A digital grading system for blood smears stained with alkaline phosphatase

Abstract

This invention relates to the field of cell morphology medical testing, specifically a digital grading system for blood smears stained with alkaline phosphatase. It includes: an image processing module to obtain images of neutrophils and reference cells; a cell segmentation module to obtain images of the cytoplasm; a feature extraction module to acquire the color and texture features of the reference cells and neutrophils; and a grading module that uses an MLP classifier to obtain the grade of each neutrophil, aggregates the scores of 100 neutrophils, and compares them with a preset threshold to obtain the grading result. Through the above technical solution, this invention solves the problems of low standardization and strong subjectivity in grading results due to the influence of operator skill level.

Description

Technical Field

[0001] This invention relates to the field of cell morphology medical testing, specifically a digital grading system for blood smears stained with alkaline phosphatase. Background Technology

[0002] Neutrophil alkaline phosphatase (NAP), a marker of granulocyte maturation, first appears in the myeloid cell stage and increases with granulocyte maturation. Elevated NAP activity is seen in bacterial infections, leukemia-like reactions, aplastic anemia, multiple myeloma, neuroblastoma, acute lymphoblastic leukemia, and acute transformation of chronic myeloid leukemia; decreased NAP activity is seen in chronic myeloid leukemia, acute myeloid leukemia, chloroma, erythroleukemia, PNH, and viral infections.

[0003] NAP staining has some reference value in the differential diagnosis of the following diseases: chronic myeloid leukemia and leukemia-like reaction, aplastic anemia and PNH, and differentiation of cell types in acute leukemia. The NAP activity score is based on the number and distribution of NAP-positive granules in the cytoplasm. The scores of each cell are added together to obtain the total NAP score. Manual NAP score determination does not require complex equipment; only routine blood smears and a microscope are needed, making the operation simple and inexpensive. However, the accuracy of the NAP score is greatly affected by the operator's skill level, and the standardization of the results is low. Furthermore, the interpretation of the results is highly subjective; judging the NAP score requires scoring based on cell morphology under a microscope, which is easily influenced by the observer's experience and other subjective factors. Summary of the Invention

[0004] In order to solve the problems in related technologies, the present invention provides a digital grading system for blood smears stained with alkaline phosphatase. This system solves the problems of low standardization of grading results and strong subjectivity in judgment due to the influence of the operator's technical level.

[0005] To solve the above problems, the following technical solutions are provided:

[0006] The present invention provides a digital grading system for blood smears stained with alkaline phosphatase, comprising:

[0007] The image processing module is used to preprocess the blood image, cropping it to obtain X1 preprocessed images of the same size, brightness, and high boundary contrast; process the X1 preprocessed images to detect and crop them to obtain X2 white blood cell images; classify the X2 white blood cell images to obtain neutrophil images and reference cell images.

[0008] The cell segmentation module performs preliminary segmentation on the neutrophil image and the reference cell image, performs background separation processing on the neutrophils and the reference cells respectively, and then performs deep segmentation on the neutrophils and the reference cells to separate the cytoplasm and the nucleus, thus obtaining cytoplasm images and nucleus images;

[0009] The feature extraction module includes color feature extraction and texture feature extraction. Color feature extraction calculates the statistical values ​​of cytoplasmic region pixels for neutrophils and reference cells in seven channels across three color spaces (RGB, HSV, and grayscale). Then, it calculates the difference between the statistical values ​​of neutrophils and reference cells in each channel to obtain a color difference feature vector. Texture feature extraction performs grayscale processing on the cytoplasmic image of neutrophils and then calculates the texture feature vector.

[0010] The grading module fuses color difference feature vectors and texture feature vectors to obtain a fused vector. This fused vector is then input into an MLP classifier to output the grade of each neutrophil. Used for all neutrophil grades The score is calculated and compared with the preset threshold to obtain the classification result.

[0011] By adopting the above scheme, the collaborative work of each module ensures the uniformity of quality between neutrophil images and reference cell images. The cell segmentation module accurately separates the cytoplasm and nucleus, and the feature extraction module extracts features from multiple dimensions such as color and texture. Combined with an MLP classifier, it performs grading, reducing interference from human factors and making the grading results more stable and consistent. The standardization level is significantly improved, thus solving the problem of low standardization of grading results and strong subjectivity in judgment due to the influence of the operator's technical level. From the processing of blood images, white blood cell detection and classification in image preprocessing, to the separation of cells from background and cytoplasm from nucleus in the cell segmentation module, and then to feature extraction and grading, the entire process is automated without manual operation, saving a lot of labor costs and improving grading efficiency.

[0012] The statistic used in the color feature extraction is the mean. Standard deviation skewness and kurtosis Taking the R channel in the RGB color space as an example, the calculation process is as follows:

[0013] ,

[0014] ,

[0015] in, This represents the number of pixels within the cytoplasm region. For pixel values, This represents the mean of the R channel;

[0016] For each statistic, the difference features between neutrophils and reference cells were calculated separately. The following is the color difference feature vector of the R channel:

[0017] Color difference feature vector of R-channel mean: ,

[0018] Color difference feature vector of R-channel standard deviation: ,

[0019] R-channel skewness color difference feature vector: ,

[0020] R-channel kurtosis color difference feature vector: ,

[0021] in, This represents the mean value of R channels in neutrophils. The mean value of R channels in reference cells.

[0022] By employing the above scheme, the overall staining intensity of different samples may vary due to factors such as staining reagent concentration and staining time during the preparation of blood smears, while neutrophils and reference cells within the same field of view exhibit similar staining degrees. By calculating the differences in mean, standard deviation, skewness, and kurtosis across seven channels in the RGB, HSV, and grayscale color spaces, systematic errors caused by fluctuations in overall staining conditions can be effectively offset, reducing the interference of non-specific differences between individual samples on color feature extraction.

[0023] In existing manual grading, operators lack quantitative standards for judging the "staining depth". This invention converts the color difference between neutrophils and reference cells into an objective value through difference calculation, avoiding errors caused by inconsistent perception of the "normal staining baseline" in subjective judgment by operators, and reducing errors in subjective judgment.

[0024] The specific process for calculating the feature map vector in the texture feature extraction is as follows:

[0025] The cytoplasm image of neutrophils was processed into a grayscale image. The gradient response plot is calculated using the Sobel operator. ,

[0026] in, , ,

[0027] Then, the Laws texture energy filter bank is used to generate five response maps, and a local energy feature map is calculated for each response map:

[0028] ,

[0029] in, For the first One Laws filter template;

[0030] Then, calculate the four eigenvectors of the gray-level co-occurrence matrix: Energy, Correlation, Homogeneity, and Contrast.

[0031] ,

[0032] ,

[0033] :

[0034] , ,

[0035] , ,

[0036] Global average pooling is performed on the multi-channel response map and the local energy feature map respectively to obtain two sets of channel feature vectors. Finally, these two sets of channel feature vectors are concatenated with the gray-level co-occurrence matrix feature vector to form the final texture feature vector.

[0037] By adopting the above scheme, the multi-channel response map is used to capture macroscopic texture features such as the edge contour and directionality of neutrophils; the local energy feature map can quantify the local aggregation intensity and distribution pattern of particles, reflecting the texture activity of microscopic regions; the gray-level co-occurrence matrix feature focuses on the spatial correlation between pixels, describing the uniformity of particle distribution, contrast and other deep texture attributes; after the three are fused, the texture information related to neutrophils is fully covered from macroscopic to microscopic and from local to global, avoiding the limitations of a single feature dimension and significantly improving the comprehensiveness and accuracy of texture feature extraction.

[0038] The specific process for obtaining the fusion vector is as follows:

[0039] The fusion vector is obtained by integrating all color difference feature vectors and texture feature vectors of the seven channels.

[0040] The calculation process of the hierarchical module is as follows:

[0041] Grading of all neutrophils in the blood image Summation, then multiplication by a coefficient That is, the scores are aggregated into 100 neutrophils to obtain the total score. :

[0042] ,

[0043] in, The total number of neutrophils in the blood image;

[0044] Total score By comparing the result with a preset threshold range, the final classification result can be obtained.

[0045] Grading results = .

[0046] By adopting the above scheme, the grades of all neutrophils are converted into a standardized total score, eliminating the influence of differences in cell number, and clarifying the threshold range (S<35 indicates reduced activity, 35≤S<100 indicates normal activity, and S≥100 indicates increased activity). This transforms the diagnostic results from subjective description to objective numerical judgment, reduces human statistical errors, and improves the accuracy of result judgment.

[0047] The specific steps for obtaining the preprocessed image in the image processing module are as follows:

[0048] A separable 5×5 Gaussian filter was applied to the magnified view of the acquired blood images.

[0049] ,

[0050] To suppress high-frequency noise and preserve cell edge details; subsequently, in HSV space... Adaptive histogram equalization is performed on the V channel.

[0051] ,

[0052] To enhance cell boundary contrast, an affine transformation matrix is ​​finally applied.

[0053] ,

[0054] By performing rotation correction and scale normalization on the image, a preprocessed image with consistent size, brightness, and prominent boundary contrast can be obtained.

[0055] By adopting the above method, we can ensure that all preprocessed images are uniform in size, angle, and scale, thereby reducing errors caused by differences in image quality.

[0056] The specific steps of the cell segmentation are as follows: Segmentation is performed on the reference cell and neutrophils respectively, and the data are fed into a UNet architecture that integrates attention gates at skip connections to generate cytoplasmic masks end-to-end in one go. With nuclear mask The Dice+BCE hybrid loss is used for training to ensure high-precision segmentation of small structures. Each cell image includes an inner black nucleus region, a middle stained cytoplasm, and an outermost black background region. Segmentation yields a cytoplasm image.

[0057] By adopting the above scheme, dual high-precision segmentation is achieved, separating cells from the background and separating cytoplasm from the nucleus. This provides accurate regional localization for subsequent feature extraction targeting only the cytoplasm region, avoiding feature contamination caused by segmentation deviations.

[0058] The reference cells are any one or both of monocytes and lymphocytes.

[0059] The above solution has the following advantages:

[0060] 1. The present invention provides a digital grading system for blood smears stained with alkaline phosphatase. Through the collaborative work of various modules, the image processing module ensures the uniformity of quality between neutrophil and reference cell images, the cell segmentation module accurately separates the cytoplasm and nucleus, and the feature extraction module extracts features from multiple dimensions such as color and texture. Combined with an MLP classifier for grading, it reduces interference from human factors, making the grading results more stable and consistent, and significantly improving the degree of standardization. This solves the problem of low standardization and strong subjectivity in grading results due to the influence of the operator's technical level. From the processing of blood images, white blood cell detection and classification in image preprocessing, to the separation of cells from background and cytoplasm from nucleus in the cell segmentation module, and then to feature extraction and grading, the entire process is automated without manual operation, saving a lot of labor costs and improving grading efficiency.

[0061] 2. By employing color feature extraction, it is possible that during the preparation of blood smears, factors such as staining reagent concentration and staining time may lead to differences in the overall staining intensity of different samples, while neutrophils and reference cells within the same field of view exhibit similar staining degrees. By calculating the differences in mean, standard deviation, skewness, and kurtosis between the two cells across seven channels in the RGB, HSV, and grayscale color spaces, systematic errors caused by fluctuations in overall staining conditions can be effectively offset, reducing the interference of non-specific differences between individual samples on color feature extraction. This invention transforms the color difference between neutrophils and reference cells into objective numerical values ​​through difference calculation, avoiding errors caused by inconsistent perceptions of the "normal staining baseline" in subjective judgment by operators, and reducing errors from subjective judgment.

[0062] 3. By employing texture feature extraction, multi-channel response maps are used to capture macroscopic texture features such as neutrophil edge contours and directionality; local energy feature maps can quantify the local aggregation intensity and distribution pattern of particles, reflecting the texture activity of microscopic regions; gray-level co-occurrence matrix features focus on the spatial correlation between pixels, describing deep texture attributes such as the uniformity and contrast of particle distribution; after the fusion of the three, the texture information related to neutrophils is comprehensively covered from macroscopic to microscopic and from local to global, avoiding the limitations of a single feature dimension and significantly improving the comprehensiveness and accuracy of texture feature extraction. Attached Figure Description

[0063] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0064] Figure 1 A flowchart of a digital grading system for blood smears stained with alkaline phosphatase;

[0065] Figure 2 This is a framework diagram of a digital grading system for blood smears stained with alkaline phosphatase. Detailed Implementation

[0066] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] In a specific embodiment, such as Figures 1-2 As shown, the present invention provides a digital grading system for blood smears stained with alkaline phosphatase, comprising:

[0068] S1: The image processing module includes an image preprocessing module, a white blood cell detection module, and a cell classification module;

[0069] S1-1: Image preprocessing module,

[0070] A separable 5×5 Gaussian filter was applied to the magnified view of the acquired blood images.

[0071] ,

[0072] To suppress high-frequency noise and preserve cell edge details; subsequently, in HSV space... Adaptive histogram equalization is performed on the V channel.

[0073] ,

[0074] To enhance cell boundary contrast, an affine transformation matrix is ​​finally applied.

[0075] ,

[0076] By performing rotation correction and scale normalization on the image, a preprocessed image with consistent X1 size, consistent brightness, and prominent boundary contrast can be obtained.

[0077] S1-2: White blood cell detection module.

[0078] The trained YOLO model was used to detect white blood cells in X1 preprocessed images. Finally, the detected cells were cropped to obtain the desired result. Image of white blood cells.

[0079] S1-3: Cell classification module

[0080] The system receives After taking images of white blood cells, a trained ResNet model was used to classify the white blood cells into two categories: neutrophils and reference cells. The reference cells were either monocytes or lymphocytes. The reference cells did not contain neutrophil alkaline phosphatase, and their cytoplasm could be used as a reference for negative results, i.e., a score of 0.

[0081] S2: Cell segmentation module

[0082] Segmentation is performed on reference cells and neutrophils separately, and the data is fed into a UNet architecture that integrates attention gates at skip connections to generate cytoplasmic masks end-to-end in a single operation. With nuclear mask The Dice+BCE hybrid loss is used for training to ensure high-precision segmentation of small structures. Each cell image includes an inner black nucleus region, a middle stained cytoplasm, and an outermost black background region. Segmentation yields a cytoplasm image.

[0083] S3: Feature extraction module, which includes color feature extraction and texture feature extraction;

[0084] S3-1: Color Feature Extraction

[0085] Color feature extraction was used to calculate the statistical measures of cytoplasmic region pixels in seven channels across three color spaces (RGB, HSV, and GRAY) for neutrophils and reference cells. The statistical measures were the mean values. Standard deviation skewness and kurtosis The following are the feature calculations for the R channel, G channel, B channel, H channel, S channel, V channel, and GRAY channel:

[0086] ,

[0087] ,

[0088] ,

[0089] ,

[0090] ,

[0091] ,

[0092] ,

[0093] ,

[0094] ,

[0095] ,

[0096] ,

[0097] ,

[0098] ,

[0099] ,

[0100] in, This represents the number of pixels within the cytoplasm region. For pixel values, This represents the mean of the R channel. The standard deviation of the R channel. The skewness of the R channel. This represents the kurtosis of the R channel; similarly, the kurtosis of the other channels can be obtained by using the corresponding letters.

[0101] For each statistic, the difference features between neutrophils and reference cells were calculated separately. The following is the color difference feature vector for the 7 channels:

[0102] Color difference feature vector of R-channel mean: ,

[0103] Color difference feature vector of R-channel standard deviation: ,

[0104] R-channel skewness color difference feature vector: ,

[0105] R-channel kurtosis color difference feature vector: ,

[0106] Color difference feature vector of G channel mean: ,

[0107] Color difference feature vector of G channel standard deviation: ,

[0108] G-channel skewness color difference feature vector: ,

[0109] G-channel kurtosis color difference feature vector: ,

[0110] Color difference feature vector of B channel mean: ,

[0111] Color difference feature vector of B-channel standard deviation: ,

[0112] B-channel skewness color difference feature vector: ,

[0113] B-channel kurtosis color difference feature vector: ,

[0114] Color difference feature vector of H channel mean: ,

[0115] H-channel standard deviation color difference feature vector: ,

[0116] H-channel skewness color difference feature vector: ,

[0117] H-channel kurtosis color difference feature vector: ,

[0118] Color difference feature vector of S-channel mean: ,

[0119] Color difference feature vector of S-channel standard deviation: ,

[0120] S-channel skewness color difference feature vector: ,

[0121] S-channel kurtosis color difference feature vector: ,

[0122] Color difference feature vector of V channel mean: ,

[0123] Color difference feature vector of V channel standard deviation: ,

[0124] V-channel skewness color difference feature vector: ,

[0125] V-channel kurtosis color difference feature vector: ,

[0126] Color difference feature vector of GRAY channel mean: ,

[0127] Color difference feature vector of GRAY channel standard deviation: ,

[0128] GRAY channel skewness color difference feature vector: ,

[0129] GRAY channel kurtosis color difference feature vector: ,

[0130] in, This represents the mean value of R channels in neutrophils. As a reference cell R channel mean, This represents the mean value of H channels in neutrophils. To reference the mean value of H channels in cells, This represents the mean value of the gray channels for neutrophils. The mean value of the gray channel of the reference cell;

[0131] S3-2: Texture Feature Extraction

[0132] The cytoplasm image of neutrophils was processed into a grayscale image. The gradient response plot is calculated using the Sobel operator. ,

[0133] in, , ,

[0134] Then, the Laws texture energy filter bank is used to generate five response maps, and a local energy feature map is calculated for each response map:

[0135] ,

[0136] in, For the first One Laws filter template;

[0137] Then, calculate the four eigenvectors of the gray-level co-occurrence matrix: Energy, Correlation, Homogeneity, and Contrast.

[0138] ,

[0139] ,

[0140] :

[0141] , ,

[0142] , ,

[0143] Global average pooling is performed on the multi-channel response map and the local energy feature map respectively to obtain two sets of channel feature vectors. Finally, these two sets of channel feature vectors are concatenated with the gray-level co-occurrence matrix feature vector to form the final texture feature vector.

[0144] S4: Hierarchical module

[0145] By integrating all color difference feature vectors and texture feature vectors from the seven channels, a fused vector is obtained. This fused vector is then input into an MLP classifier to output the rank of each neutrophil. .

[0146] Then, the grade of all neutrophils in the blood image was determined. Summation, then multiplication by a coefficient , get the total score :

[0147] ,

[0148] in, The total number of neutrophils in the blood image;

[0149] Grading results = .

[0150] Total score By comparing the result with the preset threshold range, the final classification result can be obtained.

[0151] Obviously, the above embodiments are merely examples for clear illustration and are not intended to limit the implementation. For those skilled in the art, other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all implementation methods here, and the obvious variations or modifications derived therefrom are still within the protection scope of this invention.

Claims

1. A digital grading system for blood smears stained with alkaline phosphatase, characterized in that, include: The image processing module is used to preprocess the blood image and crop it to obtain X1 preprocessed images of the same size, brightness and high edge contrast. X1 preprocessed images are processed, detected, and cropped to obtain X2 images of white blood cells; the X2 images of white blood cells are classified to obtain images of neutrophils and reference cells, wherein the reference cells are any one or both of monocytes and lymphocytes; The cell segmentation module performs preliminary segmentation on the neutrophil image and the reference cell image, performs background separation processing on the neutrophils and the reference cells respectively, and then performs deep segmentation on the neutrophils and the reference cells to separate the cytoplasm and the nucleus, thus obtaining cytoplasm images and nucleus images; The feature extraction module performs color feature extraction and texture feature extraction. Color feature extraction calculates the mean values ​​of pixels in the cytoplasm region for neutrophils and reference cells across seven channels in the RGB, HSV, and grayscale color spaces. Standard deviation skewness and kurtosis Then, the mean values ​​of neutrophils and reference cells in each channel were calculated separately. Standard deviation skewness kurtosis The difference is used to obtain the color difference feature vector; texture feature extraction is used to process the grayscale image of neutrophil cytoplasm, and then calculate the texture feature vector. The grading module fuses color difference feature vectors and texture feature vectors to obtain a fused vector. This fused vector is then input into an MLP classifier to output the grade of each neutrophil. ; Grades for receiving all neutrophils The score is calculated and compared with the preset threshold to obtain the classification result.

2. The digital grading system for blood smears stained with alkaline phosphatase as described in claim 1, characterized in that, Taking the R channel in the RGB color space as an example, the calculation process is as follows: ， ， in, This represents the number of pixels within the cytoplasm region. For pixel values, This represents the mean of the R channel; For each mean Standard deviation skewness and kurtosis The differential features between neutrophils and reference cells were calculated separately. The following is the color difference feature vector of the R channel: Color difference feature vector of R-channel mean: , Color difference feature vector of R-channel standard deviation: , R-channel skewness color difference feature vector: , R-channel kurtosis color difference feature vector: , in, This represents the mean value of R channels in neutrophils. The mean value of R channels in reference cells; The calculation methods for the G and B channels in the RGB color space, the H, S, and V channels in the HSV color space, and the GRAY channel in the grayscale color space are the same as those for the R channel.

3. The digital grading system for blood smears for alkaline phosphatase staining as described in claim 1, characterized in that, The specific process for calculating the feature map vector in the texture feature extraction is as follows: The cytoplasm image of neutrophils was processed into a grayscale image. The gradient response plot is calculated using the Sobel operator. , in, , , Then, the Laws texture energy filter bank is used to generate five response maps, and a local energy feature map is calculated for each response map: ， in, For the first One Laws filter template; Then, calculate the four eigenvectors of the gray-level co-occurrence matrix: Energy, Correlation, Homogeneity, and Contrast. ， ， ： ， ， ， ， Global average pooling is performed on the multi-channel response map and the local energy feature map respectively to obtain two sets of channel feature vectors. Finally, these two sets of channel feature vectors are concatenated with the gray-level co-occurrence matrix feature vector to form the final texture feature vector.

4. The digital grading system for blood smears for alkaline phosphatase staining as described in claim 1, characterized in that, The specific process for obtaining the fusion vector is as follows: The fusion vector is obtained by integrating all color difference feature vectors and texture feature vectors of the seven channels.

5. The digital grading system for blood smears for alkaline phosphatase staining as described in claim 1, characterized in that, The calculation process of the hierarchical module is as follows: Grading of all neutrophils in the blood image Summation, then multiplication by a coefficient , get the total score : ， in, The total number of neutrophils in the blood image; Total score By comparing the result with a preset threshold range, the final classification result can be obtained. Grading results = .

6. The digital grading system for blood smears for alkaline phosphatase staining as described in claim 1, characterized in that, The specific steps for obtaining the preprocessed image in the image processing module are as follows: A separable 5×5 Gaussian filter was applied to the magnified view of the acquired blood images. ， To suppress high-frequency noise and preserve cell edge details; subsequently, in HSV space... Adaptive histogram equalization is performed on the V channel. ， To enhance cell boundary contrast, an affine transformation matrix is ​​finally applied. ， By performing rotation correction and scale normalization on the image, a preprocessed image with consistent size, brightness, and prominent boundary contrast can be obtained.

7. The digital grading system for blood smears for alkaline phosphatase staining as described in claim 1, characterized in that, The specific steps of cell segmentation are as follows: Segmentation is performed on reference cells and neutrophils separately, and the data is fed into a UNet architecture that integrates attention gates at skip connections to generate cytoplasmic masks end-to-end in a single operation. With nuclear mask The Dice+BCE hybrid loss is used for training to ensure high-precision segmentation of small structures. Each cell image includes an inner black nucleus region, a middle stained cytoplasm, and an outermost black background region. Segmentation yields a cytoplasm image.

Images