Cell segmentation model training method, cell segmentation method and system

Through the two-stage training cell segmentation model, noise prediction and pseudo-label correction are used to solve the problem of poor subjectivity and reproducibility of cell segmentation in traditional methods, and efficient and accurate cell segmentation is achieved.

CN118736579BActive Publication Date: 2024-11-15INST OF AUTOMATION CHINESE ACAD OF SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411204424.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-30
Publication Date
2024-11-15
Estimated Expiration
2044-08-30

AI Technical Summary

Technical Problem

Traditional cell segmentation methods have strong subjectivity and poor reproducibility, making it difficult to achieve accurate and efficient cell resolution, especially when the cell types and morphology are diverse.

Method used

Using a training method of cell segmentation model, two-stage training is carried out by obtaining training sample sets and pseudo-labels. The first stage uses noise introduction to perform uncertainty prediction, corrects pseudo-labels, and further training is carried out in the second stage to improve Accuracy and efficiency of cell segmentation.

Benefits of technology

Accurate segmentation of cells is achieved, the efficiency and accuracy of cell segmentation is improved, manual intervention is reduced, and the reproducibility of results is enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118736579B_ABST
    Figure CN118736579B_ABST
Patent Text Reader

Abstract

The present disclosure relates to the field of artificial intelligence and image processing, and provides a training method, a cell segmentation method and a system for a cell segmentation model. The training method includes: obtaining a training sample set and a pseudo-label; based on the training sample set and the pseudo-label, performing a first-stage training on the cell segmentation model to obtain a first-stage cell segmentation model; based on the training result of the first stage, performing uncertainty prediction by introducing noise to obtain an uncertainty prediction result; according to the uncertainty prediction result, correcting the first label in the pseudo-label to obtain a corrected pseudo-label; based on the training sample set and the corrected pseudo-label, performing a second-stage training on the first-stage cell segmentation model to obtain a final cell segmentation model. The present disclosure can solve the problem of difficulty in improving the efficiency and accuracy of cell segmentation, and can train a cell segmentation model that can accurately segment cells, thereby improving the efficiency and accuracy of cell segmentation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the fields of artificial intelligence and image processing, and in particular to a training method for a cell segmentation model, a cell segmentation method and a system. Background Art

[0002] In traditional cell tissue analysis, operators generally use a microscope to observe cell slices with the naked eye to distinguish the cell's morphology, characteristics, etc.

[0003] However, this method is highly subjective and has poor reproducibility, making it difficult to accurately and reliably distinguish cells. In particular, whether it is animal cells or plant cells, there are many types of cells with different shapes and sizes. Manually segmenting cells in images is very challenging, with low work efficiency and accuracy. Summary of the invention

[0004] The present disclosure provides a cell segmentation model training method, a cell segmentation method and a system to at least solve the problem that it is difficult to improve the efficiency and accuracy of cell segmentation in the related art. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of the present disclosure, a training method for a cell segmentation model is provided, the training method comprising: obtaining a training sample set and a pseudo-label for annotating the training sample set, wherein each training sample in the training sample set comprises a cell slice image, and the pseudo-label comprises a first label at a pixel level generated based on a point label; based on the training sample set and the pseudo-label, performing a first-stage training on the cell segmentation model to obtain a first-stage cell segmentation model; based on the training result of the first stage, performing uncertainty prediction by introducing noise to obtain an uncertainty prediction result; according to the uncertainty prediction result, correcting the first label in the pseudo-label to obtain a corrected pseudo-label; based on the training sample set and the corrected pseudo-label, performing a second-stage training on the first-stage cell segmentation model to obtain a final cell segmentation model.

[0006] Optionally, the pseudo-label also includes a second label that is iteratively updated during the training process, wherein the first stage or the second stage training of the cell segmentation model is performed in the following manner: the training sample is input into the cell segmentation model to obtain a predicted segmentation result; the predicted segmentation result is compared with the first label and the second label respectively to obtain a prediction loss to train the cell segmentation model, wherein the prediction loss includes a first loss and a second loss, the first loss represents the loss corresponding to the first label, and the second loss represents the loss corresponding to the second label.

[0007] Optionally, the training of each stage includes multiple rounds, wherein the second label of each round is obtained in the following manner: the second label of the iteration is obtained by weighting and summing the segmentation result predicted in the previous round and the second label of the previous round respectively; the second label of the iteration is corrected by using the first label to obtain the second label of the current round, wherein the second label of the first round is the segmentation result initially predicted by the cell segmentation model, wherein the first label includes a clustering label, and the clustering label includes a first class label, a second class label and a third class label, the first class label represents a cell area, the second class label represents a background area, and the third class label represents other areas except the cell area and the background area, wherein the first label is used to correct the second label of the iteration in the following manner: for each pixel in the cell slice image, in response to the first label indicating that the pixel is the first class label or the second class label, the second label of the iteration is updated to be the same as the first label; in response to the first label indicating that the pixel is the third class label, the second label of the iteration is kept unchanged.

[0008] Optionally, the first label includes a cluster label, wherein, for any cell slice image, the cluster label is determined in the following manner: determine the target pixels in a predetermined area centered on each point label, wherein each point label represents the position of a cell in the cell slice image; for each target pixel, determine the color distance value between the target pixel and each adjacent pixel; determine the standard deviation value of the color distance values ​​for all target pixels; based on the standard deviation value, adjust the color channel value of each pixel in the cell slice image to obtain an adjusted image; and obtain the cluster label by clustering the color features of the pixels in the adjusted image.

[0009] Optionally, the uncertainty prediction result is obtained in the following manner: for the input features of the last layer of the cell segmentation model in the last round of training in the first stage, a Gaussian noise distribution is introduced to obtain a noise introduction feature; the noise introduction feature is input into the last layer of the cell segmentation model to obtain the uncertainty prediction result.

[0010] Optionally, the uncertainty prediction result includes an uncertainty value for each pixel in the cell slice image, the uncertainty value indicating the possibility that the pixel is incorrectly labeled in the pseudo-label, wherein the first label in the pseudo-label is corrected in the following manner: based on the average value of the uncertainty values ​​of the pixels in each cell region, determining a target cell region whose average value is within a preset range; selecting a region corresponding to the target cell region from the prediction results of the first stage, and generating a segmentation prediction mask; using the segmentation prediction mask to replace the corresponding region in the first label, correcting the first label, and obtaining a corrected pseudo-label.

[0011] According to a second aspect of the present disclosure, a cell segmentation method is provided, comprising: acquiring a target image, wherein the target image is a cell slice image; inputting the target image into a cell segmentation model, and using the cell segmentation model to segment cells in the target image to obtain a segmentation result, wherein the cell segmentation model is trained according to the training method of the cell segmentation model described in the present disclosure.

[0012] According to a third aspect of the present disclosure, a training system for a cell segmentation model is provided, the training system comprising: a training acquisition unit, configured to acquire a training sample set and a pseudo-label for annotating the training sample set, wherein each training sample in the training sample set comprises a cell slice image, and the pseudo-label comprises a first label at a pixel level generated based on a point label; a first-stage training unit, configured to perform a first-stage training on the cell segmentation model based on the training sample set and the pseudo-label to obtain a first-stage cell segmentation model; an uncertainty prediction unit, configured to perform uncertainty prediction by introducing noise based on the training result of the first stage to obtain an uncertainty prediction result; a denoising unit, configured to correct the first label in the pseudo-label according to the uncertainty prediction result to obtain a corrected pseudo-label; and a second-stage training unit, configured to perform a second-stage training on the cell segmentation model of the first stage based on the training sample set and the corrected pseudo-label to obtain a final cell segmentation model.

[0013] According to a fourth aspect of the present disclosure, a cell segmentation system is provided, comprising: an image acquisition unit, configured to acquire a target image, wherein the target image is a cell slice image; an image segmentation unit, configured to input the target image into a cell segmentation model, and use the cell segmentation model to segment cells in the target image to obtain a segmentation result, wherein the cell segmentation model is trained according to the training method of the cell segmentation model described in the present disclosure.

[0014] According to a fifth aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing instructions executable by the processor, wherein the instructions executable by the processor, when executed by the processor, cause the processor to execute a training method for a cell segmentation model or a cell segmentation method according to the present disclosure.

[0015] According to a sixth aspect of the present disclosure, a computer-readable storage medium is provided. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the training method of the cell segmentation model or the cell segmentation method according to the present disclosure.

[0016] According to a seventh aspect of the present disclosure, a computer program product is provided, which includes computer executable instructions, and when the computer executable instructions are executed by at least one processor, the training method of the cell segmentation model or the cell segmentation method according to the present disclosure is implemented.

[0017] The technical solution provided by the present disclosure brings at least the following beneficial effects:

[0018] According to the present disclosure, the cell segmentation model can be trained in two stages, and the uncertainty prediction results obtained based on the first stage of training can be used to correct the pseudo-labels, and then the second stage of training can be performed based on the corrected pseudo-labels, so that a cell segmentation model that can accurately segment cells can be trained, thereby improving the efficiency and accuracy of cell segmentation.

[0019] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0021] Figure 1 is a schematic flowchart of a method for training a cell segmentation model according to an exemplary embodiment of the present disclosure.

[0022] Figure 2 is a schematic diagram of two training stages in a training method for a cell segmentation model according to an exemplary embodiment of the present disclosure.

[0023] Figure 3 is a schematic flowchart of the steps of determining cluster labels in a training method for a cell segmentation model according to an exemplary embodiment of the present disclosure.

[0024] Figure 4It is a schematic flowchart of dynamically obtaining pseudo labels in a training method of a cell segmentation model according to an exemplary embodiment of the present disclosure.

[0025] Figure 5 It is a schematic block diagram of dynamically obtaining pseudo labels in a training method for a cell segmentation model according to an exemplary embodiment of the present disclosure.

[0026] Figure 6 is a schematic diagram of a cell segmentation model in a method for training a cell segmentation model according to an exemplary embodiment of the present disclosure.

[0027] Figure 7 is a schematic diagram of an example process of a method for training a cell segmentation model according to an exemplary embodiment of the present disclosure.

[0028] Figure 8 is a schematic flowchart of a cell segmentation method according to an exemplary embodiment of the present disclosure.

[0029] Fig. 9 is a schematic block diagram of a training system for a cell segmentation model according to an exemplary embodiment of the present disclosure.

[0030] Fig.10 is a schematic block diagram of a cell segmentation system according to an exemplary embodiment of the present disclosure.

[0031] Fig.11 is a block diagram of an electronic device according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0032] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the present application will clearly and completely describe the technical solutions in the embodiments of the present disclosure in conjunction with the accompanying drawings.

[0033] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0034] It should be noted that the phrase "at least one of the items" in the present disclosure includes three types of parallel situations: "any one of the items", "a combination of any number of the items", and "all of the items". For example, "including at least one of A and B" includes the following three types of parallel situations: (1) including A; (2) including B; (3) including A and B. Another example is "executing at least one of step 1 and step 2" which means the following three types of parallel situations: (1) executing step 1; (2) executing step 2; (3) executing step 1 and step 2.

[0035] As mentioned above, in the related art, there is a problem that it is difficult to accurately and reliably distinguish cells.

[0036] Cells are the basic structural and functional units of organisms. The cells mentioned here can be cells of any organism, such as animal cells, plant cells, etc. Since cells are extremely small, their shapes can generally only be observed under a microscope, and there are many types of cells with various shapes.

[0037] Taking pathological image analysis in pathology as an example, the traditional pathological analysis method is mainly for pathologists to observe the slices with the naked eye under a microscope, but this method is highly subjective and the diagnostic results are not reproducible. With the development of scanning technology, it has become a trend to convert tissue pathology slices into digital full-scan pathology images (Whole Slide Images, WSIs) and use computers for auxiliary analysis. Cell segmentation in histopathological images is an important basis and key step for cancer diagnosis and immune response prediction. Accurate cell segmentation can provide phenotypic information for clinical diagnosis, and researchers can analyze the development of tumor diseases at the cellular level based on this phenotypic information. However, there are tens of thousands of cells in WSI, and these cells have many categories and different morphologies and sizes. Manual segmentation of cells in pathological images is very challenging. Therefore, how to automatically segment cells in pathological images is crucial.

[0038] Based on traditional image segmentation algorithms, such as watershed, color thresholding, superpixel, and graph cutting, the segmentation effect is poor for crowded or contacting cells, which can easily lead to under-segmentation or over-segmentation. The deep learning-based method can avoid morphological operations and complex preprocessing operations, and achieve excellent segmentation results. Methods based on convolutional neural networks (CNNs), such as fully convolutional networks (FCNs), U-Net, and DeepLab series, have demonstrated good performance in natural image segmentation. However, these fully supervised deep learning methods rely on a large amount of accurate pixel-level annotations, which requires a huge amount of annotation work and requires professionals to complete the annotation.

[0039] In view of the above, exemplary embodiments of the present disclosure propose a training method for a cell segmentation model, a cell segmentation method, a training system for a cell segmentation model, a cell segmentation system, an electronic device, a computer-readable storage medium, and a computer program product, which can solve or at least alleviate the above problems.

[0040] In a first aspect of an exemplary embodiment of the present disclosure, a method for training a cell segmentation model is provided.

[0041] The training method of the cell segmentation model according to the exemplary embodiment of the present disclosure can be applied to scenarios where users interact with software. For example, the software can be loaded on the user terminal, and the user can input model training instructions on the user terminal. The user terminal can obtain the final cell segmentation model by executing the training method of the cell segmentation model according to the exemplary embodiment of the present disclosure.

[0042] Specifically, the user terminal may obtain a training sample set and a pseudo label for annotating the training sample set, wherein each training sample in the training sample set includes a cell slice image, and the pseudo label includes a first label at a pixel level generated based on a point label.

[0043] The user terminal can also perform the first stage training of the cell segmentation model based on the training sample set and the pseudo-label to obtain the cell segmentation model of the first stage. The user terminal can also perform uncertainty prediction based on the training result of the first stage by introducing noise to obtain the uncertainty prediction result. The user terminal can also correct the first label in the pseudo-label according to the uncertainty prediction result to obtain the corrected pseudo-label. The user terminal can also perform the second stage training of the cell segmentation model of the first stage based on the training sample set and the corrected pseudo-label to obtain the final cell segmentation model.

[0044] The above-mentioned user terminal can be, for example, a tablet computer, a laptop computer, a digital assistant, a wearable device, etc. However, the implementation scenario of the above method is only an example scenario. The training method of the cell segmentation model according to the exemplary embodiment of the present disclosure can also be applied to other application scenarios. For example, the user can request a training model from the server through the network on the user terminal (for example, a mobile phone, a desktop computer, a tablet computer, etc.). The server can complete the request by executing the training method of the cell segmentation model according to the exemplary embodiment of the present disclosure. Here, the server can be an independent server, a server cluster, a cloud computing platform, or a virtualization center.

[0045] According to the training method of the cell segmentation model of the exemplary embodiment of the present disclosure, the segmentation algorithm of digital pathology and deep learning can be used to train a cell segmentation model that can accurately segment cells, realize automatic identification of cells from images, and generate the final segmentation result, intuitively display the spatial distribution inside the tissue, and improve the efficiency and accuracy of cell segmentation.

[0046] The following will refer to Figures 1 to 7 An example of a method for training a cell segmentation model according to an embodiment of the present disclosure is described. Figure 1 As shown, the training method of the cell segmentation model may include the following steps:

[0047] In step S110, a training sample set and pseudo labels for annotating the training sample set may be obtained.

[0048] Here, each training sample in the training sample set may include a cell slice image, and the pseudo label may include a first label at a pixel level generated based on a point label.

[0049] As an example, the cell slice image may be, but is not limited to, a cell pathology slice image in medicine, for example, a plant cell slice image in plant research, a fungus image in eukaryotic research, and the like.

[0050] When preparing training samples, the cell slice image can be stained to obtain a slice stained by, for example, hematoxylin-eosin staining (H&E), and then the H&E stained slice is digitized to obtain an H&E stained image. As an example, the staining process used here can be, but is not limited to, using hematoxylin staining solution to dye the chromatin in the cell and the nucleic acid in the cytoplasm purple-blue, and using eosin staining solution to dye the components in the cytoplasm and the extracellular matrix red.

[0051] For training samples, professionals such as pathologists and plant researchers can perform pixel-level annotation for each image, outline each cell in the image in detail, and binarize the outlined labels to obtain a segmentation mask for each cell.

[0052] Point labels can be generated based on pixel-level annotation. For example, the maximum horizontal length and maximum vertical length of each cell can be calculated based on the segmentation mask, and a line can be drawn at the maximum distance (or length). The intersection of the maximum horizontal length line and the maximum vertical length line can be used as the centroid of the cell, and the position of the centroid can be marked as the point label of the cell.

[0053] Based on the above method, a data set of cell slice images can be prepared. For example, a training sample set, a validation set and a test set can be divided from the prepared data set, or a training sample set, a validation set and a test set can be prepared separately.

[0054] Here, the original point label can only provide the location information of the cell but not the shape and boundary information of the cell, so the point label can be converted into a usable pixel-level pseudo label, that is, the first label mentioned above, which can label the pixels related to the cell center point and the cell boundary.

[0055] As an example, the first label may be generated based on the location information of the cell center point represented by the point label according to the method of generating pixel-level labels in the related art.

[0056] According to an embodiment of the present disclosure, the first label may include, but is not limited to, a Voronoi label and / or a cluster label, such as a K-means cluster label.

[0057] In one example, for Voronoi labels, a boundary corresponding to each centroid point can be generated based on each centroid point marked by the point label. As an example, the boundary can be a Voronoi boundary, which can be formed by connecting a polygonal area composed of perpendicular bisectors of two adjacent centroid points, thereby forming a Voronoi boundary around each centroid point. For example, Figure 2 As shown, the Voronoi boundary can divide the image into multiple Voronoi cells, and each polygonal Voronoi cell has only one centroid point. In addition, the pixels in each Voronoi cell are closer to the centroid point in the cell than any other point annotation.

[0058] As an example, the centroid point in each cell can be used as a positive pixel (such as a cell), the boundary of the cell can be used as a negative pixel (such as the background), and the remaining pixels can be ignored in training. The Voronoi boundary can separate each cell and avoid crowded and overlapping cells that are difficult to distinguish.

[0059] In another example, for K-means clustering labels, such as Figure 3 As shown in the figure, for any cell slice image, the cluster label is determined by:

[0060] In step S310, target pixels within a predetermined area centered on each point label may be determined.

[0061] Here, as described above, each point label can characterize the position of a cell in the cell slice image.

[0062] The predetermined area can be set according to actual needs, for example, according to the size of the analyzed cell, the pixel resolution of the image itself, etc. The predetermined area can be, for example, an area that can include the entire cell. For example, the predetermined area can be L×L small areas (also called "patch") centered on the point label, and L can be set according to actual needs.

[0063] In step S320, for each target pixel, a color distance value between the target pixel and each adjacent pixel may be determined.

[0064] Here, the neighboring pixels refer to the pixels in a 9×9 pixel area surrounding the target pixel. For example, the color distance value from each pixel in each patch to the neighboring pixels in its 9×9 area can be calculated.

[0065] The color distance value can be a value that quantifies the difference between two colors. It can measure the similarity or difference between two colors by calculating the Euclidean distance in the color space. For example, taking the RGB color channel as an example, the RGB value of a pixel color can be composed into a feature vector ,in, Represents pixels The values ​​of the three RGB color channels, Represents pixels The distance value to the nearest point label.

[0066] In step S330 , a standard deviation value of the color distance values ​​for all target pixels may be determined.

[0067] When the color distance value between each target pixel and each adjacent speed limit is determined, the standard deviation value of all the obtained color distance values ​​can be calculated for the entire image area. .

[0068] In step S340, the color channel value of each pixel in the cell slice image may be adjusted based on the standard deviation value to obtain an adjusted image.

[0069] For example, taking the RGB color channel as an example, for the input RGB image Pixels in , whose color channel value is , based on the standard deviation , the color channel value can be adjusted by the following formula (1):

[0070] (1)

[0071] in, Represents the adjusted color channel value, represents a global scaling term, which may be a constant, for example, which may be determined by empirical methods or set according to actual needs.

[0072] The color channels here are only examples, and other color channels may also be used for calculation, and the principle is similar to the above method.

[0073] In step S350, cluster labels may be obtained by clustering color features of pixels in the adjusted image.

[0074] In this step, clustering can be performed based on the K-means clustering algorithm. Here, the cluster labels can include first-class labels, second-class labels, and third-class labels. The first-class labels can represent cell regions, the second-class labels can represent background regions, and the third-class labels can represent other regions except cell regions and background regions. For example, the first-class labels can indicate that the class with the largest overlapping area with the point label is the cell region, the second-class labels can indicate that the class with the smallest overlap is the background region, and the third-class labels can indicate other regions, which are not involved in calculating the loss during training.

[0075] Based on the above method, the color difference between cells and background pixels can be considered for clustering, and in the clustering process, the color channel value of each pixel is adjusted based on the global standard deviation value to obtain the image adaptive feature representation (such as the above ), thereby avoiding over-learning of the specific color distribution of a certain type of image itself during the training process, making the clustering process image-adaptive, adapting to the color distribution of each image, and improving the generalization ability of the model.

[0076] In addition, preferably, when the Voronoi labels are used, the complete shape information of the cells may not be provided well, so the boundary and shape information of the cells can be better supplemented by introducing other pseudo labels such as cluster labels.

[0077] Return to reference Figure 1 In step S120, the cell segmentation model can be trained in the first stage based on the training sample set and the pseudo labels to obtain a cell segmentation model in the first stage.

[0078] In one example, based on the training sample set and the first label, the model can be used to make predictions and calculate the loss to perform the first stage of training the model.

[0079] In another example, according to an embodiment of the present disclosure, in order to further improve the training effect, in addition to using the first label as the static pseudo label, a second label as a dynamically updated pseudo label can also be introduced. Since the quality of the second label will gradually improve during the training process, it can help improve the training effect.

[0080] Specifically, the pseudo-label may also include a second label that is iteratively updated during the training process. As an example, the first or second stage training of the cell segmentation model may be performed in the following manner: the training sample is input into the cell segmentation model to obtain a predicted segmentation result; the predicted segmentation result is compared with the first label and the second label respectively to obtain the prediction loss to train the cell segmentation model.

[0081] Here, the prediction loss may include a first loss and a second loss, the first loss may represent a loss corresponding to the first label, and the second loss may represent a loss corresponding to the second label.

[0082] Taking the example of the first label including Voronoi label and cluster label, supervised training can be performed under Voronoi label, cluster label and second label respectively, and the prediction loss is It can be expressed by the following formula (2):

[0083] (2)

[0084] in, represents the loss calculated based on the Voronoi labels, represents the loss calculated based on the cluster labels, represents the loss calculated based on the second label, Indicates the second loss Specifically, considering that the quality of the second label generated by iterative updating will gradually improve during the training process, the loss weight of the label can be increased accordingly. As an example, the weight can be set according to actual needs; as another example, considering the iterative update characteristics of the second label, the weight can be associated with the number of iterations of the second label to make the weighting of the second loss more reasonable. For example, during the training process, the weight The value of gradually increases. For example, the weight can be determined by the following formula (3): :

[0085] (3)

[0086] in, represents the total number of current epochs (i.e. training rounds), represents the total number of training epochs, It is a constant and can be set according to actual needs.

[0087] It should be noted that, although the above description is made taking the first stage training as an example, the above training process can be applied to both the first stage training and the second stage training, and the second stage training process is similar to the process described above.

[0088] As described above, the second label can be a pseudo-label that is iteratively updated. For example, the training of each stage (the first stage or the second stage) can include multiple rounds, and the second label can be determined based on the prediction results of the model obtained in each round of training. As the number of training rounds increases, the prediction of the model becomes more and more accurate, and the annotation of the second label becomes more and more accurate.

[0089] As an example, Figure 4 As shown, the second label of each round can be obtained in the following manner: step S410, the segmentation result predicted in the previous round and the second label of the previous round can be weighted and summed to obtain the iterative second label; step S420, the iterative second label can be corrected using the first label to obtain the second label of the current round. Here, the second label of the first round can be the segmentation result initially predicted by the cell segmentation model.

[0090] In step S410, Figure 5 As shown, an exponential moving average strategy (EMA) can be used to iterate the second label using the segmentation results of each round of prediction to optimize the second label used in the next round of prediction. Specifically, the following formula (4) can be used:

[0091] (4)

[0092] in, The second label representing the current round; Indicates the result after the last exponential moving average, that is, the second label of the last round; Represents the predicted output of the network obtained in the previous round of prediction; Represents the weight of EMA, which can be a given value, such as 0.1; represents the time step of the moving average, which means how many steps the prediction is averaged, that is, Indicates round.

[0093] It should be noted that, although the EMA method is described here for iteration, it is not limited to this, and other methods may also be used for iteration.

[0094] In step S420, the iterative second label can be corrected using the first label in the following manner: for each pixel in the cell slice image, in response to the first label indicating that the pixel is a first-category label or a second-category label, the iterative second label is updated to be the same as the first label; in response to the first label indicating that the pixel is a third-category label, the iterative second label is kept unchanged.

[0095] As an example, the average prediction can be used to determine the labels for the ignored regions of the cluster labels:

[0096] (5)

[0097] in, Indicates the second label of the current round Middle The label of pixels, Indicates The cluster labels of pixels, , the cluster label values ​​0, 1, and 2 represent the background, foreground, and ignored areas, respectively.

[0098] The second label after adaptive clustering label correction can be obtained by the above formula (5). Based on the second label, the Kullback-Leibler (KL) divergence can be used to calculate the joint training loss. For example, the second loss can be determined by the following formula (6):

[0099] (6)

[0100] in, represents the total number of pixels in the image that are not ignored, Represents the prediction results output by the cell segmentation model, such as a prediction probability map, Indicates i However, the method of calculating the second loss is not limited to the above formula (6), for example, other joint training loss calculation methods can also be used.

[0101] According to an embodiment of the present disclosure, although there are many annotated pixels in the cluster label, the network can better learn some cell areas near the point annotation based only on the cluster label. However, if the cluster label prediction is wrong, the network may be misled by such erroneous label information, thereby affecting the accuracy of the prediction. Taking this into account, in the above process, through iterative methods such as the exponential moving average strategy (EMAC), the interference of erroneous cluster labels can be minimized to further improve the accuracy of the prediction. In this way, according to this pseudo-label correction strategy, by adopting the method of iterative second labels and introducing cluster labels for auxiliary correction, those areas ignored by cluster labels can be found, and the predicted second labels can be calculated regularly to obtain more robust pseudo-labels.

[0102] In step S120 , in the first stage of training, two pseudo labels (eg, Voronoi labels and cluster labels) generated based on point labels and a pseudo label (eg, second label) generated by a pseudo label correction strategy are used to supervise the training of the model.

[0103] As an example, Figure 6 As shown in the figure, a U-Net-like encoder-decoder network can be used as a cell segmentation model, wherein the convolutional layer of ResUNet34 can be used to replace the encoder part of U-Net, and a multi-scale feature fusion (MFF) module is introduced in the decoder part, wherein the MFF module can include a convolutional block attention module (CBAM).

[0104] Here, we give example parameters for the training scheme and hyperparameter settings chosen in the first stage of training:

[0105] (1) Use the Stochastic Gradient Descent (SGD) algorithm as the optimization function when updating model parameters;

[0106] (2) Use Mini-Batch Gradient Descent (MBGD) as the training scheme;

[0107] (3) Set the mini-batch size of the data subset (i.e., batch-size) of the hyperparameter to 8;

[0108] (4) The number of training rounds is 100;

[0109] (5) The learning rate is set to 1e-3 (i.e., );

[0110] (6) Use the cross entropy loss function and the KL divergence loss function.

[0111] In step S130, based on the training result of the first stage, uncertainty prediction can be performed by introducing noise to obtain an uncertainty prediction result.

[0112] Here, uncertainty prediction may refer to predicting the uncertainty of a first label. The uncertainty of the first label may include, for example, accuracy, completeness, consistency, etc. of the label. Labels with high uncertainty may require more precise annotations.

[0113] As an example, the uncertainty prediction result can be obtained in the following way: for the input features of the last layer of the cell segmentation model in the last round of training in the first stage, a Gaussian noise distribution is introduced to obtain a noise introduction feature; the noise introduction feature is input into the last layer of the cell segmentation model to obtain an uncertainty prediction result.

[0114] Specifically, during the first stage of training, we can focus on the uncertainty in the data to guide the second stage of training, making the training labels in the second stage more accurate.

[0115] For example, in the first phase of training, The feature vector output before the last classification layer (such as the softmax layer) of pixels Place a Gaussian noise distribution on the input features of the last layer of the model, and the noise introduced features For example, it can be expressed by the following formula (7):

[0116] (7)

[0117] in, W represents the parameter matrix of the model, represents the simulated Gaussian noise variance, which is regarded as data uncertainty. A sampling method such as Monte Carlo sampling can be used to select a predetermined number of samples to simulate the Gaussian noise variance. Here, , T is the sampling size, Indicates that for t Normal distribution of sample , I is a constant, for example, it can be 1.

[0118] Combined with the above formula (7), the noise is introduced into the feature Input to the last layer of the cell segmentation model to obtain the uncertainty prediction result, for example, Uncertainty prediction results for pixels It can be expressed by the following formula (8):

[0119] (8)

[0120] In the above manner, uncertainty prediction can be performed to find pixels with high uncertainty in the first label, thereby guiding the optimization of the first label.

[0121] In step S140, the first label in the pseudo label may be corrected according to the uncertainty prediction result to obtain a corrected pseudo label.

[0122] In this step, the first label can be optimized to use the optimized more accurate pseudo label for the second stage of training.

[0123] As an example, the uncertainty prediction result includes an uncertainty value for each pixel in a cell slice image, which indicates the probability that the pixel is incorrectly labeled in the pseudo-label.

[0124] In this example, the first label in the pseudo-label can be corrected in the following ways: based on the average value of the uncertainty values ​​of the pixels in each cell area, the target cell area whose average value is within a preset range can be determined; the area corresponding to the target cell area can be selected from the prediction results of the first stage to generate a segmentation prediction mask; the segmentation prediction mask can be used to replace the corresponding area in the first label, and the first label can be corrected to obtain a corrected pseudo-label.

[0125] Specifically, the uncertainty prediction results can be represented by an uncertainty estimation map, where the average uncertainty value of pixels can be calculated in an area with a radius of a predetermined number of pixels (e.g., 10 pixels) centered on each cell, and the average uncertainty value is regarded as the uncertainty of the cell. Here, cell regions with high uncertainty values ​​may have large noise in the Voronoi labels or adaptive clustering labels, so such regions can be annotated more accurately.

[0126] As an example, the uncertainty values ​​of all cells can be sorted, and the first predetermined number (for example, the first 5%) of cell regions can be selected as target cell regions, and the target cell regions can be replaced with corresponding segmentation prediction masks for annotation, thereby providing more accurate pseudo labels. Here, the segmentation prediction mask can be a region corresponding to the target cell region selected from the final prediction result of the first stage, and the segmentation prediction mask can be used to replace the corresponding region in the first label, and the first label can be corrected to perform the second stage of training based on more accurate pseudo labels to improve the training effect.

[0127] For example, for Voronoi labels, the points in the target cell region can be replaced with the corresponding masks in the segmentation prediction mask, and other pixels in the Voronoi cells in the target cell region can be marked as background; for cluster labels, the cell mask in the segmentation prediction mask can be copied to the corresponding region. The optimized labels obtained in this way combine the information of point labels and mask annotations, and the spliced ​​Voronoi labels and cluster labels can be sent to the second stage for training.

[0128] In step S150, the cell segmentation model of the first stage may be trained in the second stage based on the training sample set and the corrected pseudo-labels to obtain a final cell segmentation model.

[0129] In this step, the training sample set and the corrected pseudo-labels can be used to train the model obtained in the first stage for the second stage. The corrected first labels obtained by using the uncertainty denoising strategy are sent to the second stage for model training. In the second stage, the second labels generated by the pseudo-label correction strategy are also used to jointly supervise the training of the segmentation network to obtain the final segmentation prediction result of the network. Figure 2 For example, the pseudo labels trained in the first stage can be replaced by the corrected pseudo labels (including the optimized Voronoi labels and the optimized cluster labels) to carry out the second stage of training.

[0130] The first phase of training has been described above. The second phase of training is similar to the above process. The difference is that uncertainty prediction is not required in the second phase, and the model training process is the same as the first phase. Therefore, the second phase of training will not be described here.

[0131] As an example, here are the sample parameters of the training scheme and hyperparameter settings selected in the second stage of training:

[0132] (1) Use the SGD algorithm as the optimization function when updating model parameters;

[0133] (2) Use mini-batch gradient descent (MBGD) as the training scheme;

[0134] (3) Set the hyperparameter mini-batch (batch-size) size to 8;

[0135] (4) The epoch size is set to 150;

[0136] (5) The learning rate is set to 1e-3 (i.e., );

[0137] (6) Use the cross entropy loss function and the KL divergence loss function.

[0138] Figure 7 FIG. 2 shows an example process of a method for training a cell segmentation model according to an exemplary embodiment of the present disclosure. Figure 7 As shown, taking the cell segmentation for H&E staining image as an example, in step S710, H&E staining image data can be obtained; in step S720, the obtained data can be divided into a training set, a validation set and a test set, which are used for model training, validation and testing, respectively. In step S730, a pseudo-label can be generated for the H&E staining image data (for example, the data in the training set), and the pseudo-label can include, for example, the first label described above. In step S740, the first stage of training can be performed, and in step S750, the pseudo-label can be denoised based on the training data of the first stage to obtain a corrected pseudo-label, so as to perform the second stage of training in step S760 to obtain the final cell segmentation model.

[0139] By using a cell segmentation model trained according to the embodiments of the present disclosure, it is possible to utilize digital pathology and deep learning segmentation algorithms to automatically identify cells within a tissue from a cell slice image and generate a final cell segmentation result, which intuitively displays the spatial distribution within the tissue, and helps auxiliary professionals to perform subsequent analysis and judgment based on the segmentation results, for example, it can help doctors grade patients and perform prognostic analysis.

[0140] According to the embodiments of the present disclosure, a solution for training a cell segmentation model using a weakly supervised learning method can be provided. For segmentation problems, weakly supervised learning only requires a small amount of rough annotations for training, for example, weak annotations such as image-level annotations, scribble annotations, bounding box annotations, and point annotations can be used. In the embodiments of the present disclosure, for histopathological images, since the cells are very small and densely distributed, point annotations (or point labels) are more suitable for weakly supervised cell segmentation to obtain more effective and accurate segmentation results.

[0141] In the training method according to the present disclosure, based on the exponential moving average strategy (EMAC) of cluster label correction, the interference of erroneous cluster labels can be minimized, and cluster labels are introduced for auxiliary correction, so as to determine the area ignored by the cluster labels, and then the predicted EMA can be calculated regularly to obtain more robust pseudo labels. The pseudo labels generated by such pseudo label correction strategy can be used together with Voronoi labels and adaptive cluster labels to supervise network training.

[0142] This deep learning method is used to build a model to help segment cells in images, and it can be displayed intuitively. It can segment any number of cell types, helping professionals analyze the spatial heterogeneity of cell structures inside tissues, thereby facilitating prognostic analysis in the medical field and formulating more appropriate treatment plans. It can achieve cell segmentation of images using point labels, generate pixel-level segmentation results, and intuitively display the spatial distribution of internal tissue structures.

[0143] In a second aspect of an exemplary embodiment of the present disclosure, a cell segmentation method is provided. The cell segmentation method can be applied to a scenario where a user interacts with software, for example, the software can be loaded on a user terminal, the user can input a cell segmentation instruction on the user terminal, and the user terminal can obtain the final segmented cells by executing the cell segmentation method according to the exemplary embodiment of the present disclosure.

[0144] like Figure 8As shown, the cell segmentation method may include: step S810, a target image may be acquired, wherein the target image is a cell slice image; step S820, the target image may be input into a cell segmentation model, and the cell segmentation model may be used to segment the cells in the target image to obtain a segmentation result, wherein the cell segmentation model is trained according to the training method of the cell segmentation model of the embodiment of the present disclosure.

[0145] The specific steps and details of the cell segmentation method according to the exemplary embodiment of the present disclosure can be implemented and understood similarly to the training method of the cell segmentation model described in the first aspect above, and accordingly, the cell segmentation method has the same or similar beneficial effects as the training method of the cell segmentation model described above, which will not be repeated here.

[0146] In a third aspect of the exemplary embodiments of the present disclosure, a training system for a cell segmentation model is provided, such as Fig. 9 As shown, the training system includes a training acquisition unit 910, a first-stage training unit 920, an uncertainty prediction unit 930, a denoising unit 940 and a second-stage training unit 950.

[0147] The training acquisition unit 910 is configured to acquire a training sample set and a pseudo label for annotating the training sample set, wherein each training sample in the training sample set includes a cell slice image, and the pseudo label includes a first label at a pixel level generated based on a point label.

[0148] The first stage training unit 920 is configured to perform first stage training on the cell segmentation model based on the training sample set and the pseudo labels to obtain a first stage cell segmentation model.

[0149] The uncertainty prediction unit 930 is configured to perform uncertainty prediction based on the training result of the first stage by introducing noise to obtain an uncertainty prediction result.

[0150] The denoising unit 940 is configured to correct the first label in the pseudo label according to the uncertainty prediction result to obtain a corrected pseudo label.

[0151] The second-stage training unit 950 is configured to perform a second-stage training on the cell segmentation model of the first stage based on the training sample set and the corrected pseudo-labels to obtain a final cell segmentation model.

[0152] As an example, the pseudo-label also includes a second label that is iteratively updated during the training process, wherein the first-stage training unit 920 is configured to perform the first-stage training on the cell segmentation model in the following manner or the second-stage training unit 950 is configured to perform the second-stage training on the cell segmentation model in the following manner: inputting the training sample into the cell segmentation model to obtain a predicted segmentation result; and obtaining a prediction loss by comparing the predicted segmentation result with the first label and the second label respectively to train the cell segmentation model, wherein the prediction loss includes a first loss and a second loss, the first loss representing the loss corresponding to the first label, and the second loss representing the loss corresponding to the second label.

[0153] As an example, the training of each stage includes multiple rounds, wherein the first-stage training unit 920 or the second-stage training unit 950 is configured to obtain the second label of each round in the following manner: by weighting the predicted segmentation result of the previous round and the second label of the previous round and summing them up to obtain the iterative second label; using the first label to correct the iterative second label to obtain the second label of the current round, wherein the second label of the first round is the segmentation result initially predicted by the cell segmentation model, wherein the first label includes a clustering label, the clustering label includes a first category label, a second category label and a third category label, the first category label represents the cell area, the second category label represents the background area, and the third category label represents other areas except the cell area and the background area.

[0154] As an example, the first-stage training unit 920 or the second-stage training unit 950 is configured to correct the iterative second label using the first label in the following manner: for each pixel in the cell slice image, in response to the first label indicating that the pixel is a first-category label or a second-category label, the iterative second label is updated to be the same as the first label; in response to the first label indicating that the pixel is a third-category label, the iterative second label is kept unchanged.

[0155] As an example, the first label includes a cluster label, wherein, for any cell slice image, the training acquisition unit 910 is further configured to determine the cluster label in the following manner: determine the target pixels in a predetermined area centered on each point label, wherein each point label represents the position of a cell in the cell slice image; for each target pixel, determine the color distance value between the target pixel and each adjacent pixel; determine the standard deviation value of the color distance values ​​for all target pixels; based on the standard deviation value, adjust the color channel value of each pixel in the cell slice image to obtain an adjusted image; and obtain the cluster label by clustering the color features of the pixels in the adjusted image.

[0156] As an example, the uncertainty prediction unit 930 is configured to obtain an uncertainty prediction result in the following manner: for the input features of the last layer of the cell segmentation model in the last round of training in the first stage, a Gaussian noise distribution is introduced to obtain a noise introduction feature; the noise introduction feature is input into the last layer of the cell segmentation model to obtain an uncertainty prediction result.

[0157] As an example, the uncertainty prediction result includes an uncertainty value for each pixel in the cell slice image, and the uncertainty value indicates the possibility that the pixel is incorrectly labeled in the pseudo-label, wherein the denoising unit 940 is configured to correct the first label in the pseudo-label in the following manner: based on the average value of the uncertainty values ​​of the pixels in each cell region, determine the target cell region whose average value is within a preset range; select the region corresponding to the target cell region from the prediction results of the first stage, and generate a segmentation prediction mask; use the segmentation prediction mask to replace the corresponding region in the first label, correct the first label, and obtain a corrected pseudo-label.

[0158] Regarding the device in the above-mentioned embodiment, the specific manner in which each unit performs the operation has been described in detail in the embodiment of the method. Each unit in the training device for the cell segmentation model can execute the corresponding steps in the method according to the training method for the cell segmentation model in the method embodiment of the first aspect above, and will not be elaborated here.

[0159] In a fourth aspect of the exemplary embodiments of the present disclosure, a cell segmentation system is provided, such as Fig.10 As shown, the cell segmentation system includes an image acquisition unit 1010 and an image segmentation unit 1020 .

[0160] The image acquisition unit 1010 is configured to acquire a target image, wherein the target image is a cell slice image. The image segmentation unit 1020 is configured to input the target image into a cell segmentation model, and use the cell segmentation model to segment the cells in the target image to obtain a segmentation result, wherein the cell segmentation model is trained according to the training method of the cell segmentation model described in the embodiment of the present disclosure.

[0161] Regarding the device in the above embodiment, the specific manner in which each unit performs the operation has been described in detail in the embodiment of the method, and will not be elaborated here.

[0162] Fig.11 is a block diagram of an electronic device according to an exemplary embodiment of the present disclosure. Fig.11As shown, the electronic device 10 includes a processor 101 and a memory 102 for storing processor executable instructions. Here, when the processor executable instructions are executed by the processor, the processor is prompted to execute the training method of the cell segmentation model or the cell segmentation method as described in the above exemplary embodiments.

[0163] As an example, the electronic device 10 is not necessarily a single device, but may be any collection of devices or circuits that can execute the above instructions (or instruction sets) individually or in combination. The electronic device 10 may also be part of an integrated control system or system manager, or may be configured as a server that is interconnected with a local or remote (e.g., via wireless transmission) interface.

[0164] In the electronic device 10, the processor 101 may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller or a microprocessor. As an example and not a limitation, the processor 101 may also include an analog processor, a digital processor, a microprocessor, a multi-core processor, a processor array, a network processor, etc.

[0165] The processor 101 may execute instructions or codes stored in the memory 102, wherein the memory 102 may also store data. Instructions and data may also be sent and received over a network via a network interface device, wherein the network interface device may employ any known transmission protocol.

[0166] The memory 102 may be integrated with the processor 101, for example, by placing RAM or flash memory within an integrated circuit microprocessor or the like. In addition, the memory 102 may include a separate device, such as an external disk drive, a storage array, or any other storage device that can be used by a database system. The memory 102 and the processor 101 may be operatively coupled, or may communicate with each other, such as through an I / O port, a network connection, etc., so that the processor 101 can read files stored in the memory 102.

[0167] In addition, the electronic device 10 may also include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, a mouse, a touch input device, etc.) All components of the electronic device 10 may be connected to each other via a bus and / or a network.

[0168] In an exemplary embodiment, a computer-readable storage medium may also be provided. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the training method of the cell segmentation model or the cell segmentation method as described in the above exemplary embodiments. The computer-readable storage medium may be, for example, a memory including instructions. Optionally, the computer-readable storage medium may be: a read-only memory (ROM), a random access memory (RAM), a random access programmable read-only memory (PROM), an electrically erasable programmable read-only memory (EEPROM), a dynamic random access memory (DRAM), a static random access memory (SRAM), a flash memory, a non-volatile memory, a CD-ROM, a CD-R, a CD+R, a CD-RW, a CD+RW, a DVD-ROM, a DVD-R, a DVD+R, a DVD-RW, a DVD+RW, a DVD-RAM, a BD-ROM, a BD-R, a BD-R LTH, BD-RE, Blu-ray or optical disk storage, hard disk drive (HDD), solid state drive (SSD), card storage (such as, multimedia card, secure digital (SD) card or extreme digital (XD) card), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid state disk and any other device, any other device is configured to store computer programs and any associated data, data files and data structures in a non-transitory manner and provide the computer programs and any associated data, data files and data structures to a processor or computer so that the processor or computer can execute the computer program. The computer program in the above-mentioned computer-readable storage medium can be run in an environment deployed in a computer device such as a client, a host, an agent device, a server, etc. In addition, in one example, the computer program and any associated data, data files and data structures are distributed on a networked computer system, so that the computer program and any associated data, data files and data structures are stored, accessed and executed in a distributed manner by one or more processors or computers.

[0169] According to an exemplary embodiment of the present disclosure, a computer program product may also be provided, which includes computer executable instructions, and when the computer executable instructions are executed by at least one processor, a training method for a cell segmentation model or a cell segmentation method according to an exemplary embodiment of the present disclosure is implemented.

[0170] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.

[0171] In addition, it should be noted that although several examples of each step are described above with reference to specific drawings, it should be understood that the embodiments of the present disclosure are not limited to the combinations given in the examples, and the steps appearing in different drawings may be combined, and the order of execution of the steps may be changed, which is not exhaustive here.

[0172] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for training a cell segmentation model, characterized in that: The training method comprises: Acquire a training sample set and a pseudo label for annotating the training sample set, wherein each training sample in the training sample set includes a cell slice image, and the pseudo label includes a first label at a pixel level generated based on a point label; Based on the training sample set and the pseudo-label, the cell segmentation model is trained in the first stage to obtain a cell segmentation model in the first stage; Based on the training results of the first stage, uncertainty prediction is performed by introducing noise to obtain uncertainty prediction results; According to the uncertainty prediction result, correcting the first label in the pseudo label to obtain a corrected pseudo label; Based on the training sample set and the corrected pseudo-labels, the cell segmentation model of the first stage is trained in the second stage to obtain a final cell segmentation model. Wherein, the pseudo label also includes a second label that is iteratively updated during the training process, wherein the cell segmentation model is trained in the first stage or the second stage in the following manner: Inputting the training sample into the cell segmentation model to obtain a predicted segmentation result; The predicted segmentation result is compared with the first label and the second label respectively to obtain a prediction loss to train the cell segmentation model. The prediction loss includes a first loss and a second loss, the first loss represents a loss corresponding to the first label, and the second loss represents a loss corresponding to the second label. The training of each stage includes multiple rounds, wherein the second label of each round is obtained by: weighting and summing the segmentation result predicted in the previous round and the second label of the previous round to obtain the iterative second label; using the first label to correct the iterative second label to obtain the second label of the current round, wherein the second label of the first round is the segmentation result initially predicted by the cell segmentation model.

2. The training method according to claim 1, characterized in that: The first label includes a cluster label, and the cluster label includes a first type of label, a second type of label, and a third type of label. The first type of label represents a cell area, the second type of label represents a background area, and the third type of label represents other areas except the cell area and the background area. The iterative second label is corrected by using the first label in the following manner: For each pixel in the cell slice image, in response to the first label indicating that the pixel is of the first category label or the second category label, the iterative second label is updated to be the same as the first label; in response to the first label indicating that the pixel is of the third category label, the iterative second label is kept unchanged.

3. The training method according to claim 1, characterized in that: The first label includes a cluster label, wherein, for any cell slice image, the cluster label is determined by: Determine a target pixel within a predetermined area centered on each point label, wherein each point label represents a position of a cell in the cell slice image; For each target pixel, determining a color distance value between the target pixel and each adjacent pixel; Determine a standard deviation value of the color distance values ​​for all target pixels; Based on the standard deviation value, adjusting the color channel value of each pixel in the cell slice image to obtain an adjusted image; The clustering label is obtained by clustering the color features of the pixels in the adjusted image.

4. The training method according to claim 1, characterized in that: The uncertainty prediction results are obtained by: For the input features of the last layer of the cell segmentation model in the last round of training in the first stage, a Gaussian noise distribution is introduced to obtain a noise introduction feature; The noise introduction feature is input into the last layer of the cell segmentation model to obtain the uncertainty prediction result.

5. The training method according to claim 1, characterized in that: The uncertainty prediction result includes an uncertainty value of each pixel in the cell slice image, wherein the uncertainty value indicates the possibility that the pixel is incorrectly labeled in the pseudo-label, wherein the first label in the pseudo-label is corrected in the following manner: Based on the average value of uncertainty values ​​of pixels in each cell region, determining a target cell region where the average value is within a preset range; Selecting a region corresponding to the target cell region from the prediction results of the first stage to generate a segmentation prediction mask; The segmentation prediction mask is used to replace a corresponding area in the first label, and the first label is corrected to obtain a corrected pseudo label.

6. A cell segmentation method, characterized in that: The cell segmentation method comprises: Acquiring a target image, wherein the target image is a cell slice image; The target image is input into a cell segmentation model, and the cells in the target image are segmented using the cell segmentation model to obtain a segmentation result, wherein the cell segmentation model is trained according to the training method of the cell segmentation model according to any one of claims 1 to 5.

7. A training system for a cell segmentation model, characterized in that: The training system comprises: A training acquisition unit is configured to acquire a training sample set and a pseudo label for annotating the training sample set, wherein each training sample in the training sample set includes a cell slice image, and the pseudo label includes a first label at a pixel level generated based on a point label; A first-stage training unit is configured to perform a first-stage training on the cell segmentation model based on the training sample set and the pseudo-label to obtain a first-stage cell segmentation model; An uncertainty prediction unit is configured to perform uncertainty prediction based on the training result of the first stage by introducing noise to obtain an uncertainty prediction result; A denoising unit is configured to correct the first label in the pseudo label according to the uncertainty prediction result to obtain a corrected pseudo label; The second stage training unit is configured to perform second stage training on the cell segmentation model of the first stage based on the training sample set and the corrected pseudo-labels to obtain a final cell segmentation model. The pseudo-label also includes a second label that is iteratively updated during the training process, wherein the first-stage training unit is configured to perform the first-stage training of the cell segmentation model in the following manner or the second-stage training unit is configured to perform the second-stage training of the cell segmentation model in the following manner: inputting the training sample into the cell segmentation model to obtain a predicted segmentation result; and obtaining a prediction loss by comparing the predicted segmentation result with the first label and the second label respectively, so as to train the cell segmentation model, wherein the prediction loss includes a first loss and a second loss, the first loss represents the loss corresponding to the first label, and the second loss represents the loss corresponding to the second label, Wherein, the training of each stage includes multiple rounds, wherein the first-stage training unit or the second-stage training unit is further configured to obtain the second label of each round in the following manner: by weighting and summing the segmentation result predicted in the previous round and the second label of the previous round respectively, to obtain the iterative second label; using the first label to correct the iterative second label to obtain the second label of the current round, wherein the second label of the first round is the segmentation result initially predicted by the cell segmentation model.

8. A cell segmentation system, characterized in that: The cell segmentation system comprises: An image acquisition unit, configured to acquire a target image, wherein the target image is a cell slice image; An image segmentation unit is configured to input the target image into a cell segmentation model, and use the cell segmentation model to segment the cells in the target image to obtain a segmentation result, wherein the cell segmentation model is trained according to the training method of the cell segmentation model according to any one of claims 1 to 5.

9. An electronic device, characterized in that: The electronic device comprises: processor; a memory for storing instructions executable by the processor, Wherein, when the processor executable instructions are executed by the processor, the processor is prompted to execute the training method of the cell segmentation model according to any one of claims 1 to 5 or the cell segmentation method according to claim 6.

10. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the training method of a cell segmentation model according to any one of claims 1 to 5 or the cell segmentation method according to claim 6.

11. A computer program product comprising computer executable instructions, characterized in that: When the computer executable instructions are executed by at least one processor, the training method of the cell segmentation model according to any one of claims 1 to 5 or the cell segmentation method according to claim 6 is implemented.