Construction method and system of blood cell segmentation network
By pre-processing, screening and feature analysis of blood cell images, error-free cell images are screened out and blood cell segmentation network model is constructed, which solves the problems of limited blood cell segmentation performance and insufficient recognition accuracy in the prior art, and achieves higher segmentation accuracy and model reliability.
Patent Information
- Application Number
- CN202510309150.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-17
- Publication Date
- 2025-06-20
AI Technical Summary
The existing blood cell segmentation method has limited segmentation performance when cell boundaries are blurred and cell overlapping, and lacks steps to carefully preprocess and screen images, resulting in model training being disturbed by noise and abnormal data, and lacks special design for blood cell image features, affecting the identification accuracy of similar cells.
By collecting and pretreating blood cell images, screening and segmenting to remove non-cell regions and abnormal cells, feature extraction and feature matching analysis are performed to identify similar cells, and error-free cell images are screened through error analysis and detection, and a blood cell segmentation network model is constructed for training and testing.
It improves the accuracy of blood cell image segmentation, reduces the impact of noise and abnormal data on model training, enhances the recognition ability of similar cells, and improves the accuracy and reliability of the model.
Smart Images
Figure CN120182232A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and more particularly to a method and system for constructing a blood cell segmentation network. Background Art
[0002] With the rapid development of medical imaging technology, the automatic recognition and segmentation of blood cells have become an important research direction in clinical diagnosis and disease monitoring. Traditional blood cell analysis methods mostly rely on manual microscopic observation, which is time-consuming and easily affected by human factors, resulting in inconsistent results. Therefore, the method of segmenting blood cells through deep learning technology has emerged. It mainly automatically segments blood cell images, can automatically extract features in the images, and learn complex features in cell images, adapt to different types of blood cell images, improve the accuracy of segmentation, and provide a new solution for the automatic recognition of blood cells. However, the above process still has the following disadvantages: Firstly, the existing blood cell segmentation methods are limited in segmentation performance in the case of blurred cell boundaries and cell overlaps, resulting in the problem of insufficient cell segmentation accuracy, and lacking steps for detailed preprocessing and screening of images, which may cause the model to be interfered by noise and abnormal data during the training process. Secondly, the existing blood cell segmentation methods lack special designs for the characteristics of blood cell images and lack means for further feature matching analysis of blood cell images, which may affect the recognition accuracy of similar cells in cell images. Summary of the Invention
[0003] In order to overcome the above defects of the prior art, the present invention provides a method and system for constructing a blood cell segmentation network to solve the problems existing in the above background art.
[0004] The present invention provides the following technical solutions: A method for constructing a blood cell segmentation network, comprising: S1: Collecting original blood cell images and preprocessing the collected original blood cell images; S2: Screening the preprocessed original blood cell images to screen out normal cell images, removing non-cell regions and abnormal cell images, and using the screened blood cell images as cell region images; S3: Segmenting the cell region images into independent cell images as cell images through the gray-scale difference between cells and the background, and labeling each cell in the segmented cell images; S4: Extracting features from the cell images and using a feature matching algorithm to analyze similar cells in the cell images to obtain a matching degree coefficient, and pairing the similar cells. S5: Based on the characteristic parameters of the cell image, perform error analysis on the cell image to obtain an error evaluation coefficient, and evaluate the error of the cell image through the error detection coefficient; S6: Based on the error analysis result, compare it with a preset error threshold to further detect the error of the cell image, and screen out the cell images without error; S7: Use the cell images without error as the input data for the construction and prediction of the blood cell segmentation network model, and use them for model training and testing.
[0005] Preferably, in S1, a high-resolution microscope camera is used to take a blood smear, so as to collect high-quality blood cell images, and record the relevant information of each image, including the shooting time, microscopic parameters, and cell type. The collected high-quality blood cell images are used as the original blood cell images, and the original blood cell images are subjected to image conversion, size adjustment, grayscale conversion, image contrast enhancement, image denoising, and illumination correction.
