Method for determining rat estrous cycle based on image recognition

Abstract

The present invention relates to the field of image recognition and judgment technology, and in particular to a method for judging the estrous cycle of rats based on image recognition. Its technical solution includes sample collection and preparation, image collection and optimization, cell segmentation and positioning, feature extraction and analysis, cell classification and recognition, and result verification and feedback. The present invention significantly improves the efficiency and accuracy of morphological recognition of rat vaginal exfoliated cells through systematic and precise operations. From precise sample collection and image acquisition optimization, to multi-dimensional feature extraction and integrated learning classification, and then to result verification and feedback, the scientific nature of recognition is fully guaranteed, which not only greatly shortens the detection time and reduces manual errors, but also continuously improves performance through dynamic optimization, providing efficient and reliable technical support for the study of the estrous cycle of rats.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition and judgment, and in particular to a method for judging the estrous cycle of rats based on image recognition. Background Art

[0002] In biomedical research, rats are commonly used as experimental animals, and accurate monitoring of their estrous cycles is crucial for studying estrogen-related physiological mechanisms, drug development, and reproductive biology. Estrogen can affect the estrous cycle in rats, primarily by promoting the proliferation of vaginal mucosal cells, which in turn accelerates their shedding.

[0003] Traditional methods for monitoring the estrous cycle in rats rely primarily on daily vaginal exfoliated cell examinations. The specific procedure involves first acclimatizing the rats for a week, followed by daily vaginal exfoliated cell collection between 8:00 and 9:00 a.m. for the next 10 days. During the collection process, personnel use a pipette to drip 200 μL of sterile saline into each well of a 48-well plate, matching the well number to the rat number. Afterwards, the rat is secured with the left hand, the tail is pulled to expose the vulva, and the right hand holds the pipette to draw up the saline from the corresponding well. The tip of the pipette is gently placed 2-4 mm into the rat's vaginal opening, and the cells are slowly injected and aspirated. After repeating this three times, 80 μL of the aspirated fluid is injected into the original well. Repeated pipetting allows the vaginal exfoliated cells to enter the saline solution, and a new pipette tip is required for each rat.

[0004] After sample collection is complete, the 48-well plate is left to stand for 10 minutes until the cells settle to the bottom of the wells, and then placed under an inverted microscope for observation. Staff need to determine the rat's estrous cycle based on the dynamic changes in the types and proportions of three types of cells: white blood cells, nucleated epithelial cells, and non-nucleated epithelial cells observed under the microscope. This method of judgment not only requires staff to have extensive experience and the ability to accurately identify the morphological characteristics of different cells, but also requires a deep understanding of the relationship between changes in cell proportions and the estrous cycle.

[0005] However, this traditional method of judging the estrous cycle of rats through manual microscopic examination has many disadvantages. First, the operation process is extremely cumbersome. Samples need to be collected at a fixed time every day, and the collection process involves multiple delicate operation steps, which requires high technical skills of the operator. Secondly, microscopic examination is time-consuming and labor-intensive. One microscopic examination often requires processing dozens of slides, and it usually takes a whole day to complete the observation and analysis of all samples. Furthermore, the accuracy of manual judgment depends to a large extent on the professional level and working status of the operator, and is easily affected by subjective factors. There may be differences in judgment between different operators, which makes it difficult to ensure the reliability and consistency of the results.

[0006] In summary, this application proposes a method for determining the estrous cycle of rats based on image recognition. Summary of the Invention

[0007] The purpose of the present invention is to address the problem of low efficiency in manually identifying cell morphological characteristics in the background art and to propose a method for determining the rat estrous cycle based on image recognition.

[0008] The technical solution of the present invention: a method for determining the estrous cycle of rats based on image recognition, comprising the following steps:

[0009] Sample collection and preparation: Rat vaginal exfoliated cell samples were collected, processed, and preserved according to specific methods;

[0010] Image acquisition and optimization: Use professional equipment to acquire high-resolution images and perform geometric correction, color correction, and image denoising preprocessing;

[0011] Cell segmentation and localization: A semantic segmentation model based on a convolutional neural network is used to segment cells, and the segmented cells are accurately located and labeled;

[0012] Feature extraction and analysis: Extract the shape, texture, and optical features of cells, and perform feature selection and dimensionality reduction;

[0013] Cell classification and identification: Use an integrated learning classifier to classify cells and optimize the classification results;

[0014] Result verification and feedback: Verify the classification results through multiple indicators, and provide feedback and optimization to each link based on the verification results.

