Three-dimensional cell segmentation method

Abstract

The application provides a three-dimensional cell segmentation method, and relates to the technical field of computer image processing, and comprises the following steps: S1, performing over-segmentation and screening on the cell nucleus of the adherent cell image by adopting a eukaryotic mode or a pseudo-nuclear mode, and screening the cell nucleus of a target cell under the condition that the cell nucleus of all cells is obtained; S2, cutting the adherent cell semantic segmentation result into different blocks, and the gap between the blocks comprises a cutting gap between cells and a cutting gap in a single cell; S3, fusing the different blocks into a target cell image based on the screened cell nucleus, and completing the first fusion of the target cell; and S4, performing the second fusion on the target cell based on the volume and the contact area of the irregular cell, so that the cutting gap in the single cell is fused and the cutting gap between the cells is retained, and the segmentation of the target cell is completed. The three-dimensional cell segmentation method provided by the application can obtain a good segmentation effect for irregular adherent cells of different cell images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer image processing, and particularly relates to a three-dimensional cell segmentation method. BACKGROUND

[0002] There are various types of immune cells in the liver, such as macrophages, dendritic cells, hepatic stellate cells, etc. The number of immune cells is large, and the shapes are different, and they are in contact with each other. The function of the cell is closely related to the shape. The synaptic structure of the cell has the characteristics of multi-antenna and multi-branch. Through the synaptic structure, the cells are in contact with each other, thereby improving the signal transmission efficiency. The cell instance segmentation method can segment the cells in contact with each other and obtain the accurate shape of a single cell. Through quantitative analysis of a single cell, it is helpful to study the function of the cell. At present, the cell instance segmentation method mainly includes two technical routes, one is a segmentation method based on a neural network, and the other is a segmentation method based on feature engineering.

[0003] The cell instance segmentation method based on the neural network mainly includes three steps: first, collecting large-scale image data; second, manually labeling the image data to make a training data set; and third, using the cell instance segmentation method based on the neural network for supervised learning. In the neural network training process, researchers need to make a large-scale data set so that the neural network can learn more and better. The cell instance segmentation method based on the neural network has made great breakthroughs in two-dimensional tasks. In the three-dimensional cell instance segmentation task, the time spent on labeling a three-dimensional training sample is N times that of labeling a two-dimensional sample, because a three-dimensional sample has N slices in the Z-axis plane, and in the three-dimensional space, the contact between cells is more complex than in the two-dimensional plane. In addition, the parameters of the three-dimensional neural network are usually more than those of the two-dimensional neural network, and in order to achieve better segmentation effect, the three-dimensional neural network needs more training data than the two-dimensional neural network. In order to solve the problem of insufficient data set in the three-dimensional cell instance segmentation task, the Cellpose method can reuse the model and parameters of the two-dimensional neural network without using the three-dimensional data set for neural network training. For approximately circular cells and cell nuclei, the Cellpose method can achieve satisfactory results. However, due to the more complex contact between irregular cells, the segmentation effect of the Cellpose method for irregular cells is significantly reduced. Therefore, the first difficulty of the cell instance segmentation method based on the neural network in segmenting irregular cells is the lack of three-dimensional data set, and the neural network method cannot achieve good segmentation effect.

[0004] In optical fluorescence microscopic imaging, organs and tissues with different spectral absorption properties, fluorescent signals with different expression patterns, and cells with different densities and morphologies are different application scenarios. The second difficulty of the cell instance segmentation method based on neural network is that after the neural network method achieves good results in a certain application scenario, the segmentation effect will decrease sharply when it is replaced in another obviously different scenario. This problem is called training domain bias in the field of machine learning. Due to the training domain bias, we need to re-label a new data set for the new scenario and then train again. Therefore, when the cell instance segmentation method based on neural network is used to segment irregular cells, there are problems of insufficient data set and training domain bias.

[0005] The cell instance segmentation method based on feature engineering, such as the MMQT method (used for segmenting microglial cells). The MMQT method as a whole contains two steps: the first step is to locate the nucleus of the microglial cell; the second step is to separate the microglial cell based on the nucleus. In the first step, the cell body is determined by the cell morphology, and the nucleus is labeled by the DAPI dye. The position of the nucleus can be accurately located by the spatial position relationship between the cell body and the nucleus. The cell body of the microglial cell is easy to determine, and then the position of the nucleus can be accurately located, so the MMQT method can achieve good segmentation effect.

