Intelligent diagnosis method for cervical cell atrophy level

Abstract

The application relates to an intelligent diagnosis method for cervical cell atrophy levels, and relates to the problem that intelligent discrimination of cervical cell atrophy degrees is lacking in automatic pathological diagnosis technology. Cervical vaginal squamous epithelium is divided into surface layer cells, middle layer cells and basal layer cells, and ovarian estrogen affects the growth and maturity of the cells; a decrease in estrogen level can cause symptoms such as atrophic vaginitis and osteoporosis, which need to be treated in time; at present, squamous cell atrophy diagnosis is not clear enough, and there are few reports on atrophy degree diagnosis research, so it is of great significance to establish a systematic and intelligent diagnosis process for cervical cell atrophy degree discrimination. In order to improve the problem, the application provides an intelligent diagnosis method for cervical cell atrophy levels; the method first detects the surface layer cells, the middle layer cells and the basal layer cells by using a target detection model, then segments the cell nucleus of each layer of cells detected by using an instance segmentation model, and finally calculates the cell quantity ratio, the nucleus-cytoplasm ratio and the cell crowding degree index of each layer; the indexes are input into a random forest classification model to grade the atrophy degree; it is known through sufficient experimental verification that good effects are achieved in cervical cell atrophy degree discrimination. The application is applied to cervical cell atrophy degree discrimination.

Description

Technical Field

[0001] This invention designs an intelligent diagnostic method for the degree of cervical cell atrophy. Background Technology

[0002] The squamous epithelium of the vaginal portion of the cervix is ​​divided into superficial, intermediate, and basal cells. Ovarian estrogen affects cell growth and maturation. Decreased estrogen levels can lead to symptoms such as atrophic vaginitis and osteoporosis, requiring timely treatment. Currently, the diagnosis of squamous cell atrophy is not clear enough, and there are few reports on the diagnosis of the degree of atrophy. At present, the diagnosis of the degree of atrophy relies entirely on manual operation and visual observation by doctors. This method has two major drawbacks: (1) Pathologists have a low accuracy rate in detecting formed elements in TCT images, resulting in misdiagnosis and missed diagnosis; (2) Pathological diagnosis is labor-intensive, requiring a certain level of professional knowledge from doctors, and there are few practicing pathologists, making it difficult to promote the examination on a large scale. Therefore, it is of great significance to use artificial intelligence technology to assist doctors in the detection of pathogenic microorganisms. Nowadays, with the development of artificial intelligence, intelligent diagnostic systems are widely used in cervical cytology examination. The system automatically analyzes the formed elements of cervical cytology smears through artificial intelligence and image processing, thereby screening common cellular components, helping pathologists save more than 90% of the reading time and improving diagnostic efficiency. Automated recognition of medical microscopic images has become a popular research area, but current research mainly focuses on human cell carcinogenesis, with little attention paid to the assessment of cervical atrophy levels. To address this issue, we propose an intelligent diagnostic method for cervical cell atrophy levels. This method calculates multiple evaluation indicators based on target detection and segmentation models, and then inputs these indicators into a random forest classification model to determine the degree of cervical atrophy.

[0003] The key to determining the degree of cervical atrophy lies in the detection of surface, middle, and basal cells and the segmentation of cell nuclei. Because panoramic cervical slice images are extremely large, typically containing hundreds of millions of pixels, image cropping is employed to obtain images of appropriate size. First, a target detection model is used to detect surface, middle, and basal cells. Then, an instance segmentation network model is used to segment the nuclei of each detected cell layer. Finally, the cell number ratio, average nucleocytoplasmic ratio, and cell crowding index for each layer are calculated. These indices are then input into a random forest classification model for atrophy degree classification. Extensive experimental verification has shown that this method achieves excellent results in determining the degree of cervical cell atrophy. Summary of the Invention

[0004] The purpose of this invention is to develop a systematic and intelligent diagnostic process for determining the degree of cervical cell atrophy, and to propose an intelligent diagnostic method for the level of cervical cell atrophy.

