CN108564114B - A method for automatic identification of human fecal leukocytes based on machine learning - Google Patents

Abstract

The invention discloses a human body excrement leukocyte automatic identification method based on machine learning, and relates to a method for automatically identifying and detecting leukocytes in human body excrement based on a digital image processing technology, in particular to a machine learning technology. The method has the advantages of accurate recognition, high speed, high efficiency and small calculated amount.

Description

A method for automatic identification of human fecal leukocytes based on machine learning

technical field

The invention relates to a method for automatic identification and detection of leukocytes in human feces based on digital image processing technology, in particular, machine learning technology.

Background technique

White blood cells in human feces are often detected in mucus and pus and blood, and most of them are neutral paged granulocytes. Due to cell degeneration, the cells are swollen and the structure is unclear. If the number is large, there are piles, and the cell membrane is incomplete or broken, it is also called pus cells at this time, indicating a serious infection. Different from human blood and urine, human feces contain a large amount of impurities such as food residues, resulting in a complex image background; some white blood cell membranes are no longer intact, and are adhered to some impurities. The above reasons all lead to the difficulty of dividing white blood cells. Due to cellular degeneration, leukocytes are morphologically variable and are not easily identified using morphology-based methods.

The applicant has applied for a patent for an automatic detection method for leukocytes in bronchoalveolar lavage smears in 14 years (patent number is 2014103696359) At the same time, the appearance and internal characteristics of leukocytes are used for screening, and finally leukocytes are identified. The pure image processing method is used to detect leukocytes in the sample, which requires a large amount of calculation in practical applications.

SUMMARY OF THE INVENTION

The purpose of the present invention is to design an automatic identification method for leukocytes in human feces in view of the complex background of stool microscopic images and the diversity of leukocyte morphological changes. The method is based on digital image processing technology and machine learning technology. The white blood cells are automatically identified in the image, so as to achieve the purpose of high detection accuracy and low missed detection rate, and meet the needs of clinical detection.

The technical solution provided by the present invention is a method for automatic identification of human fecal leukocytes based on machine learning, and the method comprises the following steps:

Step 1: Pretreatment of human feces by dilution, stirring and precipitation, and using a biological microscope to collect microscopic images of human feces samples;

Step 2: Grayscale the image in Step 1 and convert it to a grayscale image;

Step 3: perform local reverse binarization processing on the grayscale image in step 2 to obtain a binarized image;

Step 4: Perform a hole filling operation on the binary image in Step 3, and fill the area with an area smaller than S1, where the value range of S1 is 180-230 pixels;

Step 5: Perform a morphological expansion operation on the binary image in Step 4, using a circular template with a radius of R1; perform hole filling on the expanded binary image; perform a morphological erosion operation on the hole-filled binary image , using a circular template with a radius of R2; the value range of R1 is 2 to 5 pixels, the value range of R2 is 1 to 3 pixels, and R1 is greater than R2;

Step 6: Mark the connected area of the binary image in step 5, and calculate the area feature and circumscribed rectangle of the connected area; the reserved area is greater than S2, and the proportion of the area of the circumscribed rectangle is greater than D1; the width and height of the circumscribed rectangle is greater than H1, and smaller than the area of H2 ; The value range of S2 is 1400~1550 pixels, the value range of D1 is 60%~80%, the value range of H1 is 30~50 pixels, and the value range of H2 is 100~120 pixels;

Step 7: For the area reserved in Step 6, crop the binary image according to the circumscribed rectangle; for the cropped binary image, first calculate the center and radius of the largest inscribed circle of the binary image, and only retain the area where the radius of the inscribed circle is greater than R3 ; For further operations on the binary image, only the points whose distance from the center of the circle is less than L1 are retained; the value range of R3 is 15 to 20 pixels, and the value range of L1 is 20 to 30 pixels;

Step 8: For the binary image in Step 7, calculate the minimum circumscribed circle of the binary image, and reserve the binary image whose area of the binary area accounts for more than the circumscribed circle area D2; the value range of D2 is 72% to 90%

