A method for identifying clue cells in leucorrhea microscopic images

Abstract

The present invention provides a method for identifying clue cells in a leucorrhea microscopic image, which belongs to the technical field of clue cell identification, and comprises: firstly, pre-processing the stained leucorrhea microscopic image, including graying, medical image enhancement and noise removal. Through spectral feature analysis, a feature space is constructed and adaptive threshold segmentation is adopted to extract the cell area, and then edge detection and morphological operations are performed to accurately extract the complete outline of the cell. The morphological feature parameters such as cell area, perimeter, roundness, grayscale mean and variance are innovatively calculated, and a feature discrimination equation group is established in combination with the spectral feature vector to realize accurate classification and identification of cells. The identification results are post-processed to remove noise interference and perform connected area analysis, the number and proportion of clue cells are counted, and annotated images are generated; and the iterative optimization of the data set verified by medical experts solves the technical problem that the consistency of the diagnosis results is poor due to the subjective experience differences of doctors.

Description

Technical Field

[0001] The invention belongs to the technical field of clue cell recognition, and in particular relates to a method for identifying clue cells in a leucorrhea microscopic image. Background Art

[0002] Microscopic examination of vaginal secretions is an important means of diagnosing gynecological diseases and is widely used in clinical practice. Among them, the identification of clue cells is of great significance for the diagnosis of bacterial vaginosis. Clue cells refer to vaginal epithelial cells whose cell edges are blurred due to the attachment of a large number of bacteria. Their existence is one of the important diagnostic indicators of bacterial vaginosis. Traditional microscopic examination of vaginal secretions mainly relies on visual judgment by experienced doctors. However, this method has the following major problems: First, differences in the subjective experience of doctors lead to poor consistency in diagnostic results. Especially for young doctors with inexperience, it is difficult to accurately identify clue cells, which can easily lead to missed diagnoses or misdiagnoses, such as Figure 1 As shown in the figure, (a) is a high-power microscopic image of true clue cells after staining, and (b) is a high-power microscopic image of false clue cells after staining. They can be distinguished under high-power microscopic images, but under ordinary medical microscopes commonly used by non-laboratory doctors, due to the small magnification, it is difficult for inexperienced doctors to distinguish between true and false clue cells, and they are prone to misjudgment. Secondly, traditional manual recognition methods are inefficient. When faced with a large number of samples, it is easy to cause visual fatigue to doctors and affect the accuracy of diagnosis. In clinical practice, incorrect recognition results may lead to serious diagnostic errors. Therefore, it is of great clinical significance to establish an accurate and reliable automatic recognition method for clue cells. Summary of the invention

[0003] In view of this, the present invention provides a method for identifying clue cells in leucorrhea microscopic images, which can solve the technical problem that the consistency of diagnosis results is poor due to differences in doctors' subjective experience.

[0004] The present invention is achieved in that:

[0005] The present invention provides a method for identifying clue cells in a leucorrhea microscopic image, comprising the following steps:

[0006] S10, obtaining a stained leucorrhea microscopic image, and preprocessing the leucorrhea microscopic image, wherein the preprocessing includes grayscale processing, medical image enhancement processing, and noise removal processing to obtain a preprocessed microscopic image;

[0007] S20, performing spectral feature analysis on the preprocessed microscopic image, extracting the spectral feature vector of the microscopic image, and constructing a spectral feature space of the microscopic image;

[0008] S30, in the spectral feature space, using an adaptive threshold segmentation method to perform cell region segmentation on the spectral feature space to obtain an initial cell region segmentation map;

[0009] S40, performing edge detection processing on the initial segmentation map of the cell region, performing morphological opening operation processing and morphological closing operation processing on the segmentation map after the edge detection processing, and extracting the complete outline of the cell region;

[0010] S50, calculating morphological characteristic parameters of the cell region based on the complete outline of the cell region, wherein the morphological characteristic parameters include cell area, cell perimeter, cell roundness, cell grayscale mean, and cell grayscale variance;

[0011] S60, establishing a characteristic discriminant equation group according to the spectral characteristic vector and the morphological characteristic parameter, and classifying and identifying the cell region using the characteristic discriminant equation group;

[0012] S70, post-processing the classification and recognition results, wherein the post-processing includes noise interference removal processing and connected region analysis processing to obtain clue cell position information;

[0013] S80, counting the number of clue cells in the leucorrhea microscopic image, calculating the proportion of the clue cells in the leucorrhea microscopic image, and generating a clue cell identification result annotated image;

[0014] Optionally, the method further includes S90, performing iterative optimization processing on the feature discrimination equation group based on a verification data set annotated by a medical expert, until the accuracy and recall rate of the clue cell identification reach preset standard values.

[0015] Furthermore, the characteristic discrimination equation group includes a spectral characteristic equation, a morphological characteristic equation, a comprehensive scoring equation and a discrimination threshold equation;

[0016] The spectral feature equation is used to calculate the spectral feature score of the cell region, the input includes the spectral feature vector of the cell region and the standard spectral template, and the output is the spectral similarity score of the cell region;

[0017] The morphological characteristic equation is used to calculate the morphological characteristic score of the cell, and the input includes parameters such as the area, perimeter, roundness, grayscale mean and grayscale variance of the cell, and the output is the morphological conformity score of the cell;

[0018] The comprehensive scoring equation is used to fuse the spectral feature score and the morphological feature score, the input includes the spectral similarity score, the morphological conformity score and their respective weight coefficients, and the output is the comprehensive score value of the cell;

[0019] The discrimination threshold equation is used to determine the discrimination threshold of clue cells. The input includes the distribution of comprehensive score values ​​of expert-labeled samples, and the output is the optimized discrimination threshold.

[0020] Furthermore, the implementation process of step S20 mainly includes three core links: preparation of microscope filters, image acquisition and spectral feature extraction. In the preparation stage of microscope filters, it is necessary to select standard optical glass filters, including red (620-750nm transmission band), green (495-570nm transmission band), blue (450-495nm transmission band) and yellow (570-590nm transmission band). The specifications of these filters are 25mm in diameter to fit the microscope filter slot, the transmittance is not less than 85%, the optical flatness must reach λ / 4, and the surface quality requirement is 60 / 40 to ensure the accuracy and stability of spectral acquisition.

