A method for identifying abnormal cervical cells by fusing in-sample control information

Abstract

This invention relates to a method for identifying abnormal cervical cells by integrating intra-sample control information. In cytopathological testing, the variability between samples leads to low accuracy and high false positive rates in identifying abnormal cervical cells. Because cervical pathological states naturally differ across physiological states and age groups, the morphology and size of cervical cells vary. Deep learning methods primarily use labeled data of abnormal cervical cells from different samples to train models, rarely considering the differences between cervical cytopathological samples. This results in ambiguous identification standards, low accuracy, and high false positive rates. To improve this, this invention proposes a method for identifying abnormal cervical cells by integrating intra-sample control information. Experiments show that this method can effectively identify abnormal cells by comparing the morphological characteristics of normal cells within the sample, reducing the false positive rate and improving the accuracy, sensitivity, and specificity of sample detection. This invention is mainly applied to the identification of abnormal cells in cervical cytopathological testing.

Description

Technical Field

[0001] This invention relates to the identification of abnormal cervical cells. Background Technology

[0002] Cervical cancer is a malignant tumor that seriously endangers women's health and threatens their lives. Cervical cancer has a long incubation period and is difficult to detect in its early stages due to the lack of symptoms. The average age difference between the onset of grade 3 cervical intraepithelial neoplasia and cervical cancer is 10 years, and the cure rate for early-stage cervical cancer is as high as 91.5%. Therefore, early detection and treatment are currently the only effective means of combating cancer. Current methods for detecting cancer include tumor marker testing, gene testing, imaging examinations (ultrasound, CT scans, etc.), endoscopy, and pathological examination. Among these, only pathological examination is currently a definitive diagnostic tool. Pathological diagnosis includes cytopathological diagnosis and histopathological diagnosis. It involves collecting exfoliated cells or tissue slides, staining them, and then having a pathologist observe them under a microscope to look for abnormal cells and make a diagnosis. Cytopathological diagnosis is simple to obtain, widely used, and can provide qualitative diagnosis, making it particularly suitable for early diagnosis and screening. It deserves widespread adoption, and image-based cytopathological screening is currently the most widely used method for precancerous screening.

[0003] In recent years, with the development of machine learning theory, represented by deep learning, convolutional neural networks have achieved revolutionary breakthroughs in image recognition due to their powerful and effective feature extraction capabilities, and have been widely applied in medical image analysis fields such as cervical abnormal cell identification. Because cervical pathological states vary naturally across different physiological states and age groups, the morphology and size of cervical cells differ. Furthermore, factors such as sampling, slide preparation, and staining can influence cervical cell images, resulting in diverse colors and complex content. Deep learning methods primarily focus on the cervical cell level, rarely considering the differences between cervical pathological samples and their impact on the detection of cervical abnormal cells. Training models using labeled data of cervical abnormal cells from different samples leads to problems such as ambiguous classification standards, low accuracy, and high false positive rates. To address these issues, a method for identifying cervical abnormal cells that integrates in-sample control information is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the problem of low accuracy and high false positive rate in identifying abnormal cervical cells due to differences between cytopathological samples, and to propose a method for identifying abnormal cervical cells that integrates control information within the sample.

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

[0006] S1. Prepare training data for cervical abnormal cell detection, training data for cervical normal middle squamous epithelial cell detection, and training data for cell nucleus and cytoplasm separation.

[0007] First, a large number of cervical cytology smears diagnosed as positive and negative were collected and scanned to obtain panoramic images of the cervical cytology smears. Then, pathologists were invited to annotate abnormal cervical cells on the panoramic images of positive samples and normal squamous epithelial cells on the panoramic images of negative samples. Next, based on the pathologists' annotations, images of a certain size were randomly cropped from the panoramic images, ensuring that the annotation information was complete in each image. The cropped images from the panoramic images of positive samples and their corresponding annotations were used as training data for detecting abnormal cervical cells, and the cropped images from the panoramic images of negative samples and their corresponding annotations were used as training data for detecting normal squamous epithelial cells in the middle layer of the cervix. Finally, images of cervical cells annotated by the doctors were cropped from the sample panoramic images, and the outlines of the cell nuclei and cytoplasm in all cervical cell images were manually annotated using LabelImg software. The annotated data was used as training data for cell nucleus and cytoplasm segmentation.