[0006] Preferably, in S2, feature extraction is performed on the original blood cell images, the cell area, cell perimeter, cell roundness, and cell texture in the images are extracted, and then statistical analysis is performed on the extracted cell features, and the mean value of the cell area, the standard deviation of the cell area, the mean value of the cell roundness, and the standard deviation of the cell roundness are calculated, so as to determine the characteristic ranges of normal cells and abnormal cells, and screen out normal cells and abnormal cells according to the characteristic ranges of normal cells and abnormal cells; The specific calculation formula for the mean value of the cell area is , where represents the area of the i-th cell, and N represents the number of cells in the image; the specific calculation formula for the standard deviation of the cell area is ; The specific calculation formula for the mean value of the cell roundness is , where represents the perimeter of the i-th cell, represents pi; the specific calculation formula for the standard deviation of the cell roundness is ; According to the characteristic distribution of the cells, set the characteristic range of normal cells. If the area A range of the cells belongs to , and the roundness S range belongs to , it indicates that the cell belongs to a normal cell. Otherwise, the cell belongs to an abnormal cell. The normal cells and abnormal cells are distinguished based on the screening results. The original blood cell image is initially segmented according to the cell screening results using a U-Net grid, and the normal cells and non-cell regions are removed from the image. The screened normal cell image is used as the cell region image.
[0007] Preferably, in S3, the image is binarized by using the global threshold segmentation method to separate the cells from the background. Among them, the cell regions are marked white and the background is marked black. Then, morphological operations are used to improve the segmentation results. The specific operations include removing smaller noise points, filling small holes inside the cells, and smoothing the cell boundaries. Based on the region markings of the cells and the background, each independent cell region is segmented to form an independent cell image, and an image annotation tool is used to annotate the boundaries and attributes of each cell.
[0008] Preferably, in S4, feature extraction is performed on each independent cell image, including color, texture, shape, and geometric features, so as to construct a feature vector to represent the attributes of each cell. The specific analysis method of the matching degree coefficient is as follows: Step S411: For two cells and , the feature vector of cell is extracted as p, and the feature vector of cell is extracted as q. Step S412: Calculate the similarity between the two cells as , where represents the j-th feature vector of cell , represents the j-th feature vector of cell , and m represents the dimension of the feature vector. Step S413: Based on the similarity between the two cells, comprehensively calculate the matching degree coefficient as , where represents the maximum distance among all possible paired cells. Set a matching degree threshold , and compare the matching degree coefficient R with the matching degree threshold to determine the similarity of the cells. If the matching degree coefficient R is less than the matching degree threshold , it is determined that cell is not similar to cell , and there is no need to pair them. If the matching degree coefficient R is greater than or equal to the matching degree threshold , it is determined that cell is similar to cell are similar and pair them up; By comparing the feature vectors of all cell images pairwise, calculate the matching degree coefficient between each pair of cells. According to the matching degree coefficient and the matching degree threshold, pair up the cells that meet the similarity conditions.
[0009] Preferably, S5 is based on further feature extraction of the cell images after pairing is completed, performs error analysis on the cell images according to the feature parameters extracted, and further evaluates and detects the errors of the cell images based on the results of the error analysis; The specific analysis method of the error evaluation coefficient is as follows: Step S511: Set Q as the set of all cell images, is the k-th cell image, is the cell image 's feature vector, T is the set of true feature parameters obtained through manual annotation, is 's true feature parameter vector; Step S512: Set the error set E. E represents the error set of all cell images, and each element represents 's error vector, then the error vector of each element can be expressed as ; Step S513: Based on the comprehensive analysis of the error vectors of each element, calculate the error evaluation coefficient as , where represents the cardinality of the set Q, that is, the total number of cell images, represents the norm of the error vector , represents the norm of the true feature parameter vector ;
[0010] Preferably, S6 is based on the error analysis results of the cell images to detect the errors of the cell images. Set a preset error threshold through historical experience and experimental results. For each cell image , if , then it is considered that the error of the cell image is within the acceptable range, put it into the error cell image set, and screen out the set of cell images without errors at this time and transmit it to S7 for the construction and prediction of the blood cell segmentation network model. If , then it is considered that the cell image has a large error, and screen out the cell image at this time.