Step 9: Recalculate and correct the coordinates of the circumscribed rectangle for the binary image in Step 8; keep the area where the width and height of the circumscribed rectangle are greater than H3, and the difference between the absolute values of width and height is less than H4; the value range of H3 is 50 to 57 pixels, The value range of H4 is 12 to 18 pixels;

Step 10: For the area reserved in Step 9, according to the newly calculated bounding rectangle coordinates, crop the corresponding grayscale image and color image; calculate the variance, grayscale mean and sharpness of the grayscale image; retain the variance greater than F1, and the grayscale mean Greater than G1, the definition is greater than the area of Q1; the value range of F1 is 295~308, the value range of G1 is 55~70, and the value range of Q1 is 11~20;

The following is divided into two processes: training classifier and actual detection;

Steps 11 to 16 are the classifier training process; Steps 17 to 20 are the actual detection process;

Step 11: Compare the coordinates of the region retained in step 10 with the coordinates of the corresponding artificial frame-selected white blood cell region, remove the real white blood cells, and intercept the color image of the remaining area as the impurity of the negative sample set; the positive sample set used for training adopts the artificial frame Selected leukocytes; all parameters used in steps 4 to 10 can ensure that leukocytes in all samples enter the retained area after step 10;

Step 12: Normalize all sample sets and sample images, and then extract the Gabor features and LBP features of the three color channels respectively;

Step 13: Preprocess the features extracted from the sample set in Step 12;

Step 13-1: First, perform PCA (Principal Component Analysis) principal component analysis on the eigenvectors, and only retain 99% of the principal components;

Step 13-2: Perform 'L2' normalization on the features obtained in Step 13-1;

Step 13-3: Standardize the features obtained in Step 13-2;

Step 14: Use the k-fold cross-validation method to divide the sample set into N equal parts, N>3, use one or two of them as the training set each time, and the rest as the test set;

Step 15: For the training set in Step 14, use the SVM (Support Vector Machine) support vector machine model for training to obtain classification model 1;

Step 16: According to the classification model 1 obtained in step 15, train the classification model 2;

Step 17: crop the color image corresponding to the reserved area in step 10, extract Gabor and LBP features and combine them, perform data preprocessing according to the method of step 13, and obtain the feature vector to be measured;

Step 18: Input the feature vector to be tested in Step 17 into Classifier 1 to obtain a classification result;

Step 19: Extract the HOG (Histogram of Oriented Gradient) feature of the color image of the white blood cells as the result in Step 18, perform data preprocessing according to the method in Step 13, and send the processed feature vector to Classifier 2 to obtain the final classification result ;

Step 20: Label all the leukocytes detected in Step 19 on the color microscopic image of Step 1.

Further, the method for local area binarization in the step 3 is as follows: the processing local area is a square, the side length is 11 pixels, the threshold selection method is the gray mean value, and the threshold scaling factor is 0.98.

Further, the sharpness calculation in the step 10 adopts the method of secondary blurring, and the specific steps are:

Step 10-1: Calculate the gradient value of each pixel of the grayscale image to obtain the original gradient map.

Step 10-2: Perform Gaussian blur on the grayscale image, the template size is 9*9, and the variance is 1.5.

Step 10-3: Calculate the gradient value of each pixel for the grayscale image blurred in step 10-2 to obtain a blurred gradient map.

Step 10-4: Calculate the average absolute difference between the original gradient image obtained in step 10-1 and the blurred gradient image obtained in step 10-3, that is, to obtain a sharpness value.

Further, the specific method of the step 12 is:

Scale the image to 80*80 size, and then extract the Gabor features and LBP features of the three color channels respectively; the Gabor feature scale is 15; the direction is {0°, 45°, 90°, 135°, 180°, 225° ,270°,315°}8 directions; wavelength is π/4; spatial aspect ratio is 1.0; standard deviation is 1.5; phase offset is 0; ) feature adopts LBP equivalent pattern (Uniform Pattern), combining Gabor feature and LBP feature.