[0015] Optionally, when collecting rat vaginal exfoliated cell samples, use a high-precision pipette and draw and inject 200 μL of sterile saline each time. After collection, store the sample at an ambient temperature of 25-28°C. Before microscopic observation, add 3-4% paraformaldehyde solution as a cell fixative. The ratio of fixative to sample is 1:4, and the fixation time is 15-30 minutes.

[0016] Optionally, in the image acquisition and optimization, the geometric correction adopts a method based on polynomial fitting. For the quadratic polynomial distortion model, the formula is:

[0017]

[0018]

[0019] in are the original coordinates, are the corrected coordinates, and is the correction coefficient, and the distortion model is established by collecting images of a known standard grid.

[0020] Optionally, in the image acquisition and optimization, color correction uses a white balance algorithm to select a white reference area in the image, calculate its average RGB value, and then normalize the RGB value of the entire image. The formula is:

[0021] , ,

[0022] in 、 、 is the original RGB value, 、 、 is the average RGB value of the white reference area, 、 、 is the corrected RGB value.

[0023] Optionally, in the image acquisition and optimization, the image denoising adopts the wavelet transform denoising method, selects the db4 wavelet to perform multi-layer decomposition on the image, removes the high-frequency noise component by threshold processing, and then performs wavelet reconstruction. The threshold adopts a universal threshold according to the noise level. Determine, among which, is the noise standard deviation, is the total number of image pixels.

[0024] Optionally, in the cell segmentation and localization, a semantic segmentation model based on a convolutional neural network adopts U-Net, the network structure includes a contraction path and an expansion path, and a large number of labeled rat vaginal exfoliated cell images are used for network training, and a cross entropy loss function is adopted:

[0025]

[0026] in, is the total number of pixels, is the number of cell types, including background, white blood cells, nucleated epithelial cells and keratinized epithelial cells. is the true label of the category to which the pixel belongs, It is the prediction probability, and the training dataset is expanded using data augmentation techniques, including random rotation, flipping, scaling, and elastic deformation.

[0027] Optionally, in the cell segmentation and positioning, when positioning the segmented cells, the centroid coordinates of each cell are calculated. , the formula is:

[0028]

[0029] in, is the pixel set of the cell area, are pixel coordinates and assign a unique identifier to each cell.

[0030] Optionally, in the feature extraction and analysis, shape feature extraction includes fractal dimension, which is calculated using the box counting method, dividing the cell image into grids of different sizes, and counting the number of grids containing cell pixels. , fractal dimension:

[0031]

[0032] It also includes extracting the aspect ratio and convex hull area ratio features of cells;

[0033] On the basis of gray-level co-occurrence matrix and local binary pattern, the directional gradient histogram feature is added to divide the cell area into multiple small cell units, the gradient direction histogram in each unit is calculated, and then the histograms are connected in series as the texture feature of the cell;

[0034] A texture analysis method based on wavelet transform is used to extract wavelet coefficients at different scales as texture features to capture subtle texture changes inside cells.

[0035] The optical properties of cells under different wavelengths of light were analyzed. Multispectral imaging technology was used to collect cell images at different wavelengths of 400nm-700nm. The spectral reflectance and absorbance optical characteristics of the cells were extracted, and the optical characteristics were reduced and analyzed using the principal component analysis method.

[0036] A correlation-based feature selection method was used to screen out features with high correlation with cell types. The evaluation indicators of CFS are:

[0037]

[0038] in, is the number of features, is the average correlation between features and classes, is the average correlation between features.

[0039] Optionally, the cell classification and identification specifically includes the following steps: constructing and training a random forest, determining the number of decision trees K through a cross-validation experiment, where K is an integer greater than 100, randomly selecting a feature subset of size L% of the total number of features from all features, where L is an integer between 30 and 70, using information gain or Gini index as a splitting criterion, and stopping growth when the decision tree reaches a maximum depth H, where H is an integer greater than 5;

[0040] For SVM classifier training, the linear kernel SVM uses a sequential minimum optimization algorithm. The penalty parameter C is determined through cross-validation. The degree of the polynomial kernel function ranges from [2 to 5], and the value range of the radial basis kernel function γ ranges from [0.001 to 10]. The optimal value is determined through grid search and cross-validation. When integrating multiple SVM classifiers, the weight is determined based on the accuracy of each SVM classifier on the validation set.