[0006] When the cell body and the nucleus are not easy to locate, strict filtering of the nucleus will result in less filtered nucleus than the real nucleus, thereby causing under-segmentation; on the contrary, it will cause over-segmentation. It is difficult for the MMQT segmentation effect to balance between over-segmentation and under-segmentation.

[0007] In addition, the MMQT method must determine the spatial position of the cell through the nucleus. When the microscope cannot obtain the nucleus image or the nucleus is too dense and adheres together (such as immune cells around blood vessels), the nucleus image cannot be collected, and therefore the MMQT method cannot be used for cell instance segmentation. SUMMARY

[0008] The purpose of the present application is to overcome the defects of the prior art and provide a three-dimensional cell segmentation method which can achieve good segmentation effect for irregularly adhered cells in different cell images.

[0009] To achieve the above purpose, the following specific technical solutions are adopted in the present application:

[0010] The three-dimensional cell segmentation method provided by the present application comprises the following steps:

[0011] S1, over-segmentation and screening of the nucleus of the adhered cell image are carried out by using eukaryotic mode or pseudo-nuclear mode, and the nucleus of the target cell is screened out under the condition that the nuclei of all cells are obtained;

[0012] S2, cutting the adhesion cell semantic segmentation result into different blocks, and the gaps between the blocks including cutting gaps between cells and cutting gaps inside single cells;

[0013] S3, fusing the different blocks into a target cell image based on the screened nuclei, to complete the first fusion of the target cell;

[0014] S4, performing the second fusion of the target cell based on the volume and contact area of the irregular cell, so that the cutting gaps inside single cells are fused and the cutting gaps outside single cells are reserved, to complete the segmentation of the target cell.

[0015] Preferably, in step S1, when the fluorescent image of the nucleus is collected and the nucleus and the cell are independently matched, the eukaryotic mode is selected for the over-segmentation and screening of the nucleus; when the fluorescent image of the nucleus cannot be collected or the nucleus and the cell cannot be independently matched, the pseudo-nucleus mode is selected for the over-segmentation and screening of the nucleus.

[0016] Preferably, when the eukaryotic mode is selected, step S1 specifically includes the following steps:

[0017] S101, labeling all the nuclei of the cells using a nuclear dye;

[0018] S102, performing cell semantic segmentation and nucleus semantic segmentation on the adhesion cell image;

[0019] S102, co-localizing the cell semantic segmentation result and the nucleus semantic segmentation result to determine the nucleus corresponding to the target cell.

[0020] Preferably, when the pseudo-nucleus mode is selected, step S1 specifically includes the following steps:

[0021] S110, performing cell semantic segmentation on the adhesion cell image;

[0022] S111, defining an initial radius and constructing a three-dimensional spherical structure based on the initial radius to perform an erosion operation on each connected domain in the cell semantic segmentation result;

[0023] S112, calculating the volume of the erosion result, if the volume of the erosion result is greater than a set maximum nucleus volume threshold, increasing the radius of the three-dimensional spherical structure, performing the erosion operation on the connected domain again, and meanwhile reserving the results in the erosion result with a volume less than the maximum nucleus volume threshold;

[0024] S113, repeating step S112 until the volume of the erosion result of the connected domain is less than the maximum nucleus volume threshold.

[0025] Preferably, in step S2, the adherent cell semantic segmentation result is cut into different blocks, including the following steps:

[0026] S21, cut the connected domain into catchment basins according to a three-dimensional watershed algorithm;

[0027] S22, fuse the catchment basins into blocks, if the volume of a block is greater than a set maximum volume threshold of a cell synapse, stop the fusion of the catchment basins, otherwise continue the fusion operation until the volume of the fused block approaches the maximum volume threshold of the cell synapse;

[0028] S23, use an undirected graph to represent the spatial relationship between the blocks, use the three-dimensional center coordinates of each block as the vertices of the undirected graph, and assign a number to each block, use the contact relationship between the blocks obtained by the watershed ridge line as the edges of the undirected graph, and use the contact area between the watershed ridge line and the block to obtain the contact area between the blocks as the weight of the edges of the undirected graph.

