a detailed flowchart of step S800 of the leukocyte division method using a non-parametric probability model and texture information according to an embodiment of the present invention.
FIG. 12 is a diagram illustrating a wavelet energy map result in a white blood cell division method using a nonparametric probability model and texture information according to an embodiment of the present invention. FIG.
FIG. 13 is a view showing a leukocyte cytoplasmic region merging result in a leukocyte division method using a nonparametric probability model and texture information according to an embodiment of the present invention. FIG.
14 is a diagram illustrating a partitioning performance of a leukocyte division method using a nonparametric probability model and texture information according to an embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. However, in describing the preferred embodiment of the present invention in detail, if it is determined that the detailed description of the related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. The same or similar reference numerals are used throughout the drawings for portions having similar functions and functions.

In addition, throughout the specification, when a part is 'connected' to another part, it is not only 'directly connected' but also 'indirectly connected' with another element in between. Include. In addition, the term 'comprising' of an element means that the element may further include other elements, not to exclude other elements unless specifically stated otherwise.

1 is a flowchart of a leukocyte division method using a nonparametric probability model and texture information according to an embodiment of the present invention. As shown in FIG. 1, the leukocyte segmentation method using a nonparametric probability model and texture information according to an embodiment of the present invention includes extracting saturation and brightness values from the leukocyte nucleus region divided from blood (S100), Generating a probabilistic model for the leukocyte nucleus region using the extracted saturation and brightness values (S200), and generating a probability map representing the probability of leukocyte nuclei for each pixel in the blood cell image using the generated probability model. Step S300, separating a sub-image that is a leukocyte candidate region from the generated probability map, in step S400, and performing region clustering on the separated sub-image using mean-shift. ), Performing a merge between the cluster regions of the sub-image using the combined map combining the distance map and the probability map (S600), and finally, the merged cluster regions are finally divided. The method may include determining the leukocyte nuclear region (S700), and detecting the leukocyte cytoplasmic region (S800) using the wavelet energy map and the finally divided leukocyte nuclear region.

In step S100, saturation and lightness values are extracted from the leukocyte nucleus region divided in the blood. In blood cell imaging, white blood cells have a special color by staining material. This is distinguished from other cells such as erythrocytes and platelets, and can be used to generate a probabilistic model for the color of white blood cells to distinguish between white blood cells and non-white blood cells. In particular, since the nucleus of leukocytes has a higher saturation and a lower brightness than other regions, a probabilistic model can be generated for the saturation and brightness.

In step S200, a probability model for the leukocyte nucleus region is generated using the saturation and lightness values extracted in step S100. The step S200 will be described in detail with reference to FIG. 2.

2 is a diagram illustrating a probability density function estimation result obtained by applying a Gaussian smooth kernel in a leukocyte segmentation method using a nonparametric probability model and texture information according to an embodiment of the present invention. In general, the nuclear region of leukocytes is more prominent in brightness and saturation than in the surrounding region, but the distribution of each data is not constant and does not show a commonly used Gaussian distribution. Therefore, as shown in Figure 2, in step S200 of the white blood cell segmentation method using a non-parametric probability model and texture information according to an embodiment of the present invention, a histogram that is a non-oarametric method using saturation, brightness values The probability model can be generated more precisely by estimating the probability density by applying a Gaussian Smooth Kernel among the Pozen-windows.

In step S300, a probability map representing a probability of a leukocyte nucleus is generated for each pixel in the blood cell image using the probability model generated in step S200. In general, blood cell images include various blood cells such as red blood cells and platelets in addition to leukocytes, and leukocytes usually do not occupy more than 30% of the entire image. Therefore, performing leukocyte segmentation on all pixels of an image is very inefficient and affects the segmentation accuracy. Therefore, segmentation process should be performed only on the region where the leukocytes can be predicted to exist. Therefore, the leukocyte candidate region is divided into sub-image units, and a segmentation process is performed for each sub-image, thereby generating a probability map in order to increase efficiency in terms of time and accuracy. The step S300 will be described in detail with reference to FIG. 3.

