A method for generating weakly supervised labels for cell nucleus segmentation

Abstract

The application discloses a weak supervision label generation method for cell nucleus segmentation. The method firstly generates point labels by selecting the center pixels of existing cell nucleus labels to simulate the manual labeling of cell nucleus point labels; then, the distance information of each point and its neighbor points is acquired; and finally, an algorithm is constructed to generate the weak supervision label of the cell nucleus. The method uses the existing cell nucleus label to generate the point label of the cell nucleus, acquires the distance information of each point and its neighbor points, and then constructs an algorithm process to generate the weak supervision label of the cell nucleus, thereby greatly reducing the calculation cost, providing the weak supervision learning of the cell nucleus segmentation with the label, greatly reducing the calculation amount compared with the current weak supervision label generation method of the cell nucleus, and simplifying the label generation process without setting the hyperparameter.

Description

Technical Field

[0001] This invention belongs to the field of cell nucleus image processing technology, and specifically relates to a method for generating weakly supervised labels for cell nucleus segmentation. Background Technology

[0002] With the development of computer vision and medical image processing technologies, deep learning methods have achieved good results in cell nucleus segmentation. However, deep learning methods typically require a large amount of pixel-level labeled data for training. Furthermore, due to the extremely high pixel count of cell nucleus images—a single image may contain tens of thousands of nuclei—labeling large-scale datasets is a very time-consuming, tedious, and error-prone task. Obtaining pixel-level cell nucleus labels has severely limited the model's ability to achieve higher accuracy and better generalization. Weakly supervised learning is a method that utilizes limited or incomplete supervisory information for model training. Therefore, researchers have begun to apply weakly supervised learning methods to cell nucleus segmentation, and generating weakly supervised cell nucleus labels is a crucial step in this process.

[0003] Currently, there are several common schemes for generating weakly supervised labels for cell nuclei, such as image-level labels, region labels, and point labels. Image-level labels are used to indicate the presence of cell nuclei in the entire image. This method does not provide specific location information of the cell nuclei and is only used for image-level binary classification tasks. Region labels provide information about the overall shape and extent of the region where the cell nucleus is located, without requiring pixel-level annotation. However, they do not provide pixel-level segmentation information within the cell nucleus, and the model may produce inaccurate predictions at pixel boundaries within the region. Point labels provide point annotations of the cell nucleus, requiring only the labeling of a single pixel at the center of the cell nucleus. This method can be used to generate the distribution of cell nucleus center points. Point labels are generally used for cell nucleus detection tasks because they only contain location information of the cell nucleus. However, researchers can easily use point labels to generate more informative weakly supervised labels for cell nucleus segmentation. Some researchers have used point labels to generate Voronoi labels and clustering labels in cell nucleus segmentation. The Voronoi label contains points representing the cell nucleus and edges representing the cell nucleus boundary, used to isolate the cell nucleus. The clustering label contains the shape information of the cell nucleus pixels. The combination of the two provides more supervisory information. However, the cell nucleus boundary error in the Voronoi label is relatively large, and the clustering label has strict requirements on the parameter settings of the clustering algorithm, which consumes a lot of computation time and is difficult to generate good clustering results. Therefore, it is necessary to design a method for generating weakly supervised labels for cell nucleus segmentation to solve the above problems. Therefore, this patent, based on point labels, uses triangulation technology to generate triangular meshes, and then constructs an algorithm to generate weakly supervised labels for cell nuclei. Compared with the Voronoi label, it provides more effective cell nucleus pixel information and cell nucleus boundary information. Compared with the clustering label, it does not require attention to the algorithm parameter settings, and the computational cost is greatly reduced. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for generating weakly supervised labels for cell nucleus segmentation. The method aims to generate point labels for cell nuclei using existing cell nucleus labels, then obtain the distance information between each point and its neighboring points, and then construct an algorithm to generate weakly supervised labels for cell nuclei. This method provides labels for weakly supervised learning of cell nucleus segmentation while greatly reducing computational costs.

[0005] To achieve the above-mentioned technical effects, the technical solution adopted by the present invention is as follows:

[0006] A method for generating weakly supervised tags for cell nucleus segmentation includes the following steps:

[0007] S1, Generate point labels: Simulate manual labeling of cell nucleus points by selecting the center pixel of existing cell nucleus labels;

[0008] S2, obtain the distance information between each point and its neighboring points;

[0009] S3, constructs an algorithm to generate weakly supervised labels for cell nuclei.