[0005] The aforementioned objectives are primarily achieved through the following technical solutions:

[0006] 1. Scan the sample to obtain a panoramic slice image, and then use image cropping to crop the patient sample into an image of appropriate size as training data for the target detection network;

[0007] The key to determining the degree of cervical cell atrophy lies in the extraction of image features. Because panoramic slice images are extremely large, typically containing hundreds of millions of pixels, image cropping is used to segment the patient sample into images of appropriate sizes.

[0008] 2. An object detection network is used to detect objects, identifying cells in the top, middle, and basal layers of each image.

[0009] S21. Add circular box annotations to the data in step S1;

[0010] S22. The target detection network adopts an improved CircleNet circular target detection network. The Gaussian kernel generation method for generating heatmaps within the network is improved. The formula for calculating the Gaussian kernel radius is shown below, where O is the overlap area between the circular bounding box and the predicted bounding box, R is the radius of the circular target bounding box, r1 is the Gaussian kernel radius when the predicted bounding box contains the circular bounding box, r2 is the Gaussian kernel radius when the original bounding box contains the predicted bounding box, r3 is the Gaussian kernel radius when the predicted bounding box intersects with the original bounding box, and G... r The final chosen Gaussian kernel radius:

[0011] (πR 2 (π(R+r1)) 2 )=O (1)

[0012] (π(R-r2) 2 )(πR 2 )=O (2)

[0013]

[0014] G r =min(r1,r2,r3) (4)

[0015] S23. Pyramid convolution is introduced into the basic module of CircleNet feature extraction. Pyramid convolution uses convolution kernels of varying sizes to perform convolution operations on image feature maps. Different sizes of convolution kernels are used. Generally, small-sized convolution kernels have a smaller receptive field and obtain small targets and local detail information, while larger convolution kernels have a larger receptive field and can obtain global semantic information of large targets. Finally, the feature maps obtained from the four sizes are fused to obtain the feature map after pyramid convolution.

[0016] S24, the CircleNet object detection model can detect the number of cells in each layer of each image and the number of cells in each layer of the patient sample.

[0017] 3. The intelligent diagnostic method for cervical cell atrophy level as described in claim 1, characterized in that, the method for calculating the number and crowding degree of the three layers of cells based on the detected cells in step S3 is as follows:

[0018] S31. The number ratio of cells in each layer is calculated as follows, where N total For all cells detected in the patient sample, N surface For all surface cells detected in the patient sample, N middle For all the middle layer cells detected in the patient sample, N basal All basal cells detected in the patient sample:

[0019] N total =N surface +N middle +N basal (5)

[0020] R surface =N surface / N total (6)

[0021] R middle =N middle / N total (7)

[0022] R basal =N basal / N total (8)

[0023] S32. The crowding density of each cell layer is calculated as the ratio of the sum of the areas of each cell layer to the total area of ​​the image, as follows, where CR... surface For the crowding of surface cells in patient samples, CR middle The crowding of cells in the middle layer of the patient sample, CR basal PA represents the crowding density of basal cells in a patient sample. itotal Let PA be the area of ​​the i-th patch in the patient sample, H be the height of the patch, W be the width of the patch, PA be the area of ​​the cell, and r be the radius of the cell. ijsurface PA represents the area of ​​the j-th surface cell detected in the i-th patch of the patient sample. ijmiddle PA represents the area of ​​the j-th midlayer cell detected in the i-th patch of the patient sample. ijbasalLet m be the area of ​​the j-th basal cell detected in the i-th patch of the patient sample, m be the total number of cells in the corresponding cell layer detected in the i-th patch, and n be the total number of patches into which the patient sample is segmented.

[0024] PA itotal =H*W (9)

[0025] PA = πr 2 (10)

[0026]

[0027]

[0028]

[0029] 4. The intelligent diagnostic method for cervical cell atrophy level as described in claim 1, characterized in that, in step S4, uniformly selecting small cell images detected from the target, using them as training data for the cell segmentation model, and employing an instance segmentation network to segment the cell nucleus to obtain the cell nucleus outline of each cell, the method is as follows:

[0030] S41. Considering that different staining styles and slide preparation methods will affect the segmentation effect, surface cells, middle cells and basal cells were selected from different domains, and the same number of individual cells were selected from each layer.