[0029] Preferably, before step S21 is performed, a pre-step is further included:

[0030] According to the number of cell nuclei in the connected domain, the connected domain is classified, if the number of cell nuclei is zero, the connected domain is deleted, if the number of cell nuclei is one, the block step S21 is not performed, and if the number of cell nuclei is greater than or equal to two, the step S21 is performed.

[0031] Preferably, step S3 specifically includes the following steps:

[0032] S31, overlap each cell nucleus screened out with the blocks in three-dimensional space, find the block with the maximum overlap with each cell nucleus, define it as the seed block of the undirected graph, as the position of the cell nucleus, and define the vertex corresponding to the seed block in the graph structure as the seed vertex;

[0033] S32, perform the fusion operation of the blocks according to the undirected graph in which the seed vertex is defined.

[0034] Preferably, when the undirected graph has two seed vertices, step S32 specifically includes the following steps:

[0035] S321, calculate all connected paths between the two seed vertices according to a path search algorithm;

[0036] S322, find the edge with the minimum weight in each connected path and delete it;

[0037] S323, repeat steps S321 and S322 until there is no connected path between the two seed vertices, thereby disconnecting the undirected graph to form two subgraphs, each of which contains a seed vertex;

[0038] S324, according to the number of blocks, the blocks belonging to the same subgraph are fused to form the target cell image.

[0039] Preferably, when the number of seed vertices in the undirected graph is greater than three, step S32 specifically comprises the following steps:

[0040] S321`, the directly connected seed vertices are retained, and the non-directly connected seed vertices are removed;

[0041] S322`, all connected paths between the directly connected seed vertices are calculated according to a path search algorithm;

[0042] S323`, the edge with the minimum weight in each connected path is found and deleted;

[0043] S324`, steps S322` and S323` are repeated until there is no connected path between the directly connected seed vertices, so as to disconnect the undirected graph to form two subgraphs, each of which contains a seed vertex;

[0044] S325`, according to the number of blocks, the blocks belonging to the same subgraph are fused to form the target cell image.

[0045] Preferably, step S4 specifically comprises the following steps:

[0046] S41, the average volume of a single cell and the fusion rule based on the volume and the contact area are set; wherein, a 26-neighborhood filter is used to find the boundary connected to the target cells for the first fusion, and the number of pixel points on the boundary is counted as the contact area between the target cells.

[0047] S42, when the contact area between the synapses of the two target cells is less than the contact area between the synapses and the cell body, the volume of the synapses is less than the volume of the cell body, and the volume of the two target cells is less than the average volume of a single cell set, it is considered that the fusion rule is met, and the two target cells are fused, otherwise, the two target cells are not fused.

[0048] The present application can achieve the following technical effects:

[0049] 1. Compared with the Cellpose method, the present application can improve the segmentation effect of irregularly adherent cells with complex contact conditions;

[0050] 2. Compared with the Cellpose method, the present application can achieve good segmentation effect for different cell images;

[0051] 3. Compared with the MMQT method, the present application can reduce the dependence on the cell nucleus, and does not have the problems of over-segmentation and under-segmentation, so that the segmentation effect is balanced between over-segmentation and under-segmentation, thereby improving the segmentation accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a flowchart of a three-dimensional cell segmentation method according to an embodiment of the present application;

[0053] Figure 2 is an effect diagram of performing instance segmentation on CX3CR1 in liver, spleen and kidney using pseudo-nuclear mode according to an embodiment of the present application; +

[0054] Figure 3 is an instance segmentation flowchart of a three-dimensional cell segmentation method according to an embodiment of the present application;

[0055] Figure 4 is a diagram of comparing segmentation results of different segmentation methods according to an embodiment of the present application. DETAILED DESCRIPTION

[0056] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. In the following description, the same modules are denoted by the same reference numerals. In the case of the same reference numerals, their names and functions are also the same. Therefore, detailed descriptions thereof will not be repeated.

[0057] In order to make the objectives, technical solutions and advantages of the present application clearer, further detailed descriptions will be made to the present application in combination with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and do not constitute a limitation on the present application.