[0011] Preferably, the S7 constructs a blood cell segmentation network model by selecting a fully convolutional network, a cross-entropy loss function, and SGD optimization. It divides the error-free cell image set into a training set, a validation set, and a test set, uses the training set data to train the model, the validation set to monitor the training process and adjust the model parameters, evaluates the performance of the model through the test set, and deploys the trained model to actual applications for automatic segmentation of blood cell images.
[0012] To achieve the above object, the present invention provides the following technical solutions: A construction system for a blood cell segmentation network, implementing the above-mentioned construction method for a blood cell segmentation network, including: Image collection module: Used to collect original blood cell images and preprocess the collected original blood cell images. Image screening module: Based on the preprocessed original blood cell images, screens out normal cell images, removes non-cell regions and abnormal cell images, and uses the screened blood cell images as cell region images. Image segmentation module: Through the gray-scale difference between cells and the background, divides the cell region image into independent cell images as cell images, and labels each cell in the segmented cell images. Image pairing module: Used to extract features from cell images, analyze similar cells in cell images using a feature matching algorithm, obtain a matching degree coefficient, and pair similar cells. Error analysis module: Based on the feature parameters of cell images, conducts error analysis on cell images to obtain an error evaluation coefficient, and evaluates the error of cell images through the error detection coefficient. Error detection module: Based on the error analysis results, compares the preset error threshold with the error analysis results to further detect the error of cell images and screen out error-free cell images. Model construction module: Uses the error-free cell images as input data for the construction and prediction of the blood cell segmentation network model, and is used for model training and testing.
[0013] The technical effects and advantages of the present invention: The present invention collects original blood cell images, preprocesses the collected original blood cell images, screens the preprocessed original blood cell images to select normal cell images, removes non-cell regions and abnormal cell images, and uses the screened blood cell images as cell region images. Through the gray-scale difference between cells and the background, the cell region images are segmented into independent cell images as cell images, and each cell in the segmented cell images is labeled. By extracting features from the cell images and using a feature matching algorithm to analyze similar cells in the cell images, the similar cells are paired. Through the characteristic parameters of the cell images, error analysis is performed on the cell images to obtain an error evaluation coefficient, and the error of the cell images is evaluated. By comparing a preset error threshold with the error analysis result, the error of the cell images is further detected, and error-free cell images are selected. By using the selected error-free cell images for the construction and prediction of a blood cell segmentation network model and for model training and testing, segmentation is performed through the gray-scale difference between cells and the background, and combined with the labeling process, it helps to more accurately identify cell boundaries, can increase the segmentation accuracy of cell images. Through the steps of meticulous preprocessing and screening of the images, it helps to reduce the influence of noise and abnormal data on model training and improve the accuracy of the model. Through matching analysis and error detection of cell images, it helps to more accurately identify and analyze similar cells in cell images, is beneficial to reducing the influence of errors on model training, and thus improves the accuracy and reliability of model prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 It is a flowchart of the method steps of the present invention.
[0015] Figure 2 It is a block diagram of the system structure of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0016] Next, the technical solutions in the present invention will be clearly and completely described in conjunction with the drawings in the present invention. In addition, the forms of the various structures described in the following embodiments are merely examples, and a method and system for constructing a blood cell segmentation network according to the present invention are not limited to the various structures described in the following embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.
[0017] As Figure 1 shown, this embodiment provides a method for constructing a blood cell segmentation network, including: S1: Used to collect original blood cell images and preprocess the collected original blood cell images.
[0018] In this embodiment, S1 collects high-quality blood cell images by using a high-resolution microscope camera, records relevant information of each image, including the shooting time, microscopic parameters, and cell type, uses the collected high-quality blood cell images as the original blood cell images, and performs image conversion, size adjustment, grayscale conversion, image contrast enhancement, image denoising, and illumination correction on the original blood cell images.