Further, in the step 15:

When using the SVM support vector machine model for training, select the RBF kernel function, gamma is automatically selected, the penalty coefficient C is 2, and the weight ratio of positive and negative samples is 2:1;

The specific steps in the step 16 are:

Step 16-1: Take all the false negative samples of model 1 in step 15 as the negative sample set of model 2. The positive sample set is the artificially framed white blood cells after scaling.

Step 16-2: Extract the HOG (Histogram of Oriented Gradient) direction gradient histogram feature from the grayscale image of the sample set. The cell size is 8*8, the gradient is 9 directions; the block size is 2*2; the step size is 8. Obtain a 3600-dimensional feature vector.

Step 16-3: Use the data preprocessing method in Step 13 for the features of the sample set to obtain a 2320-dimensional feature vector.

Step 16-4: Adopt the same training method as Model 1, use the SVM model to obtain classification Model 2, and achieve the purpose of further removing falsely detected impurities on the basis of retaining white blood cells. Finally, the RBF kernel function is selected, gamma is automatically selected; the penalty coefficient C is 1; the weight ratio of positive and negative samples is 6:1, and the others are default parameters.

The invention adopts the feature extraction of sample images, uses the extracted features to train the SVM support vector machine, and then applies the trained vector machine to the identification of white blood cells. The method of the invention has the advantages of accurate identification, high speed, high efficiency, and high computational Small size advantage.

Description of drawings

FIG. 1 is a flow chart of an automatic identification method for detecting white blood cells in human feces according to the present invention.

FIG. 2 is a flowchart of the corresponding training classification model.

Figure 3 is the final automatic identification result graph.

Detailed ways

A detection method for human fecal leukocytes, the method comprises the following steps:

Step 1: Pre-processing operations such as dilution, stirring, and precipitation are performed on human feces, and microscopic images of human feces samples are collected using a biological microscope.

Step 2: Grayscale the image in Step 1 and convert it to a grayscale image.

Step 3: Perform local reverse binarization processing on the grayscale image in step 2 to obtain a binarized image. The local processing area is a square with a radius of 11; the threshold selection method is gray mean value, and the threshold scaling factor is 0.98.

Step 4: Perform the hole filling operation on the binary image in Step 3, and then remove the area with an area less than 200.

Step 5: Perform a morphological expansion operation on the binary image in Step 4, using a circular template with a radius of 4. Fill holes in the dilated binary image. The morphological erosion operation is performed on the binary image after hole filling, and a circular template with a radius of 2 is used.

Step 6: Mark the connected region of the binary image in Step 5, and calculate the area feature and circumscribed rectangle of the connected region. The reserved area is greater than 1500, and accounts for more than 70% of the area of the circumscribed rectangle; the width and height of the circumscribed rectangle are greater than 35 and less than 110.

Step 7: For the area reserved in Step 6, crop the binary image according to the circumscribing rectangle. For the cropped binary image, first calculate the center and radius of the largest inscribed circle of the binary image, and only keep the area where the radius of the inscribed circle is greater than 18. For further operations on the binary image, only the points less than 24 from the center of the circle are retained.

Step 8: For the binary image in Step 7, calculate the minimum circumscribed circle of the binary image, and reserve the binary image whose area of the binary area accounts for more than 84% of the area of the circumscribed circle.

Step 9: Recalculate and correct the coordinates of the circumscribed rectangle for the binary image in Step 8. Retain the area where the width and height of the enclosing rectangle are greater than 53, and the difference between the absolute value of the width and height is less than 15.

Step 10: For the area reserved in Step 9, according to the newly calculated coordinates of the circumscribed rectangle, crop the corresponding grayscale image and color image. Calculates the variance, gray mean, and sharpness of a grayscale image. Retain areas with variance greater than 300, gray mean value greater than 60, and sharpness greater than 15.