[0058] For the segmentation of irregularly adherent cells, the present application provides a general morphological-based three-dimensional cell segmentation method, called Crop Once Merge Twice (C1M2), which cuts multiple connected cells once and merges twice to achieve segmentation.

[0059] Figure 1 A flowchart of a three-dimensional cell segmentation method according to an embodiment of the present application is shown.

[0060] As shown in Figure 1 The three-dimensional cell segmentation method provided by the embodiments of the present application includes the following steps:

[0061] S1, perform over-segmentation and screening on the nuclei of the adherent cell image using eukaryotic mode or pseudo-nuclear mode, and screen out the nuclei of the target cell under the condition that the nuclei of all cells are obtained.

[0062] When the fluorescence image of the nucleus can be collected, and the nuclei are rarely adherent, and the nuclei and cells can be independently matched, step S1 is performed using eukaryotic mode. ​

[0063] When the fluorescent image of the cell nucleus cannot be collected or the collected fluorescent image of the cell nucleus is too dense to be matched with the cell independently, the pseudo-nucleus mode is required to perform step S1.

[0064] When the eukaryotic mode is selected, step S1 specifically includes the following steps:

[0065] S101, label all cell nuclei of cells using a cell nucleus dye.

[0066] S102, perform cell semantic segmentation and cell nucleus semantic segmentation on the adherent cell image.

[0067] S102, co-localize the cell semantic segmentation result and the cell nucleus semantic segmentation result to determine the cell nucleus corresponding to the target cell.

[0068] Co-localization refers to performing a logical AND operation on the cell semantic segmentation result and the cell nucleus semantic segmentation result, and through the operation, the cell nuclei of non-target cells can be removed, and only the cell nuclei of target cells are retained.

[0069] When the pseudo-nucleus mode is selected, step S1 specifically includes the following steps:

[0070] S110, perform cell semantic segmentation on the adherent cell image.

[0071] S111, define a small initial radius and construct a three-dimensional spherical structure based on the initial radius to perform an erosion operation on each connected domain in the cell semantic segmentation result.

[0072] S112, calculate the volume of the erosion result, if the volume of the erosion result is greater than the set maximum cell nucleus volume threshold, increase the radius of the three-dimensional spherical structure, and perform the erosion operation on the connected domain again, while retaining the result in the erosion result whose volume is less than the maximum cell nucleus volume threshold.

[0073] S113, repeat step S112 until the volume of the erosion result of the connected domain is less than the maximum cell nucleus volume threshold.

[0074] It should be noted that the above pseudo-nucleus mode is not a traditional pseudo-nucleus mode. The traditional pseudo-nucleus mode is to first create a three-dimensional spherical structure with a fixed radius, and then use the three-dimensional spherical structure to perform an erosion operation on the semantic segmentation result of the cell to screen the cell nucleus. When the morphology and volume of the connected domain of the cell differ greatly, the morphology and volume of the screened cell nucleus will also differ greatly when the erosion operation is performed using the three-dimensional spherical structure with a fixed radius, which does not conform to the morphology of the real cell nucleus, and cannot correctly screen the cell nucleus, thereby reducing the accuracy of subsequent instance segmentation. Therefore, the present application proposes a pseudo-nucleus mode based on limiting the volume of the cell nucleus.

[0075] Figure 2 Figures 2(a) to 2(f) show the results of semantic segmentation and instance segmentation of CX3CR1 in liver, spleen and kidney using pseudo-nucleus mode according to an embodiment of the present application. + The effect of instance segmentation is shown.

[0076] As shown in Figure 2 Figures 2(a) to 2(f), (a) is the semantic segmentation result of CX3CR1 in liver, red represents the semantic segmentation result of cells, and yellow represents pseudo-nucleus; (b) is the instance segmentation result of CX3CR1 in liver, red represents a single independent cell, and other colors represent instance segmentation results; (c) is the semantic segmentation result of CX3CR1 in spleen, red represents the semantic segmentation result of cells, and yellow represents pseudo-nucleus; (d) is the instance segmentation result of CX3CR1 in spleen, red represents a single independent cell, and other colors represent instance segmentation results; (e) is the semantic segmentation result of CX3CR1 in kidney, red represents the semantic segmentation result of cells, and yellow represents pseudo-nucleus; and (f) is the instance segmentation result of CX3CR1 in kidney, red represents a single independent cell, and other colors represent instance segmentation results. + + + + + +

[0077] No matter in eukaryotic mode or pseudo-nucleus mode, the number of detected nuclei exceeds the true situation, so as to ensure that all nuclei can be detected without missing any nuclei. The over-segmentation result of nuclei will not reduce the accuracy of cell instance segmentation after subsequent cutting and fusion operations.