[0021] In the image acquisition stage, the microscope must first be preheated for 20 minutes to ensure the stability of the light source, and the key parameters of the microscope must be set, including using a 40X objective, opening the aperture to the maximum, and adjusting the condenser to the optimal focal plane position; then different filters are inserted in turn for image acquisition. The acquisition sequence is to first acquire the white light image without a filter, and then acquire the images under the red, green, blue, and yellow filters in turn. Each color must be acquired 3 times and the average value must be taken to improve data reliability. The acquired images should be stored in a resolution of 1920×1080, 8-bit bit depth, and lossless TIF format to ensure the accuracy of subsequent processing.

[0022] In the spectral feature extraction stage, the images collected under each color channel need to be preprocessed, including removing image noise, background correction and grayscale normalization. Then, the response values ​​of the target area under different color channels are extracted from the processed images and a spectral feature vector is constructed. The spectral feature space is constructed after normalization of the feature vector. Optionally, special attention should be paid to quality control measures throughout the process, including regular inspection of the cleanliness and integrity of the filter, continuous monitoring of the stability of the microscope light source, minimization of ambient light interference, and ensuring that the sample remains still during image acquisition.

[0023] Wherein, the microscope spectrum analysis equation is specifically expressed as follows:

[0024] I c (x,y)=I0·T c (λ)·R(x,y,λ)+N c (x,y);

[0025] In the formula, I c(x, y) is the gray value of the image at position (x, y) when using the filter with color c; I0 is the incident light intensity of the microscope light source; T c (λ) is the spectral transmittance function of the filter with color c; R(x,y,λ) is the reflectivity of the sample at position (x,y) for light with wavelength λ; N c (x,y) is the imaging noise.

[0026] The filter transmittance function is described by the Gaussian model:

[0027]

[0028] In the formula, A c is the peak transmittance; c is the central wavelength; σ c is the half peak width; B c is the baseline transmittance.

[0029] For the target area Ω, its spectral response value under the color channel c is calculated as:

[0030]

[0031] Where, |Ω| is the number of pixels in the region; is the average gray value of the background area; is the average gray value of the whiteboard area; ∈ c is the measurement error, ranging from -0.02 to 0.02.

[0032] Normalization of spectral feature vectors:

[0033]

[0034] In the formula, is the normalized spectral response value; V is the original spectral response value vector [V red ,V green ,V blue ,V yellow ] T .

[0035] The parameter acquisition method is:

[0036] 1) Filter parameters can be obtained directly from the filter manual or calibrated by a spectrophotometer: a) measure the transmittance curve; b) fit the Gaussian model to obtain A c , c , σ c , B c ;

[0037] 2) The background and whiteboard correction values ​​are obtained by the following steps: a) Select the blank area of ​​the image to calculate b) Collect standard whiteboard image and calculate

[0038] Explanation of the equation construction principle:

[0039] 1) The basic imaging equation takes into account the combined effects of light source, filter, sample and imaging system;

[0040] 2) The Gaussian model can better describe the spectral characteristics of the filter;

[0041] 3) The response value calculation adopts regional average and background correction to improve the measurement stability;

[0042] 4) Normalization eliminates the scale differences between different channels.

[0043] Furthermore, the spectral characteristic equation is specifically expressed as follows:

[0044]

[0045] In the formula, S spec is the spectral feature score, ranging from 0 to 1; V i is the spectral response value of the i-th band of the cell to be tested; T i is the spectral response value of the i-th band of the standard template, that is, the spectral response value obtained when no sample is placed on the microscope and only the filter is set; n is the number of spectral bands; α, β are weight coefficients, and they satisfy α+β=1; γ is the correction term, and its value range is -0.1~0.1.

[0046] The morphological characteristic equation is specifically expressed as follows:

[0047]

[0048] In the formula, S morph is the morphological feature score, ranging from 0 to 1; A is the cell area; P is the cell perimeter; μ is the grayscale mean of the cell area; μ0 is the standard grayscale mean; σ is the grayscale standard deviation of the cell area; A0 is the standard cell area; σ A is the area standard deviation; ω1, ω2, ω3 are weight coefficients, and satisfy δ is the error correction term, and its value range is -0.1 to 0.1.

[0049] The comprehensive scoring equation is specifically expressed as follows:

[0050]

[0051] In the formula, S totalis the comprehensive score of cells, ranging from 0 to 1; λ1, λ2 are feature weight coefficients, and satisfy λ1+λ2=1; θ is the cross-term coefficient, ranging from 0 to 0.3; η is the comprehensive error term, ranging from -0.05 to 0.05.

[0052] The discrimination threshold equation is specifically expressed as follows:

[0053]

[0054] In the formula, T is the discrimination threshold; is the comprehensive score of the sample annotated by the i-th expert; m is the number of samples annotated by the expert; k is the confidence coefficient, ranging from 1.5 to 2.5; is the sample average score.

[0055] The parameter acquisition method is:

[0056] 1) The morphological parameters A, P, and μ are calculated by image processing algorithm:

[0057] A=∑ x,y f(x,y), where f(x,y) is the binarized cell region image;

[0058] P=∑ x,y g(x,y), where g(x,y) is the contour image after edge detection;

[0059] Where h(x,y) is the original grayscale image;

[0060] 2) Standard parameters μ0, A0, σ A Obtained through positive samples annotated by statistical experts.

[0061] Explanation of the equation construction principle:

[0062] 1) The spectral characteristic equation adopts a combination of relative error term and root mean square error term, which can consider the differences in spectral shape and amplitude at the same time;

[0063] 2) The morphological characteristic equation includes three features: circularity, grayscale distribution and area distribution, and uses an exponential function to reflect the penalty effect of parameter deviation from the standard value;

[0064] 3) The introduction of cross terms into the comprehensive scoring equation can reflect the synergistic effect of the two types of features;

[0065] 4) The discrimination threshold equation is based on statistical principles, taking into account the sample mean and standard deviation, and has good robustness.

[0066] The following is a detailed analysis of the derivation and establishment process, optimization steps and parameter design principles of each equation.

[0067] 1. Derivation and analysis of spectral characteristic equation

[0068]

[0069] Derivation steps:

[0070] 1) First consider the two key dimensions of spectral similarity:

[0071] Relative error term:

[0072] Root mean square error term:

[0073] 2) Design principle of relative error term: Use max(V i ,T i ) as a normalization factor to ensure that the error is in the range of 0-1; the errors of each band are accumulated to reflect the difference in the overall spectral shape; the absolute value is used to ensure that the error is positive;

[0074] 3) Design principle of the root mean square error term: the square term amplifies larger errors and is more sensitive to outliers; the square root operation makes the dimension of the final result consistent with the relative error term; it can reflect the deviation of the overall intensity of the spectrum;

[0075] 4) Optimization of weight coefficients α, β: Determine the optimal weight ratio through cross-validation; satisfy α+β=1 to ensure score normalization; typical values: α=0.6, β=0.4, the relative error term weight is slightly larger.

