CN109975196A - A kind of reticulocyte detection method and system thereof - Google Patents

Abstract

The invention discloses a reticulocyte detection method and a system thereof. The method comprises the following steps: extracting cell edges according to cell images, and obtaining pixel features in the cell edges; sending the pixel features into a classifier to classify the pixels to obtain Target pixel area; identify reticulocytes by the positional relationship between the target pixel area and the cell edge. Since the pixel features are classified by the classifier, and the reticulocytes are detected and identified by the positional relationship between the target pixel area and the cell edge, the precision and recall rate of detecting reticulocytes can be improved.

Description

A kind of reticulocyte detection method and system thereof

technical field

The invention relates to the technical field of reticulocyte detection, in particular to a reticulocyte detection method and a system thereof.

Background technique

Reticulocytes (reticulocytes, referred to as reticulocytes) are transitional cells that are excreted from the nucleus to the disappearance of intracellular RNA content to form mature red blood cells. Normally, red blood cells in the bone marrow are released into the blood circulation only when they are non-nucleated.

In the prior art, the detection methods of reticulocytes mainly include manual microscopy and instrumental methods based on flow cytometry. Manual microscopy is low in cost, but has low repeatability and is easily affected by subjective factors; although the instrument method based on flow cytometry overcomes the above shortcomings of manual microscopy, it is susceptible to leukocytes, platelets and other nucleated blood in the blood. Or the interference of nucleic acid substances, especially when the proportion of reticulum increases to about 20%, the detection accuracy is low.

Therefore, the existing technology still needs to be improved and developed.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a reticulocyte detection method and a system thereof, aiming at solving the problem of low reticulocyte detection accuracy in the prior art.

The technical scheme adopted by the present invention to solve the technical problem is as follows:

A method for detecting reticulocytes, comprising the following steps:

Extract the cell edge according to the cell image, and obtain the pixel features within the cell edge;

The pixel features are sent to the classifier to classify the pixels to obtain the target pixel area;

Reticulocytes are identified by the positional relationship between the target pixel area and the cell edge.

The described reticulocyte detection method, wherein, the classifier adopts the following steps to obtain:

The pixel set of the RNA-stained area in typical reticulocytes was used as the positive sample set, and the pixel set of the non-RNA-stained area and the pixel set of the non-reticulocyte interior were used as the negative sample set to extract pixel features;

A classifier is formed by supervised learning and cross-validation of pixel features that label positive and negative samples.

The method for detecting reticulocytes, wherein the step of extracting cell edges according to the cell image specifically includes:

Decouple the cell image to get the coupled background image;

According to the coupled background image, the cell image is decoupled and normalized to obtain a decoupled image;

Binarize the decoupled image to obtain a binary image;

Topological analysis was performed on the binary image, and the cell edge was obtained by tracing the outermost boundary.

The method for detecting reticulocytes, wherein the coupled background image is calculated by the following formula:

in, is the coupled background image, n represents the number of selected cell images, I _i represents the ith grayscale image, and Σ is the summation operation.

The method for detecting reticulocytes, wherein, the decoupling image is calculated by the following formula:

Among them, I′ _i is the decoupling image of the ith grayscale image, min[·] represents the minimum operation, and max[·] represents the maximum operation.

In the method for detecting reticulocytes, wherein the step of performing a binarization process on a decoupled image to obtain a binary image includes:

Based on Otsu algorithm, Niblack algorithm and Canny operator, the decoupled image is binarized to obtain the operation result graphs of Otsu, Niblack and Canny respectively;

After performing OR operation on the operation result graphs of Otsu, Niblack and Canny, a binary image is obtained by filling holes, denoising and morphological processing.

The method for detecting reticulocytes, wherein the step of identifying the reticulocytes through the positional relationship between the target pixel area and the edge of the cell specifically includes:

Determine whether the target pixel area is inside the cell by the number of times the ray passes through the cell boundary;

When there is only one target pixel area in the cell, and the area of the area is greater than the first preset area threshold, the cell is a reticulocyte;

When the target pixel area in the cell exceeds the preset number, and the area of the area is greater than the second preset area threshold, the cell is a reticulocyte.