[0078] S2, cutting the adherent cell semantic segmentation result into different blocks, and the gaps between the blocks include cutting gaps between cells and cutting gaps inside a single cell.

[0079] In step S2, the adherent cell semantic segmentation result is cut into different blocks, including the following steps:

[0080] S21, cutting the connected domain into catchment basins according to a three-dimensional watershed algorithm.

[0081] The three-dimensional watershed algorithm is prior art, and thus will not be described here.

[0082] The purpose of cutting the connected domain is to generate a large number of cutting lines, which can be distributed on the inside of the cells and the contact surface of the cells.

[0083] ​​​​​​Before the cutting operation of the connected domain is performed, the connected domain is first classified according to the number of cell nuclei in the connected domain. When the number of cell nuclei is 0, it is indicated that the connected domain is likely to be noise or an incomplete cell, and the connected domain needs to be deleted; when the number of cell nuclei is 1, it is indicated that the connected domain only contains one cell, and therefore the subsequent cutting and fusion operations do not need to be performed; when the number of cell nuclei is greater than 2, it is indicated that the connected domain contains multiple cells, and the subsequent cutting and fusion operations are performed.

[0084] S22, the catchment basin is fused into a block, and if the volume of the block is greater than a set maximum volume threshold of a cell synapse, the fusion of the catchment basin is stopped, otherwise the fusion operation is continued until the volume of the fused block approaches the maximum volume threshold of the cell synapse.

[0085] S23, the spatial relationship between the blocks is represented by using an undirected graph, the three-dimensional center coordinates of each block are taken as the vertices of the undirected graph, and each block is assigned a number, the contact relationship between the blocks obtained by using the watershed ridge line is taken as the edges of the undirected graph, and the contact area between the blocks obtained by using the contact area between the watershed ridge line and the blocks is taken as the weight of the edges of the undirected graph.

[0086] It should be noted that, since the volume of the catchment basin is small, the use of the catchment basin is beneficial to accurately finding the contact position of the connected cells. The number of catchment basins is one-to-one corresponding to the number of vertices in the undirected graph. Therefore, the increase in the number of catchment basins will lead to the increase in the number of vertices in the undirected graph, and the time complexity will sharply rise when the graph search algorithm is used subsequently. Although the fusion of the catchment basins into blocks can reduce the number of vertices in the undirected graph and shorten the calculation time, the fusion of the catchment basins at the cell contact position will reduce the accuracy of cell segmentation. In order to balance the segmentation accuracy and running efficiency, a fusion algorithm with volume limitation is adopted, that is, the catchment basins are fused into blocks with a suitable volume. When the volume of the block exceeds the pre-set maximum volume threshold of the cell synapse, the fusion of the catchment basins is stopped, otherwise the fusion operation is continued until the fused volume approaches the set maximum volume threshold of the cell synapse.

[0087] S3, different blocks are fused into a target cell image based on the screened cell nuclei, and the first fusion of the target cell is completed.

[0088] Step S3 specifically includes the following steps:

[0089] S31, each cell nucleus screened is overlapped with a block in a three-dimensional space, the block with the maximum overlap with each cell nucleus is found, and is defined as a seed block of an undirected graph, as the position of the cell nucleus; meanwhile, the vertex corresponding to the seed block in the graph structure is defined as a seed vertex.

[0090] S32, performing the fusion operation of the block according to the undirected graph defined by the seed vertices.

[0091] When the undirected graph has 2 seed vertices, the fusion operation of the block is realized by the following steps:

[0092] S321, calculating all the connected paths between the two seed vertices according to the path search algorithm.

[0093] S322, finding the edge with the minimum weight in each connected path and deleting it.