[0019] Specifically, the specific collection and preprocessing process of the original blood cell images includes: in a laboratory environment, preparing a blood smear with fresh whole blood treated with an anticoagulant, checking whether the smear quality is good, whether the cells are evenly distributed, and whether there are overlapping or aggregating phenomena, adjusting the microscope to the best observation state, including selecting a suitable objective lens, adjusting the focus and light, then placing the blood smear on the microscope stage and fixing it with a slide clamp, using a high-resolution microscope camera to take pictures of the blood smear to ensure that the images are clear and rich in details, and ensuring that different types of cells and background conditions are covered, ensuring that the images are clear and rich in details. At the same time, record the shooting time, microscopic magnification, objective lens type, and cell type in the blood smear, then convert the taken images into a unified format, adjust the image size to adapt to the size of the network input layer, and convert the color images into grayscale images, apply image enhancement technology to improve the image quality, use a filter to remove noise in the image, and at the same time, perform illumination correction on the image to eliminate the influence of uneven illumination on the image quality.
[0020] S2: Based on the screening of the preprocessed original blood cell images, normal cell images are selected, non-cell regions and abnormal cell images are removed, and the screened blood cell images are used as cell region images.
[0021] In this embodiment, S2 extracts the cell area, cell perimeter, cell circularity, and cell texture in the image by feature extraction of the original blood cell images, then performs statistical analysis on the extracted cell features, and calculates the mean value of the cell area, the standard deviation of the cell area, the mean value of the cell circularity, and the standard deviation of the cell circularity, so as to determine the characteristic ranges of normal cells and abnormal cells, and screen out normal cells and abnormal cells according to the characteristic ranges of normal cells and abnormal cells; The specific calculation formula for the mean value of the cell area is , where represents the area of the i-th cell, and N represents the number of cells in the image; the specific calculation formula for the standard deviation of the cell area is ; The specific calculation formula for the mean value of the cell circularity is , where represents the perimeter of the \(i\)-th cell, represents pi; the specific calculation formula for the standard deviation of the circularity of the cells is ; According to the characteristic distribution of the cells, set the characteristic range of normal cells. If the area \(A\) range of the cells belongs to , and the circularity \(S\) range belongs to , it means that the cell belongs to a normal cell. Otherwise, the cell belongs to an abnormal cell. Distinguish normal cells and abnormal cells based on the screening results; use the U-Net grid to preliminarily segment the original blood cell image according to the cell screening results, and remove normal cells and non-cell regions from the image. Save the screened normal cell image as the cell region image.
[0022] Specifically, by collecting the area \(A\) and circularity \(S\) data of normal cells, calculating the mean and standard deviation of the cell area and circularity, based on the data distribution and the proportion of normal cells required, select the value of \(k\), and the value of \(k\) is usually 1 or 2. Then set the area range value of normal cells as: , and the circularity \(S\) range value of normal cells as: , input the preprocessed original blood cell image into the U-Net grid model to segment the original blood cell image, obtain the segmentation image containing all cell regions, screen out normal cells according to the set range of the area and circularity of normal cells, and mark the cells that do not meet the characteristic range of normal cells as abnormal cells. Remove the regions marked as abnormal cells from the segmentation image, and then save the screened normal cell image as a new image; for the cell area, it can be calculated by the number of pixels belonging to the cells in the image; for the cell perimeter, it can be obtained by using the boundary tracking algorithm to find the cell contour and calculate it.
[0023] S3: Through the gray difference between the cells and the background, segment the cell region image into independent cell images as cell images, and label each cell in the segmented cell image.
[0024] In this embodiment, in S3, the image is binarized by using the global threshold segmentation method to separate the cells from the background. Among them, the cell regions are marked as white and the background is marked as black. Then use morphological operations to improve the segmentation results. The specific operations include removing smaller noise points, filling small holes inside the cells, and smoothing the cell boundaries. Based on the region markings of the cells and the background, each independent cell region is segmented to form independent cell images, and an image annotation tool is used to annotate the boundaries and attributes of each cell.
[0025] Specifically, the following annotations are made for each independent cell after segmentation: by using image annotation tools such as Labellmg, VGGlmageAnnotator, and CVAT, an accurate boundary contour is drawn for each cell, and the attributes of the cell are annotated, including cell type, cell state, and other specific features; the cell types include red blood cells, white blood cells, and platelets, the cell states include normal and abnormal, and other specific features include the position and size of the cell nucleus.