3 is a detailed flowchart of step S300 of the leukocyte division method using a non-parametric probability model and texture information according to an embodiment of the present invention. As shown in FIG. 3, step S300 of the white blood cell segmentation method using the non-parametric probability model and the texture information according to an embodiment of the present invention, the blood cell image is HSL (Hue Saturation) in a red green blue (RGB) color space. Lightness) the step of converting to the color space (S310), based on the probability model, the step of mapping the corresponding probability value for each pixel of the blood cell image to the pixel brightness value of each pixel (S320), and corresponding to each pixel In operation S330, a probability map may be generated from the blood cell image by using the pixel brightness value.

In step S310, the blood cell image is converted from the RGB color space to the HSL color space. Saturation and lightness are important for segmenting the white blood cell nuclei, so the saturation and lightness values need to be converted into an HSL color space that can be clearly identified.

In operation S320, a probability value corresponding to each pixel of the blood cell image is mapped to each pixel with a pixel brightness value based on the probability model. After normalizing the saturation and brightness values to have a range of 0 to 255, the probability values may be represented as pixel brightness values of 0 to 255 in the leukocyte nuclear probability model generated in step S200 according to the saturation and brightness values of each pixel. have.

In operation S330, a probability map is generated from a blood cell image by using pixel brightness values corresponding to each pixel. When each pixel is displayed according to the pixel brightness value, a pixel having a high probability of being a white blood cell nucleus has a high pixel brightness value and thus may be displayed relatively brighter than other areas. Therefore, by generating a probability map from blood cell images, it is possible to identify areas with a high probability of leukocyte nuclei.

In step S400, the sub-image, which is a white blood cell candidate region, is separated from the probability map generated in step S300. The step S400 will be described in detail with reference to FIG. 4.

4 is a detailed flowchart of step S400 of the leukocyte division method using a non-parametric probability model and texture information according to an embodiment of the present invention. As shown in FIG. 4, step S400 of the white blood cell segmentation method using the non-parametric probability model and the texture information according to an embodiment of the present invention comprises: binarizing the probability map by applying an automatic binarization algorithm (S410), The method may include removing noise by performing a morphological opening operation on the probability map (S420), and separating a candidate region remaining after the morphological opening operation into sub-images (S430).

In step S410, binarization is performed by applying an automatic binarization algorithm to the probability map. In this case, the automatic binarization algorithm may use Otsu's binarization algorithm, and Otsu's binarization algorithm assumes that the target area is classified as an area having a difference between two brightness values. As a finding method, there is an advantage that the accuracy is high and the threshold is automatically selected.

In step S420, a morphological opening operation is performed on the probability map to remove noise. Binarization using the automatic binarization algorithm can generate noise and unnecessary small areas, so that morphological opening can be performed for clear division of leukocyte nuclei.

In operation S430, candidate regions remaining after the morphological opening operation are separated into sub-images. In this case, a region including a candidate region and 1.1 to 1.5 times larger in height and width than the candidate region may be separated into the sub image so that the sub image may include the cytoplasm around the nucleus.

5 is a diagram illustrating a sub-image selected according to a leukocyte division method using a non-parametric probability model and texture information according to an embodiment of the present invention. As shown in FIG. 5, in the white blood cell segmentation method using the non-parametric probability model and the texture information according to an embodiment of the present invention, it is confirmed that the correct leukocyte candidate region is selected as the sub image d.

In step S500, region clustering is performed on the sub-image separated in step S400 using mean-shift. The step S500 will be described in detail with reference to FIG. 6.

FIG. 6 is a diagram illustrating a region clustering using Mean-Shift for a sub image according to a non-parametric probability model and a leukocyte segmentation method using texture information according to an embodiment of the present invention. The leukocyte segmentation method using a nonparametric probability model and texture information according to an embodiment of the present invention is different from the kernel size and the minimum region size for the nucleation and cytoplasmic division to obtain a clustering result suitable for the nucleus and cytoplasm of leukocytes. Raised. Since the nucleus appears to be different from the surrounding area, the kernel size is 15 for image space, the kernel size for color space is 10, and the minimum area size per cluster is set to 20, which is good enough to perform Mean-Shift. Results were obtained (b). However, in the case of the cytoplasm, the color values similar to those of the surrounding erythrocytes are often high, so in order to prevent the cytoplasmic region from clustering into some red blood cells or vice versa, the kernel size for the image space is 10. The mean-shift is performed so that the kernel size for the color space is 8 and the minimum area size per cluster is set to 12 to be divided into areas smaller than the nucleus (c). For reference, Mean-shift is a non-parametric feature-space analysis technique, which specifies the size and position of the initial search region and then uses the color cluster generated by repeated color segmentation calculation. A method of extracting an object of interest by determining a boundary based on an initially designated color gamut.