The following is divided into two processes: training the classifier and actual detection. Steps 11 to 16 are the classifier training process. Steps 17 to 20 are the actual detection process.

Step 11: Compare the coordinates of the region retained in Step 10 with the coordinates of the corresponding artificially framed white blood cell region, remove the real white blood cells, and intercept the color image of the remaining region as impurities in the negative sample set. The positive sample set used for training uses artificially framed white blood cells. All parameters used in steps 4 to 10 can ensure that the leukocytes in all samples enter the retained area after step 10.

Step 12: For all sample sets, scale the image to 80*80 size, and then extract the Gabor features and LBP features of the three color channels respectively. Among them, the Gabor feature scale is 15; the direction is {0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°} 8 directions; the wavelength is π/4; the spatial aspect ratio is 1.0; The standard deviation is 1.5; the phase offset is 0. At the same time, 8 times downsampling is used to reduce the feature dimension. A total of 2400 dimensional features were obtained. Among them, the LBP (Local Binary Pattern) feature adopts the LBP equivalent pattern (Uniform Pattern), and a total of 78-dimensional features are extracted. Combine Gabor features and LBP features to obtain a 2478-dimensional feature vector.

Step 13: Preprocess the features extracted from the sample set in Step 12 to obtain a 1342-dimensional feature vector.

Step 14: Using the k-fold cross-validation method, the sample set is divided into 5 equal parts, 1 part is used as the training set each time, and the remaining 4 parts are used as the test set.

Step 15: Use the SVM (Support Vector Machine) support vector machine model for training on the training set in step 14 to obtain a classification model 1. According to the results of cross-validation, the RBF kernel function is finally selected, gamma is automatically selected; the penalty coefficient C is 2; the weight ratio of positive and negative samples is 2:1, and the others are default parameters.

Step 16: According to the classification model 1 obtained in step 15, the classification model 2 is trained.

Step 17: Crop the color image corresponding to the reserved area in step 10, extract Gabor and LBP features and combine them, perform data preprocessing according to the method of step 13, and obtain the feature vector to be measured.

Step 18: Input the feature vector to be tested in Step 17 to Classifier 1 to obtain a classification result.

Step 19: Extract the HOG (Histogram of Oriented Gradient) feature of the color image of the white blood cells as the result in Step 18, perform data preprocessing according to the method in Step 13, and send the processed feature vector to Classifier 2 to obtain the final classification result .

Step 20: Label all the leukocytes detected in Step 19 on the color microscopic image of Step 1.

The sharpness calculation in the step 10 adopts the method of secondary blurring, and the specific steps are:

Step 10-1: Calculate the gradient value of each pixel of the grayscale image to obtain the original gradient map.

Step 10-2: Perform Gaussian blur on the grayscale image, the template size is 9*9, and the variance is 1.5.

Step 10-3: Calculate the gradient value of each pixel for the grayscale image blurred in step 10-2 to obtain a blurred gradient map.

Step 10-4: Calculate the average absolute difference between the original gradient image obtained in step 10-1 and the blurred gradient image obtained in step 10-3, that is, to obtain a sharpness value.

The specific steps in the step 13 are:

Step 13-1: First, perform PCA (Principal Component Analysis) principal component analysis on the 2478-dimensional feature vector, retain only 99% of the principal components, and obtain a 1342-dimensional feature vector.

Step 13-2: Perform 'L2' normalization on the features obtained in Step 13-1.

Step 13-3: Standardize the features obtained in Step 13-2.

The specific steps in the step 16 are:

Step 16-1: Take all the false negative samples of model 1 in step 15 as the negative sample set of model 2. The positive sample set is the artificially framed white blood cells after scaling.

Step 16-2: Extract the HOG (Histogram of Oriented Gradient) direction gradient histogram feature from the grayscale image of the sample set. The cell size is 8*8, the gradient is 9 directions; the block size is 2*2; the step size is 8. Obtain a 3600-dimensional feature vector.