The method for detecting reticulocytes, wherein the slope of the ray is k:

Among them, x _c , y _c are the horizontal and vertical coordinates of the centroid of the target pixel area, respectively, and x _j , y _j are the horizontal and vertical coordinate vectors of the edge coordinates of a single cell, respectively.

In the reticulocyte detection method, the pixel features include: color space conversion, color features and texture features.

A reticulocyte detection system, comprising: a processor, and a memory connected to the processor,

The memory stores a reticulocyte detection program, and when the reticulocyte detection program is executed by the processor, the following steps are implemented:

Extract the cell edge according to the cell image, and obtain the pixel features within the cell edge;

The pixel features are sent to the classifier to classify the pixels to obtain the target pixel area;

Reticulocytes are identified by the positional relationship between the target pixel area and the cell edge.

Beneficial effects: Since the pixel features are classified by the classifier, and the reticulocytes are detected and identified by the positional relationship between the target pixel area and the cell edge, the precision and recall rate of detecting the reticulocytes can be improved.

Description of drawings

FIG. 1 is a flow chart of a preferred embodiment of the method for detecting reticulocytes of the present invention.

Figure 2 is a first image of reticulocytes in the present invention.

Figure 3 is a second image of reticulocytes in the present invention.

Figure 4 is a third image of reticulocytes in the present invention.

Fig. 5 is a fourth image of reticulocytes in the present invention.

FIG. 6 is a functional principle block diagram of a preferred embodiment of the reticulocyte detection system of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer and clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

Please refer to FIG. 1 to FIG. 5 at the same time, the present invention provides some embodiments of a method for detecting reticulocytes.

As shown in Figure 1, a reticulocyte detection method of the present invention comprises the following steps:

Step S100 , extract the cell edge according to the cell image, and obtain pixel features within the cell edge.

In the present invention, alkaline substances (glorious tar blue, new methylene blue, etc.) are used to stain the RNA of cells, the residual RNA substances in reticulocytes are stained blue or blue-green, while there is no residual RNA in normal red blood cells, will not be dyed. After staining, a camera was used to capture the cell image, and the cell edge contour was extracted, which mainly included processes such as decoupling, binarization, and edge extraction.

Specifically, step S100 includes the following steps:

Step S110, decoupling the cell image to obtain a coupled background image.

Decoupling is the use of mathematical methods to separate and deal with factors that affect each other. The collected cell images have problems such as lens contamination, over- or under-exposure, etc. Considering that the images collected in consecutive blood samples have close imaging conditions, the decoupling process of the present invention is as follows:

The coupled background image is calculated using the following formula:

in, is the coupled background image, n represents the number of selected cell images, I _i represents the ith grayscale image, and Σ is the summation operation.

Step S120 , decouple the cell image according to the coupled background image and then normalize it to obtain a decoupled image.

Compared with the cell image, the decoupled image has mostly removed exposure unevenness and lens contamination, and also enhanced the image cell foreground and background contrast, which helps to improve the effect of binarization.

The decoupling image is calculated using the following formula:

Among them, I′ _i is the decoupling image of the ith grayscale image, min[·] represents the minimum operation, and max[·] represents the maximum operation.

Step S130 , performing binarization processing on the decoupled image to obtain a binary image.

The binarization operation in cell segmentation mostly adopts the combined method. Here, the global and local binarization is used to determine the cell region, and it is further combined with the cell edge to make the cell edge contour smoother and more accurate. Based on the good segmentation effect of Otsu and Niblack algorithm in global and local thresholding, and the excellent performance of Canny operator [8] in edge detection.

Step S130 specifically includes:

Step S131 , based on the Otsu algorithm, the Niblack algorithm and the Canny operator, perform binarization processing on the decoupled image to obtain operation result graphs of Otsu, Niblack and Canny, respectively.

Step S132: After performing OR operation on the operation result graphs of Otsu, Niblack and Canny, a binary image is obtained through filling holes, denoising processing and morphological processing.