[0031] S42. First, use threshold segmentation to coarsely segment the data, and then manually confirm the segmentation labels, correct erroneous labels and fill in any gaps to obtain the training dataset for the instance segmentation model.

[0032] S43. Due to factors such as low estrogen levels in the patient or the resulting dense cell stacking, a single cell detected by CircleNet may contain multiple nuclei. Therefore, instance segmentation is used to segment out individual cell nuclei, and the cell with the larger nucleus outline area is selected as the cell nucleus.

[0033] 5. The intelligent diagnostic method for cervical cell atrophy level as described in claim 1, characterized in that, in step S5, the method for calculating the nucleocytoplasmic ratio of each cell based on the cell nuclear contour and statistically analyzing the average nucleocytoplasmic ratio of surface cells, intermediate cells, and basal cells is as follows:

[0034] S51. The circular prediction box detected by CircleNet is close to the outline of the cell, so the area of ​​the prediction box is used as the area of ​​the cell. The area of ​​the cell nucleus outline segmented from the instance can be calculated, which is the area of ​​the cell nucleus.

[0035] S52. The nucleus-to-cytoplasm ratio of a cell is calculated by dividing the area of ​​the nucleus by the area of ​​the cell, using the following formula: Nisurface is the nuclear area of ​​the i-th surface cell, Cisurface is the cell area of ​​the i-th surface cell, Nimiddle is the nuclear area of ​​the i-th middle cell, Cimiddle is the cell area of ​​the i-th middle cell, Nibasal is the nuclear area of ​​the i-th basal cell, and Cibasal is the cell area of ​​the i-th basal cell.

[0036]

[0037]

[0038]

[0039] 6. The intelligent diagnostic method for cervical cell atrophy level as described in claim 1, characterized in that the index features calculated in steps S2-S5 of step S6 are input into a random forest model to obtain the prediction results of the index feature values, and finally the atrophy degree is determined for the sample. The principle is as follows:

[0040] S61. Random forest is an algorithm that integrates multiple trees using the idea of ​​ensemble learning. Its basic unit is a decision tree. Each tree is constructed by selecting a fixed number of samples and a fixed number of features from the entire training sample set. Each decision tree is a classifier. For an input sample, N trees will have N classification results. Random forest integrates all the classification voting results and designates the class with the most votes as the final output.

[0041] S62. Select samples with different degrees of shrinkage, calculate the above indicators respectively, and use sampling with replacement to select samples to form an intermediate dataset. Then, randomly select multiple features from all features in this intermediate dataset as the final dataset, thus constructing multiple datasets. In each round of sampling, the probability that sample x can be selected is 1 / m. Therefore, after m rounds of sampling, the probability that sample x has not been selected is:

[0042]

[0043] S63. Build a fully split decision tree for each dataset. Use CART to build a fully split decision tree for each dataset without pruning, and finally obtain multiple CART decision trees.

[0044] S64. In the CART decision tree algorithm, the Gini index is used to measure the impurity or uncertainty of the data. It is also used to determine the optimal binary split for class variables. The optimal split point for a feature is determined by the Gini index; that is, it is only necessary to ensure that the Gini index value at a certain point is minimized after the split. This is the key to class variable splitting in the CART decision tree algorithm. In a classification problem, assuming there are K classes, the probability that a sample point belongs to the kth class is P. k The Gini exponent of the probability distribution is defined as:

[0045]

[0046] If a sample set D is divided into two parts, D1 and D2, according to a certain index A, then the Gini index of set D, under the condition of index A, is defined as follows:

[0047] gini(D,A)=D1 / Dgini(D1)+D2 / Dgini(D2) (19)

[0048] S65. Predict new data: Calculate the predicted value of the new data based on the results of each decision tree. Random forest integrates all classification voting results and designates the category with the most votes as the final output, which is the predicted shrinkage level.