Step 16-3: Use the data preprocessing method in Step 13 for the features of the sample set to obtain a 2320-dimensional feature vector.

Step 16-4: Adopt the same training method as Model 1 and use the SVM model to obtain Classification Model 2, so as to achieve the purpose of further removing falsely detected impurities on the basis of retaining white blood cells. Finally, the RBF kernel function is selected, gamma is automatically selected; the penalty coefficient C is 1; the weight ratio of positive and negative samples is 6:1, and the others are default parameters.

Claims (5)

1. A method for automatic identification of human fecal leukocytes based on machine learning, the method comprising the following steps: Step 1: Pretreatment of human feces by dilution, stirring and precipitation, and using a biological microscope to collect microscopic images of human feces samples; Step 2: Grayscale the image in Step 1 and convert it to a grayscale image; Step 3: perform local reverse binarization processing on the grayscale image in step 2 to obtain a binarized image; Step 4: Perform a hole filling operation on the binary image in Step 3, and fill the area with an area smaller than S1, where the value range of S1 is 180-230 pixels; Step 5: Perform a morphological expansion operation on the binary image in step 4, using a circular template with a radius of R1; perform hole filling on the expanded binary image; perform a morphological erosion operation on the hole-filled binary image , using a circular template with a radius of R2; the value range of R1 is 2 to 5 pixels, the value range of R2 is 1 to 3 pixels, and R1 is greater than R2; Step 6: Mark the connected area of the binary image in step 5, and calculate the area feature and circumscribed rectangle of the connected area; the reserved area is greater than S2, and the proportion of the area of the circumscribed rectangle is greater than D1; the width and height of the circumscribed rectangle is greater than H1, and smaller than the area of H2 ; The value range of S2 is 1400~1550 pixels, the value range of D1 is 60%~80%, the value range of H1 is 30~50 pixels, and the value range of H2 is 100~120 pixels; Step 7: For the area reserved in step 6, crop the binary image according to the circumscribed rectangle; for the cropped binary image, first calculate the center and radius of the largest inscribed circle of the binary image, and only retain the area where the radius of the inscribed circle is greater than R3 ; For further operations on the binary image, only the points whose distance from the center of the circle is less than L1 are retained; the value range of R3 is 15 to 20 pixels, and the value range of L1 is 20 to 30 pixels; Step 8: For the binary image in Step 7, calculate the minimum circumscribed circle of the binary image, and reserve the binary image whose area of the binary area accounts for more than the circumscribed circle area D2; the value range of D2 is 72% to 90%; Step 9: Recalculate and correct the coordinates of the circumscribed rectangle for the binary image in Step 8; keep the area where the width and height of the circumscribed rectangle are greater than H3, and the difference between the absolute values of width and height is less than H4; the value range of H3 is 50 to 57 pixels, The value range of H4 is 12 to 18 pixels; Step 10: For the area reserved in step 9, according to the newly calculated bounding rectangle coordinates, crop the corresponding grayscale image and color image; calculate the variance, grayscale mean and sharpness of the grayscale image; retain the variance greater than F1, and the grayscale mean Greater than G1, the definition is greater than the area of Q1; the value range of F1 is 295~308, the value range of G1 is 55~70, and the value range of Q1 is 11~20; The following is divided into two processes: training classifier and actual detection; Steps 11 to 16 are the classifier training process; Steps 17 to 20 are the actual detection process; Step 11: Compare the coordinates of the region retained in step 10 with the coordinates of the corresponding artificial frame-selected leukocyte region, remove the real leukocytes, and intercept the color image of the remaining area as the impurity of the negative sample set; the positive sample set used for training adopts the artificial frame Selected leukocytes; all parameters used in steps 4 to 10 can ensure that leukocytes in all samples enter the retained area after step 10; Step 12: Normalize all sample sets and sample images, and then extract the Gabor features and LBP features of the three