[0026] S4: Used to extract features from cell images and analyze similar cells in the cell images using a feature matching algorithm to obtain a matching degree coefficient, and pair the similar cells.
[0027] In this embodiment, S4 constructs a feature vector to represent the attributes of each cell by extracting features from each independent cell image, including color, texture, shape, and geometric features; The specific analysis method of the matching degree coefficient is as follows: Step S411: For two cells and , the feature vectors of cell are extracted as p, and the feature vector of cell is extracted as q; Step S412: Calculate the similarity between the two cells as , where represents the j-th feature vector of cell , represents the j-th feature vector of cell , and m represents the dimension of the feature vector; Step S413: Based on the similarity between the two cells, comprehensively calculate the matching degree coefficient as , where represents the maximum distance among all possible paired cells; Set a matching degree threshold , compare the matching degree coefficient R with the matching degree threshold to determine the similarity of the cells. If the matching degree coefficient R is less than the matching degree threshold , it is determined that cell and cell are not similar and do not need to be paired. If the matching degree coefficient R is greater than or equal to the matching degree threshold , it is determined that cell and cell are similar and pair them; By comparing the feature vectors of all cell images pairwise, calculating the matching degree coefficient between each pair of cells, and based on the matching degree coefficient and the matching degree threshold, pair the cells that meet the similarity conditions.
[0028] S5: Based on the feature parameters of the cell images, perform error analysis on the cell images to obtain an error evaluation coefficient, and evaluate the error of the cell images through the error detection coefficient.
[0029] In this embodiment, S5 is based on further feature extraction of the cell images after pairing is completed, performs error analysis on the cell images according to the feature parameters extracted from the features, and further evaluates and detects the error of the cell images based on the results of the error analysis; The specific analysis method of the error evaluation coefficient is as follows: Step S511: Set Q as the set of all cell images, is the k-th cell image, is the cell image 's feature vector, T is the set of true feature parameters obtained through manual annotation, is 's true feature parameter vector; Step S512: Set the error set E, E represents the error set of all cell images, and each element in it represents 's error vector, then the error vector of each element can be expressed as ; Step S513: Based on the comprehensive analysis of the error vector of each element, and calculate the error evaluation coefficient as , where represents the cardinality of the set Q, that is, the total number of cell images, represents the norm of the error vector , represents the norm of the true feature parameter vector ;
[0030] S6: Based on the results of the error analysis, compare the preset error threshold with the results of the error analysis to further detect the error of the cell images and screen out the cell images without error.
[0031] In this embodiment, S6 is based on the results of the error analysis of the cell images, detects the error of the cell images, and sets the preset error threshold through historical experience and experimental results , for each cell image , if , then it is considered that the cell image If the error is within the acceptable range, put it into the error cell image set, and screen out the cell image set without error at this time and transmit it to S7 for the construction and prediction of the blood cell segmentation network model. If , it is considered that the cell image has a large error, and the cell image at this time is screened out.
[0032] S7: Use the error-free cell images as the input data for the construction and prediction of the blood cell segmentation network model, and use them for model training and testing.
[0033] In this embodiment, S7 constructs a blood cell segmentation network model by selecting a fully convolutional network, a cross-entropy loss function, and SGD optimization. Divide the error-free cell image set into a training set, a validation set, and a test set. Use the training set data to train the model, the validation set to monitor the training process and adjust the model parameters, and evaluate the performance of the model through the test set. Deploy the trained model to actual applications to automatically segment blood cell images.
[0034] As Figure 2 shown, this embodiment provides an implementation system corresponding to an item anti-counterfeiting method based on local feature visual information, including an image collection module, an image screening module, an image segmentation module, an image pairing module, an error analysis module, an error detection module, and a model construction module. The image collection module is connected to the image screening module, the image screening module is connected to the image segmentation module, the image segmentation module is connected to the image pairing module, the image pairing module is connected to the error analysis module, the error analysis module is connected to the error detection module, and the error detection module is connected to the model construction module.