[0010] Preferably, the specific method in step S1 is as follows:

[0011] S101, Read the existing cell nucleus labels and use a rectangle to frame each cell nucleus;

[0012] S102, calculate the minimum row coordinates, maximum row coordinates, minimum column coordinates, and maximum column coordinates of the rectangle containing the cell nucleus to determine the position information of the rectangle for each cell nucleus. Then, obtain the position (x, y) of the center pixel of the cell nucleus rectangle using the following formula, set its pixel value to 255, and construct the cell nucleus point label:

[0013]

[0014] In the formula, the variables are the minimum row coordinates r. min Maximum row coordinate r max Minimum column coordinate c min and the maximum column coordinate c max .

[0015] Preferably, the specific method of step S2 is as follows:

[0016] S201, using triangulation to obtain neighbor points:

[0017] First, construct a super triangle to hold all the points, then insert each point into the triangulation, and finally delete the super triangle.

[0018] S202, for each inserted point P(x) p ,y p Determine triangle ABC, where point P is located, and then connect this triangle with the insertion point to form a new triangle;

[0019] S203, ensuring that the circumcircles of any two adjacent triangles do not intersect, meaning that no point in the subdivision should lie inside the circumcircle of a triangle. When point P lies inside the circumcircle of triangle ABC, it satisfies the formula (x p -x o ) 2 +(y p -y o ) 2 <R 2 , where (x o ,y o () are the coordinates of the circumcenter of triangle ABC, and R is the radius of the circumcircle;

[0020] S204. If point P is located inside the circumcircle of triangle ABC, a local correction is required.

[0021] Preferably, the specific method of step S3 is as follows:

[0022] S301, Calculate the minimum Euclidean distance between each vertex and its neighbors: By traversing each vertex, i.e., traversing the set of points in the point labels, we can obtain the current point P(x). p ,y p The neighboring point Q(x) of ) q ,y q The index of ) is used to compute P(x) p ,y p ) and Q(x q ,y q Record the Euclidean distance d between the two pairs of points and the shortest Euclidean distance d. min ,in:

[0023]

[0024] S302 requires calculating the radius range of the cell nucleus. Based on the existing cell nucleus dataset, calculate the minimum radius r of the cell nucleus. min The average radius r of the cell nucleus avg The maximum radius r of the cell nucleus max ;

[0025] S303, Generate foreground information for weakly supervised labels:

[0026] Iterate through the point set in the point label again to obtain the midpoint P(x) from the previous step. p ,y p The shortest Euclidean distance d obtained min If d min Smaller than the average radius r of the cell nucleus avg This indicates that the area of ​​the cell nucleus represented by the current point is small, but it does not overlap with the adjacent cell nuclei or only overlaps with the adjacent cell nuclei. In order to separate the two cell nuclei, point P(x) is used. p ,y p With ) as the center, and d min Draw a circle with half the radius of d, and consider the pixels within the circle as the foreground pixels of the cell nucleus; if d min Greater than the average radius r of the cell nucleus avg This indicates that the cell nucleus represented by the current point is within its radius r. avg Within the range, it is highly likely to be a cell nucleus pixel, therefore point P(x) is used. p ,y p With ) as the center and the average radius r of the cell nucleus avg Draw a circle with radius , and treat the pixels inside the circle as the foreground pixels of the cell nucleus;

[0027] S304, Background information for generating weakly supervised labels:

[0028] With point P(x) p ,y p With ) as the center and the maximum radius r of the cell nucleus as the radius, max Draw a circle and divide the average nuclear radius r obtained from S303. avg If we record it as r, then we can know that the distance from r to r max The pixels between these pixels cannot be determined to belong to either the foreground cell kernel pixels or the background pixels, so the loss calculation for these pixels is ignored during training.

[0029] The beneficial effects of this invention are as follows:

[0030] 1. Compared with current methods for generating weakly supervised labels in cell nuclei, this scheme significantly reduces the computational load;

[0031] 2. This solution does not require setting hyperparameters, simplifying the tag generation process. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the algorithm flow of the present invention;

[0033] Figure 2 This is a schematic diagram of an existing cell nuclear label in this invention;

[0034] Figure 3 This is a schematic diagram of the dot labels generated in this invention;

[0035] Figure 4 This is a schematic diagram showing the neighboring points of each point in an embodiment of the present invention;

[0036] Figure 5 This is a schematic diagram illustrating the effect of weak supervision tags on cell nuclei in an embodiment of the present invention;

[0037] Figure 6 This is a schematic diagram illustrating the effect of integrating a neighbor point display map and a weak supervision label of the cell nucleus in an embodiment of the present invention. Detailed Implementation

[0038] Example 1:

[0039] like Figure 1 As shown, a method for generating weakly supervised tags for cell nucleus segmentation includes the following steps:

[0040] S1, Generate point labels: Simulate manual labeling of cell nucleus points by selecting the center pixel of existing cell nucleus labels;

[0041] S2, obtain the distance information between each point and its neighboring points;

[0042] S3, constructs an algorithm to generate weakly supervised labels for cell nuclei.

[0043] Preferably, the specific method in step S1 is as follows:

[0044] S101, Read the existing cell nucleus labels and use a rectangle to frame each cell nucleus;

[0045] S102, calculate the minimum row coordinates, maximum row coordinates, minimum column coordinates, and maximum column coordinates of the rectangle containing the cell nucleus to determine the position information of the rectangle for each cell nucleus. Then, obtain the position (x, y) of the center pixel of the cell nucleus rectangle using the following formula, set its pixel value to 255, and construct the cell nucleus point label:

[0046]

[0047] In the formula, the variables are the minimum row coordinates r. min Maximum row coordinate r max Minimum column coordinate c min and the maximum column coordinate c max .

[0048] Preferably, the specific method of step S2 is as follows:

[0049] S201, using triangulation to obtain neighbor points:

[0050] First, construct a super triangle to hold all the points, then insert each point into the triangulation, and finally delete the super triangle.

[0051] S202, for each inserted point P(x) p ,y p Determine triangle ABC, where point P is located, and then connect this triangle with the insertion point to form a new triangle;

[0052] S203, ensuring that the circumcircles of any two adjacent triangles do not intersect, meaning that no point in the subdivision should lie inside the circumcircle of a triangle. When point P lies inside the circumcircle of triangle ABC, it satisfies the formula (x p -x o ) 2 +(y p -y o ) 2 <R 2 , where (x o ,y o () are the coordinates of the circumcenter of triangle ABC, and R is the radius of the circumcircle;

[0053] S204. If point P is located inside the circumcircle of triangle ABC, a local correction is required.

[0054] Preferably, the specific method of step S3 is as follows:

[0055] S301, Calculate the minimum Euclidean distance between each vertex and its neighbors: By traversing each vertex, i.e., traversing the set of points in the point labels, we can obtain the current point P(x). p ,y p The neighboring point Q(x) of ) q ,y q The index of ) is used to compute P(x) p ,y p ) and Q(x q ,y q Record the Euclidean distance d between the two pairs of points and the shortest Euclidean distance d. min ,in:

[0056]

[0057] S302 requires calculating the radius range of the cell nucleus. Based on the existing cell nucleus dataset, calculate the minimum radius r of the cell nucleus. min The average radius r of the cell nucleus avg The maximum radius r of the cell nucleus max ;

[0058] S303, Generate foreground information for weakly supervised labels:

[0059] Iterate through the point set in the point label again to obtain the midpoint P(x) from the previous step. p ,y p The shortest Euclidean distance d obtained min If d min Smaller than the average radius r of the cell nucleus avg This indicates that the area of ​​the cell nucleus represented by the current point is small, but it does not overlap with the adjacent cell nuclei or only overlaps with the adjacent cell nuclei. In order to separate the two cell nuclei, point P(x) is used. p ,y p With ) as the center, and d min Draw a circle with half the radius of d, and consider the pixels within the circle as the foreground pixels of the cell nucleus; if d min Greater than the average radius r of the cell nucleus avg This indicates that the cell nucleus represented by the current point is within its radius r. avg Within the range, it is highly likely to be a cell nucleus pixel, therefore point P(x) is used. p ,y p With ) as the center and the average radius r of the cell nucleus avg Draw a circle with radius , and treat the pixels inside the circle as the foreground pixels of the cell nucleus;

[0060] S304, Background information for generating weakly supervised labels:

[0061] With point P(x) p ,y p With ) as the center and the maximum radius r of the cell nucleus as the radius, max Draw a circle and divide the average nuclear radius r obtained from S303. avg If we record it as r, then we can know that the distance from r to r max The pixels between these pixels cannot be determined to belong to either the foreground cell kernel pixels or the background pixels, so the loss calculation for these pixels is ignored during training.

[0062] Example 2:

[0063] like Figure 2 As shown, the left side is an image of the cell nucleus, and the right side is the label. The white pixels in the label store cell nucleus information. However, these labels are only suitable for fully supervised learning and lack weakly supervised labels for weakly supervised learning. Therefore, the center pixel of the existing cell nucleus labels is selected to simulate manually labeled cell nucleus points. The method in step S1 is used to simulate manually labeled cell nucleus points, and the generated point labels are as follows. Figure 3 As shown, since each point has only one pixel, the original image cannot display the pixel information, so the image has been cropped and enlarged for display.

[0064] like Figure 4 As shown, the points representing the cell nucleus are connected by the sides of a triangle, thus obtaining the distance information between each point and its neighbors. Based on this, weakly supervised labels for cell nucleus segmentation can be generated using prior knowledge of the cell nucleus radius.

[0065] like Figure 5 As shown, in step S304, the position from r to r max The pixels between these pixels cannot be determined whether they belong to the foreground cell nucleus or the background; therefore, the loss calculation for these pixels should be ignored during training. Figure 5 The black portion in the image likely contains background pixels, such as... Figure 5 The red part in the middle.

[0066] like Figure 6 As shown, Figure 6 The labels are shown to be generated based on neighboring points.

[0067] The above embodiments are merely preferred technical solutions of this invention and should not be considered as limitations on this invention. The embodiments and features described in this application can be arbitrarily combined without conflict. The scope of protection of this invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of this invention.

Claims

1. A method for generating weakly supervised tags for cell nuclear segmentation, characterized in that: Includes the following steps: S1, Generate point labels: Simulate manual labeling of cell nucleus points by selecting the center pixel of existing cell nucleus labels; S2, obtain the distance information between each point and its neighboring points; the specific method is as follows: S201, using triangulation to obtain neighbor points: First, construct a super triangle to hold all the points, then insert each point into the triangulation, and finally delete the super triangle. S202, for each insertion point Determine the point P Find the triangle ABC where the insertion point is located, and then connect this triangle with the insertion point to form a new triangle; S203 ensures that the circumcircles of any two adjacent triangles do not intersect, meaning that no point in the subdivision should lie inside the circumcircle of a triangle. If point P lies inside the circumcircle of triangle ABC, then the formula is satisfied. ,in R is the coordinate of the circumcenter of triangle ABC, and R is the radius of the circumcircle. S204, if point P Located inside the circumcircle of triangle ABC, local corrections are required; S3, constructing an algorithm to generate weakly supervised labels for cell nuclei, the specific method is as follows: S301, Calculate the minimum Euclidean distance between each vertex and its neighbors: By traversing each vertex, i.e., traversing the set of points in the point labels, the current point can be obtained. Neighboring points The index, thus calculating and Euclidean distance between And record the value of the shortest Euclidean distance. ,in: S302, it is necessary to calculate the radius range of the cell nucleus; based on the existing cell nucleus dataset, calculate the minimum radius of the cell nucleus. The average radius of the cell nucleus The maximum radius of the cell nucleus ; S303, Generate foreground information for weakly supervised labels: Iterate through the point set in the point label again to get the midpoint from the previous step. The shortest Euclidean distance obtained ;if Smaller than the average radius of the cell nucleus This indicates that the area of ​​the nucleus represented by the current point is small, but it does not overlap with the adjacent nuclei, or it only overlaps with the adjacent nuclei. In order to separate the two nuclei, the point... With the center as the center, Draw a circle with half the radius as its radius, and consider the pixels within the circle as the foreground pixels of the cell nucleus; if Larger than the average radius of the cell nucleus This indicates that the cell nucleus represented by the current point is within its radius. The area within which is most likely a cell nucleus pixel is used, therefore, it is considered as a point. Centered on the average radius of the cell nucleus Draw a circle with radius , and treat the pixels inside the circle as the foreground pixels of the cell nucleus; S304, Background information for generating weakly supervised labels: With point Centered on the largest radius of the cell nucleus Draw a circle and measure the average radius of the cell nuclei obtained from S303. Record as Then we can know that, in arrive The pixels between these pixels cannot be determined to belong to either the foreground cell kernel pixels or the background pixels, so the loss calculation for these pixels is ignored during training.

2. The method for generating weakly supervised tags for cell nucleus segmentation according to claim 1, characterized in that: The specific method in step S1 is as follows: S101, Read the existing cell nucleus labels and use a rectangle to frame each cell nucleus; S102, calculate the minimum row coordinates, maximum row coordinates, minimum column coordinates, and maximum column coordinates of the rectangle containing the cell nucleus, thereby determining the position information of the rectangle for each cell nucleus. Then, obtain the position of the center pixel of the cell nucleus rectangle sequentially using the following formula. And set its pixel value to 255 to construct cell nucleus point labels: In the formula, the variables are the minimum row coordinates. Maximum row coordinate Minimum column coordinates and the largest column coordinate .

Images