[0049] The application of the cell atrophy diagnostic method described in step S7 to the intelligent pathological diagnostic system is as follows: Based on the cell atrophy degree judgment obtained in step S6, the overall cell atrophy degree of the cervical liquid-based panoramic cytology slide image is graded; a cell atrophy degree of 0 indicates "no atrophy", a cell atrophy degree of 1 indicates "mild atrophy", a cell atrophy degree of 2 indicates "moderate atrophy", and a cell atrophy degree of 3 indicates "severe atrophy"; in the intelligent diagnostic system, slides with a high degree of atrophy should be given high priority.

[0050] Invention Effects

[0051] The characteristic information of cervical cells is of great reference value for doctors and automated diagnostic systems. In traditional diagnostic methods, it relies entirely on doctors' manual operation and visual observation of slides. This method leads to low accuracy in detecting formed elements in TCT images by pathologists, resulting in misdiagnosis and missed diagnosis. In addition, pathological diagnosis is labor-intensive, requires a certain level of professional knowledge from doctors, and there are relatively few pathologists, hindering the widespread adoption of this examination. Therefore, it is of great significance to use artificial intelligence technology to assist doctors in pathogen detection. This algorithm first uses target detection to identify surface cells, middle cells, and basal cells. Then, it uses a strength segmentation model to segment the nuclei of each layer of cells. Finally, it calculates the cell number ratio, nucleocytoplasmic ratio, and cell crowding index for each layer, and inputs these indices into a random forest classification model to classify the degree of cell shrinkage. This provides strong support for automated diagnostic systems, reduces misdiagnosis caused by differences in the professional knowledge of pathologists, and effectively improves the efficiency and accuracy of system diagnosis. Attached Figure Description

[0052] Figure 1 Flowchart of the method implemented in this article

[0053] Figure 2 CircleNet object detection flowchart;

[0054] Figure 3 Random Forest Classification Flowchart; Specific implementation methods Specific implementation method one:

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] like Figure 1 The flowchart of the method presented in this paper is shown below:

[0058] The model training includes the following steps:

[0059] S1. The scanner scans the sample to obtain a panoramic slice image, and then uses image cropping to obtain an image of appropriate size as training data for object detection.

[0060] S2. Use an object detection network to detect cells, identifying surface cells, middle cells, and basal cells in each image.

[0061] S3. Based on the detected cells, the percentage of cells in the three cell layers and the crowding degree of cells in the three cell layers are statistically analyzed.

[0062] S4. Uniformly select small cell images detected from the target to serve as training data for the cell segmentation model. Use the instance segmentation network to segment the cell nucleus and obtain the cell nucleus outline of each cell.

[0063] S5. Calculate the nucleocytoplasmic ratio of each cell based on the cell nuclear outline, and statistically analyze the average nucleocytoplasmic ratio of surface cells, middle cells, and basal cells.

[0064] S6. Input the index features calculated in S2-S5 into the random forest model to obtain the prediction results of the index feature values, and finally determine the degree of shrinkage of the sample.

[0065] S7. Apply cell atrophy diagnostic methods to intelligent pathological diagnostic systems.

[0066] The embodiments of the present invention will be described in detail below:

[0067] The present invention is implemented using the method of the present invention as follows.

[0068] like Figure 1 The intelligent diagnosis of cervical cell atrophy level shown includes the following steps:

[0069] 1. Scan the sample with a scanner to obtain panoramic slice images, and then use image cropping to obtain 1500 images of size 1024*1024, which will be used as training data for CircleNet;

[0070] The key to determining the degree of cervical cell atrophy lies in the extraction of image features. Since panoramic slice images are extremely large, typically containing hundreds of millions of pixels, image cropping is used to obtain a 1024*1024 image.