Specifically, the result diagram of the OR operation of the three (Otsu, Niblack and Canny) is as follows:

I _{B =} I _o |IN | _IC

Among them, I _o , IN and I _C are the operation result graphs of Otsu, Niblack and Canny in turn, and _{I B is} the result graph of the three OR operations. Otsu, Niblack and Canny's operation result graphs have edge gaps of the same cell at different positions. After OR operation and hole filling, denoising, and morphological processing, a binary image is obtained. The cells in the binary image can be well closed, and the edges are more precise.

Step S140 , perform topological analysis on the binary image, and track the outermost boundary to obtain the cell edge.

The more accurate the binary image is, the simpler the topological analysis is, and the cell edge contour can be obtained only by tracing the outermost boundary.

Step S150, extracting pixel features within the cell edge.

Pixel feature extraction based on binary images helps to improve the calculation speed and reduce the interference of extracellular staining. Pixel features include: color space conversion, color features and texture features.

(1) Color space conversion.

In practical applications, it is difficult to find a single color space that can distinguish all colors well. The RGB channels are highly correlated, and the luminance and chrominance components are also highly mixed; the chrominance and luminance in HSI are completely separated, and achieve the best results in the target search for various colors; LUV is perceptually uniform, and in the The best results are achieved in multi-color space comparisons. Therefore, RGB, HSI and LUV are selected here as the comparison objects of the present invention. The following color features are extracted in the three spaces respectively, and then the classification effect is compared alone or in combination with the following texture features, and finally the color features are extracted in the LUV space.

(2) Color feature extraction.

According to the local spatial similarity model, when measuring the influence of spatial factors on the central pixel, the present invention will use the Gaussian function commonly used in image filtering and smoothing to measure.

a. Create a partial window.

With pixel m as the center, create a window of size d × d. Taking 5×5 as an example, the red frame i is the center of the window, and n is the pixel located at any position within the window. Their coordinates are (x _m , y _m ), (x _n , y _n ), and the grayscale values are g _m , g _n in sequence.

b. Calculate local spatial features.

The local spatial feature mainly considers the distance factor, that is, the influence of the distance between the pixel point and the target pixel point on the target pixel point. In a two-dimensional image, it is expressed as the relationship between pixel coordinates.

The spatial feature sf _mn is measured by using a wide range of Gaussian functions in image processing (as shown in the following equation):

Among them, σ _s represents the standard deviation of the same value in the x direction and the y direction, and also plays the role of weighting the pixels according to the distance.

c. Calculate local gray level features.

The local gray level feature gf _mn is used to measure the inhomogeneity of local intensity. In the image, it is expressed as the functional relationship between the gray values of each pixel:

where _λg is the global scale factor, is the density function used to reflect the local gray inhomogeneity.

d. Generate pixel-level features.

The final pixel-level color feature pcf _m is:

Among them, N _m represents the number of pixels in the center window, and Σ is the summation operation.

(3) Texture feature extraction.

Texture features are one of the commonly used methods in image segmentation, and are usually combined with color features to achieve better results. In image processing, texture features are embodied as a combination of a repeating pattern and the frequency of this pattern. The present invention will select one signal processing method and two statistical methods for comparison.

a, color space conversion.

Since the reticulum dyed area collected by the present invention is blue, according to the color space model of YCbCr, the texture feature will be calculated on the Cr channel here. Before the calculation, the Y channel is first processed with a nonlinear median filter to suppress the fine texture caused by grayscale changes in the image, thereby enhancing the sparse subband coefficients.

b. A measure of the subband coefficient energy.

Gabor filters measure this energy by generating directional bands in any direction. The Gabor filter is a filter G(x,y) that can be rotated in any direction formed by the linear combination of the base filters, and is expressed as follows:

where b _k (θ) is the interpolation function of an arbitrary rotation angle θ, which controls the orientation of the filter; H _K (x,y) is the rotated version of the impulse response of the base filter at θ. Edges in the image can then be detected by a base filter in the corresponding direction.

