Semi-automatic intelligent marking method of cells in digital images

Abstract

The application provides a kind of semi-automatic intelligent calibration method of cell in digital image.The method comprises: preparing the image dataset to be labeled containing target category cell;Manual ellipse marking is carried out on the target cell in the to-be-labeled dataset;Ellipse marking is taken as a reference, and a suitable mask generation function is designed according to the shape prior knowledge of the target cell;Ellipse marking is taken as a reference, and a suitable clustering method is selected to cluster the to-be-labeled image;Take the intersection of the generated mask and the clustering result as the primary pseudo-label of cell labeling;Relying on the ellipse marking, a loss function is constructed, and the pseudo-label is used to train the deep learning segmentation network, the false positive connected domain outside the ellipse marking is removed, and the remaining result is used as the new pseudo-label of cell labeling;Experts correct the pseudo-label generated by deep learning, remove the obvious false positive and supplement the labeling of the obvious missed connected domain to obtain more accurate pseudo-label;Repeat the first two steps for several times until the expert is satisfied with the label generation effect, save the intelligent labeling model and the labeling result.The scheme of the application combines the shape prior of the target cell, only needs manual ellipse marking of the cell, greatly reduces the cost of manual labeling under the premise of ensuring the accuracy of label generation, can be applied to the labeling task of various morphologies of cells under various modalities, and has the advantages of convenient and efficient, high precision, wide application range, good generalization, etc. in the field of cell labeling.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a semi-automatic intelligent labeling method of cells in a digital image. BACKGROUND

[0002] With the vigorous development of deep learning technology, artificial intelligence assisted medical diagnosis and biological analysis have become a hot topic. Since cell analysis is one of the important bases for explaining the changes of biological signs and behaviors, it has become one of the important needs to assist cell analysis with artificial intelligence technology. For example, the positioning and quantification of diseased cells and the diagnosis of diseases, the formulation of surgical plans, prognosis and other important links are closely related, therefore, cell segmentation is one of the most popular research directions in the field of intelligent medical image processing. However, the training of deep learning based algorithm is often driven by large data with fine annotation, and the annotation of cells is time-consuming and labor-intensive, which greatly limits the data size and hinders the rapid progress of related technologies. The present application takes the reduction of time cost of artificial annotation as the starting point, starts from the semi-automatic mode of human-computer interaction, and proposes a cell intelligent labeling method that balances the labeling time cost and the labeling accuracy required by the algorithm, and combines the prior knowledge of the shape of the target cell, which greatly improves the scope of application of the method, while ensuring the labeling accuracy and improving the cell labeling efficiency and labeling scale. SUMMARY

[0003] The present application provides a semi-automatic intelligent labeling method of cells in a digital image, characterized in that the method comprises:

[0004] Step 1, preparing a to-be-labeled image data set containing target category cells;

[0005] Step 2, manually labeling the target cells in the to-be-labeled data set with ellipses;

[0006] Step 3, referring to the ellipse label, designing a suitable mask generation function according to the shape prior knowledge of the target cell;

[0007] Step 4, referring to the ellipse label, selecting a suitable clustering method to cluster the to-be-labeled image;

[0008] Step 5, taking the intersection of the generated mask and the clustering result as the cell labeling primary pseudo label;

[0009] Step 6, relying on the ellipse label to construct a loss function, training a deep learning segmentation network with the pseudo label, removing false positive connected domains outside the ellipse label, and taking the remaining result as a new cell labeling pseudo label;

[0010] Step 7, correcting the pseudo label generated by the deep learning by an expert, removing obvious false positives and supplementing the labeling of obvious missed connected domains to obtain a more accurate pseudo label;

[0011] Step 8, repeat steps 6 and 7 multiple times until the expert is satisfied with the label generation effect, save the intelligent calibration model and calibration results.

[0012] Specifically, in step 1, the preparation of the image data set to be labeled containing target category cells includes: preparing several biological samples containing target category cells, including but not limited to hematoxylin-eosin stained histopathological images, immunohistochemical images, liquid-based cell images, electron microscope images and other modalities, scanning into high-resolution digital images W after slicing, selecting the area containing target cells by experts and cutting several image blocks with side length l The image block contains a varying number of target category cells, and all image blocks containing target category cells collectively form the image data set S to be labeled raw .

[0013] Specifically, in step 2, the target cells in the to-be-labeled data set are manually labeled with ellipses, including: for the target category cells x in all image blocks P in the to-be-labeled data set S raw in step 1, manually label the outside of the cell with an ellipse, denoted as x e .

[0014] Specifically, in step 3, the ellipse label is taken as a reference, and a suitable mask generation function is designed according to the shape prior knowledge of the target cell, including: taking the center point χ e of the ellipse label x p in step 2 as the center reference point, combining the shape rules of the target cell in biology, and designing a mask generation function G(x p ), the function value change law can reflect the cell shape and the distance from the ellipse center point, which is not limited to Gaussian function, radial basis function, affine function and other forms.

[0015] Specifically, in step 4, the ellipse label is taken as a reference, and a suitable clustering method is selected to cluster the to-be-labeled image, including: selecting a suitable clustering method to cluster all image blocks P in the to-be-labeled data set S raw in step 1, eliminate false positive regions outside the ellipse label x e described in step 2, if the same ellipse label contains multiple connected domains, the center points of the two connected domains are connected with a straight line, so that each ellipse label x e only corresponds to one cluster connected domain K(x e ).

[0016] Specifically, in step 5, the intersection of the generated mask and the clustering result is taken as the primary pseudo-label of cell annotation, including: taking the part of the mask generation function greater than the threshold t in step 3 as a positive sample, and taking the intersection of the mask positive sample and the clustering connected domain K(x e ) in step 4 as the primary pseudo-label Y(x e ) of the ellipse marker χ e ), that is, Y(χ e ) = [G(χ p |χ e )>t]∩K(χ e ).

[0017] Specifically, in step 6, the loss function is constructed based on the ellipse marker to train the deep learning segmentation network with the pseudo-label, and the false positive connected domain outside the ellipse marker is removed, and the remaining result is taken as the new cell annotation pseudo-label, including: constructing a suitable loss function based on the ellipse marker, including but not limited to the ellipse field loss where L c is a common segmentation loss such as cross-entropy, F(x i ,y j ;c χ ,c y ) is the weight value of the point (x i ,y j ) inside the ellipse marker x e with the center at (c χ ,c y ), and the weight value inside the ellipse marker x e is gradually changed; select a deep learning segmentation network, and use the pseudo-label Y(χ e ) generated in step 5 as supervision information to constrain the convergence of the model, and obtain an automatic labeling model M after iterative optimization; then, use the labeling model to infer and label the data set S raw in step 1, remove the false positive region outside the ellipse marker, and if the ellipse contains multiple connected domains, connect them into one connected domain through a straight line, so that the ellipse marker χ e has a one-to-one correspondence with the newly generated labeling connected domain .

[0018] Specifically, in step 7, the pseudo-label generated by the deep learning is corrected by an expert, and the annotation of the obviously false positive and the obviously missed connected domain is supplemented to obtain a more accurate pseudo-label, including: the doctor reviews the new label generated in step 6 , removes the obviously false positive region, and supplements the obviously false negative labeling, and does not modify the labeling without large errors to form a relatively more accurate labeling

[0019] Specifically, in step 8, the steps 6 and 7 are repeated multiple times until the expert is satisfied with the label generation effect, the intelligent labeling model and the labeling result are saved, including: taking the labeling generated in step 7 as the input pseudo label of step 6, and repeating steps 6 and 7 until the labeling generated in step 7 reaches the standard of manual labeling The generated labeling result is saved for subsequent development of supervised information of deep learning algorithm, and the intelligent labeling model is saved for direct generation of labeling by other similar data.

[0020] The present application is directed to the problem that the current cell fine labeling is high in cost, and the rough labeling limits the subsequent development of deep learning model, and a semi-automatic intelligent labeling method for cells in digital images is proposed, which greatly reduces the labeling cost under the premise of ensuring accuracy by means of artificial game strategy, and can adapt to cells of various modalities under various modalities, and has the advantages of wide application range, convenient and efficient operation, high precision and the like. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only a part of the embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0022] Figure 1 The flowchart of the semi-automatic intelligent labeling method for cells in digital images is shown in the figure.

[0023] Figure 2 The flowchart of the semi-automatic intelligent labeling method for signet ring cells in pathological images is shown in the figure.

[0024] Figure 3 An example of a definition form of the elliptical field loss function is shown in the figure. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0026] The development of deep learning algorithms at the present stage is limited by the number of accurately labeled data, and the accurate labeling of cells is time-consuming and labor-intensive. In view of the above problems, the present application proposes a semi-automatic intelligent labeling method for digital images of cells. This method only requires manual ellipse marking, and the labeling cost is much lower than that of fine labeling by outlining the edge, and is suitable for cell labeling tasks of various shapes under various modal images. The specific steps are as follows:

[0027] S101: Prepare a to-be-labeled image data set containing target category cells.

[0028] Specifically, prepare several biological samples containing target category cells, including but not limited to hematoxylin-eosin stained histopathology images, immunohistochemistry images, liquid-based cell images, electron microscope images and other modalities. After slicing, scan into high-resolution digital images W, select the area containing target cells by experts and crop several image blocks with side length l The image blocks contain different numbers of target category cells, and all image blocks containing target category cells together form a to-be-labeled image data set S raw .

[0029] S102: Manually mark an ellipse on the target cell in the to-be-labeled data set.

[0030] Specifically, for the target category cells χ in all image blocks P in the to-be-labeled data set S raw in step S101, manually mark an ellipse outside the cell, denoted as χ e .

[0031] S103: Refer to the ellipse mark, and design a suitable mask generation function according to the shape prior knowledge of the target cell.

[0032] Specifically, taking the center point χ e of the ellipse mark χ p in step S102 as the center reference point, combining the shape rules of the target cell in biology, designing a mask generation function G(χ p ), the function value change law can reflect the cell shape and the distance from the ellipse center point, which is not limited to Gaussian function, radial basis function, affine function and other forms.

[0033] S104: Refer to the ellipse mark, and select a suitable clustering method to cluster the to-be-labeled image.

[0034] Specifically, select a suitable clustering method to cluster all image blocks P in the to-be-labeled data set S raw in step S101, and eliminate the ellipse mark χ eThe false positive regions outside the ellipse marker are connected by a straight line if the ellipse marker contains multiple connected domains, to ensure that each ellipse marker x e corresponds to only one cluster connected domain K(x e ).

[0035] S105: Take the intersection of the generated mask and the clustering result as the primary pseudo-label of cell labeling.

[0036] Specifically, the part of the mask generation function greater than the threshold t in step S103 is taken as the positive sample, and the intersection of the mask positive sample and the cluster connected domain K(x e ) in step S104 is taken as the primary pseudo-label Y(x e ) of the ellipse marker x e ), that is, Y(x e ) = [G(x p |x e )>t]∩K(x e ).

[0037] S106: Construct a loss function based on the ellipse marker to train a deep learning segmentation network with the pseudo-label, remove the false positive connected domains outside the ellipse marker, and the remaining results are taken as the new cell labeling pseudo-label.

[0038] Specifically, a suitable loss function is constructed based on the ellipse marker, including but not limited to the ellipse field loss where L c is a common segmentation loss such as cross-entropy, and F(x i ,y j ; c x ,c y ) is the weight value of the point (x x ,y y ) inside the ellipse marker x e with the center (c i ,c j ), and the weight value inside the ellipse marker χ e is gradually changing; select a deep learning segmentation network, use the pseudo-label Y(x e ) generated in step S105 as the supervision information to constrain the convergence of the model, and obtain the automatic calibration model M after iterative optimization; then use the calibration model to infer and calibrate the data set S raw described in step S101, remove the false positive regions outside the ellipse marker, and connect multiple connected domains inside the ellipse by a straight line if there are multiple connected domains, so as to ensure that the ellipse marker x e has a one-to-one correspondence with the newly generated calibration connected domain .

[0039] S107: The expert corrects the pseudo-label generated by deep learning, removes obvious false positives, and supplements the labeling of obvious missed connected domains to obtain more accurate pseudo-labels.

[0040] Specifically, the expert reviews the new labels generated in step S106 to remove obvious false positive regions and supplement obvious false negative labeling, and does not modify the labeling without major errors to form relatively more accurate labeling

[0041] S108: Repeat steps S106 and S107 multiple times until the expert is satisfied with the label generation effect, and save the intelligent labeling model and labeling results.

[0042] Specifically, the labeling generated in step S107 is used as the input pseudo-label of step S106, and steps S106 and S107 are repeated in a loop until the labeling generated in step S107 reaches the standard of manual labeling, saves the generated labeling results as supervision information for subsequent deep learning algorithm development, and saves the intelligent labeling model to facilitate direct labeling generation for other similar data.

[0043] In order to facilitate the understanding of the present application, Figure 2 a flowchart of a semi-automatic intelligent labeling method for signet ring cells in pathological images is shown, which illustrates the implementation approach of the semi-automatic labeling generation method for pathological cells in actual application by taking signet ring cells as an example.

[0044] As Figure 2 shown, it is a flowchart of a semi-automatic intelligent labeling method for signet ring cells in pathological images, specifically comprising:

[0045] S201: Prepare a patient sample of signet ring cell carcinoma.

[0046] Specifically, prepare several patient samples of signet ring cell carcinoma, embed and section, stain, and then scan into high-resolution digital images W.

[0047] S202: The doctor selects the area containing signet ring cells, crops the image blocks, and constructs the labeling data set.

[0048] Specifically, the doctor selects the area containing signet ring cells and crops several image blocks with a side length of 1000 pixels The image blocks contain varying numbers of signet ring cells, and all the cropped image blocks collectively form the labeling pathological image data set S raw .

[0049] S203: The doctor labels the signet ring cells in the labeling data set with cell nucleus and cytoplasm ellipses, respectively.

[0050] Specifically, for the target class cell x in all image blocks P in the data set S to be labeled in step S202 raw , an artificial oval mark is made outside the cell nucleus and marked as x ne , the center point of which is marked as x np , an artificial oval mark is made outside the cytoplasm and marked as x ce , the center point of which is marked as x cp .

[0051] S204: Generate a Gaussian mask for the cell nucleus and cytoplasm respectively with different parameters with the oval mark center point as the reference.

[0052] Specifically, the center points χ np and x cp in step S203 are taken as reference points, and since the signet ring cell as a whole presents a spheroid shape, a two-dimensional Gaussian function is constructed to generate a mask function, wherein the mask function of the cell nucleus is and the mask function of the cytoplasm is The function value is maximum at the center point and decreases outwardly, and the function value change rule can reflect the cell shape and the distance to the center point.

[0053] S205: Use different algorithms to cluster the image and remove the connected domain outside the oval mark, and take the union set of different clustering methods as the clustering connected domain of the oval mark.

[0054] Specifically, the K-means clustering method is used to cluster all image blocks P in the data set S to be labeled in step S202 raw , and the connected domain outside the oval mark x ne and x ce in step S203 is removed, if multiple connected domains are contained in the same oval mark, only one connected domain is retained or the connected domains in the same oval are connected by a straight line, so as to ensure that each oval mark corresponds to only one clustering connected domain K n1 (x) or K c1 (x); similarly, the mixed Gaussian EM algorithm is used to repeat the clustering step to obtain the cell nucleus clustering connected domain set K n2 (x) and the cytoplasm clustering connected domain set K c2 (x); take the union set of different clustering methods as the clustering result, that is, the cell nucleus clustering connected domain set is K n (x) = K n1 (x)∪K n2 (x), and the cytoplasm clustering connected domain set is K c (x) = K c1 (x)∪K c2 (x).

[0055] S206: Take the intersection of the mask and the clustered connected domain as the generated label of the ellipse marker, merge the generated labels of the cell nucleus and cytoplasm of the same cell as the generated pseudo-label of the whole cell to form a pseudo-label set.

[0056] Specifically, the part of the mask generation function greater than the threshold t in step S204 is intercepted as a positive sample, and the intersection of the mask positive sample and the clustered connected domain in step S205 is taken as the preliminary pseudo-label of the ellipse marker, and the pseudo-labels Y(x) of the cytoplasm and nucleus of the same cell are merged.

[0057] S207: Supervised training of a deep learning segmentation network based on the pseudo-label set to obtain an automatic labeling model.

[0058] Specifically, the pseudo-label Y(x) generated in step S206 is used as supervision information, and a deep learning segmentation network such as U-Net is trained by constraining loss functions such as cross-entropy loss and ellipse field loss, and an automatic labeling model M is obtained after iterative optimization; then the labeling model is used to label the data set S raw Inference labeling, removing false positive connected domains outside the ellipse marker, connecting multiple connected domains inside the ellipse, ensuring that the ellipse marker is connected to the newly generated labeled connected domain There is a one-to-one correspondence.

[0059] S208: Generate a label for the image to be tested through the automatic labeling model, and manually remove obvious false positive areas and supplement large area missed area labeling by a doctor.

[0060] Specifically, the doctor reviews the new label generated in step S207 Remove the obvious false positive area, and supplement the obvious false negative labeling, and do not modify the labeling without large errors to form a relatively more accurate labeling

[0061] S209: Repeat S207 and S208 in a loop until the automatic labeling effect meets the doctor's manual labeling standard.

[0062] Specifically, the label generated in step S208 is used as the input pseudo-label of step S207, and steps S207 and S208 are repeated in a loop until the label generated in step S208 Reaches the standard of manual labeling.

[0063] S210: Use the intelligent labeling model according to specific needs, which can be used for direct positioning and quantification of signet ring cells, and can also be used to provide batch labeling of signet ring cells for subsequent deep learning tasks.

[0064] Specifically, the generated calibration result and the automatic calibration model are saved and applied according to specific requirements; the calibration result can be used as supervision information for subsequent development of deep learning algorithms; and the saved intelligent calibration model can be used for direct calibration of other similar data, and can be used for direct positioning and quantification of signet ring cells, and can also be used for batch generation of calibration to facilitate subsequent development of deep learning algorithms.

[0065] In order to facilitate the understanding of the present application, Figure 3 A definition diagram of the elliptical field loss is shown, which is one of the implementation ways according to the present application.

[0066] As Figure 3 shown, it is an example of a definition form of the elliptical field loss function, which specifically includes:

[0067] S301: Prepare a medical image to be labeled, and manually annotate an ellipse on a target cell.

[0068] Specifically, prepare a medical image such as a pathological image, scale it to an appropriate size, and manually draw an ellipse mark outside the target cell.

[0069] S302: Establish a rectangular coordinate system based on the ellipse mark.

[0070] Specifically, a rectangular coordinate system is established with the center of the ellipse as the coordinate origin, the straight line where the long axis of the ellipse is located as the y-axis, and the straight line where the short axis of the ellipse is located as the x-axis, and the long axis of the ellipse is denoted as 2a, the short axis is denoted as 2b, and the distance between the two foci is denoted as 2c.

[0071] S303: Rotate the rectangular coordinate system so that the x-axis is in the horizontal direction.

[0072] Specifically, the rectangular coordinate system described in S302 is rotated in the plane so that the x-axis of the coordinate system is in the horizontal direction and the y-axis is in the vertical direction.

[0073] S304: Calculate the weight value of any point M in the ellipse.

[0074] Specifically, it is assumed that the weight value on the edge of the ellipse is 1, and the weight value gradually decreases from outside to inside until the line segment connecting the foci is the minimum weight value; for any point M in the ellipse, an ellipse with the same foci as the ellipse mark is constructed, and the corresponding weight is set as the ratio of the long axis of the newly generated ellipse to the long axis of the ellipse mark, that is,

[0075] S305: Adjust the value range of the weight.

[0076] Specifically, the value range of the weight corresponding to any point M in the ellipse mark in S304 is [c / a, 1], and the value range is mapped and adjusted to [0, 1], so that the weight calculation formula of point M is transformed into

[0077] S306: Rotate the coordinate system to the true angle.

[0078] Specifically, the weight of any point M within the elliptical marker calculated in S305 is rotated to the true angle along with the coordinate system, that is, transformed back to the initial angle according to the rotation angle in S303, to obtain (c x ,c y The ellipse marked in the center of x e Any point (x) inside i ,y j The weights F(x) i ,y i c x ,c y ).

[0079] S307: Example of the definition of elliptic field loss.

[0080] Specifically, the weighted cross-entropy loss L calculated in S306 c The elliptic field loss L is obtained. E A definition form This is merely one embodiment of elliptic field loss as defined in this invention, used to assist in the automatic calibration of models to converge quickly and achieve excellent performance.

[0081] This invention addresses the problem of excessively high costs in current cell labeling techniques by proposing a semi-automatic intelligent cell labeling method for digital images with human-computer interaction. This method is applicable to cell labeling tasks of various shapes and modes, significantly reducing labeling costs while ensuring accuracy. It is convenient, efficient, and highly generalizable, and can be directly used for cell localization and quantization in practical applications, as well as large-scale cell labeling tasks before the development of deep learning algorithms.

Claims

1. A semi-automatic intelligent method of marking cells in a digital image, characterized in that, The method comprises the following steps: Step 1, preparing a to-be-labeled image dataset containing target category cells; Step 2, manually marking the target cells in the to-be-labeled dataset with ellipses; Step 3, taking the ellipse mark as a reference, designing a suitable mask generation function according to the shape prior knowledge of the target cells; Step 4, taking the ellipse mark as a reference, selecting a suitable clustering method to cluster the to-be-labeled image; Step 5, taking the intersection of the generated mask and the clustering result as the preliminary pseudo-label of cell labeling; Step 6, constructing a loss function based on the ellipse label to train the deep learning segmentation network with pseudo labels, eliminating false positive connected domains outside the ellipse label, and the remaining results as new cell label pseudo labels; Specifically, wherein the suitable loss function is constructed based on the ellipse label, including but not limited to ellipse field loss wherein L C is a common cross-entropy segmentation loss, F(x i ,y i ; c x ,c y ) is the weight value of the point (x x ,y y ) inside the ellipse label x e with the center at (c i ,c j ), and the weight value inside the ellipse label x e is gradually changed; Step 7, expert corrects the pseudo-label generated by deep learning, eliminates obvious false positives and supplements the labeling of obvious missed connected domains, to obtain more accurate pseudo-labels; Step 8, repeat steps 6 and 7 multiple times until the expert is satisfied with the label generation effect, save the intelligent calibration model and the calibration result.

2. The method of semi-automatic intelligent cell labeling in digital images of claim 1, wherein, In step 1, the preparation of the image data set to be labeled containing target category cells, specifically: prepare several biological samples containing target category cells, including but not limited to hematoxylin-eosin stained histopathological images, immunohistochemical images, liquid-based cell images, electron microscope images of various modalities, and scan into high-resolution digital images W after slicing. Experts select areas containing target cells and crop several image blocks with side length l The image blocks contain varying numbers of target category cells, and all image blocks containing target category cells collectively form the image data set S to be labeled raw .

3. The method of semi-automatic intelligent cell segmentation in digital images as claimed in claim 1, wherein, In step 2, the target cells in the to-be-labeled data set are manually labeled with an ellipse, specifically: for the target category cells x in all image blocks P in the to-be-labeled data set S raw in step 1, manually label an ellipse outside the cells, denoted as x e .

4. The method of claim 1, wherein the method further comprises: In step 3, the mask generation function is designed according to the shape prior knowledge of the target cell, specifically: taking the center point x e of the ellipse mark x p described in step 2 as the center reference point, combining the shape rules of the target cell in biology, and designing the mask generation function G(x p ). The function value change rule can reflect the cell shape and the distance from the ellipse center point, and is not limited to Gaussian function, radial basis function, affine function, and various forms.

5. The method of semi-automated intelligent cell segmentation in digital images of claim 1, wherein, In step 4, the image to be labeled is clustered by selecting a suitable clustering method with the ellipse mark as a reference, specifically: selecting a suitable clustering method to cluster all image blocks P in the data set S to be labeled in step 1, eliminating false positive regions outside the ellipse mark x in step 2, and if multiple connected domains are included in the same ellipse mark, connecting the center points of the two connected domains with a straight line to ensure that each ellipse mark x corresponds to only one clustering connected domain K(x e ). raw e e e ​​ 6. The method of semi-automatic intelligent cell segmentation in digital images as claimed in claim 1, wherein, In step 5, the intersection of the mask and the clustering result is taken to generate the primary pseudo label of the cell label, specifically: the part of the mask generation function greater than the threshold y in step 3 is taken as the positive sample, and the intersection of the mask positive sample and the clustering connected domain K(x e ) in step 4 is taken as the primary pseudo label Y(x e ) of the ellipse marker x e ), that is, Y(x e )=[G(x p |x e )>t]∩K(x e ) 7. The method of semi-automated intelligent cell segmentation in digital images of claim 1, wherein, In step 6, the loss function is constructed based on the ellipse label to train the deep learning segmentation network with pseudo labels, and the false positive connected domains outside the ellipse label are removed, and the remaining results are used as new cell label pseudo labels, specifically: a suitable loss function is constructed based on the ellipse label, including but not limited to ellipse field loss Where L c is the cross-entropy common segmentation loss, F(x i ,y j ; c x ,c y ) is the weight value of the point (x x ,y y ) inside the ellipse label x e with the center at (c i ,c j ), and the weight value inside the ellipse label x e is gradually changed; a deep learning segmentation network is selected, and the pseudo label Y(x e ) generated in step 5 is used as the supervision information to constrain the convergence of the model, and the automatic labeling model M is obtained after iterative optimization; then the labeling model is used to infer and label the data set S raw in step 1, and the false positive regions outside the ellipse label are removed, and if multiple connected domains are contained in the ellipse, they are connected into one connected domain through straight lines, so as to ensure that the ellipse label x e has a one-to-one correspondence with the newly generated labeling connected domain .

8. The method of claim 1, wherein, In step 7, the expert corrects the pseudo-label generated by deep learning, removes obvious false positives, and supplements the labeling of obvious missed connected domains to obtain more accurate pseudo-labels, specifically: doctors review the new labels generated in step 6 Remove obvious false positive regions and supplement obvious false negative labeling. Do not modify the labeling without major errors to form a relatively more accurate labeling 9. The method of semi-automated intelligent cell segmentation in digital images of claim 1, wherein, In step 8, the steps 6 and 7 are repeated multiple times until the expert is satisfied with the label generation effect, and the intelligent labeling model and the labeling result are saved, specifically: the labeling generated in step 7 is used as the input pseudo label of step 6, and the steps 6 and 7 are repeated in a loop until the labeling generated in step 7 reaches the standard of manual labeling The generated labeling result is saved for subsequent development of supervised information of deep learning algorithm, and the intelligent labeling model is saved to facilitate direct generation of labeling for other similar data.

Images