[0041] The classification results were optimized by setting criteria for determining uncertain samples. The uncertain samples were manually labeled and added to the training set. The number of training set samples increased by no more than T1% of the original number of training set samples each time they were retrained, where T1 was an integer between 10 and 30. Results with a confidence level lower than U were revalidated, where U was a value between 0.3 and 0.5. Feature dimensions such as cell grayscale mean, grayscale variance, and edge gradient mean were added. A deep neural network was used for reclassification, and the network structure parameters were determined experimentally.

[0042] Optionally, the result verification and feedback specifically include the following steps:

[0043] The accuracy, recall, F1-score, Matthews correlation coefficient, and area under the receiver operating characteristic curve of the classification results were calculated. Stratified sampling was used to extract 20% of the samples as a validation set to ensure that the proportions of the three cell types in the validation set were consistent with the actual distribution of the original samples.

[0044] If the classification accuracy of a certain cell type is lower than 80%, focus on analyzing the sample characteristics of that cell type, adjust the feature extraction method or feature selection strategy, and retrain the ensemble learning classifier until the classification accuracy of each cell type reaches above 80%.

[0045] Compared with the prior art, this application has at least one of the following beneficial technical effects:

[0046] The present invention greatly reduces the time cost of manual microscopic examination through the automatic recognition method. The microscopic examination work that traditional methods take one day to complete can be quickly processed by this application, significantly improving the detection efficiency.

[0047] Furthermore, the use of multiple image processing and analysis techniques, along with an integrated learning classifier, reduces subjective errors associated with manual judgment and improves the accuracy of cell type identification and estrous cycle determination. The entire identification process is based on clear algorithms and models, unaffected by operator subjective factors, ensuring reliable and consistent results. By adhering to established procedures and parameter settings, similar results can be obtained across time and by different operators, facilitating replication and validation of research.

[0048] Extracting cell features from multiple dimensions, including shape, texture, and optical characteristics, and conducting comprehensive analysis can more comprehensively describe cell characteristics and improve classification results. Through active learning strategies and confidence assessment mechanisms, classifier performance can be continuously optimized to adapt to sample changes and ensure long-term stable detection results. The parameters in the scheme are mostly determined through experimentation and can be flexibly adjusted based on actual sample conditions and research needs to adapt to different experimental conditions and data characteristics.

[0049] Through systematic and precise operations, the present invention significantly improves the efficiency and accuracy of morphological identification of rat vaginal exfoliated cells. From precise sample collection and image acquisition optimization to multi-dimensional feature extraction and integrated learning classification, and then to result verification and feedback, the scientific nature of identification is fully guaranteed. It not only greatly shortens the detection time and reduces human errors, but also continuously improves performance through dynamic optimization, providing efficient and reliable technical support for the study of the rat estrous cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The figure is a flow chart of the method for determining the rat estrous cycle based on image recognition. DETAILED DESCRIPTION

[0051] The technical solutions of the present disclosure are described clearly and completely below with reference to specific embodiments. Obviously, the embodiments described are only a portion of the embodiments of the present disclosure, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present disclosure without creative effort are also within the scope of protection of the present disclosure.

[0052] like Figure 1 As shown, the method for determining the rat estrous cycle based on image recognition proposed in the present invention includes sample collection and preparation, image collection and optimization, cell segmentation and positioning, feature extraction and analysis, cell classification and identification, and result verification and feedback. Each step is described in detail below.

[0053] 1. Sample Collection and Preparation

[0054] Sample collection should strictly follow the prescribed rat vaginal exfoliated cell collection method to ensure the independence and accuracy of each rat sample. During the collection process, pay attention to gentle manipulation to avoid unnecessary damage to the cells. Use a high-precision pipette and sterile saline solution to ensure the precise amount of liquid each time to aspirate and inject to ensure sample consistency.