[0071] 2. Cell detection is performed using an object detection network to detect surface cells, middle cells, and basal cells in each image;

[0072] S21. Add circular box annotations to the data in step S1;

[0073] S22. The target detection network adopts an improved CircleNet circular target detection network. The Gaussian kernel generation method for generating heatmaps within the network is improved. The formula for calculating the Gaussian kernel radius is shown below, where O is the overlap area between the circular bounding box and the predicted bounding box, R is the radius of the circular target bounding box, r1 is the Gaussian kernel radius when the predicted bounding box contains the circular bounding box, r2 is the Gaussian kernel radius when the original bounding box contains the predicted bounding box, r3 is the Gaussian kernel radius when the predicted bounding box intersects with the original bounding box, and G... rThe final chosen Gaussian kernel radius:

[0074] (πR 2 ) / (π(R+r1) 2 )=O (1)

[0075] (π(R-r2) 2 ) / (πR 2 )=O (2)

[0076]

[0077] G r =min(r1,r2,r3) (4)

[0078] S23. Pyramid convolution is introduced into the basic module of CircleNet feature extraction. Pyramid convolution uses convolution kernels of varying sizes to perform convolution operations on image feature maps. The convolution kernel sizes used are 1*1, 3*3, 5*5, and 7*7. Generally, smaller convolution kernels have smaller receptive fields and obtain small targets and local detail information, while larger convolution kernels have larger receptive fields and can obtain global semantic information of large targets. Finally, the feature maps obtained from the four sizes are fused to obtain the feature map after pyramid convolution.

[0079] S24, the CircleNet object detection model can detect the number of cells in each layer of each image and the number of cells in each layer of the patient sample.

[0080] 3. Based on the detected cells, the number and crowding of cells in each of the three layers were statistically analyzed;

[0081] S31. The number ratio of cells in each layer is calculated as follows, where N total For all cells detected in the patient sample, N surface For all surface cells detected in the patient sample, N middle For all the middle layer cells detected in the patient sample, N basal All basal cells detected in the patient sample:

[0082] N total =N surface +N middle +N basal (5)

[0083] R surface =N surface / N total (6)

[0084] R middle =N middle / N total (7)

[0085] R basal =N basal / N total (8)

[0086] S32. The crowding density of each cell layer is calculated as the ratio of the sum of the areas of each cell layer to the total area of ​​the image, as follows, where CR... surface For the crowding of surface cells in patient samples, CR middle The crowding of cells in the middle layer of the patient sample, CR basal PA represents the crowding density of basal cells in a patient sample. itotal Let PA be the area of ​​the i-th patch in the patient sample, H be the height of the patch, W be the width of the patch, PA be the area of ​​the cell, and r be the radius of the cell. ijsurface PA represents the area of ​​the j-th surface cell detected in the i-th patch of the patient sample. ijmiddle PA represents the area of ​​the j-th midlayer cell detected in the i-th patch of the patient sample. ijbasal Let m be the area of ​​the j-th basal cell detected in the i-th patch of the patient sample, m be the total number of cells in the corresponding cell layer detected in the i-th patch, and n be the total number of patches into which the patient sample is segmented.

[0087] PA itotal =H*W (9)

[0088] PA = πr 2 (10)

[0089]

[0090]

[0091]

[0092] 4. Uniformly select small cell images detected from the target to serve as training data for the cell segmentation model. Use the instance segmentation network to segment the cell nucleus and obtain the cell nucleus outline of each cell.

[0093] S41. Considering that different staining styles and slide preparation methods will affect the segmentation effect, surface cells, middle cells, and basal cells were selected from different domains. 1000 single cells with clearly visible nuclei were selected from each cell layer, for a total of 3000 images.

[0094] S42. The strength segmentation model uses the Mask R-CNN network. First, threshold segmentation is used to coarsely segment the data. Then, the segmentation labels are manually confirmed, erroneous labels are corrected, and omissions are filled to obtain the training dataset for Mask R-CNN.

[0095] S43. Due to factors such as low estrogen levels in the patient or the resulting dense cell stacking, a single cell detected by CircleNet may contain multiple nuclei. Therefore, we use instance segmentation MaskRcnn to segment individual cell nuclei and select the cell with the larger nucleus outline area as the cell nucleus.

[0096] 5. Calculate the nucleocytoplasmic ratio of each cell based on the nuclear outline, and statistically analyze the average nucleocytoplasmic ratio of the surface cells, middle cells, and basal cells.

