CN111311628A

CN111311628A - Full-automatic high-performance leukocyte segmentation method

Info

Publication number: CN111311628A
Application number: CN202010087468.4A
Authority: CN
Inventors: 张云超; 邱天
Original assignee: Individual
Current assignee: Individual
Priority date: 2020-02-11
Filing date: 2020-02-11
Publication date: 2020-06-19

Abstract

The invention provides a full-automatic high-performance leukocyte segmentation method. In clinical diagnosis and treatment of patients, it is very important to classify and count leukocytes on blood smears. However, manual white blood cell sorting and counting by a hematologist is time consuming and prone to error. The algorithm provided by the invention firstly carries out a series of processing such as segmentation and the like on the nucleus region through the zoomed saturation image, and then carries out image processing such as difference and the like on the green image and the blue image normalized by the red image, thereby achieving the purpose of automatically and accurately segmenting the cytoplasm region in the subimage.

Description

Full-automatic high-performance leukocyte segmentation method

Technical Field

The invention relates to the field of image processing, in particular to a full-automatic high-performance leukocyte segmentation method.

Background

Leukocytes are important components of the human immune system, and the change of the morphology and the number of the leukocytes is generally used as a reference basis for diagnosis clinically.

In most of the past, the skilled hematology staff manually identifies and classifies the white blood cells by using an optical microscope, then the types of the white blood cells are distinguished according to information such as the cell nucleus, cytoplasm and chromosome color in the white blood cells, and the number of the types of the white blood cells is recorded. According to hospital statistics, about 20% of blood routine specimens need manual microscopic examination and reexamination. The manual mode has long time for focusing, observing and identifying each time, and the hematology staff responsible for the artificial microscopic examination of the leucocytes spends a large amount of time, thereby wasting medical resources.

At present, because of the appearance of digital image processing technology, the white blood cells are segmented and identified by a machine, so that the time and the labor are saved, and the accuracy is over 95 percent. The machine identification of leukocytes in microscopic images typically goes through several processes: preprocessing of images, segmentation of image cells, feature extraction and classification of the image cells. The segmentation of the image cells is the most important step, because the accuracy of the subsequent feature extraction and classification is premised on the accuracy of cell segmentation.

Leukocytes play an important role in the diagnosis of different diseases. The hematologist obtains more accurate and more accurate results by analyzing the white blood cell images formed by the liquid smears. The total number of leukocytes in normal adults is (4.0-10.0) × 10⁹L, but it varies with time and the functional status of different organisms. Numerical and data changes in each category have significant clinical significance. By analyzing the number and morphological changes of different types of leukocytes, human health can be diagnosed. The specific changes in leukocytes are as follows:

(1) the white blood cell count increased. When the total number of leukocytes exceeds 10,000, it is considered to be leukocytosis. Increased white blood cell count is often seen in acute infections, tissue injuries, major surgery, leukemia, etc.

(2) The number of leukocytes was reduced. When the total number of leukocytes is less than 4000, it is considered to be a decrease in leukocytes. The decrease in the number of leukocytes is usually observed in typhoid and paratyphoid, malaria, X-ray and radionuclide exposure.

It can be seen that the increase and decrease of leukocytes can provide a basis for differential diagnosis of disease. Image segmentation of leukocytes is a very critical step in a fully automated leukocyte counting process, since the accuracy of subsequent feature extraction and classification depends on whether the leukocytes can be correctly segmented. However, segmentation and extraction of leukocyte images presents a number of challenges and difficulties, each of which is listed below.

1.) the shape, size, edges and location of the cells are not uniform and vary. Also, due to the non-uniformity of the blood coating thickness and the imbalance of illumination, the image contrast between the cell boundary and the background is not stable during image acquisition.

2.) the staining method and the staining agent differ among different reagents. Thus, the digital images from different media differ in hue.

3.) the cytoplasmic region is very similar to the background, so it is difficult to segment the cytoplasmic region correctly.