[0008] S2, Training models for detecting abnormal cervical cells, detecting normal cervical squamous epithelial cells in the middle layer of the cervix, and segmenting cell nuclei and cytoplasm;

[0009] The detection model structure is basically the same as the YOLO v5 model structure, mainly composed of a backbone network, a feature fusion network, and a detection network. The ordinary convolutional layers in the backbone network are modified into deformable convolutional layers. The deformable convolutional layers are composed of deformable convolution, batch normalization, and activation functions connected in sequence. The backbone network is formed by connecting the deformable convolutional layers with residual blocks composed of cross-stage local fusion networks. The feature fusion network uses spatial pyramid pooling to increase the receptive field, and then uses a path aggregation network to collect feature maps from different stages and perform feature fusion. Finally, the CIoU loss function is used for refined classification and regression. The detection network is trained with training data for detecting abnormal cervical cells and normal squamous epithelial cells, respectively, to obtain detection models for abnormal cervical cells and normal squamous epithelial cells. The original Solo v2 network structure is used as the segmentation model structure. The nucleus and cytoplasm segmentation training data are trained with the segmentation network to obtain nucleus and cytoplasm segmentation models.

[0010] S3. Prepare training data for cervical abnormal cell classification that integrates in-sample control information.

[0011] First, the panoramic image of the positive sample is cropped sequentially to a certain size. The cropped images are then predicted using a normal cervical squamous epithelial cell detection model to obtain images of normal squamous epithelial cells. Simultaneously, abnormal cells marked by the doctor in the positive sample are cropped out to obtain images of abnormal cells. Next, the images of normal and abnormal cells from each sample are fed into a nucleus and cytoplasm segmentation model to obtain the outline information of the nucleus and cytoplasm for each cell image. Then, cell characteristic indicators are calculated using the outline information, including nuclear area, nuclear integrated optical density, nuclear roundness, and nucleocytoplasmic ratio. The calculation formulas are as follows:

[0012]

[0013] In the formula, A nu S represents the area of ​​the cell nucleus. nu For the cell nucleus region, p i Let be a pixel in the cell nucleus region, and the area of ​​the cell nucleus is the sum of the number of pixels in the cell nucleus region;

[0014] Y (i,j) =0.299R (i,j) +0.587G (i,j) +0.114B (i,j) (2)

[0015]

[0016] In the formula, Y (i,j) R represents the brightness at position (i,j) in the image. (i,j) G (i,j) B (i,j) These represent the R, G, and B channel pixel values ​​at position (i,j) in the image, respectively, and the Y... μ The average brightness of the cytoplasm is represented by IOD, which is the integrated optical density of the nucleus, reflecting the chromatin content within the nucleus.

[0017]

[0018] In the formula, R nu For the roundness of the cell nucleus, L nu This refers to the perimeter of the cell nucleus, i.e., the length of the cell nucleus's outline.

[0019]

[0020] In the formula, N c-ratio For the nucleus-to-cytoplasm ratio, A cyThe cytoplasmic area is the sum of the number of pixels within the cytoplasmic region. Then, the average of all normal squamous epithelial cell characteristic indicators in a single sample is taken as the baseline characteristic indicator for that sample. Dividing each abnormal cell characteristic indicator by the corresponding baseline characteristic indicator yields the in-sample control characteristic indicator for each abnormal cell image. Finally, the image within the cytoplasmic contour information is retained, and the area outside the cytoplasmic contour is filled with black, thus obtaining a background-free abnormal cell image. Each background-free abnormal cell image and its corresponding in-sample control characteristic indicator are combined to form cervical abnormal cell classification training data, with category labels provided by the pathologist.

[0021] S4. Construct and train a cervical abnormal cell classification model that integrates in-sample control information;

[0022] First, a ResNet50 network structure is used as the cervical cell image feature learning module, and two fully connected layers are used as the non-image feature learning module for in-sample control information. Then, the image feature vectors output by the image feature learning module and the non-image feature vectors output by the non-image feature learning module are concatenated to form a feature fusion vector. Next, two fully connected layers and Softmax are used as the classification head to obtain the cervical abnormal cell classification results. Finally, the cervical abnormal cell classification model is trained using cervical abnormal cell classification training data with fused in-sample control information.

[0023] S5. Use trained detection, segmentation, and classification models to identify abnormal cervical cells in a panoramic view of cervical cell pathology.

[0024] First, the panoramic image of cervical cell pathology to be identified is cropped to a certain size. Then, the cropped images are predicted using normal and abnormal cell detection models to obtain sets of normal and abnormal cell images for the sample. Next, each cell image in the sets of normal and abnormal cell images is segmented using a segmentation model to obtain the outline information of the cell nucleus and cytoplasm for each cell image. Then, according to S3, cell feature indicators, background-free abnormal cell images, and corresponding in-sample control feature indicators are calculated using the outline information to obtain the input data for the classification model. Finally, the background-free abnormal cell images and corresponding in-sample control features are input into the classification model to predict the category information of the abnormal cell images.

[0025] Invention Effects

[0026] This invention proposes a method for identifying abnormal cervical cells by fusing in-sample control information. Cervical abnormal cell identification using deep learning methods typically requires a large amount of training data. However, this data inevitably involves abnormal cervical cells from different samples for model training, naturally lacking in-sample control information for positive and abnormal cells, leading to low identification accuracy and a high false positive rate. This invention proposes first using a target detection model to identify and locate positive and abnormal cells in a panoramic image. Then, a segmentation model is used to predict the contour information of the cell nucleus and cytoplasm, obtaining in-sample control information for positive and abnormal cells and images of abnormal cells without background interference, quantifying cervical cell pathological diagnostic indicators. Finally, a cervical abnormal cell image classification model fusing in-sample control information is used to classify cervical abnormal cells, simulating a doctor's comparison of the morphology of normal cells within the sample to identify abnormal cells, achieving in-sample control identification of positive and abnormal cervical cells. Experiments show that this method effectively reduces the false positive rate of cervical abnormal cell identification, improves sample detection accuracy, sensitivity, and specificity, and can better assist pathologists in diagnosis, reducing their workload. Attached Figure Description

[0027] Figure 1 A schematic diagram of the main algorithm flow;

[0028] Figure 2 Structural diagram of a cervical normal and abnormal cell detection model;

[0029] Figure 3 Image showing the separation of the cell nucleus and cytoplasm;

[0030] Figure 4 A structural diagram of a cervical abnormal cell classification model that integrates in-sample control information; Specific implementation methods Specific implementation method one:

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in 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.

[0033] like Figure 1 The image shows a method for identifying abnormal cervical cells that integrates in-sample control information, provided by the present invention, which includes data preparation, model training, and identification of abnormal cervical cells under panoramic images.

[0034] The data preparation and model training include the following steps:

[0035] S1. Prepare training data for cervical abnormal cell detection, training data for cervical normal middle squamous epithelial cell detection, and training data for cell nucleus and cytoplasm separation.

[0036] S2, Training models for detecting abnormal cervical cells, detecting normal cervical squamous epithelial cells in the middle layer of the cervix, and segmenting cell nuclei and cytoplasm;

[0037] S3. Prepare training data for cervical abnormal cell classification that integrates in-sample control information.

[0038] S4. Construct and train a cervical abnormal cell classification model that integrates in-sample control information;

[0039] The identification of abnormal cervical cells under panoramic imaging includes the following steps:

[0040] S5. Use trained detection, segmentation, and classification models to identify abnormal cervical cells in a panoramic view of cervical cell pathology.

[0041] This invention first prepares training data for a cervical abnormal cell detection model, a cell nucleus and cytoplasm segmentation model, and an abnormal cell classification model, and constructs and trains the detection, segmentation, and classification models. Then, the detection model is used to identify and locate normal and abnormal cells in a panoramic image, and the segmentation model is used to predict the contour information of the cell nucleus and cytoplasm, obtaining control information of normal and abnormal cells within a single sample and abnormal cell images without background interference. Finally, a cervical abnormal cell image classification model that integrates control information within the sample is used to achieve fine classification of cervical abnormal cells.

[0042] The embodiments of the present invention will be described in detail below:

[0043] Data preparation and model training include the following steps:

[0044] S1. Prepare training data for cervical abnormal cell detection, training data for cervical normal middle squamous epithelial cell detection, and training data for cell nucleus and cytoplasm separation.

[0045] First, a large number of cervical cytology smears diagnosed as positive and negative were collected and scanned to obtain panoramic images of the cervical cytology smears. Then, pathologists were invited to annotate abnormal cervical cells on the panoramic images of positive samples and normal squamous epithelial cells on the panoramic images of negative samples. Next, based on the pathologists' annotations, images of 1024×1024 pixels were randomly cropped from the panoramic images, ensuring that the annotation information was complete in each image. The cropped images from the panoramic images of positive samples and their corresponding annotations were used as training data for detecting abnormal cervical cells, and the cropped images from the panoramic images of negative samples and their corresponding annotations were used as training data for detecting normal squamous epithelial cells in the middle layer of the cervix. Finally, images of cervical cells annotated by the doctor were cropped from the sample panoramic images, and the outlines of the cell nuclei and cytoplasm in all cervical cell images were manually annotated using LabelImg software. The annotated data was used as training data for cell nucleus and cytoplasm segmentation.

[0046] S2, Training models for detecting abnormal cervical cells, detecting normal cervical squamous epithelial cells in the middle layer of the cervix, and segmenting cell nuclei and cytoplasm;

[0047] The detection model structure is basically the same as the YOLO v5 model structure, such as Figure 2 As shown, the system mainly consists of a backbone network, a feature fusion network, and a detection network. Ordinary convolutional layers in the backbone network are modified into deformable convolutional layers. These deformable convolutional layers are constructed by sequentially connecting deformable convolutions, batch normalization, and activation functions. The backbone network is formed by connecting these deformable convolutional layers with residual blocks created by a cross-stage local fusion network. The feature fusion network utilizes spatial pyramid pooling to increase the receptive field, then uses a path aggregation network to collect feature maps from different stages and perform feature fusion. Finally, the CIoU loss function is used for refined classification and regression. Training data for detecting abnormal cervical cells and normal squamous epithelial cells are input into the detection network and trained for 200 epochs to obtain detection models for abnormal cervical cells and normal squamous epithelial cells. The original Solo v2 network structure is used as the segmentation model structure. Training data for nucleus and cytoplasm segmentation are input into the segmentation network and trained for 200 epochs to obtain nucleus and cytoplasm segmentation models.

[0048] S3. Prepare training data for cervical abnormal cell classification that integrates in-sample control information.

[0049] First, the panoramic images of positive samples were sequentially cropped to 1024×1024 pixels. The cropped images were then predicted using a normal cervical squamous epithelial cell detection model to obtain images of normal squamous epithelial cells. Simultaneously, abnormal cells marked by the doctor in the positive samples were cropped out to obtain images of abnormal cells. Then, the images of normal and abnormal cells from each sample were fed into a nucleus and cytoplasm segmentation model to obtain the nucleus and cytoplasm contour information for each cell image. The segmentation results are as follows: Figure 3 As shown; next, cell characteristic indicators are calculated using contour information, including nuclear area, nuclear integrated optical density, nuclear roundness, and nucleocytoplasmic ratio, using the following formulas:

[0050]

[0051] In the formula, A nu S represents the area of ​​the cell nucleus. nu For the cell nucleus region, p i Let be a pixel in the cell nucleus region, and the area of ​​the cell nucleus is the sum of the number of pixels in the cell nucleus region;

[0052] Y (i,j) =0.299R (i,j) +0.587G (i,j) +0.114B (i,j) (2)

[0053]

[0054] In the formula, Y (i,j) R represents the brightness at position (i,j) in the image. (i,j) G (i,j) B (i,j) These represent the R, G, and B channel pixel values ​​at position (i,j) in the image, respectively, and the Y... μ The average brightness of the cytoplasm is represented by IOD, which is the integrated optical density of the nucleus, reflecting the chromatin content within the nucleus.

[0055]

[0056] In the formula, R nu For the roundness of the cell nucleus, L nu This refers to the perimeter of the cell nucleus, i.e., the length of the cell nucleus's outline.

[0057]

[0058] In the formula, N c-ratio For the nucleus-to-cytoplasm ratio, A cyThe cytoplasmic area is the sum of the number of pixels within the cytoplasmic region. Then, the average of all normal squamous epithelial cell characteristic indicators in a single sample is taken as the baseline characteristic indicator for that sample. Dividing each abnormal cell characteristic indicator by the corresponding baseline characteristic indicator yields the in-sample control characteristic indicator for each abnormal cell image. Finally, the image within the cytoplasmic contour information is retained, and the area outside the cytoplasmic contour is filled with black, thus obtaining a background-free abnormal cell image. Each background-free abnormal cell image and its corresponding in-sample control characteristic indicator are combined to form cervical abnormal cell classification training data, with category labels provided by the pathologist.

[0059] S4. Construct and train a cervical abnormal cell classification model that integrates in-sample control information;

[0060] Classification model network structure as follows Figure 4 As shown, firstly, a ResNet50 network structure is used as the cervical cell image feature learning module, and two fully connected layers are used as the non-image feature learning module for in-sample control information. Then, the image feature vectors output by the image feature learning module and the non-image feature vectors output by the non-image feature learning module are concatenated to form a feature fusion vector. Next, two fully connected layers and Softmax are used as the classification head to obtain the cervical abnormal cell classification results. Finally, the cervical abnormal cell classification model is trained using cervical abnormal cell classification training data that incorporates in-sample control information.

[0061] The identification of abnormal cervical cells under panoramic imaging includes the following steps:

[0062] S5. Use trained detection, segmentation, and classification models to identify abnormal cervical cells in a panoramic view of cervical cell pathology.

[0063] First, the panoramic image of cervical cell pathology to be identified is cropped to a 1024×1024 image. Then, the cropped images are predicted using normal and abnormal cell detection models to obtain sets of normal and abnormal cell images for the sample. Next, each cell image in the sets of normal and abnormal cell images is segmented using a segmentation model to obtain the outline information of the cell nucleus and cytoplasm for each cell image. Then, according to S3, cell feature indicators, background-free abnormal cell images, and corresponding in-sample control feature indicators are calculated using the outline information to obtain the input data for the classification model. Finally, the background-free abnormal cell images and corresponding in-sample control features are input into the classification model to predict the category information of the abnormal cell images.

[0064] 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 method of recognizing abnormal cervical cells by fusing sample-internal control information, characterized by, The method comprises the following steps: S1, preparing cervical abnormal cell detection training data, cervical normal intermediate squamous epithelial cell detection training data and cell nucleus and cytoplasm segmentation training data; S2, training a cervical abnormal cell detection model, a cervical normal intermediate squamous epithelial cell detection model and a cell nucleus and cytoplasm segmentation model based on a solov2 network structure; S3, obtaining normal cells by using the detection model trained in step S2 and extracting cell morphological parameters by using the segmentation model, and fusing the morphological parameters with the detection training data in step S1 to prepare cervical abnormal cell classification training data fused with sample-intra-control information; S4, constructing and training a cervical abnormal cell classification model fused with sample-intra-control information; S5, identifying cervical abnormal cells under a cervical cell pathological panorama by using the trained detection model, segmentation model and classification model. The detection model trained in step S2 is used to obtain normal cells, and the segmentation model is used to extract cell morphological parameters, and the morphological parameters are fused with the detection training data in step S1 to obtain cervical abnormal cell classification data fused with sample-intra-control information, and the specific steps are as follows: Firstly, the positive sample panorama is cut into pictures of a certain size in sequence, and the cervical normal intermediate squamous epithelial cell detection model is used to predict the cut pictures to obtain normal intermediate squamous epithelial cell pictures, and the abnormal cells in the positive sample marked by doctors are cut out to obtain abnormal cell pictures; then, the normal intermediate squamous epithelial cell pictures and the abnormal cell pictures of each sample are sent to the cell nucleus and cytoplasm segmentation model to obtain the cell nucleus and cytoplasm contour information of each cell picture; then, the cell feature indexes are calculated through the contour information, including cell nucleus area, cell nucleus integral optical density, cell nucleus roundness and nucleus-cytoplasm ratio, and the calculation formula is as follows: （1） In the formula, is the nucleus area, is the nucleus region, is a certain pixel point in the nucleus region, and the nucleus area is the sum of the pixel points in the nucleus region. （2） （3） wherein is the picture is the luminance at the position, is the picture is the position is the three-channel pixel value, is the average luminance of the cytoplasmic region, is the nuclear integrated optical density, reflecting the chromatin content within the nucleus; （4） wherein is the nuclear roundness, is the nuclear perimeter, i.e. the length of the nuclear contour; （5） In the formula, is the nucleus-cytoplasm ratio, is the cytoplasm area, i.e. the sum of pixel points in the cytoplasm region; then, the average value of all normal mid-layer squamous epithelial cell characteristic indexes in a single sample is taken as the sample reference characteristic index of the sample, and the sample-intra control characteristic index of each abnormal cell picture is obtained by dividing each abnormal cell characteristic index by the corresponding sample reference characteristic index; finally, the image within the cytoplasm contour information is retained, and the area outside the cytoplasm contour is filled with black, thereby obtaining the abnormal cell picture without background, and each abnormal cell picture without background and the corresponding sample-intra control characteristic index are combined to become the cervical abnormal cell classification training data, and the class label is provided by a pathologist.

2. The method for identifying abnormal cervical cells by fusing in-sample control information as described in claim 1, characterized in that, In step S1, the cervical abnormal cell detection, cervical normal intermediate squamous epithelial cell detection, cell nucleus and cytoplasm segmentation training data are prepared, and the specific steps are as follows: Firstly, a large number of cervical cell pathological smears diagnosed as positive and negative are collected, and a scanner is used to obtain cervical cell pathological panoramas; then, a pathologist marks cervical abnormal cells on the positive sample panorama and marks normal intermediate squamous epithelial cells on the negative sample panorama; then, pictures are cut from random positions of the panoramas, and it is ensured that the cut pictures contain complete marked information, the positive sample cut picture is used as cervical abnormal cell detection training data, and the negative sample cut picture is used as cervical normal intermediate squamous epithelial cell detection training data; finally, the doctor-marked cervical cell pictures are cut from the sample panoramas, and the cell nucleus and cytoplasm contours are marked by using LabelImg software as segmentation model training data.

3. The method of claim 1, wherein the method further comprises: identifying the abnormal cell in the sample by using the fusion sample internal control information. In step S2, the cervical abnormal cell detection model, the cervical normal intermediate squamous epithelial cell detection model and the cell nucleus and cytoplasm segmentation model are trained, and the specific steps are as follows: ​ The detection model adopts an improved YOLO v5 structure: in the backbone network, replace the ordinary convolution layer with a deformable convolution layer, obtain more rich gradient fusion information through the deformable convolution layer and the cross-stage local network to constitute a residual block, use the CIoU loss function for fine classification and frame regression, and train the detection model with abnormal cell and normal cell detection training data respectively; the segmentation model adopts the original structure of the Solo v2 network, and is trained end to end using the cell nucleus and cytoplasm segmentation training data.

4. The method of claim 1, wherein the method further comprises: identifying the abnormal cell in the sample by using the information of the fusion sample internal control. In step S3, the detection model trained in step S2 is used to obtain the normal cells and the segmentation model to extract the cell morphological parameters, and the morphological parameters are fused with the detection training data in step S1 to prepare the cervical abnormal cell classification training data with the fusion of the intra-sample control information, and the specific steps are as follows: ​ Firstly, the positive sample panoramic image is cut into a certain size of picture in turn, and the cervical normal intermediate squamous epithelial cell detection model is used to predict the cut picture to obtain the normal intermediate squamous epithelial cell picture, and the abnormal cells labeled by doctors in the positive sample are cut out to obtain the abnormal cell picture; then, the normal intermediate squamous epithelial cell picture and the abnormal cell picture of each sample are sent into the cell nucleus and cytoplasm segmentation model to obtain the cell nucleus and cytoplasm contour information of each cell picture; then, the cell feature indexes are calculated through the contour information, including the cell nucleus area, the cell nucleus integral optical density, the cell nucleus roundness and the nuclear-cytoplasmic ratio, and the calculation formula is as follows: （1） In the formula, is the nucleus area, is the nucleus region, is a certain pixel point in the nucleus region, and the nucleus area is the sum of the pixel points in the nucleus region. （2） （3） wherein is the picture the luminance at the position, is the picture at the position the three-channel pixel value, is the average luminance of the cytoplasmic region, is the nuclear integrated optical density, reflecting the chromatin content within the nucleus; （4） wherein is the nuclear roundness, is the nuclear perimeter, i.e. the length of the nuclear contour; （5） In the formula, is the nucleus-cytoplasm ratio, is the cytoplasm area, i.e. the sum of pixel points in the cytoplasm region; then, the average value of all normal mid-layer squamous epithelial cell characteristic indexes in a single sample is taken as the sample reference characteristic index of the sample, and the sample-intra control characteristic index of each abnormal cell picture is obtained by dividing each abnormal cell characteristic index by the corresponding sample reference characteristic index; finally, the image within the cytoplasm contour information is retained, and the area outside the cytoplasm contour is filled with black, thereby obtaining the abnormal cell picture without background, and each abnormal cell picture without background and the corresponding sample-intra control characteristic index are combined to become the cervical abnormal cell classification training data, and the class label is provided by a pathologist.

5. The method for identifying abnormal cervical cells by fusing in-sample control information as described in claim 1, characterized in that, In step S4, the cervical abnormal cell classification model with the fusion of the intra-sample control information is constructed and trained, and the specific steps are as follows: Firstly, the ResNet50 network structure is used as the cervical cell image feature learning module, and two fully connected layers are used as the intra-sample control information non-image feature learning module; Then, the image feature vector output by the image feature learning module and the non-image feature vector output by the non-image feature learning module are connected into a feature fusion vector; then, two fully connected layers and Softmax are used as the classification head to obtain the cervical abnormal cell classification result; finally, the cervical abnormal cell classification model is trained through the cervical abnormal cell classification training data with the fusion of the intra-sample control information.

6. The method for identifying abnormal cervical cells by fusing in-sample control information as described in claim 1, characterized in that, In step S5, the trained detection model, segmentation model and classification model are used to identify the cervical abnormal cells in the cervical cell pathology panoramic image, and the specific steps are as follows: Firstly, the cervical cell pathology panoramic image to be identified is cut into pictures of the same size; then, the normal cell picture set and the abnormal cell picture set of the sample are obtained by using the normal cell detection model and the abnormal cell detection model to predict the cut pictures respectively; then, the cell nucleus and cytoplasm contour information of each cell picture in the normal cell picture set and the abnormal cell picture set is obtained by using the segmentation model; then, the intra-sample control feature indexes and the background-free pictures of each abnormal cell picture are obtained by using the known contour information; finally, the class information of the abnormal cell picture is predicted by inputting the background-free abnormal cell picture and the corresponding intra-sample control features into the classification model.

Images