[0055] Sample processing and storage: After collection, samples should be immediately stored in a stable environment at 25-28°C in a 48-well plate to prevent changes in cell morphology over a short period of time. Before microscopic observation, add an appropriate amount of cell fixative (4% paraformaldehyde solution) at a 1:4 ratio of fixative to sample to ensure that the cells maintain their original morphology during subsequent processing. The fixation time should be set to 15-30 minutes.

[0056] 2. Image Acquisition and Optimization

[0057] For high-resolution image acquisition, use a professional inverted microscope equipped with a high-resolution, high-sensitivity digital camera. Select a microscope magnification between 100x and 400x, depending on cell size and morphology. Adjust the intensity and angle of the microscope light source to achieve uniform illumination with moderate contrast. Light intensity can be adjusted between 50% and 80%, and optimal lighting conditions should be determined through multiple trials.

[0058] Image preprocessing, geometric correction: Due to the possible distortion of the microscope lens, the collected images are geometrically corrected. A polynomial fitting method is used to build a distortion model by collecting images of a known standard grid. For the quadratic polynomial distortion model, the formula is:

[0059]

[0060]

[0061] in are the original coordinates, are the corrected coordinates, and is the correction factor.

[0062] Color correction: To eliminate color deviation caused by uneven lighting, a white balance algorithm is used for color correction. A white reference area (such as a blank background area) is selected in the image, its average RGB value is calculated, and then the RGB value of the entire image is normalized using the formula:

[0063]

[0064] in is the original RGB value, is the average RGB value of the white reference area, is the corrected RGB value.

[0065] In this embodiment, image denoising is performed: in addition to median filtering, wavelet transform denoising method can also be used. Select the appropriate wavelet basis function db4 wavelet, perform multi-layer decomposition on the image, remove high-frequency noise components through threshold processing, and then perform wavelet reconstruction. The threshold adopts a general threshold according to the noise level. Determine, among which, is the noise standard deviation, is the total number of image pixels.

[0066] 3. Cell Segmentation and Localization

[0067] Deep learning-based cell segmentation uses a convolutional neural network (CNN)-based semantic segmentation model (such as U-Net) to segment cells. The U-Net network structure includes a contraction path (for feature extraction) and an expansion path (for upsampling and feature fusion). The network is trained using a large number of labeled rat vaginal exfoliated cell images. The cross-entropy loss function is used:

[0068]

[0069] in, is the total number of pixels, is the number of cell types, including background, white blood cells, nucleated epithelial cells and keratinized epithelial cells. is the true label of the category to which the pixel belongs, is the predicted probability.

[0070] In order to improve the segmentation accuracy, data augmentation techniques such as random rotation, flipping, scaling and elastic deformation are used to expand the training dataset.

[0071] Cell positioning and labeling: accurately locate and label the segmented cells. Calculate the centroid coordinates of each cell. , the formula is:

[0072]

[0073] in, is the pixel set of the cell area, is the pixel coordinate, which assigns a unique identifier to each cell to facilitate subsequent feature extraction and classification.

[0074] 4. Feature Extraction and Analysis

[0075] Multi-dimensional feature extraction

[0076] Shape features, in addition to basic area, perimeter, and circularity, also extract the shape complexity of cells (such as fractal dimension). The fractal dimension is calculated using the box counting method, which divides the cell image into grids of different sizes and counts the number of grids containing cell pixels. , fractal dimension:

[0077]

[0078] The aspect ratio and convex hull area ratio features of the cells are extracted to further describe the shape characteristics of the cells.

[0079] Texture features are enhanced with histograms of oriented gradients (HOG) features based on gray-level co-occurrence matrices (GLCMs) and local binary patterns (LBPs). The cell region is divided into multiple small cell units, and the gradient orientation histogram within each unit is calculated. These histograms are then concatenated as the cell's texture features. A wavelet-based texture analysis method is used to extract wavelet coefficients at different scales as texture features, capturing subtle texture variations within the cell.

[0080] Optical characteristics: Analyze the optical properties of cells under different wavelengths of light. Using multispectral imaging technology, we capture cell images at different wavelengths between 400nm and 700nm and extract the spectral reflectance and absorbance optical characteristics of the cells. We then perform dimensionality reduction and analysis of these optical characteristics using methods such as principal component analysis (PCA).