[0097] S51. The circular prediction boxes detected by CircleNet are close to the outline of the cell, so the area of ​​the prediction box is used as the area of ​​the cell. The area of ​​the cell nucleus outline segmented by MaskRcnn can be calculated by OpenCV library, and the area of ​​the outline and the area of ​​the cell nucleus can be calculated.

[0098] S52. To ensure the calculated nucleocytoplasmic ratio of each layer is more representative, the calculated nucleocytoplasmic ratios for each cell layer are first sorted. Then, the average of the middle 3 / 5 is taken as the nucleocytoplasmic ratio for each layer. The nucleocytoplasmic ratio of a cell is calculated by dividing the area of ​​the cell nucleus by the area of ​​the cell, as shown in the following formula, where N... isurface Let C be the area of ​​the nucleus of the i-th surface cell. isurface Let N be the cell area of ​​the i-th surface cell. imiddle Let C be the area of ​​the nucleus of the i-th middle layer cell. imiddle Let N be the cell area of ​​the i-th middle layer cell. ibasal Let C be the area of ​​the nucleus of the i-th basal cell. ibasal Let i be the cell area of ​​the i-th basal cell.

[0099]

[0100]

[0101]

[0102] 6. Input the index features calculated in S2-S5 into the random forest classification model to obtain the prediction results of the index feature values, and finally determine the degree of shrinkage of the sample.

[0103] S61. Random forest is an algorithm that integrates multiple trees using the idea of ​​ensemble learning. Its basic unit is a decision tree. Each tree is constructed by selecting a fixed number of samples and a fixed number of features from the entire training sample set. Each decision tree is a classifier (assuming that the problem is classification). For an input sample, N trees will have N classification results. Random forest integrates all the classification voting results and designates the class with the most votes as the final output.

[0104] S62. Select 150 samples with different degrees of shrinkage, calculate the above 9 indicators respectively, and use sampling with replacement to select 150 samples to form an intermediate dataset. Then, randomly select 3 features from all features in this intermediate dataset as the final dataset, thus constructing multiple datasets. In each round of sampling, the probability that sample x can be selected is 1 / m. Therefore, after m rounds of sampling, the probability that sample x has not been selected is:

[0105]

[0106] S63. Build a fully split decision tree for each dataset. Use CART to build a fully split decision tree for each dataset without pruning, and finally obtain multiple CART decision trees.

[0107] S64. In the CART decision tree algorithm, the Gini index is used to measure the impurity or uncertainty of the data. It is also used to determine the optimal binary split for class variables. The optimal split point for a feature is determined by minimizing the Gini index at a given point after the split. This is the key to class variable splitting in the CART decision tree algorithm. In a classification problem, assuming there are K classes, the probability of a sample point belonging to the kth class is P. k The Gini exponent of the probability distribution is defined as:

[0108]

[0109] If a sample set D is divided into two parts, D1 and D2, according to a certain index A, then the Gini index of set D, under the condition of index A, is defined as follows:

[0110] gini(D,A)=D1 / Dgini(D1)+D2 / Dgini(D2) (19)

[0111] S65. Predict new data: Calculate the predicted value of the new data based on the results of each decision tree. Random forest integrates all classification voting results and designates the category with the most votes as the final output, which is the predicted shrinkage level.

[0112] 7. Apply cell atrophy diagnostic methods to intelligent pathological diagnostic systems;

[0113] Based on the degree of cell atrophy obtained in step S6, the overall degree of cell atrophy in the cervical liquid-based panoramic cytology slice image is graded; a degree of cell atrophy of 0 indicates "no atrophy", a degree of cell atrophy of 1 indicates "mild atrophy", a degree of cell atrophy of 2 indicates "moderate atrophy", and a degree of cell atrophy of 3 indicates "severe atrophy". In the intelligent diagnostic system, slices with a high degree of atrophy should be given high priority.

[0114] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the scope of this invention.

Claims

1. A smart diagnostic method for the grade of cervical cell atrophy, characterized in that, It includes the following steps: S1. The scanner scans the sample to obtain a panoramic slice image, and then uses image cropping to crop the patient sample into an image of an appropriate size as training data for target detection. S2. Cell detection is performed using an object detection network to detect surface cells, middle cells, and basal cells in each image; the method is as follows: S21. Add circular box annotations to the data in step S1; S22. The target detection network adopts an improved CircleNet circular target detection network. The Gaussian kernel generation method for generating heatmaps within the network is improved. The formula for calculating the Gaussian kernel radius is shown below, where O is the overlap area between the circular bounding box and the predicted bounding box, R is the radius of the circular target bounding box, r1 is the Gaussian kernel radius when the predicted bounding box contains the circular bounding box, r2 is the Gaussian kernel radius when the original bounding box contains the predicted bounding box, r3 is the Gaussian kernel radius when the predicted bounding box intersects with the original bounding box, and G... r The final chosen Gaussian kernel radius: (πR 2 ) / (π(R+r1) 2 )=O (1) (π(R-r2) 2 ) / (πR 2 )=O (2) G r =min(r1,r2,r3) (4) S23. Pyramid convolution is introduced into the basic module of CircleNet feature extraction. Pyramid convolution uses convolution kernels of different sizes to perform convolution operations on image feature maps. By using convolution kernels of different sizes, the feature maps obtained from the four sizes are finally fused to obtain the feature map after pyramid convolution. S24, the CircleNet object detection model can detect the number of cells in each layer of each image and the number of cells in each layer of the patient sample; S3. Based on the detected cells, the percentage of cells in the three cell layers and the crowding degree of cells in the three cell layers are statistically analyzed. S4. Uniformly select small cell images detected from the target to serve as training data for the cell segmentation model. Use the instance segmentation network to segment the cell nucleus and obtain the cell nucleus outline of each cell. S5. Calculate the nucleocytoplasmic ratio of each cell based on the cell nuclear outline, and statistically analyze the average nucleocytoplasmic ratio of surface cells, middle cells, and basal cells. S6. Input the index features calculated in S2-S5 into the random forest model to obtain the prediction results of the index feature values, and finally determine the degree of shrinkage of the sample. S7. Apply cell atrophy diagnostic methods to intelligent pathological diagnostic systems.

2. The intelligent diagnostic method for cervical cell atrophy level as described in claim 1, characterized in that, The method for calculating the number and crowding of cells in each of the three layers based on the detected cells in step S3 is as follows: S21. The number ratio of cells in each layer is calculated as follows, where N total For all cells detected in the patient sample, N surface For all surface cells detected in the patient sample, N middle For all the middle layer cells detected in the patient sample, N basal All basal cells detected in the patient sample: N total =N surface +N middle +N basal (5) R surface =N surface / N total (6) R middle =N middle / N total (7) R basal =N basal / N total (8) S22. The crowding density of each cell layer is calculated as the ratio of the sum of the areas of each cell layer to the total area of ​​the image, as follows, where CR surface For the crowding of surface cells in patient samples, CR middle The crowding of cells in the middle layer of the patient sample, CR basal PA represents the crowding density of basal cells in a patient sample. itotal Let PA be the area of ​​the i-th patch in the patient sample, H be the height of the patch, W be the width of the patch, PA be the area of ​​the cell, and r be the radius of the cell. ijsurface PAi represents the area of ​​the j-th surface cell detected in the i-th patch of the patient sample. jmiddle PA represents the area of ​​the j-th midlayer cell detected in the i-th patch of the patient sample. ijbasal Let m be the area of ​​the j-th basal cell detected in the i-th patch of the patient sample, m be the total number of cells in the corresponding cell layer detected in the i-th patch, and n be the total number of patches into which the patient sample is segmented. PA itotal =H*W (9) PA=πr 2 (10) 3. The intelligent diagnostic method for cervical cell atrophy level as described in claim 1, characterized in that, In step S4, the small cell images detected by the target are uniformly selected and used as training data for the cell segmentation model. The cell nucleus is segmented using an instance segmentation network to obtain the cell nucleus outline of each cell. The method is as follows: S31. Select surface cells, middle cells, and basal cells from different domains, and select the same number of individual cells from each layer. S32. First, use threshold segmentation to coarsely segment the data, and then manually confirm the segmentation labels, correct erroneous labels and fill in any gaps to obtain the training dataset for the instance segmentation model. S33. Use instance segmentation to segment individual cell nuclei, and select the cell nucleus with the larger outline area as the cell nucleus.

