Cervical TCT image cell segmentation method based on improved U-Net

Abstract

The invention relates to a cervical TCT image cell segmentation method based on improved U-Net, and belongs to the field of image processing technology and medical technology. The problems that complex morphological characteristics of cells are not fully considered and fine cell segmentation is insufficient in cervical cancer intelligent screening research based on cervical TCT images are solved. Comprising the following steps: preprocessing a cervical cell TCT image; constructing a segmentation network based on the EGE-UNet; evaluating the cell image segmentation model based on qualitative and quantitative evaluation; and obtaining a cervical cell segmentation result based on the evaluated model. According to the method, high-quality cervical TCT cell image segmentation and fine segmentation of cervical cell nucleuses and cytoplasm can be realized by using the EGE-UNet model, and related research of subsequent cervical cell intelligent diagnosis is facilitated; based on an accurate cell segmentation result, the method can assist in extracting various cell morphological characteristics, has clinical guiding significance, and provides more interpretable quantitative data for subsequent cervical cell research.

Description

Technical Field

[0001] The invention relates to a cell segmentation method based on an improved U-Net, and belongs to the fields of image processing technology and medical technology. Background Art

[0002] Cervical cancer is one of the common malignant tumors in women. Early detection and intervention are crucial to reducing its morbidity and mortality. The existing cervical cancer screening methods are mainly cytology and HPV testing. With the development of modern cytology technology, thin layer cytology (TCT) has been widely used in early screening of cervical cancer in China. The cervical TCT examination collects exfoliated epithelial cells from the cervix, prepares and stains them, and professional pathologists read the cell morphology under a microscope and judge according to the Bethesda system (TBS) standards. However, TCT examinations rely on manual film reading, and doctors need to spend a lot of time to traverse all cells, which is inevitably subjective. There are few professionals, high training costs, and long-term and intensive film reading work by doctors, which is prone to problems such as missed detection and misjudgment, increasing the uncertainty of TCT examinations.

[0003] With the advancement of digital whole slide image scanning equipment, accurate and efficient computer-assisted cervical cancer screening has become possible. Whole slide images (WSI) are obtained by scanning a slide using an optical microscope. In recent years, data-driven deep learning methods have developed rapidly and have shown the potential to surpass experts in intelligent auxiliary diagnosis of diseases. However, unlike the continuous lesion area in histopathology, cytopathology WSI usually has thousands of cells with almost no structural association with each other, and even a single isolated lesion cell may be regarded as an abnormal sample. Most of the current research focuses on improving the effect of the classification model itself, and fails to fully consider the complexity of cervical cytology. A single deep learning model is still insufficient in dealing with the complex morphological characteristics of cervical cells. For example: The publication number is CN117788478A, and the name of the invention is a medical image segmentation method, system, device and storage medium. Its technical solution proposes a general medical image segmentation method, which does not involve cervical cell TCT images, affecting the segmentation effect on specific cervical cell images.

[0004] Therefore, it is urgent to propose a cell segmentation method for cervical TCT images based on improved U-Net to solve the above technical problems. Summary of the invention

[0005] In order to solve the problem that the complex morphological characteristics of cells and the fine segmentation of cells are not fully considered in the intelligent screening research of cervical cancer based on cervical TCT images, a cell segmentation method of cervical TCT images based on improved U-Net is provided. A brief overview of the present invention is given below to provide a basic understanding of certain aspects of the present invention. It should be understood that this overview is not an exhaustive overview of the present invention. It is not intended to determine the key or important parts of the present invention, nor is it intended to limit the scope of the present invention.

[0006] The technical solution of the present invention:

[0007] A cell segmentation method for cervical TCT images based on improved U-Net, comprising the following steps:

[0008] Step 1: Preprocessing the cervical cell TCT image;

[0009] Step 2: Construct a segmentation network based on EGE-UNet to achieve fine segmentation of cervical cell nuclei and cytoplasm;

[0010] Step 3: Cell image segmentation model based on qualitative and quantitative evaluation;

[0011] Step 4: Obtain cervical cell segmentation results based on the evaluated model.

[0012] Preferably, the step of preprocessing the cervical cell TCT image in step 1 includes:

[0013] Step 1 (i) select whole slide images (WSI) of cervical TCT of multiple cases to establish a data set, sort and classify, and rename the WSI images of each case according to the rule of "category name + case serial number" to ensure the balance of the number of categories;

[0014] Step 1 (2), considering that the size of WSI is too large to load, select a suitable image size, crop and divide the image, and adjust the image blocks in batches to a unified format of 1024×1024;

[0015] Step 1 (iii) filter out useless image blocks, such as boundary traces, unclear preparation, impurities, staining overlap, etc.; standardize the data in accordance with actual clinical needs.

[0016] Preferably, the fine segmentation step of the cervical cell nucleus and cytoplasm in step 2 includes:

[0017] Step 2 (a) divide the data set into training set, validation set and test set at a ratio of 7:2:1; perform online data enhancement on the training data, including random contrast, brightness and saturation adjustment;

[0018] Step 2 (ii), build the EGE-UNet network: design a U-shaped symmetrical structure encoder and decoder based on the U-Net architecture, with a short circuit connection between the encoder and the decoder. In EGE-UNet, the GHPA attention mechanism and the GAB module are introduced;

[0019] Step 2 (3), construct the target loss function; using deep supervision to generate prediction masks of different scales, it is necessary to calculate the loss functions of different stages;

[0020] Step 2 (iv), network model training: The constructed EGE-UNet network is trained on the preprocessed data set, and the parameters of the model are adjusted to achieve the optimal model, which can accurately segment the nucleus and cytoplasm of cervical cells.

[0021] Preferred: In step 2 (ii), select BCE loss and Dice loss as the loss function of the network, and the loss weight λ for each stage i The values ​​are as follows:

[0022]

[0023] Preferably: In step 2 (3), the loss functions at different stages are calculated and defined as follows:

[0024]

[0025] In formula (1) and formula (2), Bce represents BCE loss, Dice represents Dice loss, and y represents the label value of the pixel. represents the predicted value of the pixel, l i represents the loss of each i-th stage, l represents the target loss function, and λ i Represents the weights of different stages.

[0026] Preferably, the step of evaluating the cell image segmentation model in step 3 includes:

[0027] Step 3 (a) Qualitative evaluation of segmentation results: Compare the model trained in step 2 with the commonly used segmentation network, and output the performance on the test set as a mask for intuitive evaluation;

[0028] Step 3 (ii), quantitative evaluation of segmentation evaluation indicators: For the predicted segmentation masks generated by the trained model in step 2 and the commonly used segmentation network, the mean intersection over union (MIoU), specificity, and sensitivity are calculated respectively, and quantitative indicators are compared.

[0029] Preferably: In step 3 (ii), for the predicted segmentation mask generated by the segmentation network, the mean intersection over union (MIoU), specificity, and sensitivity are calculated respectively, and the evaluation indicators are defined as follows:

[0030] MIoU is the intersection of the predicted area and the actual area divided by the union of the predicted area and the actual area. The IoU under a single category is calculated, and then this algorithm is repeated to calculate the IoU of other categories, and then their average is calculated; it can be expressed as a formula:

[0031]

[0032] In formula (4), p ij represents the number of pixels that belong to class i but are predicted to be class j, that is, the true number; p ij represents the number of pixels misclassified as category j, p ji represents the number of pixels that should be classified as class j but are mistakenly excluded;

[0033] Based on the confusion matrix, the specificity and sensitivity are defined as follows:

[0034]

[0035]

[0036] In formula (5) and formula (6), TP represents the number of correctly classified positive samples; FP represents the number of negative samples that are incorrectly labeled as positive samples; TN represents the number of correctly classified negative samples; FN represents the number of positive samples that are incorrectly labeled as negative samples;

[0037] Compare the EGE-UNet network with the most commonly used segmentation networks Mask R-CNN and Swin-UNet.

[0038] Preferably, the step of obtaining the cervical cell segmentation result based on the evaluated model in step 4 includes:

[0039] Step 4 (a), load and call the trained and evaluated EGE-UNet model, and input the tested cervical TCT image into the model;

[0040] Step 4 (ii): perform cell nucleus and cytoplasm segmentation on the specified cervical TCT cell image to obtain a segmentation result.

[0041] The present invention has the following beneficial effects:

[0042] 1. The present invention uses the EGE-UNet model to achieve high-quality cervical TCT cell image segmentation and fine segmentation of cervical cell nuclei and cytoplasm, which is conducive to subsequent research on intelligent diagnosis of cervical cells;

[0043] 2. Based on accurate cell segmentation results, the present invention can assist in extracting a variety of cell morphological features, has clinical guidance significance, and provides more interpretable quantitative data for subsequent cervical cell research;

[0044] 3. The present invention can efficiently and automatically perform cell segmentation in cervical TCT images, significantly improving the automation level of cervical cancer screening and providing technical support for large-scale cervical cancer screening;

[0045] 4. The present invention relates to the field of image processing technology and the medical field, and has a positive role in promoting intelligent assisted medical pathology analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The figure is a flowchart of a cell segmentation method for cervical TCT images based on an improved U-Net.

[0047] Figure 2 It is a schematic diagram of the preprocessing of TCT images of cervical cells;

[0048] Figure 2 (a) is a schematic diagram of a TCT whole slide image of cervical cells;

[0049] Figure 2 (b) is a schematic diagram of the cell viewing angle that can be adjusted by overlapping segmentation;

[0050] Figure 2 (c) is a schematic diagram of some images in the dataset.

[0051] Figure 3 It is a schematic diagram of the network structure of each part of the EGE-UNet model for cervical cell segmentation;

[0052] Figure 3 (a) is a schematic diagram of the EGE-UNet network structure;

[0053] Figure 3 (b) is a schematic diagram of the GAB module.

[0054] Figure 4 It is a schematic diagram of the qualitative evaluation and comparison of cervical cell segmentation models.

[0055] Figure 5 is a schematic diagram of the cervical cell segmentation results obtained based on the evaluated model. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is described below by the specific embodiments shown in the accompanying drawings. However, it should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0057] Specific implementation method 1: Combination Figure 1-5This embodiment is described. This embodiment is a cell segmentation method for cervical TCT images based on an improved U-Net. This patent is only for cervical cells. It starts with collecting cervical cell TCT images (a type of cytopathology image), including processing and data set establishment of cervical cell whole slide images WSI, to microscopic level cervical cell segmentation. It is an integrated cervical cell segmentation method, including the following steps:

[0058] Step 1: Preprocessing the cervical cell TCT image;

[0059] The steps of preprocessing the cervical cell TCT image in step 1 include:

[0060] Step 1 (i) select whole slide images (WSI) of cervical TCT of multiple cases to establish a data set, sort and classify, and rename the WSI images of each case according to the rule of "category name + case serial number" to ensure the balance of the number of categories;

[0061] Step 1 (2), considering that the size of WSI is too large to load, select a suitable image size, crop and divide the image, and adjust the image blocks in batches to a unified format of 1024×1024;

[0062] Step 1 (iii) Screen out useless image blocks, such as boundary traces, unclear film, impurities, staining overlap, etc.; standardize the data in accordance with actual clinical needs; for example: use "Several Methods for Standardization of Clinical Medical Data" to scale the data so that it falls into a small specific interval, thereby removing the data unit restriction, enhance the overall quality of the image through methods such as Z-Score standardization, and ensure the normalization of different feature vector channels, and use "Sharing of Big Data Serving Clinical Practice" to integrate multi-source heterogeneous data and perform standardization;

[0063] Step 2: Construct a segmentation network based on EGE-UNet to achieve fine segmentation of cervical cell nuclei and cytoplasm;

[0064] The detailed segmentation steps of cervical cell nuclei and cytoplasm in step 2 include:

[0065] Step 2 (a) divide the data set into training set, validation set and test set at a ratio of 7:2:1; perform online data enhancement on the training data, including random contrast, brightness and saturation adjustment;

[0066] Step 2 (ii), build the EGE-UNet network: design a U-shaped symmetrical structure encoder and decoder based on the U-Net architecture, with a short circuit connection between the encoder and the decoder. In EGE-UNet, the GHPA attention mechanism and the GAB module are introduced;

[0067] In step 2 (ii), BCE loss and Dice loss are selected as the loss function of the network, and the loss weight λ for each stage i To obtain the value;

[0068] Step 2 (3), construct the target loss function; using deep supervision to generate prediction masks of different scales, it is necessary to calculate the loss functions of different stages;

[0069] In step 2 (3), the loss functions at different stages are calculated and defined as follows:

[0070]

[0071] In formula (1) and formula (2), Bce represents BCE loss, Dice represents Dice loss, and y represents the label value of the pixel. represents the predicted value of the pixel, l i represents the loss of each i-th stage, l represents the target loss function, and λ i Indicates the weights of different stages;

[0072] Step 2 (iv), network model training: training the constructed EGE-UNet network on the preprocessed data set, adjusting the parameters of the model to achieve the optimal model, which can accurately segment the nucleus and cytoplasm of cervical cells;

[0073] Step 3: Cell image segmentation model based on qualitative and quantitative evaluation;

[0074] The steps to evaluate the cell image segmentation model in step 3 include:

[0075] Step 3 (a) Qualitative evaluation of segmentation results: Compare the model trained in step 2 with the commonly used segmentation network, and output the performance on the test set as a mask for intuitive evaluation;

[0076] Step 3 (ii), quantitative evaluation of segmentation evaluation indicators: For the predicted segmentation masks generated by the trained model in step 2 and the commonly used segmentation network, the mean intersection over union (MIoU), specificity, and sensitivity are calculated respectively, and quantitative indicators are compared;

[0077] In step 3 (ii), the mean intersection over union (MIoU), specificity, and sensitivity are calculated for the predicted segmentation mask generated by the segmentation network. The evaluation indicators are defined as follows:

[0078] MIoU is the intersection of the predicted area and the actual area divided by the union of the predicted area and the actual area. The IoU under a single category is calculated, and then this algorithm is repeated to calculate the IoU of other categories, and then their average is calculated; it can be expressed as a formula:

[0079]

[0080] In formula (4), k represents the total number of categories, p ii represents the number of pixels that belong to class i and are predicted to be class i, that is, the true number; p ij represents the number of pixels misclassified as category j, p ji represents the number of pixels that should be classified as class j but are mistakenly excluded;

[0081] Based on the confusion matrix, the specificity and sensitivity are defined as follows:

[0082]

[0083]

[0084] In formula (5) and formula (6), TP represents the number of correctly classified positive samples; FP represents the number of negative samples that are incorrectly labeled as positive samples; TN represents the number of correctly classified negative samples; FN represents the number of positive samples that are incorrectly labeled as negative samples;

[0085] Compare the EGE-UNet network with the most commonly used segmentation networks Mask R-CNN and Swin-UNet;

[0086] Step 4: Obtain cervical cell segmentation results based on the evaluated model;

[0087] The steps of obtaining the cervical cell segmentation results based on the evaluated model in step 4 include:

[0088] Step 4 (a), load and call the trained and evaluated EGE-UNet model, and input the tested cervical TCT image into the model;

[0089] Step 4 (ii), performing cell nucleus and cytoplasm segmentation on the specified cervical TCT cell image to obtain a segmentation result;

[0090] The accurate segmentation of cervical cells in this invention can provide detailed cytological features with clinical significance, thus further supporting fine-grained cervical cell identification. Incorporating medical features to guide the model is expected to further improve model performance in the future and improve the interpretability of deep learning in the field of cervical cell-assisted screening;

[0091] The present invention utilizes U-Net, which is a convolutional neural network (CNN) architecture, and is applied to various medical image analysis fields. It also provides an improved U-Net network, the efficient group enhanced UNet (EGE-UNet), which significantly reduces the number of parameters and the computational load by introducing a group aggregation bridging (GAB) module, and achieves excellent performance while being able to better capture the boundary features between the cytoplasm and the nucleus. It can be applied to the segmentation task of the cervical cell nucleus and cytoplasm, and solves the problem that there are certain difficulties in the fine segmentation of the nucleus and cytoplasm of cervical cells with small targets and diverse and complex morphologies.

[0092] The patented method lists four steps in detail, including image preprocessing, segmentation network construction, model evaluation and segmentation result acquisition; each step not only achieves fine segmentation, but also ensures the systematization and standardization of the method; through qualitative and quantitative evaluation, the patented method can ensure the reliability and stability of the segmentation model, so that it is more practical in application; the network structure adopted in the prior art "End-to-End Cervical Cell Segmentation with Fusion of Multi-Scale Convolution" Modern Computer-Volume 29, Issue 2 (Total Issue 770) relatively increases the computational complexity, while the network structure adopted by this patented method is more streamlined and lightweight, and is more suitable for efficient and automated batch screening tasks.

[0093] Embodiment 1:

[0094] A cell segmentation method for cervical TCT images based on improved U-Net is implemented through the following steps:

[0095] Step 1: Preprocessing the cervical cell TCT image;

[0096] The preprocessing method in step 1 is:

[0097] Based on the OpenSlide library, write WSI preprocessing code. Directly divide the image into blocks, extract image blocks from a WSI folder and save them to the specified directory; set 4 threads to perform segmentation operations simultaneously, and get inspiration from the reading method of pathology experts, use scanning column segmentation to obtain folders saved in batches, and each image block (patch) is named in the form of case number_batch number_patch number;

[0098] In this embodiment, in order to adjust the position of the target cells, obtain cell images from multiple angles, expand the data volume, and reduce screening omissions, the preprocessing operation also sets overlapping segmentation; some images in the data set are shown in the following example. Figure 2 As shown, the film production effect is good, the image quality is high, and the cells and background are clear;

[0099] Step 2: Construct a segmentation network based on EGE-UNet to achieve fine segmentation of cervical cell nuclei and cytoplasm;

[0100] The steps for segmenting cervical cells in step 2 are:

[0101] Step 2 (a) Use the pre-processed image as the training data set of the EGE-UNet network to produce the corresponding label data. Use several methods such as random adjustment of brightness, saturation, and contrast to enhance the data of the cervical cell image, and normalize the pixel value range;

[0102] Step 2 (ii), build a specific EGE-UNet network. Select BCE loss and Dice loss as the loss function of the network, and the loss weight λ for each stage i The values ​​of are as follows:

[0103]

[0104] Step 2 (3), training the EGE-UNet network using the training set to obtain a cervical TCT cell segmentation model;

[0105] Step 3: Cell image segmentation model based on qualitative and quantitative evaluation;

[0106] The steps to evaluate the cell image segmentation model in step 3 are:

[0107] Step 3 (a) Compare the model trained in step 2 with the commonly used segmentation network, and output the performance on the test set as a mask for intuitive evaluation, such as Figure 4 As shown;

[0108] Step 3 (ii), for the predicted segmentation mask generated by the model trained in step 2 and the commonly used segmentation network, calculate the mean intersection over union (MIoU), specificity, and sensitivity, respectively, and compare the quantitative indicators;

[0109] In this implementation, for the predicted segmentation mask generated by the segmentation network, the mean intersection over union (MIoU), specificity, and sensitivity are calculated, and the evaluation indicators are defined as follows:

[0110] MIoU is the intersection of the predicted area and the actual area divided by the union of the predicted area and the actual area. The calculated IoU (intersection over union ratio between the predicted mask and the true mask) for a single category is then repeated to calculate the IoU for other categories and then calculate their average. It can be expressed as the formula:

[0111]

[0112] In formula (4), p ij represents the number of pixels that belong to class i but are predicted to be class j, that is, the true number; p ijrepresents the number of pixels misclassified as class j (false positives), p ji represents the number of pixels that should be classified as class j but are mistakenly excluded (false negatives);

[0113] Based on the confusion matrix, the specificity (Spe) and sensitivity (Sen) are defined as follows:

[0114]

[0115] In formula (5) and formula (6), TP represents the number of correctly classified positive samples; FP represents the number of negative samples that are incorrectly labeled as positive samples; TN represents the number of correctly classified negative samples; FN represents the number of positive samples that are incorrectly labeled as negative samples;

[0116] The EGE-UNet network is compared with the most commonly used segmentation networks Mask R-CNN and Swin-UNet. The performance of each model on the test set and the specific index results are shown in Table 1;

[0117] Table 1 Index evaluation results

[0118]

[0119] Depend on Figure 4 Combined with the data in the table, it can be seen that EGE-UNet has the highest MIoU and the best overall segmentation effect; it has the highest specificity, which means that the model rarely mistakenly identifies background or non-target as target objects, reducing the occurrence of false detection; the segmentation algorithm proposed in the present invention has better cervical cell segmentation results;

[0120] Step 4: Obtain cervical cell segmentation results based on the evaluated model;

[0121] The steps of obtaining the cervical cell segmentation results based on the evaluated model in step 4 include:

[0122] Step 4 (a), load and call the trained and evaluated EGE-UNet model, and input the tested cervical TCT image into the model;

[0123] Step 4 (ii), performing cell nucleus and cytoplasm segmentation on the specified cervical TCT cell image to obtain a segmentation result;

[0124] The method adopted by the present invention can efficiently and automatically perform cell segmentation in cervical TCT images, providing more interpretable quantitative data for subsequent cervical cell research, thereby playing a better auxiliary role in intelligent and medical pathology analysis through in vitro cells.

[0125] This patent only targets cervical cells. It starts with collecting cervical cell TCT images (a type of cytopathology image), including the processing of cervical cell whole slide images WSI and the establishment of data sets, to microscopic-level cervical cell segmentation. It is an integrated cervical cell segmentation method. The prior art proposes a broad medical image segmentation method, which does not involve cervical cell TCT images, affecting the segmentation effect on specific cervical cell images.

[0126] It should be noted that in the above embodiments, as long as the technical solutions are not contradictory, they can be arranged and combined, and those skilled in the art can exhaust all possibilities based on the mathematical knowledge of arrangement and combination. Therefore, the present invention will no longer describe the technical solutions after arrangement and combination one by one, but it should be understood that the technical solutions after arrangement and combination have been disclosed by the present invention.

[0127] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A cell segmentation method for cervical TCT images based on improved U-Net, characterized by: The following steps are involved: Step 1: Preprocessing the cervical cell TCT image; Step 2: Build a segmentation network based on EGE-UNet; Step 3: Cell image segmentation model based on qualitative and quantitative evaluation; Step 4: Obtain cervical cell segmentation results based on the evaluated model.

2. According to claim 1, a cell segmentation method for cervical TCT images based on improved U-Net is characterized in that: Step one includes: Step 1 (i) Select the whole slide images of cervical TCT of several cases to establish a data set, sort and classify the WSI images of each case and rename them according to the rule of "category name + case serial number"; Step 1 (2), select the appropriate image size; Step 1 (iii): filter out useless image blocks and standardize the data.

3. According to claim 1, a cell segmentation method for cervical TCT images based on improved U-Net is characterized in that: Step 2 includes: Step 2 (a), divide the data set; Step 2 (ii), build the EGE-UNet network: design a U-shaped symmetrical structure encoder and decoder based on the U-Net architecture, with a short circuit connection between the encoder and the decoder. In EGE-UNet, the GHPA attention mechanism and the GAB module are introduced; Step 2 (3), construct the target loss function; using deep supervision to generate prediction masks of different scales, it is necessary to calculate the loss functions of different stages; Step 2 (iv), network model training: The constructed EGE-UNet network is trained on the preprocessed data set, and the parameters of the model are adjusted to achieve the optimal model, which can accurately segment the nucleus and cytoplasm of cervical cells.

4. The cell segmentation method of cervical TCT images based on improved U-Net according to claim 3, characterized in that: In step 2 (ii), BCE loss and Dice loss are selected as the loss function of the network, and the loss weight λ for each stage i Take the value.

5. According to claim 4, a cell segmentation method for cervical TCT images based on improved U-Net is characterized in that: In step 2 (3), the loss functions at different stages are calculated and defined as follows: In formula (1) and formula (2), Bce represents BCE loss, Dice represents Dice loss, and y represents the label value of the pixel. represents the predicted value of the pixel, l i represents the loss of each i-th stage, l represents the target loss function, and λ i Represents the weights of different stages.

6. The cell segmentation method of cervical TCT images based on improved U-Net according to claim 1, characterized in that: Step three includes: Step 3 (a) Qualitative evaluation of segmentation results: Compare the model trained in step 2 with the commonly used segmentation network, and output the performance on the test set as a mask for intuitive evaluation; Step 3 (ii), quantitative evaluation of segmentation evaluation indicators: For the predicted segmentation masks generated by the trained model in step 2 and the commonly used segmentation network, the mean intersection over union (MIoU), specificity, and sensitivity are calculated respectively, and quantitative indicators are compared.

7. The cell segmentation method of cervical TCT images based on improved U-Net according to claim 6, characterized in that: In step 3 (ii), the mean intersection over union (MIoU), specificity, and sensitivity are calculated for the predicted segmentation mask generated by the segmentation network. The evaluation indicators are defined as follows: MIoU is the intersection of the predicted area and the actual area divided by the union of the predicted area and the actual area. The IoU under a single category is calculated, and then this algorithm is repeated to calculate the IoU of other categories, and then their average is calculated; it can be expressed as a formula: In formula (4), k represents the total number of categories, p ii represents the number of pixels that belong to class i and are predicted to be class i, that is, the true number; p ij represents the number of pixels misclassified as category j, p ji represents the number of pixels that should be classified as class j but are mistakenly excluded; Based on the confusion matrix, the specificity and sensitivity are defined as follows: In formula (5) and formula (6), TP represents the number of correctly classified positive samples; FP represents the number of negative samples that are incorrectly labeled as positive samples; TN represents the number of correctly classified negative samples; FN represents the number of positive samples that are incorrectly labeled as negative samples; Compare the EGE-UNet network with the most commonly used segmentation networks Mask R-CNN and Swin-UNet.

8. The cell segmentation method of cervical TCT images based on improved U-Net according to claim 1 or 3, characterized in that: Step 4 includes: Step 4 (a), load and call the trained and evaluated EGE-UNet model, and input the tested cervical TCT image into the model; Step 4 (ii): perform cell nucleus and cytoplasm segmentation on the specified cervical TCT cell image to obtain a segmentation result.

Images