[0081] Feature selection and dimensionality reduction: Correlation-based feature selection (CFS) is used to select features that are highly correlated with cell types. The evaluation indicators of CFS are:

[0082]

[0083] in, is the number of features, is the average correlation between features and classes, is the average correlation between features. For high-dimensional feature data, linear discriminant analysis (LDA) or principal component analysis (PCA) is used for dimensionality reduction. PCA calculates the eigenvalues ​​and eigenvectors of the covariance matrix and projects the original features into a new, lower-dimensional space, preserving the data's variance.

[0084] 5. Cell Classification and Identification

[0085] 1. Ensemble Learning Classifier

[0086] Random forest construction and training: The number of decision trees in the random forest is determined to be K, where K is an integer greater than 100. The optimal number is determined through cross-validation experiments to ensure a balance between integration effect and computational efficiency. For each decision tree, a feature subset is randomly selected from all features. The size of the feature subset is L% of the total number of features, where L is an integer between 30 and 70. The optimal ratio is determined through experiments to increase the diversity of the decision tree. When training each decision tree, information gain or Gini index is used as the splitting criterion. The information gain calculation formula is:

[0087]

[0088] in, is a dataset, It is a feature, It is a feature The number of values ​​of It is a feature The value is A subset of samples,

[0089]

[0090] The first The proportion of class samples; the Gini index calculation formula is:

[0091]

[0092] The optimal splitting features and splitting points are selected through these criteria to construct a decision tree.

[0093] When the decision tree reaches the maximum depth H, it stops growing, where H is an integer greater than 5. The appropriate depth is determined through experiments to prevent overfitting.

[0094] SVM classifier training and integration, for SVM with linear kernel function, the optimization algorithm used during training is the sequential minimal optimization (SMO) algorithm, which determines the model parameters by solving the following optimization problem:

[0095]

[0096]

[0097] in, is the sample size, is the Lagrange multiplier, is the category label of the sample, is the kernel function, for the linear kernel , is the penalty parameter, and its value is determined by cross-validation.

[0098] For the polynomial kernel function ,in is the degree of the polynomial, and its value range is [2,5]. The optimal degree is determined through experiments to adapt to different data distributions.

[0099] For the radial basis kernel function , The value range is [0.001, 10], and the optimal value is determined through grid search and cross-validation to control the complexity and generalization ability of the model.

[0100] When integrating multiple SVM classifiers, a weighted average method is used. The weight is determined according to the accuracy of each SVM classifier on the validation set. The calculation formula is:

[0101]

[0102] in, is the number of SVM classifiers. In this way, the weights are dynamically adjusted to optimize the performance of the integrated classifier.

[0103] (2) Optimization of classification results

[0104] The active learning strategy is implemented, and the criteria for determining samples with uncertain classification are set as follows: for the random forest classifier, when the proportion of votes for the highest-voted category in the voting results of a sample in all decision trees is less than [R]%, the sample is determined to be classified as uncertain, where R is a value between 50 and 70, determined through experiments; for the ensemble SVM classifier, when the highest confidence score of a sample in the prediction results of each SVM classifier is less than S, the sample is determined to be classified as uncertain, where S is a value between 0.5 and 0.7, determined through experiments.

[0105] Manual labeling is performed on samples with uncertain classification. The labelers need to have professional knowledge of cell morphology. After labeling, these samples are added to the training set and the ensemble learning classifier is retrained. Each time the training is retrained, the number of training set samples increases by no more than T1% of the number of original training set samples, where T1 is an integer between 10 and 30, to avoid overfitting the model to the newly added samples and to ensure that the model can continuously learn new information.

[0106] The confidence assessment mechanism is established and validated twice. For the random forest classifier, the confidence calculation formula is: where is the number of votes for the category with the most votes in all decision trees, and is the total number of votes in all decision trees. For the ensemble SVM classifier, the confidence is the average confidence of the prediction results of each SVM classifier for the sample.

[0107] When the confidence is less than U, it is determined as a classification result with low confidence, where U is a value between 0.3 and 0.5, which is determined through experiments.

[0108] A secondary verification was performed on the classification results with lower confidence. The feature dimensions added during the secondary verification included the grayscale mean, grayscale variance, and edge gradient mean of the cells. The specific calculation formula was: grayscale mean, where is the grayscale value of the pixels in the cell area, and is the number of pixels in the cell area; grayscale variance; edge gradient mean, where is the set of cell edge pixels, and is the gradient value of the edge pixels.