For the Gabor filter, calculating the texture features in the window means performing convolution operation in the window. The present invention adopts the mean value of the convolution operation results in all directions as the feature value of the center pixel of the window.

c. Texture features based on statistical methods.

Both Gray Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP) are cases of gray level changes between pixels within a statistical window. The difference is that GLCM is a statistical relationship between pixel pairs (ie two pixels). That is, GLCM represents the ratio of the number of occurrences of gray-level pixel pairs formed along a certain direction to the total number of gray-level pixel pairs that may appear in the image gray level, namely:

Among them, (g _m , g _n ) is the gray level of any two pixels in the window, and t (g _m , g _n ) represents the number of occurrences of the pair of gray levels. T represents the total number of all possible gray-level pairs in the current image.

And LBP is to count the difference between the center point pixel and its neighbor pixels. LBP takes the gray value of the central pixel as a threshold, and then compares all pixels with a certain distance from the central pixel to form a binary sequence, namely:

At this time, g _m is only the center element of the window, and g _n is the real existence or calculated pixel gray value at a certain distance. u-LBP (Uniform LBP) is that the number of adjacent two-digit 0-1 and 1-0 transitions in the LBP binary sequence is not more than 2 times, and the number of times not more than 2 times is used as the local code of the window, which is more traditional LBP is more widely used.

The present invention will directly use the local code of u-LBP as the texture feature of the center pixel of the window. However, GLCM results cannot directly represent texture features, which are more sensitive to areas with drastic changes in intensity (such as edge contours, etc.).

Step S200, the pixel features are sent to the classifier to classify the pixels to obtain the target pixel area.

The classifier is preset. Specifically, the classifier is obtained by the following steps:

S10. The pixel set of the RNA-stained area in a typical reticulocyte is used as a positive sample set, and the pixel set of the non-RNA-stained area and the pixel set of the non-reticulocyte interior are used as a negative sample set, and pixel features are extracted.

Since the present invention adopts pixel-level feature classification, from the perspective of the entire image, it will inevitably lead to large and redundant samples, and the final calculation is time-consuming or even difficult to calculate. Therefore, typical reticulum and non-reticulocyte sets (Cell set) are selected as training sets here. Of course, typical reticulocytes and non-reticulocytes here are determined artificially.

Extract multiple pixel features to be selected. Convert three color spaces on the Cell set, and extract color features in each color space. In addition, six texture features are extracted on the Cr channel, and the specific steps can be found in step S100.

S20. Perform supervised learning on the pixel features of the labeled positive and negative samples and cross-validate to form a classifier.

The features to be selected are cross-validated and selected. At this time, for fast calculation, the penalty coefficient of the SVM and the expansion constant of the kernel function are set to 1. 10-fold was used to find the mean. After comparison, select the color space and texture features with the best effect.

Cross-validation is performed on the selected feature combination. At this time, the cross-validation is to adjust the penalty coefficient of the SVM and the expansion constant of the kernel function RBF. In order to speed up the adjustment speed, the present invention firstly adopts the repeated downsampling-granular SVM (RU-GSVM), that is, the edge information of the sample is preserved to the greatest extent, and the majority class is downsampled. Then, the parameters are adjusted by using the coarse and fine double-layer grid cross-validation method and 10-fold cross-validation. Finally, use the sampled samples and parameters to set the parameters and train the SVM to form a classifier (SVM Classifier).

Step S300 , identifying reticulocytes according to the positional relationship between the target pixel area and the cell edge.

Step S300 specifically includes:

Step S310: Determine whether the target pixel area is within the cell by the number of times the rays pass through the cell boundary.

The slope of the ray is k:

Among them, x _c , y _c are the horizontal and vertical coordinates of the centroid of the target pixel area, respectively, and x _j , y _j are the horizontal and vertical coordinate vectors of the edge coordinates of a single cell, respectively.

Specifically, whether the RNA-stained region is inside the cell is determined by determining the number of times the ray passes through the cell boundary, if the number of times of passing is odd, it is inside the cell, and if it is even, it is outside the cell.

Step S320: When there is only one target pixel area in the cell, and the area of the area is greater than the first preset area threshold, the cell is a reticulocyte.