As shown in (b) and (c) of FIG. 6, the mean-shift applied image (c) for cytoplasmic detection has more cluster regions than the mean-shift applied image (b) for nuclear detection. You can check it. At this time, since the nucleus and cytoplasmic regions after region clustering are not represented as a single single region, they are divided into several small regions, and thus a process of dividing them into nucleus or cytoplasm and merging them is required.

In operation S600, merging is performed between each cluster region of the sub-image using a combined map combining the distance map and the probability map. The operation S600 will be described in detail with reference to FIGS. 7 to 9.

7 is a detailed flowchart of step S600 of the leukocyte division method using a non-parametric probability model and texture information according to an embodiment of the present invention. As shown in FIG. 7, in step S600 of the white blood cell segmentation method using the non-parametric probability model and the texture information according to an embodiment of the present invention, a distance map is obtained from a sub-image and a distance map and a probability map are obtained. Generating a binding map by combining (S610), detecting a seed region of the leukocyte nucleus in the sub-image using the binding map (S620), and each cluster according to a condition previously preset based on the seed region. It may include a step of performing a merge between the regions (S630).

In operation S610, the distance map is obtained from the sub-images, and the combined map is generated by combining the distance map and the probability map. The operation S610 will be described in detail with reference to FIG. 8.

8 is a diagram illustrating a combined map of a leukocyte division method using a nonparametric probability model and texture information according to an embodiment of the present invention. Since the probability map generated in step S300 does not mean the nuclear probability in the unit of area in step S400, it means the probability of the leukocyte nucleus in each pixel unit. If you do, the overall probability map will have to be lowered. In this case, the cluster may be determined as a region other than the leukocyte nucleus in the seed selection process or the merging process. Therefore, in order to prevent the main areas of the leukocyte nucleus from being excluded from the leukocyte nucleus due to noise, a distance map (b) is obtained from the sub-image (a) and combined with the probability map (c) to generate a binding map (d). In addition, binding maps can be used to estimate initial seed regions for merging leukocyte nuclear regions.

In operation S620, the seed region of the white blood cell nucleus is detected in the sub-image using the binding map. In operation S630, merging between the cluster regions is performed according to a predetermined condition based on the seed region. Steps S620 and S630 will be described in detail with reference to FIG. 9.

9 is a detailed flowchart of steps S620 and S630 of the white blood cell division method using a non-parametric probability model and texture information according to an embodiment of the present invention. As shown in FIG. 9, step S620 of the white blood cell segmentation method using the non-parametric probabilistic model and the texture information according to the embodiment of the present invention may be performed for each cluster region by using the combined map and the number of pixels of the cluster region. Computing the average of the combined map value (S621), if the average of the combined map value and the number of pixels of the cluster region is more than a predetermined value, selecting the cluster region as the seed region (S622), the inspection is completed for all cluster regions The method may include repeating steps S621 and S622 (S623), and when no seed region is selected, selecting a region having the largest combined map value as the seed region (S624).

In addition, step S630 of the white blood cell segmentation method using the non-parametric probability model and texture information according to an embodiment of the present invention, if the saturation and brightness is above a predetermined threshold value for the cluster region in contact with the seed region, the contacting cluster region Merging into the seed region (S631), and repeating the step S631 until the inspection is completed for all cluster regions (S632).

In step S621, using the combined map and the number of pixels in the cluster region, the average of the combined map values for each cluster region is calculated according to the following equation (1).

Where CM is the combined map, r _i is the cluster region, N _ri is the number of pixels in the cluster region, μ _ri is the average of the combined map values, and m is the cluster region number.