[0109] A more complex classification algorithm is used for reclassification, such as a deep neural network (DNN). The network structure of the DNN includes an input layer, multiple hidden layers, and an output layer. The number of hidden layers is [V1] layers, where [V1] is an integer between 3 and 5, determined by experiments. The number of neurons in each layer is [W1], where [W1] is an integer between 64 and 256, determined by experiments. The activation function uses the ReLU function, and the stochastic gradient descent (SGD) algorithm is used for training. The initial value of the learning rate is [X1], and the value range is [0.001, 0.1]. It is determined by experiments that the learning rate decays to [Z1] times the original value after every [Y1] epochs, where [Y1] is an integer between 10 and 20, and [Z1] is a decimal between 0.5 and 0.9, determined by experiments, in order to improve the accuracy of the secondary verification.

[0110] 6. Result Verification and Feedback

[0111] Multi-indicator result verification

[0112] Calculate the accuracy, recall, and F1-score of the classification results, and also calculate the Matthews correlation coefficient (MCC). The formula is:

[0113]

[0114] in, represents the number of true positive samples, represents the number of true negative samples, represents the number of false positive samples, In addition, the area under the receiver operating characteristic curve (ROC) (AUC) was calculated to comprehensively evaluate the accuracy of the classification results through these indicators.

[0115] Using stratified sampling, 20% of the total sample size was selected as the validation set. This ensured that the proportions of the three cell types—leukocytes, nucleated epithelial cells, and keratinized epithelial cells—in the validation set were consistent with the actual distribution in the original sample, thereby avoiding bias in the evaluation results due to data imbalance.

[0116] Feedback and Optimization

[0117] If the classification accuracy of a cell type is found to be less than 80% during the validation process, the sample features corresponding to that cell type will be analyzed in detail to check for insufficient feature extraction or unreasonable feature selection.

[0118] Based on the analysis results, the feature extraction method or feature selection strategy is adjusted accordingly. If the shape features of a certain cell type are found to be poorly distinguishable, new shape descriptors are introduced. If redundant features are found, feature screening is repeated. The ensemble learning classifier is then retrained and revalidated until the classification accuracy for each cell type reaches above 80%.

[0119] The present invention greatly reduces the time cost of manual microscopic examination through automatic recognition methods. The traditional method requires one day to complete the microscopic examination work, but this solution can be quickly processed, significantly improving the detection efficiency. With the help of various image processing and analysis technologies, as well as integrated learning classifiers, the subjective error of manual judgment is reduced, and the accuracy of cell type identification and estrous cycle judgment is improved. The entire recognition process is based on clear algorithms and models and is not affected by subjective factors of the operator, ensuring the reliability and consistency of the results. As long as the established process and parameter settings are followed, similar results can be obtained at different times and by different operators, facilitating the duplication and verification of research.

[0120] It is worth noting that extracting cell features from multiple dimensions, including shape, texture, and optical characteristics, and conducting a comprehensive analysis can more comprehensively describe cell characteristics and improve classification results. Through active learning strategies and confidence assessment mechanisms, classifier performance can be continuously optimized to adapt to sample changes and ensure long-term stable detection results. Most of the parameters in the scheme are determined through experiments and can be flexibly adjusted according to actual sample conditions and research needs to adapt to different experimental conditions and data characteristics.