Step S330 , when the target pixel area in the cell exceeds the preset number and the area area is greater than the second preset area threshold, the cell is a reticulocyte.

As shown in Figure 2-Figure 5, considering that there is a single connected domain in the reticulum RNA staining area or the area of multiple connected domains is small, three thresholds are set here: the first preset area threshold, the preset area threshold Set the number and the second preset area threshold. When there is only one target staining domain in the cell, the first preset area threshold is used to determine whether the cell is reticulum; the number of target staining domains in the cell is not less than the preset number and the existing area is greater than the second preset area threshold It is reticulum red when it is in the staining domain.

It is worth noting that one color feature is extracted in the three color spaces of RGB, HSI and LUV, and six texture features including Gabor feature, gray level co-occurrence matrix and local contrast are extracted on the Cr channel of YCbCr, and then the color feature is extracted. Compare with the recognition effect of each feature combination of texture, and select Gabor texture feature combined with LUV color feature. Then, the SVM classifier is used to classify the pixel-level features, detect the RNA-stained area, and use the location, number, and area of the area to determine whether the target cell is a reticulocyte. The precision of the method for reticulocytes was 98.4% and the recall was 98.0%.

The present invention also provides a preferred embodiment of the reticulocyte detection system:

As shown in FIG. 6 , a reticulocyte detection system according to an embodiment of the present invention includes: a processor, and a memory connected to the processor,

The memory stores a reticulocyte detection program, and when the reticulocyte detection program is executed by the processor, the following steps are implemented:

Extract the cell edge according to the cell image, and obtain the pixel features within the cell edge;

The pixel features are sent to the classifier to classify the pixels to obtain the target pixel area;

Reticulocytes are identified by the positional relationship between the target pixel area and the cell edge, as described above.

When the reticulocyte detection program is executed by the processor, the following steps are also implemented:

The pixel set of the RNA-stained area in typical reticulocytes was used as the positive sample set, and the pixel set of the non-RNA-stained area and the pixel set of the non-reticulocyte interior were used as the negative sample set to extract pixel features;

A classifier is formed by supervised learning and cross-validation of pixel features labeling positive and negative samples, as described above.

When the reticulocyte detection program is executed by the processor, the following steps are also implemented:

Decouple the cell image to get the coupled background image;

According to the coupled background image, the cell image is decoupled and normalized to obtain a decoupled image;

Binarize the decoupled image to obtain a binary image;

Topological analysis is performed on the binary image, and cell edges are obtained by tracing the outermost boundary, as described above.

In this embodiment, the coupling background image is calculated by the following formula:

in, is the coupled background image, n represents the number of selected cell images, I _i represents the ith grayscale image, and Σ is the summation operation, as described above.

In this embodiment, the decoupling image is calculated by the following formula:

Among them, I′ _i is the decoupling image of the ith grayscale image, min[·] represents the minimum operation, and max[·] represents the maximum operation, as described above.

When the reticulocyte detection program is executed by the processor, the following steps are also implemented:

Based on Otsu algorithm, Niblack algorithm and Canny operator, the decoupled image is binarized to obtain the operation result graphs of Otsu, Niblack and Canny respectively;

After the OR operation is performed on the operation result graphs of Otsu, Niblack and Canny, a binary image is obtained by filling holes, denoising and morphological processing, as described above.

When the reticulocyte detection program is executed by the processor, the following steps are also implemented:

Determine whether the target pixel area is inside the cell by the number of times the ray passes through the cell boundary;

When there is only one target pixel area in the cell, and the area of the area is greater than the first preset area threshold, the cell is a reticulocyte;

When the target pixel area in the cell exceeds the preset number, and the area area is greater than the second preset area threshold, the cell is a reticulocyte, which is specifically as described above.

In this embodiment, the slope of the ray is k:

Wherein, x _c , y _c are the horizontal and vertical coordinates of the centroid of the target pixel area, respectively, and x _j , y _j are the horizontal and vertical coordinate vectors of the edge coordinates of a single cell, respectively, as described above.