4.) previous people have tried several methods in blood smear digital image segmentation including edge and boundary detection, region growing, filtering, morphological analysis, intensity thresholding, edge analysis, area growing, deformable model fitting and watershed clustering algorithms, etc. For example, riter et al propose a fully automated method for segmentation and boundary recognition of all objects in an image acquired from a peripheral blood smear that do not overlap the boundary. CellaVision developed software and equipment for leukocyte segmentation, the method of which had reached about 89% accuracy in 5-class classification.

Disclosure of Invention

To solve the above problems, an embodiment of the present invention provides a fully automatic high-performance white blood cell segmentation method.

The technical scheme adopted by the embodiment of the invention for solving the problems is as follows:

a full-automatic high-performance leukocyte segmentation method comprises the following steps:

converting the RGB image into an image of an HSL color space;

detecting nuclei from the scaled saturation image;

marking a kernel region on the full resolution sub-image;

the cell paste fractions were divided and their area calculated.

Further, the detecting cell nuclei from the scaled saturation image specifically includes the following steps:

calculating a first histogram from the saturation;

obtaining a first binary threshold from a first histogram;

if the saturation is greater than the first binary threshold, the ScaledMask (x, y) is nucleolus color, and if the saturation is less than the first binary threshold, the ScaledMask (x, y) is background;

performing image corrosion algorithm on the ScaledMask (x, y) by using a morphological analysis method in image processing to remove small noise;

clearing cells near the image boundaries to ensure the integrity of the white blood cells;

if (Green (x, y) -Blue (x, y))/(Red (x, y) +1) < ═ 0, then ScaledMask (x, y) is set to the background;

etching the ScaledMask (x, y) again to remove small noise;

tracing the connecting region of the ScaledMask (x, y);

filtering the connected regions, and when a condition (area array [ i ]. Size > maxnecolusaccream | | area array [ i ]. Size < minneolumsbeckacreage)) is satisfied, setting the region [ i ] as a background, wherein maxnecolumascrage and minneolumsbeckacreage are constant empirical values;

connecting and classifying areas which are relatively close to each other into the same cell;

the center of a set of connected regions classified as the same cell is calculated as the final leukocyte center, and the coordinates are converted to coordinate positions in the non-zoomed image.

Further, the marking of the nucleus region on the full-resolution sub-image specifically includes the following steps:

clipping a sub-image group of the white blood cells by using the center coordinates of the white blood cells, wherein the sub-image group comprises a red image, a blue image, a green image, a saturation image and a brightness image;

calculating a second image histogram for the sub-image group of the white blood cells according to the saturation;

acquiring a second binary threshold value from the second image histogram;

if the saturation is larger than the second binary threshold, Mask (x, y) is the color of the cell nucleus, and if the saturation is smaller than the second binary threshold, Mask (x, y) is background.

Further, the segmenting the cytoplasm part and calculating the area of the cytoplasm part specifically comprises the following steps:

applying a gaussian low pass filter to the set of sub-images of the white blood cells;

(x, y) the color of the cytoplasm if (Green (x, y) -Blue (x, y))/(Red (x, y) +1) > Tab and Mask (x, y) back, otherwise Mask (x, y) back, where Tab is an empirical constant, the color of the cytoplasm is a constant integer between [1,255], and back 0;

acquiring the average value of the green sub-images;

etching Mask (x, y) to remove noise;

the red image of the RGB sub-image is replaced with Mask (x, y) and the final image is saved to facilitate comparison of the segmentation results from a manual inspection perspective.

The full-automatic high-performance leukocyte segmentation method has the following beneficial effects: the automatic identification degree is high, the performance is good, and the white blood cells can be well identified.

Drawings

The invention is further illustrated with reference to the following figures and examples.

Fig. 1 is a blood cell image including 3 white blood cells. The image was a digital image of a blood smear captured with Wright staining and magnified 500 times;

FIG. 2 is a scaled saturation image;

FIG. 3 is a scaled, binarized saturation image;

FIG. 4 is a green sub-image cut from a full size image of a detected white blood cell;

FIG. 5 is a mask of sub-images automatically generated by the algorithm;

FIG. 6 is a graph showing the results of the detection.