[0094] Since the weight of the edge represents the contact area between the blocks, the edge with the minimum weight represents the weakest connection between the two cells, which is considered as the connection position of the two cells, so the edge with the minimum weight needs to be deleted.

[0095] S323, repeating step S321 and step S322 until there is no connected path between the two seed vertices, thereby disconnecting the undirected graph to form two subgraphs, each of which contains one seed vertex.

[0096] When there is no connected path between the two seed vertices, it means that the connected two seed vertices are completely disconnected. After disconnecting all the connected seed vertices, the undirected graph is disconnected to form two subgraphs, each of which contains one seed vertex.

[0097] S324, according to the block number, merging the blocks belonging to the same subgraph to form the target cell image.

[0098] According to the block number, find the blocks corresponding to each subgraph, and merge the blocks belonging to one subgraph to finally form two independent connected domains, and each of these independent connected domains is allocated a screened nucleus.

[0099] When the number of seed vertices in the undirected graph is greater than 3, it is necessary to first determine the direct connection relationship between the seed vertices (i.e. there is no other seed vertex on the shortest path between the seed vertices), and then use steps S321 to S324 to fuse each pair of seed vertices. In order to determine the direct connection relationship between the seed vertices, the first step is to use the shortest path search algorithm to calculate the shortest path between each seed vertex and other seed vertices, and the second step is to judge whether the shortest path contains other seed vertices in addition to the two seed vertices; if the shortest path contains other seed vertices, it means that this pair of seed vertices does not belong to the direct connection relationship and needs to be discarded; if the shortest path does not contain other seed vertices, it means that this pair of seed vertices belongs to the direct connection relationship. By traversing all the seed vertices, all the direct connection relationships between the seed vertices are calculated, and then the first round of cell fusion is performed.

[0100] S4, based on the volume and contact area of the irregular cells, the second fusion of the target cells is performed, so that the cutting gap inside the single cell is fused and the cutting gap outside the single cell is retained, and the segmentation of the target cells is completed.

[0101] The first round of fusion is based on the screened cell nucleus. Since the number of cell nuclei exceeds the number of real cells, the over-segmentation result of the first round output is solved by using a second round of fusion operation. In the second round of fusion, a 26-neighbor filter is first used to find the boundary connected with the first round of fusion cells, and the number of pixel points on the boundary is counted as the contact area between the target cells.

[0102] Irregular cells are connected through cell synapses, so the contact area between cells is usually smaller than the contact area inside the cell. Based on this assumption, the application sets a fusion rule based on volume and contact area for the second round of fusion, so as to fuse the cells with small cell volume and large cell contact area again.

[0103] In the second round of fusion, two parameters need to be set: the first is the average volume of a single cell, which estimates the number of cells contained in the connected domain. When the number of isolated cells in the image is large, the isolated cells can be automatically calculated to automatically set; the second parameter is the fusion rule based on volume and contact area. When cells are connected through synapses, there are two obvious characteristics: one is that the contact area between synapses is smaller than the contact area between synapses and cell bodies; the other is that the volume of synapses is smaller than the volume of cell bodies. Therefore, the application sets a fusion rule based on the contact area and volume of cells. When the contact area between the synapses of the two target cells is smaller than the contact area between the synapses and the cell bodies, the volume of the synapses is smaller than the volume of the cell bodies, and the volume of the two target cells is smaller than the average volume of a single cell set, the fusion rule is satisfied, and it is considered that the two cells are one cell, and the two cells need to be fused, otherwise, the two target cells are not fused.

[0104] The second parameter setting is relatively complex, so the application provides a visual parameter setting platform to adjust the second parameter by observing the fusion results of some cells. When the cell type and microscope mode are changed, the above two parameters need to be adjusted again to obtain good segmentation accuracy.

[0105] Figure 3 An example segmentation process of the three-dimensional cell segmentation method provided by the embodiment of the application is shown.