[0121] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art may make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for determining the estrous cycle of rats based on image recognition, characterized in that: The following steps are involved: Sample collection and preparation: collection, processing and preservation of rat vaginal exfoliated cell samples; Image acquisition and optimization, acquisition of high-resolution images, and geometric correction, color correction and image denoising preprocessing; Cell segmentation and positioning: Use a semantic segmentation model based on convolutional neural networks to segment cells, and accurately locate and label the segmented cells; Feature extraction and analysis: extracting cell shape features, texture features, and optical features, and performing feature selection and dimensionality reduction; Cell classification and identification: using ensemble learning classifiers to classify cells and optimize the classification results; Result verification and feedback: Verify the classification results through multiple indicators, and provide feedback and optimization for each link based on the verification results; In the feature extraction and analysis, shape feature extraction includes fractal dimension, which is calculated using the box counting method. The cell image is divided into grids of different sizes and the number of grids containing cell pixels is counted. , fractal dimension: It also includes extracting the aspect ratio and convex hull area ratio features of cells; On the basis of gray-level co-occurrence matrix and local binary pattern, the directional gradient histogram feature is added to divide the cell area into multiple small cell units, the gradient direction histogram in each unit is calculated, and then the histograms are connected in series as the texture feature of the cell; A texture analysis method based on wavelet transform is used to extract wavelet coefficients at different scales as texture features to capture subtle texture changes inside cells. Analyze the optical properties of cells under different wavelengths of light. Use multispectral imaging technology to collect cell images at different wavelengths of 400nm-700nm, extract the spectral reflectance and absorbance optical characteristics of the cells, and perform dimensionality reduction and analysis of the optical characteristics through principal component analysis. A correlation-based feature selection method was used to screen out features with high correlation with cell types. The evaluation indicators of CFS are: in, is the number of features, is the average correlation between features and classes, is the average correlation between features; The cell classification and identification method specifically includes the following steps: constructing and training a random forest, determining the number of decision trees K through a cross-validation experiment, where K is an integer greater than 100, randomly selecting a feature subset of size L% of the total number of features from all features, where L is an integer between 30 and 70, using information gain or Gini index as a splitting criterion, and stopping the growth of the decision tree when it reaches a maximum depth H, where H is an integer greater than 5; The classification results were optimized by setting criteria for determining uncertain samples. The uncertain samples were manually labeled and added to the training set. The number of training set samples increased by no more than T1% of the original number of training set samples each time they were retrained, where T1 was an integer between 10 and 30. Results with a confidence level lower than U were revalidated, where U was a value between 0.3 and 0.

5. The feature dimensions of cell grayscale mean, grayscale variance, and edge gradient mean were increased. A deep neural network was used for reclassification, and the network structure parameters were determined experimentally.

2. The method for determining the estrous cycle of rats based on image recognition according to claim 1, wherein In the image acquisition and optimization, geometric correction adopts a method based on polynomial fitting. For the quadratic polynomial distortion model, the formula is: in are the original coordinates, are the corrected coordinates, is the correction coefficient, and the distortion model is established by collecting images of a known standard grid.

3. The rat estrous cycle judgment method based on image recognition according to claim 1, characterized in that, In the image acquisition and optimization, the color correction adopts the white balance algorithm, selects the white reference area in the image, calculates its average RGB value, and then normalizes the RGB value of the entire image. The formula is: in is the original RGB value, is the average RGB value of the white reference area, is the corrected RGB value.

4. The method for determining the estrous cycle of rats based on image recognition according to claim 1, wherein In the image acquisition and optimization, the image denoising adopts the wavelet transform denoising method, selects the db4 wavelet to perform multi-layer decomposition on the image, removes the high-frequency noise component through threshold processing, and then performs wavelet reconstruction. The threshold adopts the general threshold according to the noise level. Determine, among which, is the noise standard deviation, is the total number of image pixels.

5. The rat estrous cycle judgment method based on image recognition according to claim 1, characterized in that, In the cell segmentation and localization, a semantic segmentation model based on a convolutional neural network is used, and the network structure includes a contraction path and an expansion path. A large number of labeled rat vaginal exfoliated cell images are used for network training, and the cross entropy loss function is used: in, is the total number of pixels, is the number of cell types, including background, white blood cells, nucleated epithelial cells and keratinized epithelial cells. is the true label of the category to which the pixel belongs, It is the prediction probability, and the training dataset is expanded using data augmentation techniques, including random rotation, flipping, scaling, and elastic deformation.

6. The method for determining the estrous cycle of rats based on image recognition according to claim 1, wherein The result verification and feedback specifically include the following steps: The accuracy, recall, F1-score, Matthews correlation coefficient, and area under the receiver operating characteristic curve of the classification results were calculated. Stratified sampling was used to extract 20% of the samples as a validation set to ensure that the proportions of the three cell types in the validation set were consistent with the actual distribution of the original samples. If the classification accuracy of a certain cell type is lower than 80%, focus on analyzing the sample characteristics of that cell type, adjust the feature extraction method or feature selection strategy, and retrain the ensemble learning classifier until the classification accuracy of each cell type reaches above 80%.

Images