Detailed Description

Referring to fig. 1, in order to improve the accuracy and adaptability of the white blood cell segmentation algorithm and solve or partially solve the above problems, the present invention first collects a digital image of a blood smear stained with Wright and magnifies it by 500 times, and then subdivides white blood cells using the following flow chart:

step S100 converts the RGB image into an image in an HSL (Hue, Saturation, brightness) color space. The model of the HSL color space corresponds to a double cone and a sphere (white is at an upper vertex, black is at a lower vertex, and the center of a circle of a maximum cross section is half-way gray) in a cylindrical coordinate system, and the image of the HSL color space comprises three sub-images of hue, saturation and brightness;

s200, detecting cell nucleuses from the zoomed saturation images;

step S300, marking a cell nucleus area on the full-resolution sub-image;

step S400, the cytoplasm part is divided and the area of the cytoplasm part is calculated.

Further, images with 0.5 scale factors for red, green, and blue in the scaled saturation map may increase the computation speed. The scaled saturation image is referred to in fig. 2. In step S200, detecting cell nuclei from the scaled saturation image, specifically including the following steps:

step S201, calculating a first histogram according to the saturation;

step S202, acquiring a first binary threshold value from a first histogram; the principle is to minimize the sum of the deviations of the background and foreground regions; the binary threshold method has stronger robustness; referring to fig. 3, fig. 3 is a binarized image;

step S203, if the saturation is greater than the first binary threshold, the ScaledMask (x, y) is nucleolus color, and if the saturation is less than the first binary threshold, the ScaledMask (x, y) is backsground;

step S204, performing image corrosion algorithm on the scaledMask (x, y) by using a morphological analysis method in image processing to remove small noise;

step S205, cleaning cells near the image boundary to ensure the integrity of the white blood cells;

step S206, if { Green (x, y) -Blue (x, y) }/(Red (x, y) +1) < ═ 0, setting the ScaledMask (x, y) as the background;

step S207, corroding the scaledMask (x, y) and removing the small noise again;

step S208, tracking a connection area of the scaledMask;

step S209, filtering the connected areas, and setting the area [ i ] as a background when a condition (area [ i ]. Size > MaxNeoclusAcreage | | area [ i ]. Size < MinNeoclusBlockAcreed)) is met, wherein the area [ i ] is a constant experience value;

s210, communicating and classifying areas which are relatively close to each other into the same cell;

and step S211, calculating a group of connected region centers of the same cell as a final leukocyte center, and converting the coordinates into coordinate positions in the non-zooming image.

Further, in step S300, marking a nuclear region on the full-resolution sub-image specifically includes the following steps:

step S301, a leukocyte subimage is intercepted by using the central coordinate of the leukocyte, wherein the central coordinate of the leukocyte is obtained from step S211. A red image, a blue image, a green image, a saturation image, and a luminance image. Referring to fig. 4, a green sub-image cut out from the full-size image of the first detected white blood cell;

step S302, calculating a second image histogram according to the saturation degree to the sub-image group of the white blood cells;

step S303, acquiring a second binary threshold value from the second image histogram; the principle is to minimize the sum of the deviations of the background and foreground regions; the binary threshold method has stronger robustness;

step S304, if the saturation is greater than the second binary threshold, Mask (x, y) is the color of the cell nucleus, and if the saturation is less than the second binary threshold, Mask (x, y) is background.

Further, in step S400, the steps of segmenting the cytoplasm portion and calculating the area thereof specifically include the following steps:

step S401, applying a Gaussian low-pass filter to the group of sub-images of the white blood cells obtained in step S301;

step S402, if (Green (x, y) -Blue (x, y))/(Red (x, y) +1) > Tab and Mask (x, y) ═ background, Mask (x, y) ═ cytoplasm color, otherwise Mask (x, y) ═ background, where Tab is an empirical constant, cytoplasm color is a constant integer between [1,255], background is 0; referring to fig. 5, fig. 5 is a generated Mask (x, y);

step S403, obtaining an average value of the green sub-images;

step S404, etching Mask (x, y) to eliminate small noise;

step S405, replacing the red image of the RGB subimage with Mask (x, y), and storing the final image to facilitate the comparison of whether the segmentation result is correct from the angle of manual inspection; referring to fig. 6, fig. 6 is a final output result.

Claims

1. A full-automatic high-performance leukocyte segmentation method is characterized by comprising the following steps: