Unsupervised cervical cell instance segmentation method based on visual attention

Abstract

The application relates to a visual attention-based unsupervised cervical cell instance segmentation method, and relates to the problems of missing labeled data and accurate segmentation of cervical cells in intelligent auxiliary diagnosis technology of cervical cancer. Computer intelligent auxiliary diagnosis technology is widely applied, and cell segmentation technology is the basis of various downstream tasks. A deep learning model needs a large amount of labeled data for training, pixel-level labeling is time-consuming and labor-consuming, and there are impurities such as bacteria, white blood cells and bubbles caused by physiological reasons and film reasons, in addition, there are problems such as overlapping adhesion and visual inseparability of cervical cell images. In order to improve these problems, a visual attention-based unsupervised cervical cell instance segmentation method is provided. Experiments show that the method can effectively improve the accuracy of segmentation and reduce the missing detection problems caused by the interference of impurities and incomplete labels. The application is applied to accurate segmentation of cervical cells under the condition of no label.

Description

TECHNICAL FIELD

[0001] The present application relates to the instance segmentation problem of cytoplasm and nucleus of cervical cells. BACKGROUND

[0002] Cervical cancer is a malignant tumor that seriously endangers women's health, and timely and effective screening is of great significance for the prevention and treatment of cervical cancer. However, the traditional manual screening method has the disadvantages of time-consuming, labor-intensive and strong subjectivity, which limits the large-scale application of cervical cancer screening. The computer-aided diagnosis method based on deep learning can effectively solve this pain point. However, deep learning methods usually need a large amount of high-precision labeled data for training to obtain accurate and generalizable results, and the acquisition of labeled data is a tedious and expensive task. At the same time, the segmentation step in computer-aided diagnosis is the premise of various downstream tasks, and the calculation of cytoplasm and nucleus indicators through segmentation is an important means of quantifying the degree of cell abnormalities, providing strong explainability for cancer grading and diagnosis.

[0003] The cervical cell instance segmentation method needs to segment each cell on the one hand, and needs to segment the cytoplasm and nucleus on the other hand. Due to the existence of a large number of neutrophils and impurities in the cervical image, the scale difference between the cytoplasm and the nucleus is large, the image style varies from person to person, and there are a large number of overlapping cluster cells, which all pose challenges to unsupervised instance segmentation. This paper proposes an unsupervised cervical cell segmentation method based on feature pyramid network (FPN), which consists of three modules: pseudo-label generation stage, model training stage and self-training stage. This application proposes a new cell region feature enhancement module based on visual attention mechanism to simulate human cell recognition behavior. Specifically, the cell attention feature is fused with the multi-scale feature to guide the network to segment the cell and greatly reduce the interference of complex background, and improve the segmentation accuracy; at the same time, due to the shortcomings of incomplete pseudo-labels, a new loss function is proposed in the target detection stage, so that the model can find more instances in the training process and improve the detection recall rate. SUMMARY

[0004] The purpose of the present application is to solve the high-quality instance segmentation task of cervical cells without manual labeling, and thus proposes an unsupervised cervical cell instance segmentation method based on visual attention.

[0005] The above application purpose is mainly realized by the following technical scheme:

[0006] S1, collect cell images with diversity and perform data enhancement;

[0007] A large number of cell images are obtained by an automatic scanner, images of different patients, different institutions and different hospitals are selected, the images are cropped into small images with a size of , and are divided into a training set, a test set and a validation set, each cytoplasm and nucleus contour of the test set and the validation set is marked by a professional using a marking tool as a basis for measuring the accuracy of the model; and all images are subjected to different degrees of data enhancement, including contrast, saturation, color and random horizontal and vertical flipping.

[0008] S2, obtaining labels of unlabeled images by using a pseudo-label generation algorithm, and selecting high-quality target frame labels as pseudo-labels of the data set;

[0009] There are two methods of pathological staining: Papanicolaou staining and Hemotoxylin & Eosin (HE) staining. Papanicolaou staining has clear cytoplasm and nucleus, the nucleus is dark blue, and the cytoplasm is pink or orange; HE staining has clear cytoplasm and nucleus, the nucleus is purple blue, and the cytoplasm is light red; it can be seen that the cytoplasm and nucleus of the cell image have color priori, shape priori, and the cell image has the structural priori that the cytoplasm contains the nucleus; therefore, the pseudo-label generation algorithm that fuses priori knowledge fully considers the various priori knowledge of cells to obtain pseudo-labels of cytoplasm and nucleus, and the specific method includes the following sub-steps:

[0010] S2-1, using an iterative threshold segmentation algorithm and fusing color priori and shape priori, obtaining a binary mask map of the nucleus, which is expressed by the formula:

[0011]

[0012]

[0013] In the formula, represents all the nucleus pseudo-labels, represents a binary image obtained by threshold segmentation with as the threshold value, is the lower limit of the threshold value, is the upper limit of the threshold value, is the area of the contour, and the calculation formula is (3), is the roundness of the contour, and the calculation formula is (4), is the threshold value of the contour area, is the threshold value of the contour roundness;

[0014]

[0015]

[0016] wherein, is the width of, is the height of, is the perimeter of the contour, is the balance factor;

[0017] S2-2, the best cytoplasm segmentation threshold is obtained by Otsu algorithm, denoted as ;

[0018] S2-3, considering that when is larger, the image is lighter, that is, the actual best cytoplasm segmentation threshold should be larger, so a mapping function is introduced to obtain the actual best cytoplasm segmentation threshold, the formula is as follows:

[0019]

[0020] wherein, denotes the segmentation threshold obtained by Otsu algorithm, denotes the new cytoplasm segmentation threshold, is the balance coefficient;

[0021] S2-4, the connected regions are segmented, first, the watershed algorithm is used to segment the single cells and simple clustered cells, wherein the threshold is the best cytoplasm segmentation threshold obtained in S2-3; then the level set algorithm is used to segment the complex clustered cells, taking the nucleus mask generated in S2-1 as the seed point, and finding the boundary points of the clustered cells through the evolution function, so as to obtain the mask map of the cytoplasm;

[0022] S2-5, since there is a structural prior between cells, that is, the cytoplasm is contained in the nucleus, this characteristic can be used to remove neutrophils, impurities and other substances, so as to obtain more accurate cytoplasm and nucleus labels, and the cytoplasm and nucleus that satisfy formula (6) are selected:

[0023]

[0024] wherein, and denote the center distance of the cytoplasm, and denote the center distance of the nucleus, the calculation of the center distance of the contour, denotes the radius of the cytoplasm contour.

[0025] S3, build an instance segmentation model, and train the instance segmentation model through the data set built in S1 and the pseudo label generated in S2;

[0026] S3-1, the backbone part in the instance segmentation model adopts ResNet for feature extraction, and since the scale difference between the cytoplasm and the nucleus is large, a feature pyramid network (FPN) is used to extract multi-scale semantic information, and 5 feature layers are used to detect objects of different sizes, and the feature layers are P2, P3, P4, P5 and P6 from large to small;

[0027] S3-2, the visual attention mechanism in the instance segmentation model attempts to simulate the process of the human visual system when segmenting cells, randomly selects a point of each high-quality nucleus, obtains a nucleus attention image centered on the nucleus point through an attention encoder, fuses the multi-scale semantic information, and fuses into a semantic decoder to obtain a semantic attention map, so that the model pays attention to the cells and reduces the interference of the background and impurities;

[0028] S3-3, the overall loss of the instance segmentation model includes a target detection loss and a segmentation loss, wherein the labels of the model training are pseudo labels generated by the pseudo label generation algorithm in step S2, and since the pseudo labels are not complete, in order to reduce the missed detection defects of the model, a class-wise IoU-weighted loss (CIWL) function is proposed in the target detection stage, so that the model does not completely rely on the label and explores more instances, wherein the design of the class-wise enables the model to pay attention to the nucleus inside the cytoplasm, and the specific loss function is:

[0029]

[0030]

[0031] In the formula, N represents the number of classes, N represents the number of predicted instances, N represents the maximum overlap degree of each predicted region and the label, N represents the overlap threshold of the prediction and the label, N represents the class loss, including a bounding box regression loss and a center point regression loss;

[0032]

[0033] In the formula, N represents the predicted region, representing the true label, representing with intersection of pixel points, representing with union of pixel points;

[0034] In the segmentation stage, the projection loss and color similarity loss are used to segment the foreground and background under the condition of only target frame label.

[0035] S4, using label updating algorithm to obtain more accurate pseudo label;

[0036] According to the reanalysis of the model on the training set, the threshold is greater than , the predicted label greater than Replace the label in the original position, update the label through multiple self-training, and then obtain more accurate and comprehensive label, wherein the threshold selection rule is:

[0037]

[0038] In the formula, representing the first round of self-training predicted label threshold, representing the threshold attenuation coefficient, representing the self-training round, representing the predicted label threshold of this self-training.

[0039] S5, through multiple self-training, obtain model parameters more fitting image distribution;

[0040] S5-1, select the model parameters that do not completely fit the data from the last training stage, and take the early self-training method to avoid the dilemma of local optimum of the model;

[0041] S5-2, replace the model training label with the updated pseudo label obtained in S4;

[0042] S5-3, change the hyperparameters of self-training, including pre-training model, learning rate, re-enter the training stage, and repeat the training process in S3, S4 and S5, until the model parameters better fit the data.

[0043] Inventive effect

[0044] This invention provides an unsupervised cervical cell instance segmentation method based on visual attention. The algorithm first selects a large amount of unlabeled data and generates pseudo-labels for the dataset using a pseudo-label generation algorithm. Then, it trains an instance segmentation model to model the boundaries of the cytoplasm and nucleus. A novel visual attention mechanism is proposed to make the network focus more on cells, reducing the influence of various impurities in the image and improving segmentation accuracy. In the object detection stage, a new category-aware weighted loss is proposed, taking into account the characteristics of unsupervised tasks. This allows the network to focus on more instances, reducing missed detections. The multi-scale features enable the model to focus on objects of different scales, making it more suitable for cell images with large scale differences between the cytoplasm and nucleus. In the segmentation stage, projection loss and color similarity loss are used to enable the network to accurately distinguish between foreground and background even with only bounding box annotations. Experiments show that this invention can accurately segment the cytoplasm and nucleus of each cell and can be effectively applied to automated computer-aided diagnostic systems. Attached Figure Description

[0045] Figure 1 This is an overall structural diagram of an unsupervised cervical cell instance segmentation method based on visual attention.

[0046] Figure 2 Algorithm diagram for generating pseudo-labels;

[0047] Figure 3 A diagram of the network structure for unsupervised cervical cell instance segmentation based on visual attention;

[0048] Figure 4 This is a rendering of the model. Specific implementation methods Specific implementation method one:

[0050] 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.

[0051] like Figure 1 As shown, this paper presents an unsupervised cervical cell instance segmentation method based on visual attention, which mainly includes the following steps:

[0052] S1. Acquire diverse cell images and perform data augmentation;

[0053] S2. Use a pseudo-label generation algorithm to obtain labels for unlabeled images, and select high-quality bounding box labels as pseudo-labels for the dataset.

[0054] S3. Construct an instance segmentation model and train the instance segmentation model using the dataset constructed in S1 and the dataset generated in S2 with pseudo-labels.

[0055] S4. Use the label update algorithm to obtain more accurate pseudo-labels;

[0056] S5. Through multiple self-training sessions, model parameters that better fit the image distribution are obtained.

[0057] In this embodiment of the invention, a diverse dataset is first collected and augmented to varying degrees to serve as the dataset for the unsupervised method. Then, pseudo-labels are obtained using multiple prior knowledge of cells. An instance segmentation model is constructed, and a human visual perception module is added to mitigate interference from complex backgrounds and improve segmentation accuracy. Simultaneously, this invention proposes a CIWL loss function to reduce the penalty for false negative samples in the prediction, thereby improving the model's recall. Finally, more accurate pseudo-labels are obtained through self-training, further enhancing model performance.

[0058] The embodiments of the present invention will be described in detail below:

[0059] like Figure 1 The implementation of the algorithm shown includes the following steps:

[0060] S1. Acquire diverse cell images and perform data augmentation;

[0061] A large number of cell images were acquired using an automated scanner. Images from different patients, institutions, and hospitals were selected and cropped into 1024×1024 small images. These images were then divided into training, testing, and validation sets in a 6:2:2 ratio. Professionals used annotation tools to mark the outlines of each cytoplasm and nucleus in the testing and validation sets, which served as a basis for measuring the accuracy of the model. All images were augmented to varying degrees, including by adjusting contrast, saturation, color, and random horizontal and vertical flipping.

[0062] S2. Use a pseudo-label generation algorithm to obtain labels for unlabeled images, and select high-quality bounding box labels as pseudo-labels for the dataset.

[0063] There are two methods for pathological staining: Papanicolaou staining and hematoxylin & eosin (HE) staining. Papanicolaou staining clearly distinguishes the nucleus and cytoplasm, with the nucleus appearing dark blue and the cytoplasm pink or orange. HE staining also clearly distinguishes the nucleus and cytoplasm, with the nucleus appearing purplish-blue and the cytoplasm pale red. Therefore, it is evident that cell images possess innate color and shape priors for the cytoplasm and nucleus, and that cell images also possess a structural prior of the cytoplasm containing the nucleus. Algorithms for pseudo-label generation that integrate this prior knowledge include... Figure 2As shown, the cytoplasm and the nucleus are pseudo-labeled by fully considering various prior knowledge of the cells, and the specific method comprises the following sub-steps:

[0064] S2-1, an iterative threshold segmentation algorithm is used, and color prior and shape prior are fused to obtain a binary mask map of the nucleus, which is expressed by the formula:

[0065]

[0066]

[0067] In the formula, represents all the nucleus pseudo-labels, represents a binary map obtained by threshold segmentation with as the threshold value, is the lower limit of the threshold value, is the upper limit of the threshold value, is the area of the contour, and the calculation formula is (13), is the roundness of the contour, and the calculation formula is (14), is the threshold value of the contour area, is the threshold value of the contour roundness, and in the embodiment, is 50, is 0.5;

[0068]

[0069]

[0070] In the formula, is the width of , is the height of , is the perimeter of the contour, is the balance factor, and in the embodiment, is ;

[0071] S2-2, the best cytoplasm segmentation threshold value is obtained by Otsu algorithm, and is recorded as ;

[0072] S2-3, considering that when is large, the image staining is lighter, that is, the actual best cytoplasm segmentation threshold value should be larger, therefore a mapping function is introduced to obtain the actual best cytoplasm segmentation threshold value, and the formula is as follows:

[0073]

[0074] In the formula, represents the segmentation threshold obtained by the Otsu algorithm, is a balance coefficient, in the embodiment, is 1.2, is 20, is 230;

[0075] S2-4, the connected regions are segmented, first, the single cells and simple clustered cells are segmented using the watershed algorithm, wherein the threshold is the optimal cytoplasm segmentation threshold obtained in S2-3; then, the complex clustered cells are segmented using the level set algorithm, taking the nucleus mask generated in S2-1 as a seed point, and finding the boundary points of the clustered cells through an evolution function, so as to obtain a cytoplasm mask graph;

[0076] S2-5, since there is a structural prior between cells, that is, the cytoplasm is contained in the nucleus, this characteristic can be used to remove neutrophils, impurities and other substances, so as to obtain more accurate cytoplasm and nucleus labels, and the cytoplasm and the nucleus satisfying formula (16) are selected:

[0077]

[0078] In the formula, and represent the center distance of the cytoplasm, and represent the center distance of the nucleus, and the calculation of the center distance of the contour, represents the radius of the cytoplasm contour.

[0079] S3, an instance segmentation model is constructed, and the instance segmentation model is trained through the data set constructed in S1 and the data set pseudo-label generated in S2;

[0080] S3-1, the model structure diagram is as shown in Figure 3 , wherein the backbone part in the instance segmentation model adopts ResNet for feature extraction, since the scale difference between the cytoplasm and the nucleus is large, and the FPN structure is used to extract multi-scale semantic information, 5 feature layers are adopted to detect objects of different sizes, and the feature layers are P2, P3, P4, P5 and P6 from large to small.

[0081] S3-2, the visual attention mechanism in the instance segmentation model attempts to imitate the process of the human visual system when segmenting cells, randomly selects a point of each high-quality nucleus, obtains a nucleus attention image centered on the nucleus point through the attention encoder, fuses with multi-scale semantic information, and fuses into a semantic decoder to obtain semantic attention, so that the model strengthens the attention to the cells and reduces the interference of the background and impurities;

[0082] S3-3, the overall loss in the instance segmentation model includes a target detection loss and a segmentation loss, wherein the label of the model training is generated by the pseudo label generation algorithm in step S2, and since the pseudo label is not complete, in order to alleviate the missed detection defect of the model, a CIWL function is proposed in the target detection stage, so that the model does not completely rely on the label and explores more instances, wherein the design of the class awareness enables the model to focus on the nucleus inside the cytoplasm, and the specific loss function is:

[0083]

[0084]

[0085] In the formula, represents the class number, represents the prediction result, represents the maximum overlap degree of each prediction region and the label, represents the overlap degree threshold of the prediction and the label, represents the class loss, includes a bounding box regression loss and a center point regression loss, in the embodiment, is 0.01;

[0086]

[0087] In the formula, represents the prediction region, represents the true label, represents and the intersection of the pixel points, represents and the union of the pixel points;

[0088] In the segmentation stage, projection loss and color similarity loss are used, so that the foreground and background can be segmented under the condition that only the target frame label is available.

[0089] S4, a label updating algorithm is used to obtain more accurate pseudo labels;

[0090] According to the reanalysis of the training set by the model, the threshold is greater than 、 greater than The predicted label replaces the original label, and the label is updated by multiple self-training to obtain more accurate and comprehensive labels, wherein the threshold selection rule is:

[0091]

[0092] In the formula, represents the first round of self-training predicted label threshold, represents the threshold attenuation coefficient, represents the self-training round, represents the self-training predicted label threshold, in this embodiment, 0.6, 0.05, 0.7.

[0093] S5, through multiple self-training, the model parameters more fitting the image distribution are obtained;

[0094] S5-1, the model parameters that do not completely fit the data are selected from the last training stage, and the early self-training mode is adopted to avoid the dilemma of local optimum of the model;

[0095] S5-2, the model training label is replaced by the updated pseudo label obtained in S4;

[0096] S5-3, the hyperparameters of self-training are changed, including pre-training model, learning rate, re-entering the training stage, and repeating the training process in S3, S4 and S5 until the model parameters better fit the data.

[0097] The final implementation effect is shown in Figure 4 As can be seen from the figure, the method is suitable for different staining style cervical cell images, and has good accuracy and generalization.

[0098] The present application also has other various embodiments, and those skilled in the art can make various corresponding changes and modifications according to the present application without departing from the spirit and essence of the present application, but these corresponding changes and modifications should belong to the scope of the present application.

Claims

1. An unsupervised cervical cell instance segmentation method based on visual attention, characterized in that, Includes the following steps: S1. Acquire diverse cell images and perform data augmentation; S2. Use a pseudo-label generation algorithm to obtain labels for unlabeled images, and select high-quality bounding box labels as pseudo-labels for the dataset. S3. Construct an instance segmentation model and train it using the dataset constructed in S1 and the dataset generated in S2 with pseudo-labels. The backbone of the instance segmentation model uses ResNet for feature extraction and the feature pyramid structure FPN is used to extract multi-scale semantic information. Five feature layers are used to detect objects of different sizes. The feature layers from largest to smallest are P2, P3, P4, P5, and P6. The visual attention mechanism in the instance segmentation model mimics the process of the human visual system when segmenting cells. It randomly selects a point from each high-quality cell nucleus, obtains a cell nucleus attention image centered on the cell nucleus point through an attention encoder, fuses it with multi-scale semantic information, and integrates it into a semantic decoder to obtain a semantic attention map, so that the model pays attention to the cell and reduces the interference of background and impurities. The overall loss in the instance segmentation model includes object detection loss and segmentation loss. The labels used for model training are pseudo-labels generated by the pseudo-label generation algorithm described in step S2. Since these pseudo-labels are incomplete, a novel category-aware weighted loss function (CIWL) is proposed in the object detection stage to mitigate the model's missed detection defects. This allows the model to explore more instances rather than relying entirely on labels. The category-aware design enables the model to focus on the cell nucleus within the cytoplasm. The specific loss function is as follows: In the formula, Indicates the category number, Indicates the prediction result. This represents the maximum overlap between each predicted region and the label. This represents the threshold for the overlap between the prediction and the label. Indicates category loss. This includes bounding box regression loss and center point regression loss; In the formula, Indicates the prediction area. Indicates the true label, express and The intersection of pixels express and Union of pixels; In the segmentation stage, projection loss and color similarity loss are used to segment the foreground and background even when only the bounding box labels are available. S4. Use the label update algorithm to obtain more accurate pseudo-labels; S5. Through multiple self-training sessions, model parameters that better fit the image distribution are obtained.

2. The unsupervised cervical cell instance segmentation method based on visual attention as described in claim 1, characterized in that, The diversity data collection and data augmentation in step S1 are explained below: A large number of cell images were acquired using an automated scanner. Images from different patients, institutions, and hospitals were selected and cropped to a specific size. Small images were collected and divided into training, testing, and validation sets. Professionals used annotation tools to mark the outlines of each cytoplasm and nucleus in the testing and validation sets as a basis for measuring the accuracy of the model. Data augmentation was performed on all images to varying degrees, including contrast, saturation, color, and random horizontal and vertical flipping.

3. The unsupervised cervical cell instance segmentation method based on visual attention as described in claim 1, characterized in that, The pseudo-label generation algorithm in step S2 is as follows: There are two methods for pathological staining: Papanicolaou staining and hematoxylin-eosin (HE) staining. Papanicolaou staining clearly distinguishes the nucleus from the cytoplasm, with the nucleus appearing deep blue and the cytoplasm pink or orange. HE staining also clearly distinguishes the nucleus from the cytoplasm, with the nucleus appearing purplish-blue and the cytoplasm pale red. Therefore, it is evident that cell images possess innate color and shape priors for the cytoplasm and nucleus, and that the cytoplasm contains the nucleus structurally. Thus, the pseudo-label generation algorithm fully considers these multiple prior knowledge points to obtain pseudo-labels for the cytoplasm and nucleus. The specific method includes the following sub-steps: S2-1. Using an iterative thresholding algorithm and fusing color and shape priors, a binary mask image of the cell nucleus is obtained, expressed by the formula: In the formula, This represents all nuclear pseudotags. Indicated by Binary images obtained by thresholding based on a threshold value. This is the lower limit of the threshold. The upper limit of the threshold, The area of ​​the outline is calculated using formula (6). The roundness of the outline is calculated using formula (7). The threshold value for the contour area. The threshold for the roundness of the outline; In the formula, for width, for height, Let be the perimeter of the outline. As a balance factor; S2-2, Calculate the optimal cytoplasmic segmentation threshold using the Otsu algorithm, denoted as . ; S2-3. Introduce a mapping function to obtain the actual optimal cytoplasmic segmentation threshold, as shown in the following formula: In the formula, This represents the segmentation threshold obtained by the Otsu algorithm. This is the balance coefficient; S2-4. Separate the connected regions. First, use the watershed algorithm to segment individual cells and simple clusters of cells, where the threshold is the optimal cytoplasmic segmentation threshold obtained in S2-3. Then, use the level set algorithm to segment complex clusters of cells. Using the cell nucleus mask generated in S2-1 as the seed point, find the boundary points of clusters of cells through the evolution function to obtain the cytoplasmic mask map. S2-5. Due to a structural priori existence between cells, namely that the cytoplasm is contained within the nucleus, this property can be used to remove neutrophils and impurities, thereby obtaining more accurate cytoplasmic and nuclear labels. Select cytoplasm and nucleus that satisfy formula (9): In the formula, and Indicates the distance between the centers of the cytoplasm. and This represents the calculation of the center distance between cell nuclei and cell outlines. The radius representing the outline of the cytoplasm.

4. The unsupervised cervical cell instance segmentation method based on visual attention as described in claim 1, characterized in that, The label update algorithm in step S4 is as follows: Based on the model's reanalysis of the training set, select thresholds greater than [a certain threshold]. , Greater than The predicted label replaces the label in the original position. Through multiple self-training iterations, the label is continuously updated to obtain a more accurate and comprehensive label. The threshold selection rule is as follows: In the formula, This represents the threshold for predicting labels during the first round of self-training. Indicates the threshold attenuation coefficient. Indicates the number of training rounds. This indicates the threshold for predicting labels during this self-training process.

5. The unsupervised cervical cell instance segmentation method based on visual attention as described in claim 1, characterized in that, The self-training steps in step S5 are as follows: S5-1. Select model parameters from the previous training stage that did not fully fit the data, and adopt a self-training approach in advance to avoid the model getting stuck in local optima. S5-2, Replace the model training labels with the updated pseudo-labels obtained in S4; S5-3. Change the hyperparameters to self-training parameters, including the pre-trained model and learning rate, re-enter the training phase, and repeat the training process in S3, S4, and S5 until the model parameters fit the data well.

Images