In this embodiment, the pixel features include: color space conversion, color features, and texture features, as described above.

To sum up, the present invention provides a reticulocyte detection method and system. The method includes the following steps: extracting cell edges according to cell images, and obtaining pixel features in the cell edges; sending pixel features into classification The device classifies the pixels to obtain the target pixel area; the reticulocytes are identified by the positional relationship between the target pixel area and the cell edge. Since the pixel features are classified by the classifier, and the reticulocytes are detected and identified by the positional relationship between the target pixel area and the cell edge, the precision and recall rate of detecting reticulocytes can be improved.

It should be understood that the application of the present invention is not limited to the above examples. For those of ordinary skill in the art, improvements or transformations can be made according to the above descriptions, and all these improvements and transformations should belong to the protection scope of the appended claims of the present invention.

Claims (10)

1. A reticulocyte detection method is characterized by comprising the following steps:

extracting a cell edge according to the cell image, and obtaining pixel characteristics in the cell edge;

sending the pixel characteristics into a classifier to classify the pixels to obtain a target pixel area;

and identifying the reticulocytes through the position relation of the target pixel area and the cell edge.

2. The reticulocyte detection method of claim 1, wherein the classifier is obtained by:

adopting a pixel set of an RNA staining area in a typical reticulocyte as a positive sample set, and adopting a pixel set of a non-RNA staining area and an internal pixel set of the non-reticulocyte as a negative sample set, and extracting pixel characteristics;

and carrying out supervised learning and cross validation on the pixel characteristics of the marked positive and negative samples to form a classifier.

3. The method for detecting reticulocytes of claim 1, wherein the step of extracting cell edges from the cell image comprises:

decoupling the cell image to obtain a coupling background image;

decoupling the cell image according to the coupling background image and then normalizing to obtain a decoupling image;

carrying out binarization processing on the decoupling image to obtain a binary image;

and performing topology analysis on the binary image, and tracking the outermost boundary to obtain the cell edge.

4. The method according to claim 3, wherein the coupled background map is calculated by the following formula:

wherein,for the background image, n represents the number of selected cell images, I_iThe ith gray scale map is represented, and Σ is the summation operation.

5. The method of claim 3, wherein the decoupled image is calculated using the following formula:

wherein, I'_iIs a decoupled image of the ith gray scale map, min [ ·]Denotes a minimization operation, max ·]Indicating a max operation.

6. The method for detecting reticulocytes of claim 3, wherein the step of binarizing the decoupled image to obtain a binary image comprises:

carrying out binarization processing on the decoupling image based on an Otsu algorithm, a Niblack algorithm and a Canny operator to respectively obtain operation result graphs of Otsu, Niblack and Canny;

and after performing OR operation on the operation result graphs of Otsu, Niblack and Canny, filling the hole, performing denoising processing and morphological processing to obtain a binary image.

7. The method according to claim 1, wherein the step of identifying reticulocytes by the positional relationship between the target pixel region and the cell edge specifically comprises:

determining whether the target pixel region is within the cell by the number of times the ray crosses the cell boundary;

when only one target pixel region is arranged in the cell and the area of the region is larger than a first preset area threshold value, the cell is a reticulocyte;

when the number of target pixel regions in the cell exceeds a preset number and the area of the region is larger than a second preset area threshold value, the cell is a reticulocyte.

8. The method according to claim 7, wherein the slope of the ray is k:

wherein,x_c、y_crespectively the abscissa and ordinate, x, of the centroid of the target pixel region_j、y_jRespectively, are the abscissa and ordinate vectors of the edge coordinates of a single cell.

9. The method of claim 1, wherein the pixel characteristics comprise: color space conversion, color features, and texture features.

10. A reticulocyte detection system, comprising: a processor, and a memory coupled to the processor,

the memory stores a reticulocyte detection program that when executed by the processor implements the steps of:

extracting a cell edge according to the cell image, and obtaining pixel characteristics in the cell edge;

sending the pixel characteristics into a classifier to classify the pixels to obtain a target pixel area;

and identifying the reticulocytes through the position relation of the target pixel area and the cell edge.