color channels respectively; Step 13: Preprocess the features extracted from the sample set in Step 12; Step 13-1: First perform PCA principal component analysis on the eigenvectors, and only retain 99% of the principal components; Step 13-2: Perform 'L2' normalization on the features obtained in Step 13-1; Step 13-3: Standardize the features obtained in Step 13-2; Step 14: Use the k-fold cross-validation method to divide the sample set into N equal parts, N>3, use one or two of them as the training set each time, and the rest as the test set; Step 15: For the training set in Step 14, use the SVM support vector machine model for training to obtain classification model 1; Step 16: According to the classification model 1 obtained in step 15, train the classification model 2; Step 16-1: Take all the false negative samples of Model 1 in Step 15 as the negative sample set of Model 2, and the positive sample set is the scaled artificial white blood cells; Step 16-2: Extract the HOG direction gradient histogram feature from the grayscale image of the sample set; the cell size is 8*8, the gradient is 9 directions, the block size is 2*2, and the step size is 8 to obtain a 3600-dimensional feature vector ; Step 16-3: Use the data preprocessing method in Step 13 for the features of the sample set to obtain a 2320-dimensional feature vector; Step 16-4: Adopt the same training method as model 1, use the SVM model to obtain classification model 2, and achieve the purpose of further removing falsely detected impurities on the basis of retaining white blood cells; finally select the RBF kernel function, gamma is automatically selected; penalty coefficient C takes 1; the weight ratio of positive and negative samples is 6:1, and the others are default parameters; Step 17: crop the color image corresponding to the reserved area in step 10, extract Gabor and LBP features and combine them, perform data preprocessing according to the method of step 13, and obtain the feature vector to be measured; Step 18: Input the feature vector to be tested in Step 17 into Classifier 1 to obtain a classification result; Step 19: Extract the HOG feature of the color image of the white blood cells as the result in Step 18, perform data preprocessing according to the method of Step 13, and send the processed feature vector to Classifier 2 to obtain the final classification result; Step 20: Label all the leukocytes detected in Step 19 on the color microscopic image of Step 1. 2. a kind of human fecal leukocyte automatic identification method based on machine learning as claimed in claim 1, it is characterized in that in described step 3, the method for local area binarization is: the area of processing local is square, and side length is 11 pixels, the threshold selection method is gray mean value, and the threshold scaling factor is 0.98. 3. a kind of human fecal leukocyte automatic identification method based on machine learning as claimed in claim 1 is characterized in that the sharpness calculation in described step 10 adopts the method of secondary blurring, and concrete steps are: Step 10-1: Calculate the gradient value of each pixel of the grayscale image to obtain the original gradient map; Step 10-2: Perform Gaussian blur on the grayscale image, the template size is 9*9, and the variance is 1.5; Step 10-3: Calculate the gradient value of each pixel on the grayscale image blurred in step 10-2 to obtain a blurred gradient map; Step 10-4: Calculate the average absolute difference between the original gradient image obtained in step 10-1 and the blurred gradient image obtained in step 10-3, that is, to obtain a sharpness value. 4. a kind of human fecal leukocyte automatic identification method based on machine learning as claimed in claim 1 is characterized in that the concrete method of described step 12 is: Scale the image to 80*80 size, and then extract the Gabor features and LBP features of the three color channels respectively; the Gabor feature scale is 15; the direction is {0°, 45°, 90°, 135°, 180°, 225° ,270°,315°}8 directions; wavelength is π/4; spatial aspect ratio is 1.0; standard deviation is 1.5; phase offset is 0; Valence mode, combining Gabor features and LBP features. 5. a kind of human fecal leukocyte automatic identification method based on machine learning as claimed in claim 1 is characterized in that in described step 15: When using the SVM support vector machine model for training, select the RBF kernel function, gamma is automatically selected, the penalty coefficient C is 2, and the weight ratio of positive and negative samples is 2:1.

Images