[0035] The image collection module is used to collect original blood cell images and preprocess the collected original blood cell images; The image screening module screens the preprocessed original blood cell images, screens out normal cell images, removes non-cell regions and abnormal cell images, and uses the screened blood cell images as cell region images; The image segmentation module divides the cell region image into independent cell images through the gray-scale difference between cells and the background as cell images, and labels each cell in the segmented cell images; The image pairing module is used to extract features from cell images and use a feature matching algorithm to analyze similar cells in the cell images to obtain a matching degree coefficient, and pair the similar cells; The error analysis module performs error analysis on cell images based on the feature parameters of the cell images to obtain an error evaluation coefficient, and evaluates the error of the cell images through the error detection coefficient; Based on the error analysis result, the error detection module further detects the error of the cell image by comparing the preset error threshold with the error analysis result, and screens out the cell images without error; The model construction module uses the error-free cell images as the input data for the construction and prediction of the blood cell segmentation network model, and is used for model training and testing.
[0036] Finally, the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
[0037] The above is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in this application, and all should be covered by the protection scope of this application. Therefore, the protection scope of this application shall be subject to the protection scope of the claims.
Claims
1. A method for constructing a blood cell segmentation network, characterized in that: include: collecting original blood cell images, and preprocessing the collected original blood cell images; Screening the pre-processed original blood cell image to screen out normal cell images, removing non-cell regions and abnormal cell images, and using the screened blood cell image as a cell region image; By using the grayscale difference between cells and background, the cell region image is segmented into independent cell images as the cell image, and each cell in the segmented cell image is labeled; Extract features from cell images, and use feature matching algorithms to analyze similar cells in cell images, obtain matching coefficients, and pair similar cells; Based on the characteristic parameters of the cell image, the error analysis of the cell image is performed to obtain the error evaluation coefficient, and the error of the cell image is evaluated by the error detection coefficient; Based on the error analysis results, the errors of the cell images are further detected by comparing the preset error threshold with the error analysis results, and the error-free cell images are screened out; The error-free cell images are used as input data for building and predicting the blood cell segmentation network model and for model training and testing.
2. The method for constructing a blood cell segmentation network according to claim 1, characterized in that: The S1 collects high-quality blood cell images by photographing blood smears using a high-resolution microscope camera, and records relevant information of each image, including the photographing time, microscopic parameters, and cell types. The collected high-quality blood cell images are used as original blood cell images, and the original blood cell images are converted, resized, gray-scale converted, contrast enhanced, denoised, and corrected for illumination.
3. The method for constructing a blood cell segmentation network according to claim 1, characterized in that: The S2 extracts the cell area, cell perimeter, cell circularity and cell texture in the image by performing feature extraction on the original blood cell image, and then performs statistical analysis on the extracted cell features, and calculates the mean value of the cell area, the standard deviation of the cell area, the mean value of the cell circularity and the standard deviation of the cell circularity, so as to determine the feature range of normal cells and abnormal cells, and screen out normal cells and abnormal cells according to the feature range of normal cells and abnormal cells; The specific calculation formula for the mean value of the cell area is: ,in, represents the area of the ith cell, N represents the number of cells in the image; the specific calculation formula for the standard deviation of the cell area is: ; The specific calculation formula for the mean value of the cell circularity is: ,in, represents the perimeter of the ith cell, represents pi; the specific calculation formula for the standard deviation of the cell circularity is: ; According to the characteristic distribution of cells, the characteristic range of normal cells is set. If the area A of the cell belongs to , and the circularity S range belongs to , it means that the cell is a normal cell, otherwise, the cell is an abnormal cell, and normal cells and abnormal cells are distinguished based on the screening results; the original blood cell image is preliminarily segmented according to the cell screening results by using the U-Net grid, and the normal cells and non-cell areas are removed from the image, and the screened normal cell image is used as the cell area image.
4. The method for constructing a blood cell segmentation network according to claim 1, characterized in that: The S3 uses a global threshold segmentation method to binarize the image, thereby separating the cells from the background, wherein the cell area is marked as white and the background is marked as black, and then morphological operations are used to improve the segmentation results. The specific operations include removing smaller noise points, filling small holes inside the cells, and smoothing cell boundaries. Based on the regional marking of cells and background, each independent cell area is segmented to form an independent cell image, and the boundaries and attributes of each cell are annotated using an image annotation tool.
5. The method for constructing a blood cell segmentation network according to claim 1, characterized in that: The S4 extracts features of each independent cell image, including color, texture, shape and geometric features, thereby constructing a feature vector to represent the attributes of each cell; The specific analysis method of the matching degree coefficient is as follows: Step S411: For two cells and , respectively, extract the cells The eigenvector of the cell is p. The eigenvector of is q; Step S412: Calculate the similarity between two cells as ,in, Represents cells The jth eigenvector of Represents cells The j-th eigenvector of , m represents the dimension of the eigenvector; Step S413: Based on the similarity between the two cells, the matching coefficient is calculated as ,in, represents the maximum distance among all possible paired cells; Set a matching threshold , the matching degree coefficient R and the matching degree threshold Compare to determine the similarity of cells. If the matching coefficient R Matching threshold , then determine the cell With cells If they are not similar, there is no need to pair them. If the matching coefficient R Matching threshold , then determine the cell With cells are similar, and pair them up; By comparing the feature vectors of all cell images pairwise, the matching coefficient between each pair of cells is calculated, and the cells that meet similar conditions are paired according to the matching coefficient and matching threshold.
6. The method for constructing a blood cell segmentation network according to claim 1, characterized in that: S5 further extracts features from the paired cell images, performs error analysis on the cell images according to feature parameters extracted from the features, and further evaluates and detects the errors of the cell images based on the results of the error analysis; The specific analysis method of the error evaluation coefficient is: Step S511: Set Q to be the set of all cell images, is the kth cell image, For cell images The feature vector of , T is the real feature parameter set obtained by manual annotation, for The true characteristic parameter vector of Step S512: Set the error set E, where E represents the error set of all cell images, and each element express The error vector of each element is It can be expressed as ; Step S513: Based on a comprehensive analysis of the error vector of each element, the error evaluation coefficient is calculated as ,in represents the cardinality of set Q, i.e. the total number of cell images, represents the error vector The norm of represents the true feature parameter vector The norm of .
7. The method for constructing a blood cell segmentation network according to claim 1, characterized in that: The S6 detects the error of the cell image based on the error analysis result of the cell image, and sets the preset error threshold value according to historical experience and experimental results. , for each cell image ,like , then the cell image If the error is within the acceptable range, it is put into the error cell image set, and the error-free cell image set is selected and transmitted to S7 for the construction and prediction of the blood cell segmentation network model. , then the cell image There is a large error, so the cell image at this time is screened out.
8. The method for constructing a blood cell segmentation network according to claim 1, characterized in that: The S7 constructs a blood cell segmentation network model by selecting a fully convolutional network, a cross entropy loss function and SGD optimization, divides the error-free cell image set into a training set, a validation set and a test set, uses the training set data to train the model, uses the validation set to monitor the training process, adjusts the model parameters, evaluates the performance of the model through the test set, deploys the trained model to practical applications, and automatically segments blood cell images.
9. A blood cell segmentation network construction system, implementing a blood cell segmentation network construction method as claimed in any one of claims 1 to 8, characterized in that: include: Image collection module: used for collecting original blood cell images and preprocessing the collected original blood cell images; Image screening module: based on screening the pre-processed original blood cell image, normal cell images are screened out, non-cell areas and abnormal cell images are removed, and the screened blood cell images are used as cell area images; Image segmentation module: The cell region image is segmented into independent cell images as cell images based on the grayscale difference between cells and background, and each cell in the segmented cell image is labeled; Image pairing module: used to extract features from cell images, and use feature matching algorithms to analyze similar cells in cell images, obtain matching coefficients, and pair similar cells; Error analysis module: Based on the characteristic parameters of the cell image, the error analysis of the cell image is performed to obtain the error evaluation coefficient, and the error of the cell image is evaluated through the error detection coefficient; Error detection module: Based on the error analysis results, the error of the cell image is further detected by comparing the preset error threshold with the error analysis results, and the error-free cell images are screened out; Model building module: The error-free cell images are used as input data for building and predicting the blood cell segmentation network model, and are used for model training and testing.