[0076] 2. Derivation and analysis of morphological characteristic equations

[0077]

[0078] Derivation steps:

[0079] 1) Circularity feature item: Based on the principle of equal perimeter area ratio, the value is 1 for a perfect circle, and the more irregular the shape, the smaller the value;

[0080] 2) Grayscale distribution feature items: The Z-score standardization form is adopted, μ0 is obtained through statistics of expert-labeled samples, and σ reflects the discrete degree of grayscale distribution;

[0081] 3) Area distribution characteristics: Using the Gaussian function form, the area is close to the standard value A0 and the value is close to 1, σ A Control the allowable deviation range;

[0082] 3. Derivation and analysis of comprehensive scoring equation

[0083]

[0084] Derivation steps:

[0085] 1) Linear combination term: λ1S spec +λ2S morph , reflects the independent contribution of the two types of features, and the weight coefficient is obtained through machine learning optimization;

[0086] 2) Cross terms: The geometric mean form is used to reflect the synergistic effect between features, and the coefficient θ is limited to the range of 0 to 0.3 to avoid excessive influence;

[0087] 4. Derivation and analysis of the discrimination threshold equation

[0088]

[0089] Derivation steps:

[0090] 1) Sample mean term: Reflect the overall rating level and reduce the impact of random fluctuations;

[0091] 2) Standard deviation term: It reflects the degree of discreteness of the score, and the coefficient k controls the strictness of the judgment;

[0092] Innovation:

[0093] 1. The spectral characteristic equation considers both relative error and absolute error, which improves the feature expression ability;

[0094] 2. The morphological characteristic equation adopts multi-feature fusion, and each feature has a clear physical meaning;

[0095] 3. The comprehensive scoring equation introduces cross terms to better describe the interaction between features;

[0096] 4. The discrimination threshold equation is based on statistical principles and has strong adaptability.

[0097] Compared with the prior art, the beneficial effect of the method for identifying clue cells in leucorrhea microscopic images provided by the present invention is: a multi-feature fusion strategy is adopted, spectral features and morphological features are combined, and an adaptive threshold adjustment mechanism is introduced to achieve high-precision identification of clue cells.

[0098] Firstly, the method preprocesses the acquired leucorrhea microscopic images, including grayscale, image enhancement and noise filtering, to improve the image quality and lay the foundation for subsequent analysis. Then, it uses multi-band spectral analysis technology to extract the spectral features of the image and construct the spectral feature space. On this basis, the adaptive threshold segmentation algorithm is used to segment the cell area in the feature space to obtain preliminary cell detection results.

[0099] Next, the method performs edge detection and morphological processing on the segmentation results to extract the complete outline of the cell area. Based on these outline information, the morphological characteristic parameters such as the cell area, perimeter, roundness, grayscale mean and grayscale variance are calculated. The spectral features and morphological features are integrated to construct a comprehensive feature discrimination equation group to classify and identify the cell area.

[0100] To further improve the recognition accuracy, this method uses post-processing techniques such as noise interference filtering and connected region analysis, and finally obtains reliable clue cell location information. At the same time, it also counts the proportion of clue cells in the entire leucorrhea sample and marks the location of clue cells on the original microscopic image.

[0101] Therefore, the present invention solves the technical problem that differences in doctors' subjective experience lead to poor consistency in diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] Figure 1 It is a high-power microscopic image of true and false clue cells in the background technology.

[0103] Figure 2 A flow chart of the method provided by the present invention;

[0104] Figure 3 is a grayscale image in the embodiment;

[0105] Figure 4 is the enhanced grayscale image in the embodiment;

[0106] Figure 5 is a grayscale image after denoising in the embodiment. DETAILED DESCRIPTION

[0107] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0108] like Figure 2 As shown, it is a flow chart of a method for identifying clue cells in a leucorrhea microscopic image provided by the present invention, and the specific implementation of each step is described in detail as follows:

[0109] The specific implementation method of step S10 is to pre-process the acquired leucorrhea microscopic image. First, the leucorrhea microscopic image is grayed to convert the color image into a grayscale image. Next, the grayscale image is subjected to medical image enhancement processing, including histogram equalization processing and contrast enhancement processing. These processing can improve the clarity of the image and enhance the visual effect of the cell area. Then, the image after the enhancement processing is subjected to noise removal processing, including median filtering processing and Gaussian filtering processing. These filtering operations can effectively remove noise interference in the image and further improve the image quality. Finally, the image after the noise removal processing is output as the pre-processing result.

[0110] The specific implementation of step S20 is to perform spectral feature analysis on the pre-treated microscopic image. Specifically, the specific implementation of step S20 mainly includes three core links: preparation of microscope filters, image acquisition and spectral feature extraction. In the microscope filter preparation stage, it is necessary to select standard optical glass filters, including red (620-750nm transmission band), green (495-570nm transmission band), blue (450-495nm transmission band) and yellow (570-590nm transmission band) four filters. The specifications of these filters are required to be 25mm in diameter to fit the microscope filter slot, the transmittance is not less than 85%, the optical flatness must reach λ / 4, and the surface quality requirement is 60 / 40 to ensure the accuracy and stability of spectral acquisition.

[0111] In the image acquisition stage, the microscope must first be preheated for 20 minutes to ensure the stability of the light source, and the key parameters of the microscope must be set, including using a 40X objective, opening the aperture to the maximum, and adjusting the condenser to the optimal focal plane position; then different filters are inserted in turn for image acquisition. The acquisition sequence is to first acquire the white light image without a filter, and then acquire the images under the red, green, blue, and yellow filters in turn. Each color must be acquired 3 times and the average value must be taken to improve data reliability. The acquired images should be stored in 1920×1080 resolution, 8-bit depth, and lossless TIF format to ensure the accuracy of subsequent processing.

[0112] In the spectral feature extraction stage, the image collected under each color channel needs to be preprocessed, including removing image noise, background correction and grayscale normalization. Then, the response value of the target area under different color channels is extracted from the processed image and the spectral feature vector is constructed. The spectral feature space is constructed after normalization of the feature vector. Optionally, in the whole process, special attention should be paid to quality control measures, including regular inspection of the cleanliness and integrity of the filter, continuous monitoring of the stability of the microscope light source, minimization of ambient light interference, and ensuring that the sample remains still during image acquisition. The steps of constructing the spectral feature vector specifically include: first, multi-band decomposition of the preprocessed microscopic image is performed to obtain the spectral response value of the image in different bands. Then, the spectral response values ​​of different bands are combined into a spectral feature vector. Next, the spectral feature vector is normalized to eliminate the dimension difference of the spectral response value. Finally, the spectral feature space is constructed based on the normalized spectral feature vector to provide a basis for subsequent cell region segmentation.