In step S622, if the average of the combined map values and the number of pixels in the cluster region are equal to or greater than a predetermined threshold value, the cluster region is selected as the seed region. This may be represented by the following equation (2).

Where S is the seed region set, R is the entire cluster region set, and th ₁ and th ₂ represent the predetermined threshold values, respectively.

In step S623, steps S621 and S622 are repeated until the inspection is completed for all cluster areas. In step S624, when none of the seed areas is selected, the area having the largest combined map value is determined according to the following equation (3). Select the seed region.

In step S631, the cluster region that is in contact with the seed region is merged with the seed region when the saturation and brightness are greater than or equal to a predetermined threshold. This can be expressed by the following equation (4).

Here, th ₃ represents a predetermined threshold value, Sat _ri represents saturation of r _i , Lum _ri represents brightness of r _i , and Seed _j represents a seed region.

In step S632, step S631 is repeated until the inspection is completed for all cluster regions. As a result, all cluster regions judged to be leukocyte nucleus regions can be merged.

In step S700, the merged cluster region is determined as the finally divided leukocyte nucleus region. The step S700 will be described in detail with reference to FIG. 10.

FIG. 10 is a diagram illustrating a result of merging leukocyte nuclear regions in a leukocyte division method using a nonparametric probability model and texture information according to an embodiment of the present invention. FIG. As shown in FIG. 10, the seed region set remaining in the sub-image is passed through the step S700 of the leukocyte segmentation method using the nonparametric probability model and the texture information according to an embodiment of the present invention. b) can be judged.

In step S800, the cytoplasmic region of the white blood cells is detected using the wavelet energy map and the finally divided leukocyte nuclear region. The step S800 will be described in detail with reference to FIG. 11.

11 is a detailed flowchart of step S800 of the leukocyte division method using a nonparametric probability model and texture information according to an embodiment of the present invention. As shown in FIG. 11, step S800 of the white blood cell segmentation method using the non-parametric probability model and the texture information according to an embodiment of the present invention may include vertical, horizontal, and diagonal components that can be obtained through wavelet transformation in a sub-image. Combining the single wavelet energy map into a single wavelet energy map (S810), classifying the cluster region into a background region set (S820); Computing the average of the wavelet energy map value for each cluster region using the number of pixels of the cluster region (S830), if the average of the wavelet energy map value is a predetermined value or more, selecting the cluster region as the cytoplasmic region ( S840), repeat steps S820, S830, and S840 until the inspection is completed for all cluster regions. In operation S850, when the average of the wavelet energy map values is greater than or equal to a predetermined value for the cluster region that is in contact with the cytoplasmic region, merging the contacted cluster region with the cytoplasmic region in operation S860, and all the cluster regions may be inspected. It may include the step (S870) to repeat the step S860 until.

In step S810, the vertical, horizontal, and diagonal components obtained through the wavelet transform in the blood cell image are combined into one single wavelet energy map. Background, erythrocytes, and cytoplasm remain in the sub-images of the leukocyte nucleus region, and in particular, it is difficult to distinguish between cytoplasm and erythrocytes having similar color ranges. At this time, the color of the cytoplasm and erythrocytes shows a similar distribution, but in the case of erythrocytes, since there is little noise inside, almost no texture component appears, whereas in the cytoplasm, many noises of color similar to the nucleus are found inside the texture component. This appears a lot. Thus, wavelet transform can be used to extract cellular components from the cytoplasm and erythrocytes to segment the cytoplasmic region. The operation S810 will be described in detail with reference to FIG. 12.

12 is a diagram illustrating a wavelet energy map result in a white blood cell division method using a nonparametric probability model and texture information according to an embodiment of the present invention. As shown in FIG. 12, in the leukocyte segmentation method using the nonparametric probability model and the texture information according to an embodiment of the present invention, the wavelet energy map (b) is an image excluding the leukocyte nuclear region divided in step S700. It can be seen that the cytoplasmic region has a higher energy value than the red blood cell region.

In step S820, if the edge length of the cluster area that meets the edge of the sub-image is equal to or greater than a predetermined threshold value, the cluster area is included in the background area set. This may be represented by the following equation (5).