4. The intelligent diagnostic method for cervical cell atrophy level as described in claim 1, characterized in that, The method described in step S5 for calculating the nucleocytoplasmic ratio of each cell based on the nuclear outline and for statistically analyzing the average nucleocytoplasmic ratio of surface cells, middle cells, and basal cells is as follows: S41. Treat the area of ​​the predicted bounding box as the area of ​​the cell. The area of ​​the cell nucleus outline segmented from the instance can be calculated, which is the area of ​​the cell nucleus. S42. The nucleus-to-cytoplasm ratio of a cell is calculated by dividing the area of ​​the nucleus by the area of ​​the cell, using the following formula, where N... isurface Let C be the area of ​​the nucleus of the i-th surface cell. isurface Let N be the cell area of ​​the i-th surface cell. imiddle Let C be the area of ​​the nucleus of the i-th middle layer cell. imiddle Let N be the cell area of ​​the i-th middle layer cell. ibasal Let C be the area of ​​the nucleus of the i-th basal cell. ibasal Let i be the cell area of ​​the i-th basal cell.

5. The intelligent diagnostic method for cervical cell atrophy level as described in claim 1, characterized in that, The index features calculated in steps S2-S5 of step S6 are input into the random forest model to obtain the prediction results of the index feature values, and finally the degree of shrinkage is determined for the sample. The principle is as follows: S51. Random forest is an algorithm that integrates multiple trees using the idea of ​​ensemble learning. Its basic unit is a decision tree. Each tree is constructed by selecting a fixed number of samples and a fixed number of features from the entire training sample set. Each decision tree is a classifier. For an input sample, N trees will have N classification results. Random forest integrates all the classification voting results and designates the class with the most votes as the final output. S52. Select samples with different degrees of shrinkage, calculate the above indicators respectively, and use sampling with replacement to select samples to form an intermediate dataset. Then, randomly select multiple features from all features in this intermediate dataset as the final dataset, thus constructing multiple datasets. In each round of sampling, the probability that sample x can be selected is 1 / m. Therefore, after m rounds of sampling, the probability that sample x has not been selected is: S53. Build a fully split decision tree for each dataset. Use CART to build a fully split decision tree for each dataset without pruning, and finally obtain multiple CART decision trees. S54. In the CART decision tree algorithm, the Gini index is used to measure the impurity or uncertainty of the data. It is also used to determine the optimal binary split for class variables. The optimal split point for a feature is determined by minimizing the Gini index at a given point after the split. This is the key to class variable splitting in the CART decision tree algorithm. In a classification problem, assuming there are K classes, the probability of a sample point belonging to the kth class is P. k The Gini exponent of the probability distribution is defined as: If a sample set D is divided into two parts, D1 and D2, according to a certain index A, then the Gini index of set D, under the condition of index A, is defined as follows: gini(D,A)=D1 / Dgini(D1)+D2 / Dgini(D2) (19) S55. Predict new data: Calculate the predicted value of the new data based on the results of each decision tree. Random forest integrates all classification voting results and designates the category with the most votes as the final output, which is the predicted shrinkage level.

6. The intelligent diagnostic method for cervical cell atrophy level as described in claim 1, characterized in that, The application of the cell atrophy diagnostic method described in step S7 to the intelligent pathological diagnostic system is as follows: Based on the cell atrophy degree judgment obtained in step S6, the overall cell atrophy degree of the cervical liquid-based panoramic cytology slide image is graded; a cell atrophy degree of 0 indicates "no atrophy", a cell atrophy degree of 1 indicates "mild atrophy", a cell atrophy degree of 2 indicates "moderate atrophy", and a cell atrophy degree of 3 indicates "severe atrophy"; in the intelligent diagnostic system, slides with a high degree of atrophy should be given high priority.

Images