[0113] The specific implementation of step S30 is to segment the spectral feature space into cell regions using an adaptive threshold segmentation method. First, the histogram distribution of the spectral feature space is calculated, and the initial threshold is determined based on the histogram information. Then, the initial threshold is used in the spectral feature space for iterative optimization, and the optimal segmentation threshold is finally obtained. Finally, the spectral feature space is segmented using this optimal threshold to obtain an initial segmentation map of the cell region. This method can adaptively determine the optimal segmentation threshold according to the image content, thereby improving the accuracy of segmentation.

[0114] The specific implementation method of step S40 is to perform edge detection and morphological processing on the initial segmentation map to extract the complete outline of the cell area. First, the initial segmentation map is edge detected to obtain a cell edge map. Then, the edge map is subjected to morphological opening operation processing to remove small noise interference. Next, the image after the opening operation processing is subjected to morphological closing operation processing to fill the gaps on the edge. Finally, the complete outline of the cell area is extracted from the image after the closing operation processing. These processes can ensure that accurate cell boundary information is obtained.

[0115] The specific implementation of step S50 is to calculate various morphological characteristic parameters based on the extracted cell area contour. First, the cell area and cell perimeter are calculated based on the cell area contour. Then, the roundness index of the cell is calculated using the area and perimeter. Next, the grayscale value inside the cell area is extracted, and the mean and variance of the grayscale value are calculated. These morphological characteristic parameters can describe the size, shape and grayscale distribution characteristics of the cell, providing a basis for subsequent cell identification.

[0116] The specific implementation method of step S60 is to establish a feature discrimination equation group to classify and identify the cell area. First, construct a spectral feature discrimination equation, and use the spectral feature vector of the cell area and the standard spectral template to calculate the spectral similarity score. Secondly, establish a morphological feature discrimination equation, and calculate the morphological conformity score according to the area, perimeter, roundness, grayscale mean and grayscale variance of the cell. Then, construct a comprehensive scoring equation, fuse the spectral feature score and the morphological feature score, and obtain the comprehensive score of the cell. Finally, establish a discrimination threshold equation, and determine the optimal discrimination threshold based on the comprehensive score distribution of the expert-annotated samples. Through these feature discrimination equations, clue cells can be effectively identified.

[0117] The specific implementation method of step S70 is to post-process the classification and recognition results to further improve the accuracy of recognition. First, the classification results are subjected to morphological filtering to remove possible noise interference. Then, the image after filtering is subjected to connected region analysis to mark each connected region. Next, the area and perimeter parameters of each connected region are calculated. Finally, based on the area and perimeter characteristics of the connected region, the region that meets the clue cell characteristics is screened out to obtain the final clue cell location information. These post-processing steps can effectively suppress noise interference and extract more reliable clue cell detection results.

[0118] The specific implementation method of step S80 is to count the number of clue cells and their proportion in the leucorrhea microscopic image, and mark the position of the clue cells on the original image. First, count the number of clue cells in the clue cell position information obtained in the previous step. Then, calculate the ratio of the number of clue cells to the total number of cells in the leucorrhea microscopic image to obtain the proportion of clue cells in the image. Finally, mark the position information of the clue cells on the original leucorrhea microscopic image, and generate a clue cell identification result image with annotations. Through these steps, the distribution of clue cells in the entire leucorrhea sample can be intuitively reflected.

[0119] The specific implementation method of step S90 is to use the verification data set annotated by medical experts to iteratively optimize the feature discriminant equation group. First, a verification data set annotated by medical experts containing more than 1,000 images is established to ensure the diversity and representativeness of the data. Then, the verification data set is used to calculate the clue cell recognition accuracy and recall rate of the current feature discriminant equation group. Next, according to the evaluation results of the accuracy and recall rate, various parameters in the feature discriminant equation group, such as weight coefficients, correction terms, etc., are adjusted. Finally, the evaluation of the accuracy and recall rate and the process of parameter adjustment are repeated until the two indicators simultaneously reach the preset high standard values, such as the accuracy rate is not less than 90% and the recall rate is not less than 85%. This iterative optimization mechanism based on large-scale verification data can continuously improve the accuracy and robustness of clue cell recognition.

[0120] Among them, the validation data set annotated by medical experts needs to follow strict construction standards. First of all, the diversity and representativeness of the data should be ensured in the sample collection process, which includes collecting leucorrhea microscopic images of patients of different ages, different time periods, different lighting conditions and different magnifications. The total size of the validation data set should be no less than 1,000 images, and the resolution of each image is recommended to be no less than 2048×1536 pixels. The ratio of positive samples to negative samples should be maintained at about 1:4 to ensure the rationality of data distribution. In terms of image quality control, blurred and severely polluted images need to be eliminated, while ensuring the consistency of image acquisition parameters, and recording the acquisition environment parameters in detail to establish a complete image quality scoring system.

[0121] In the data annotation process, a double-blind cross-annotation method is used to ensure the objectivity and accuracy of the annotation results. The annotation content mainly includes two levels: one is the precise annotation of the clue cell position, using a rectangular box to mark and record the coordinate information (x, y, width, height), and annotating the integrity and clarity level of the cell; the second is the quantitative recording of cell characteristics, including key parameters such as the average area A0, grayscale mean μ0, and area standard deviation σA of standard clue cells. These parameters will be directly used for the construction and optimization of the feature discriminant equation group. The validation data set is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The training set is used for parameter training of the feature discriminant equation group, the validation set is used for parameter optimization and model selection, and the test set is strictly isolated for the final performance evaluation.

[0122] In order to ensure the continuous improvement of system performance, a complete iterative optimization mechanism has been established. Clear target values ​​are set for system performance indicators, including accuracy not less than 90%, recall not less than 85%, F1 score not less than 87%, and area under the ROC curve not less than 0.92. These indicators are continuously monitored and evaluated through the validation data set, and new validation samples are regularly added to adapt to new situations that may arise in practical applications. Through this systematic verification and optimization mechanism, it is ensured that the feature discriminant equation group can achieve stable and reliable clue cell recognition effects in practical applications.

[0123] The construction and application process of the entire validation data set has formed a closed-loop technical system, which provides a solid data foundation and reliable verification means for the intelligent identification of clue cells through scientific data collection, strict quality control, systematic parameter optimization and continuous performance evaluation. This method based on large-scale annotated data not only ensures the accuracy and reliability of the recognition algorithm, but also through continuous data accumulation and optimization, the system has good adaptability and scalability, and can continuously improve the recognition performance to meet the needs of actual medical applications.

[0124] The innovation of the whole technical solution lies in:

[0125] 1. The introduction of error correction terms improves the adaptability of the model. Correction terms are introduced in various characteristic calculation equations to compensate for the deviation of some parameters from the standard values ​​and enhance the applicability of the algorithm in practical applications.

[0126] 2. The multi-feature fusion strategy is adopted to improve the recognition accuracy. This method combines two key indicators, spectral features and morphological features, and introduces the synergistic terms between features in the comprehensive scoring equation, giving full play to the synergistic effect between different features, thereby greatly improving the recognition accuracy of clue cells.

[0127] 3. The threshold adaptive adjustment mechanism enhances the practicality of the method. The discrimination threshold equation is adaptively determined according to the comprehensive score distribution of expert-labeled samples. It has good robustness and can adapt to the recognition needs under different sample conditions, which enhances the applicability of the method in practical applications.

[0128] Specifically, the principle of the present invention is to construct a multi-level, adaptive image analysis framework. This framework realizes intelligent analysis of the entire process from image acquisition to the output of the final recognition result through a systematic processing flow. Although researchers have tried to apply image processing technology to the analysis of leucorrhea microscopic images. For example, some studies have used color feature-based segmentation methods for cell recognition, and some studies have used morphological feature extraction to classify cells. However, these methods still have some limitations: on the one hand, a single feature extraction method is difficult to accurately describe the complex characteristics of clue cells, especially in cases where the cell edges are blurred or overlapped, the recognition accuracy is low; on the other hand, existing methods generally lack adaptability and are difficult to cope with image changes in different batches and under different staining conditions, resulting in limited practicality of the algorithm.

[0129] The present invention constructs a multi-level, adaptive image analysis framework, the specific principles are as follows:

[0130] First, the method effectively improves the image quality by preprocessing the leucorrhea microscopic image, including grayscale, image enhancement and noise filtering. This lays a good foundation for the subsequent spectral feature extraction and cell region segmentation. Spectral features are important indicators for describing the biochemical composition of cells. Compared with traditional morphological features, they can provide richer and more detailed cell information. Therefore, the present invention uses multi-band spectral analysis technology to construct the spectral feature space of cells.

[0131] In the spectral feature space, this method uses an adaptive threshold segmentation algorithm to perform preliminary segmentation of the cell area. Compared with fixed threshold segmentation, the adaptive threshold can be dynamically adjusted according to the changes in image content, thereby improving the accuracy and stability of segmentation. This step lays the foundation for subsequent morphological feature extraction.

[0132] Next, the method uses edge detection and morphological processing techniques to extract the complete outline of the cell area from the preliminary segmentation results. Based on this outline information, the morphological characteristic parameters such as the cell area, perimeter, roundness, grayscale mean and grayscale variance are calculated. These features can reflect the size, shape and structural characteristics of the cell and are an important basis for distinguishing different cell types.

[0133] By integrating spectral features and morphological features, the present invention constructs a set of comprehensive feature discrimination equations. Among them, the spectral feature equation uses a combination of relative error and root mean square error, and takes into account the spectral shape and amplitude differences at the same time; the morphological feature equation includes multiple features such as circularity, grayscale distribution and area distribution, and uses an exponential function to reflect the penalty effect of parameter deviation from the standard. On this basis, the comprehensive scoring equation integrates spectral and morphological features, and introduces cross terms between features to reflect the synergistic effect of the two types of features. Finally, the discrimination threshold equation is based on statistical principles, combined with the sample mean and standard deviation to determine the optimal threshold, which has good robustness.

[0134] Through these characteristic discriminant equations, this method can effectively classify and identify cell regions and achieve accurate detection of clue cells. To further improve the recognition effect, this method also uses post-processing techniques such as noise interference filtering and connected region analysis to remove possible interference and obtain more reliable clue cell location information.

[0135] The key to the solution of the present invention being able to effectively solve the problem of clue cell identification lies in its systematicity and adaptability. The systematicity is reflected in the construction of a complete processing and analysis chain, and each link has been optimized and designed to ensure the improvement of overall performance. Adaptability is reflected in multiple levels: adaptive enhancement in the image preprocessing stage, adaptive threshold selection in the segmentation stage, multi-dimensional feature combination in the feature extraction stage, and the final iterative optimization mechanism. This design enables the system to adapt to image changes under different conditions and maintain stable recognition performance.

[0136] The following is a specific embodiment 1 of the present invention. The specific implementation of each step in this embodiment 1 is described in detail as follows: The specific implementation of step S10 is to pre-process the acquired leucorrhea microscopic image. First, the leucorrhea microscopic image is grayed to convert the color image into a grayscale image:

[0137] Igray (x,y)=0.299×I R (x,y)+0.587×I G (x,y)+0.114×I B (x,y);

[0138] Among them, I R (x,y),I G (x,y) and I B (x, y) represent the pixel values ​​of the original color image in the red, green, and blue channels respectively. gray (x,y) is the converted grayscale image pixel value.

[0139] Next, the grayscale image is subjected to medical image enhancement processing, including histogram equalization processing and contrast enhancement processing. The histogram equalization formula is as follows:

[0140]

[0141] Among them, I eq (x, y) is the pixel value after equalization, L is the grayscale level (usually 256), N is the total number of pixels in the image, n i is the number of pixels of the i-th gray level.

[0142] The contrast enhancement is performed using the following formula:

[0143]

[0144] Then, the enhanced image is subjected to noise removal processing, including median filtering and Gaussian filtering. The median filtering formula is:

[0145] I med (x,y)=median{I enh (xk,yl)|k,l∈W};

[0146] Where W is the m×n filter window, and median represents the median of the pixel values ​​in the window. The Gaussian filter formula is:

[0147]

[0148] σ is the standard deviation of the Gaussian kernel and determines the smoothness of the filter.

[0149] Finally, the image after noise removal is output as the preprocessing result.

[0150] The specific implementation of step S20 is to perform spectral feature analysis on the pre-processed microscopic image. First, multi-band decomposition is performed on the pre-processed microscopic image to obtain the spectral response values ​​V of the image in different bands.i Then, the spectral response values ​​of different bands are combined into a spectral feature vector V = [V1, V2, ..., V n ] T , where n is the number of spectral bands. Next, the spectral feature vector is normalized:

[0151]

[0152] Get the normalized spectral feature vector Finally, the spectral feature space is constructed based on the normalized spectral feature vector.

[0153] The specific implementation of step S30 is to use an adaptive threshold segmentation method to segment the spectral feature space into cell regions. First, the histogram distribution of the spectral feature space is calculated:

[0154]

[0155] Among them, δ(x) is the Dirac delta function. Then, the initial threshold T0 is determined based on the histogram distribution:

[0156] T0=argmax T {h(T)};

[0157] Next, the initial threshold T0 is used for iterative optimization in the spectral feature space to obtain the optimal threshold T * :

[0158]

[0159] Finally, using T * The spectral feature space is segmented to obtain the initial segmentation map of the cell area. This method can adaptively determine the optimal segmentation threshold according to the image content, thus improving the accuracy of segmentation.

[0160] The specific implementation of step S40 is to perform edge detection and morphological processing on the initial segmentation map to extract the complete outline of the cell area. First, perform Canny edge detection on the initial segmentation map to obtain a cell edge map:

[0161]

[0162] in, and is the gradient of the image in the x and y directions, θ is the gradient direction, T1 and T2 are high and low thresholds. Then, for the edge map I edge Perform morphological opening operation to remove small noise interference:

[0163]

[0164] in, and Denote the erosion and dilation operations respectively, and B_1 and B_2 are structural elements. Next, perform a morphological closing operation on the image after the opening operation to fill the gaps on the edges:

[0165]

[0166] Finally, from the image I after the closed operation close The complete outline of the cell area was extracted.

[0167] The specific implementation of step S50 is to calculate various morphological characteristic parameters based on the extracted cell area contour. First, the cell area A and cell perimeter P are calculated based on the cell area contour:

[0168] A=∑ x,y f(x,y);

[0169] P=∑ x,y g(x,y);

[0170] Among them, f(x, y) is the cell area image after binarization, and g(x, y) is the contour image after edge detection. Then, the cell roundness C is calculated using the area A and perimeter P:

[0171]

[0172] Next, extract the grayscale value h(x,y) inside the cell area and calculate the mean μ and variance σ of the grayscale value 2 :

[0173]

[0174] These morphological characteristic parameters can describe the size, shape and grayscale distribution characteristics of cells, providing a basis for subsequent cell identification.

[0175] The specific implementation method of step S60 is to establish a characteristic discrimination equation group to classify and identify the cell area. First, construct the spectral characteristic discrimination equation:

[0176]

[0177] in, is the normalized spectral response value of the i-th band of the cell to be tested, is the normalized spectral response value of the i-th band of the standard template, α and β are weight coefficients, and γ is the correction term.

[0178] Then, the morphological feature discriminant equation is constructed:

[0179]

[0180] Among them, A is the cell area, P is the cell perimeter, μ is the grayscale mean of the cell area, μ0 is the standard grayscale mean, σ is the grayscale standard deviation of the cell area, A0 is the standard cell area, σ A is the area standard deviation, ω1, ω2 and ω3 are weight coefficients, and δ is the correction term.

[0181] Next, construct the comprehensive scoring equation:

[0182]

[0183] Among them, λ1 and λ2 are feature weight coefficients, θ is the cross-term coefficient, and η is the comprehensive error term.

[0184] Finally, construct the discrimination threshold equation:

[0185]

[0186] in, is the comprehensive score of the sample annotated by the i-th expert, m is the number of samples, k is the confidence coefficient, is the sample average score.

[0187] Through these characteristic discriminant equations, clue cells can be effectively identified.

[0188] The specific implementation of step S70 is to post-process the classification and recognition results to further improve the recognition accuracy. First, the classification results are processed by morphological filtering to remove possible noise interference:

[0189]

[0190] Among them, I rec is the classification recognition result, B5 and B6 are the structural elements used for opening and closing operations. Then, the filtered image I filtered Perform connected region analysis and mark each connected region:

[0191] R i ={(x,y)|I filtered (x,y)=i};

[0192] Next, calculate each connected region R i Area A i and perimeter P i :

[0193]

[0194] Among them, N(x,y) is the 8-neighborhood of the (x,y) pixel. Finally, according to the area and perimeter characteristics of the connected region, the region that meets the characteristics of the clue cell is screened out to obtain the final clue cell location information.

[0195] The specific implementation of step S80 is to count the number of clue cells and their proportion in the leucorrhea microscopic image, and mark the location of the clue cells on the original image. First, count the number of clue cells N in the clue cell location information obtained in the previous step. c Then, calculate the number of clue cells N c The total number of cells in the leucorrhea microscopic image N t The ratio of the clue cells in the image is obtained. t :

[0196]

[0197] Finally, the location information of clue cells is marked on the original leucorrhea microscopic image to generate an annotated clue cell identification result image.

[0198] The specific implementation of step S90 is to use the verification data set annotated by medical experts to iteratively optimize the feature discriminant equation group. First, a verification data set annotated by medical experts containing more than 1,000 images is established. Then, the clue cell recognition accuracy P and recall rate R of the current feature discriminant equation group are calculated using the verification data set:

[0199]

[0200] Among them, TP is the number of true positives, FP is the number of false positives, and FN is the number of false negatives. Next, according to the evaluation results of precision P and recall R, adjust the various parameters in the feature discriminant equation group, such as weight coefficients α, β, ω i ,λ i , correction terms γ, δ, η, as well as the confidence coefficient k and the cross-term coefficient θ. Finally, the process of evaluating the accuracy and recall rate and adjusting the parameters is repeated until both indicators reach the preset high standard values ​​at the same time, such as the accuracy rate is not less than 90% and the recall rate is not less than 85%.

[0201] The innovation of the whole technical solution lies in:

[0202] 1. The error correction term is introduced to improve the adaptability of the model. In various characteristic calculation equations, the correction terms γ, δ, and η are introduced to compensate for the deviation of some parameters from the standard values ​​and enhance the applicability of the algorithm in practical applications.

[0203] 2. The multi-feature fusion strategy is adopted to improve the recognition accuracy. This method combines two key indicators, spectral features and morphological features, and introduces cross terms between features in the comprehensive scoring equation, giving full play to the synergy between different features, thereby greatly improving the recognition accuracy of clue cells.

[0204] 3. The threshold adaptive adjustment mechanism enhances the practicality of the method. The discrimination threshold equation is adaptively determined according to the comprehensive score distribution of expert-labeled samples. It has good robustness and can adapt to the recognition needs under different sample conditions, which enhances the applicability of the method in practical applications.

[0205] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: The Department of Obstetrics and Gynecology of a certain hospital adopts the method of the present invention to automatically detect and analyze the leucorrhea specimens of visiting patients, aiming to improve the accuracy of clue cell identification and provide a more reliable basis for clinical diagnosis.

[0206] Sample collection and pretreatment

[0207] The hospital uses conventional vaginal discharge sampling methods to collect vaginal discharge samples from patients. The collected samples are thinly smeared and stained to make microscope specimens. For these stained vaginal discharge microscopic images, grayscale preprocessing is first performed. Specifically, the original color image is converted into grayscale according to I gray (x,y)=0.299×I R (x,y)+0.587×I G (x,y)+0.114×I B The formula for (x,y) is converted to a grayscale image (such as Figure 3 shown).

[0208] Next, the grayscale image is enhanced. First, the histogram equalization technique is used to improve the image contrast. The formula is: Where L = 256 is the gray level, N is the total number of pixels, n i is the number of pixels at the i-th gray level. Then, the visual effect of the cell area is further enhanced by contrast enhancement processing, and the formula is Get the image as Figure 4 As shown (for a clearer representation Figure 3 and Figure 4 The difference is shown in the attached figure. Figure 3 The size is converted to Figure 4 same size).

[0209] Finally, the enhanced image is subjected to noise removal. First, the median filter technique is used to eliminate small interference spots. med(x,y)=median{I enh (xk,yl)|k,l∈W}, where W is a 5×5 filter window. Then, Gaussian filtering is applied to further smooth the image. Where σ = 1.0. The image is as follows Figure 5 Through the above preprocessing steps, a good quality leucorrhea microscopic image was obtained, laying the foundation for the subsequent spectral feature analysis.

[0210] Spectral feature extraction and cell region segmentation: For the preprocessed microscopic images, this method uses multi-band spectral analysis technology to extract the spectral features of cells. Thus, the spectral response value V of each pixel in different bands is obtained. i . These spectral response values ​​are combined into a spectral feature vector V = [V1, V2, ..., V 11 ] T In order to eliminate the dimension difference, the spectral feature vector is normalized. in

[0211] Based on the normalized spectral feature vector This method constructs the spectral feature space. First, the histogram distribution h(v) of the feature space is calculated, and then the initial threshold T0=argmax is determined. T {h(T)}. Next, the adaptive threshold optimization algorithm is used to segment the cell region in the feature space to obtain the optimal threshold The optimal threshold is used to segment the feature space and obtain the initial segmentation map of the cell area.

[0212] Morphological feature extraction and cell recognition: For the cell area obtained by preliminary segmentation, this method further extracts morphological features. First, the Canny operator is used to perform edge detection on the segmentation image to obtain the cell outline image I edge Then, morphological opening and closing operations are performed on the edge map to remove noise interference and fill the edge gaps to obtain the complete outline of the cell area.

[0213] Based on these contour information, the area A, perimeter P, and circularity of each cell region are calculated. Grayscale mean μ and grayscale variance σ 2 Morphological characteristic parameters.

[0214] The above spectral features and morphological features are integrated to construct a comprehensive feature discrimination equation group. The spectral feature equation is:

[0215]

[0216] in, is the normalized spectral response value of the i-th band of the standard template. The morphological characteristic equation is:

[0217]

[0218] The comprehensive scoring equation is:

[0219]

[0220] The discrimination threshold equation is:

[0221]

[0222] in, is the comprehensive score of the sample labeled by the i-th expert is the sample average score.

[0223] According to the above characteristic discriminant equation group, the cell area is classified and identified. After obtaining the preliminary clue cell detection results, this method also performs post-processing:

[0224] 1. Morphological filtering, Remove noise interference;

[0225] 2. Connected region analysis, marking each cell region R i ;

[0226] 3. Calculate the area A of each connected region i and perimeter P i ;

[0227] 4. According to A i and P i Features, filter out areas that meet the characteristics of clue cells.

[0228] Finally, reliable clue cell location information was obtained.

[0229] Sample verification and performance evaluation: To ensure the stability and reliability of the data processing method in practical applications, the hospital established a rigorous verification data set. From August 2022 to July 2023, a total of 500 leucorrhea microscopic images of different patients were collected, covering different age groups, different physiological states, and different collection times. These images were double-blindly annotated by experts, and the precise location information of clue cells and standard parameters such as cell area A0=80±25 and grayscale mean μ0=16±8 were obtained.

[0230] The validation data set is divided into training set, validation set and test set in a ratio of 6:2:2. The training set is used for parameter optimization of the feature discriminant equation group, the validation set is used for model selection and performance evaluation, and the test set is used as the final performance verification. The specific indicator requirements are as follows:

[0231] 1. The accuracy of clue cell identification P ≥ 90%;

[0232] 2. The recall rate of clue cell recognition R ≥ 85%;

[0233] 3. F1 score

[0234] 4. Area under the ROC curve AUC ≥ 0.92;

[0235] By repeatedly evaluating on the validation set, the parameters in the feature discrimination equation group are continuously optimized, such as spectral feature weights α=0.6, β=0.4, morphological feature weights ω1=0.4, ω2=0.3, ω3=0.3, comprehensive score weights λ1=0.6, λ2=0.4, θ=0.2, and various correction terms γ=0.05, δ=0.03, η=0.02, k=2, etc. Finally, the method was verified on the test set, and the clue cell recognition accuracy of the method reached 91.3%, the recall rate reached 86.7%, the F1 score was 88.9%, and the area under the ROC curve was 0.94, which fully met the preset performance goals.

[0236] The hospital has widely used the data processing method proposed in this invention in its daily work. The following are some recent typical cases:

[0237] Case 1: A 32-year-old woman came to the hospital with complaints of abnormal vaginal discharge for 2 weeks. After collecting the specimen, the method automatically analyzed the microscopic image of the vaginal discharge and detected that the proportion of clue cells reached 28%, which was much higher than the normal level. Based on this result, the doctor diagnosed it as bacterial vaginitis and gave targeted antibacterial treatment. After a follow-up examination 2 weeks later, the proportion of clue cells dropped to 5%, and the patient's symptoms were relieved.

[0238] Case 2: A 45-year-old woman came to the hospital with complaints of vaginal discomfort and lower abdominal pain. The results of the leucorrhea analysis showed that the proportion of clue cells was 12%, indicating the possibility of moderate vaginitis. The doctor prescribed anti-inflammatory drugs based on this result and asked for a follow-up examination in 2 weeks. During the follow-up examination, the proportion of clue cells dropped to 7%, and the symptoms improved significantly.

[0239] Case 3: A 28-year-old woman was found to have abnormal vaginal discharge during a physical examination, and the doctor prescribed a vaginal discharge examination. This method analyzed the microscopic images of vaginal discharge and detected that the proportion of clue cells was only 3%, which was within the normal range. Combined with other indicators, the doctor determined that the patient did not need special treatment and recommended regular follow-up.

[0240] From the above cases, it can be seen that the data processing method proposed by the present invention can accurately identify clue cells in leucorrhea and provide doctors with objective and reliable diagnosis basis. Compared with the traditional manual interpretation method, this method has higher accuracy and repeatability, greatly improving the diagnosis efficiency. After adopting this method, the hospital not only shortened the diagnosis cycle, but also significantly reduced the risk of misdiagnosis.

[0241] It should be noted that the variables involved in the description of the present invention are explained as shown in Table 1 below.

[0242] Table 1 Variable explanation table

[0243]

[0244] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for identifying clue cells in leucorrhea microscopic images, characterized in that: The following steps are involved: S10, obtaining a stained leucorrhea microscopic image, and preprocessing the leucorrhea microscopic image, wherein the preprocessing includes grayscale processing, medical image enhancement processing, and noise removal processing to obtain a preprocessed microscopic image; S20, performing spectral feature analysis on the preprocessed microscopic image, extracting the spectral feature vector of the microscopic image, and constructing a spectral feature space of the microscopic image; S30, in the spectral feature space, using an adaptive threshold segmentation method to perform cell region segmentation on the spectral feature space to obtain an initial cell region segmentation map; S40, performing edge detection processing on the initial segmentation map of the cell region, performing morphological opening operation processing and morphological closing operation processing on the segmentation map after the edge detection processing, and extracting the complete outline of the cell region; S50, calculating morphological characteristic parameters of the cell region based on the complete outline of the cell region, wherein the morphological characteristic parameters include cell area, cell perimeter, cell roundness, cell grayscale mean, and cell grayscale variance; S60, establishing a characteristic discriminant equation group according to the spectral characteristic vector and the morphological characteristic parameter, and classifying and identifying the cell region using the characteristic discriminant equation group; S70, post-processing the classification and recognition results, wherein the post-processing includes noise interference removal processing and connected region analysis processing to obtain clue cell position information; S80, counting the number of clue cells in the leucorrhea microscopic image, calculating the proportion of the clue cells in the leucorrhea microscopic image, and generating a clue cell identification result annotated image.

2. The method according to claim 1, characterized in that The step S10 specifically includes: Step 101, performing grayscale preprocessing on the leucorrhea microscopic image to convert the leucorrhea microscopic image into a grayscale image; Step 102: performing medical image enhancement processing on the grayscale image, including histogram equalization processing and contrast enhancement processing; Step 103, performing noise removal processing on the image after the medical image enhancement processing, including median filtering processing and Gaussian filtering processing; Step 104: output the image after the noise removal process as a pre-processed microscopic image.

3. The method according to claim 1, characterized in that The step S20 specifically includes: Step 201, performing multi-band decomposition on the preprocessed microscopic image to obtain spectral response values ​​of the preprocessed microscopic image in different bands; Step 202: Combining the spectral response values ​​of the different wavebands to form a spectral feature vector; Step 203, normalizing the spectral feature vector to obtain a normalized spectral feature vector; Step 204: construct a spectral feature space based on the normalized spectral feature vector.

4. The method according to claim 1, characterized in that: The step S30 specifically includes: Step 301, calculating the histogram distribution of the spectral feature space; Step 302: determining an initial threshold based on the histogram distribution; Step 303: performing iterative optimization using the initial threshold in the spectral feature space to obtain an optimal threshold; Step 304: segment the spectral feature space using the optimal threshold to obtain an initial segmentation map of the cell region.

5. The method according to claim 1, characterized in that: The step S40 specifically includes: Step 401, performing edge detection processing on the initial segmentation map of the cell region to obtain a cell edge map; Step 402, performing morphological opening operation on the cell edge map to remove small interferences; Step 403, performing morphological closing operation processing on the image after the morphological opening operation processing to fill the edge gaps; Step 404: extract the complete outline of the cell region in the image after the morphological closing operation.

6. The method according to claim 1, characterized in that The step S50 specifically includes: Step 501, calculating the cell area and cell perimeter according to the complete outline of the cell region; Step 502, calculating the cell roundness according to the cell area and the cell perimeter; Step 503, extracting the gray value of the cell region according to the complete outline of the cell region; Step 504: Calculate the mean and variance of the grayscale value of the cell region.

7. The method according to claim 1, characterized in that The step S60 specifically includes: Step 601, establishing a spectral feature discrimination equation, and calculating the spectral feature score of the cell area; Step 602: Establish a morphological feature discrimination equation to calculate the morphological feature score of the cell region; Step 603: Establish a comprehensive scoring equation to integrate the spectral feature score and the morphological feature score; Step 604: Establish a discrimination threshold equation to determine the threshold for clue cell classification and identification.

8. The method according to claim 1, characterized in that: The step S70 specifically includes: Step 701: Perform morphological filtering on the classification and recognition results to remove noise interference; Step 702: marking connected regions of the image after the morphological filtering process; Step 703: Calculate the area and perimeter of the connected region; Step 704: Filter and obtain clue cell location information according to the area and perimeter of the connected region.

9. The method according to claim 1, characterized in that: The step S80 specifically includes: Step 801, counting the number of clue cells in the clue cell position information; Step 802, calculating the ratio of the number of clue cells to the total number of cells in the leucorrhea microscopic image; Step 803, marking the location of the clue cells in the leucorrhea microscopic image; Step 804: Generate a clue cell recognition result image with annotation information.

10. The method according to claim 1, characterized in that The method further includes S90, performing iterative optimization processing on the feature discrimination equation group based on the verification data set annotated by the medical expert until the accuracy and recall rate of the clue cell identification reach a preset standard value, specifically including: Step 901: Establish a verification data set annotated by medical experts; Step 902: Calculate the accuracy and recall rate of clue cell recognition using the verification data set; Step 903: adjusting the parameters in the feature discrimination equation group according to the accuracy and the recall rate; Step 904: Repeat step 902 and step 903 until the accuracy and the recall rate reach preset standard values.

Images