Where r _i is a cluster area, B _P (r _i ) is an edge length of r _i , th ₄ is a predetermined threshold value, B is a background area set, and R is an entire cluster area set.

In step S830, using the wavelet energy map and the number of pixels of the cluster region, the average of the wavelet energy map values for each cluster region is calculated according to the following equation (6).

Here, WM denotes a wavelet energy map, N _ri denotes the number of pixels in the cluster region, μ _ri denotes an average of the wavelet energy map values, and m denotes a cluster region number.

In step S840, if the average of the wavelet energy map values is equal to or greater than a predetermined threshold value, the cluster region is selected as the cytoplasmic region. This may be represented by the following equation (7).

Here, CR represents a set of cytoplasmic regions, th ₅ is a predetermined threshold value, and is an experimental value and may vary depending on blood characteristics.

In step S850, steps S820, S830, and S840 are repeated until the inspection is completed for all cluster areas, and in step S860, the average of the wavelet energy map values for the cluster area that is in contact with the cytoplasmic area is equal to or greater than a predetermined threshold value. In this case, the abutted cluster region is merged into the cytoplasmic region. This can be expressed by the following equation (8).

Here, cr _j represents a cytoplasmic region, th ₆ is a predetermined threshold value, and is an experimental value and may vary depending on blood characteristics.

In step S870, step S860 is repeated until the inspection is completed for all cluster regions. Step S870 will be described in detail with reference to FIG. 13.

FIG. 13 is a diagram illustrating a result of merging leukocyte cytoplasmic regions in a leukocyte division method using a nonparametric probability model and texture information according to an embodiment of the present invention. FIG. As shown in FIG. 13, the white blood cell segmentation method using the nonparametric probability model and the texture information according to an embodiment of the present invention confirms that the cytoplasm is accurately divided when the cytoplasmic region merger results are shown in black lines. Can be.

14 is a diagram illustrating the segmentation performance of the leukocyte division method using a nonparametric probability model and texture information according to an embodiment of the present invention. At this time, the division accuracy was evaluated by the following equation (9).

Here, S _U represents a low division error ratio, S _O represents an over division error ratio, and AVG_S represents an average division accuracy, respectively.

The low division error ratio indicates the ratio of the white blood cell region to be divided, but cannot be divided, and the hyperdivision error ratio represents the ratio of the region which is not the actual white blood cell but divided into the nucleus or cytoplasm of the white blood cell. Average splitting accuracy means higher error value and less error. As shown in FIG. 14, the leukocyte splitting method using the nonparametric probability model and the texture information according to an embodiment of the present invention showed an extremely high accuracy of up to 98%, and in the case of cytoplasm, more than 70% accuracy. Showed. In the case of the leukocyte nucleus, the ratio of low-dividing and over-dividing errors is about 1% and the error rate is very low. In the case of the cytoplasm, wavelet energy values of some regions appeared and merged into the background region rather than into the cytoplasm.

The present invention described above may be variously modified or applied by those skilled in the art, and the scope of the technical idea according to the present invention should be defined by the following claims.

S100: extracting saturation and lightness values from the leukocyte nuclear region divided in the blood
S200: generating a probabilistic model for the leukocyte nucleus region using the saturation and brightness values
S300: generating a probability map representing a probability of leukocyte nuclei for each pixel in the blood cell image using the probability model
S400: separating a sub-image which is a white blood cell candidate region from the probability map
S500: performing area clustering on the sub-images using mean-shift
S600: performing merging between each cluster region of the sub-image using a combined map combining the distance map and the probability map
S700: determining the merged cluster region as the final divided leukocyte nucleus region

Claims (10)

(1) extracting saturation and brightness values from the leukocyte nucleus region divided in blood;
(2) generating a probabilistic model for the leukocyte nucleus region using the extracted chroma and brightness values;
(3) generating a probability map representing a probability of leukocyte nuclei for each pixel in the blood cell image by using the generated probability model;
(4) separating a sub-image which is a leukocyte candidate region from the generated probability map;
(5) performing region clustering on the separated sub-images using mean-shift;
(6) performing merging between respective cluster regions of the separated sub-images using a combined map combining a distance map and the generated probability map; And
And (7) determining the cluster region in which the merging is completed as the finally divided leukocyte nucleus region.
The method of claim 1,
And (8) detecting the leukocyte cytoplasmic region using the wavelet energy map and the finally divided leukocyte nuclear region.
The method of claim 1, wherein in step (2),
The probability is constructed by constructing a histogram, a non-oarametric method, using the extracted saturation and brightness values, and estimating probability density by applying a Gaussian Smooth-Kernel of the Parzen-window. Leukocyte segmentation method using a non-parametric probability model and texture information characterized in that to generate a model.
2. The method of claim 1, wherein step (3)
(3a) converting the blood cell image from a red green blue (RGB) color space to a hues saturation lightness (HSL) color space;
(3b) mapping the probability value corresponding to each pixel of the blood cell image to the pixel brightness value based on the generated probability model; And
And (3c) generating the probability map from the blood cell image by using the pixel brightness value corresponding to each pixel, using the non-parametric probability model and texture information.
The method of claim 1, wherein step (4) comprises
(4a) binarizing the generated probability map by applying an automatic binarization algorithm;
(4b) removing noise by performing a morphological opening operation on the generated probability map; And
And (4c) separating the candidate regions remaining after the morphological opening operation into sub-images.
The method according to claim 5, wherein in step (4c),
A non-parametric probability model and texture information comprising the sub-image is divided into sub-images including the candidate region and 1.1 to 1.5 times the height and width of the candidate region so that the sub-image can include the cytoplasm around the nucleus. Leukocyte splitting method using the.
The method of claim 1, wherein in step (5),
Considering that the color of the leukocyte nucleus is different from the surrounding region, the mean-shift is performed by setting the kernel size for the image space to 15, the kernel size for the color space to 10, and the minimum region size per cluster to 20, or In order for the white blood cell cytoplasm to be clustered clearly without being confused with the surrounding red blood cells, the mean-shift is performed by setting the kernel size for the image space, the kernel size for the color space to 8, and the minimum region size per cluster to 12. Leukocyte segmentation method using a non-parametric probability model and texture information characterized in that.
The method of claim 1, wherein step (6) comprises
(6a) obtaining a distance map from the separated sub-image, and generating a combined map by combining the obtained distance map with the generated probability map;
(6b) detecting a seed region of a leukocyte nucleus in the separated sub-image using the generated binding map; And
(6c) a method of merging white blood cells using non-parametric probability model and texture information, comprising merging between cluster regions according to a predetermined condition based on the detected seed region.
The method of claim 8, wherein step (6b)
(6ba) calculating an average of the combined map values for each cluster region using the generated combined map and the number of pixels of the cluster region;
(6bb) selecting the cluster region as a seed region if the average of the calculated combined map values and the number of pixels of the cluster region are equal to or greater than a predetermined value;
(6bc) repeating steps 6ba and 6bb until inspection is complete for all cluster regions; And
(6bd) if none of the seed regions is selected, selecting a region having the largest combined map value as a seed region,
The step (6c) is,
(6ca) merging the abutted cluster region with the seed region if the saturation and lightness of the cluster region abutting the detected seed region are greater than or equal to a predetermined value; And
(6cb) A method for white blood cell division using nonparametric probability model and texture information, comprising repeating step (6ca) until the test is completed for all cluster regions.
The method of claim 2, wherein step (8) comprises
(8a) combining vertical, horizontal and diagonal components obtained through wavelet transform in the separated sub-image into one single wavelet energy map;
(8b) classifying the cluster area into a background area set if an edge length of the cluster area that meets the edge of the separated sub-image is equal to or greater than a predetermined value;
(8c) calculating an average of wavelet energy map values for each cluster region by using the wavelet energy map and the number of pixels of the cluster region;
(8d) selecting the cluster region as a cytoplasmic region if the average of the wavelet energy map values is equal to or greater than a predetermined value;
(8e) repeating steps (8b), (8c), (8d) until the inspection is completed for all cluster regions;
(8f) merging the abutted cluster region into the cytoplasmic region if the mean of the wavelet energy map values for the cluster region abutting the cytoplasmic region is greater than or equal to a predetermined value; And
And (8g) repeating step (8f) until the test is completed for all cluster regions.

Images