[0106] As shown in Figure 3 (a) is a three-dimensional liver immunofluorescence image, blue represents DAPI (cell nucleus dye), red represents F4 / 80 + (cell antibody), and green represents CD11c+ Cells; (b) is F4 / 80 + (c) shows the semantic segmentation results of cells; (d) shows the semantic segmentation results of cell nuclei; and (e) shows the F4 / 80 cells selected using the eukaryotic model. + Cells; (e) F4 / 80 cells screened using pseudonucleus pattern. + (f) is a connected component containing multiple cells in the semantic segmentation result; (g) is the segmentation of the connected component into blocks of different sizes using the 3D watershed algorithm; (h) is the representation of the contact relationship and contact area between blocks using an undirected graph, where nodes in the undirected graph represent the spatial location of the blocks, edges in the undirected graph represent the contact relationship between blocks, and non-blue edges represent the minimum contact area between seed vertices; (i) is the first round of fusion, which transforms an undirected graph into multiple subgraphs by deleting non-blue edges; (j) is the fusion of blocks based on the seed vertices in the subgraphs; (k) is the second round of fusion based on cell volume and fusion area, where each color represents a cell; (l) is the final segmentation result, where red represents non-adhesive cells and other colors represent adhesive cells.

[0107] Figure 4 The following diagram illustrates the segmentation results of different segmentation methods provided according to embodiments of the present invention.

[0108] like Figure 4 As shown, (a) demonstrates the use of the MMQT method, CellPose method, and C1M2 method to optimize F4 / 80. + Cell segmentation results; in the original image, green represents F4 / 80. + Cells, red represents DAPI; in the segmentation results of the MMQT method, CellPose method and C1M2 method, red represents isolated cells, and other colors represent segmented connected cells; (b) shows connected F4 / 80 cells. + Cells; (c) demonstrates the use of the MMQT, CellPose, and C1M2 methods to target CD11c cells. + Cell segmentation results; in the original image, green represents CD11c. + Cells, red represents DAPI; in the segmentation results of the MMQT, CellPose, and C1M2 methods, red represents isolated cells, and other colors represent segmented connected cells; (d) shows connected CD11c cells. + cell.

[0109] from Figure 4 As can be seen, the accuracy of the C1M2 algorithm is significantly higher than that of the traditional MMQT and CellPose methods. At the same time, the C1M2 algorithm can achieve good segmentation results for different images.

[0110] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, different embodiments or examples described in the specification and the features of different embodiments or examples can be combined and combined by those skilled in the art without contradiction.

[0111] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

[0112] The specific embodiments of the application described above do not constitute a limitation on the scope of protection of the present application. Any various other corresponding changes and modifications made in accordance with the technical concept of the present application shall be included in the scope of protection of the claims of the present application.

Claims

1. A three-dimensional cell image segmentation method, characterized by, Comprise the following steps: S1, adopt eukaryotic model or pseudo-nuclear model to carry out over segmentation and screening to the nucleus of the adherent cell image, in the case of obtaining the nucleus of all cells, screen out the nucleus of the target cell; In step S1, when the fluorescent image of the nucleus is collected, and the nucleus and the cell are independently matched, the eukaryotic model is selected for over segmentation and screening of the nucleus; When the fluorescent image of the nucleus cannot be collected or the nucleus and the cell cannot be independently matched, the pseudo-nuclear model is selected for over segmentation and screening of the nucleus; When the eukaryotic model is selected, step S1 specifically comprises the following steps: S101, label the nucleus of all cells using nuclear dye; S102, carry out cell semantic segmentation and nucleus semantic segmentation on the adherent cell image; S102, co-localize the cell semantic segmentation result and the nucleus semantic segmentation result to determine the nucleus corresponding to the target cell; When the pseudo-nuclear model is selected, step S1 specifically comprises the following steps: S110, carry out cell semantic segmentation on the adherent cell image; S111, define an initial radius and construct a three-dimensional spherical structure based on the initial radius to carry out erosion operation on each connected domain in the cell semantic segmentation result; S112, calculate the volume of the erosion result, if the volume of the erosion result is greater than the set maximum volume threshold of the nucleus, increase the radius of the three-dimensional spherical structure, and carry out erosion on the connected domain again, while retaining the results in the erosion result whose volume is less than the maximum volume threshold of the nucleus; S113, repeat step S112 until the volume of the erosion result of the connected domain is less than the maximum volume threshold of the nucleus; S2, cut the adherent cell semantic segmentation result into different blocks, and the gap between the blocks includes the cutting gap between the cells and the cutting gap inside the single cell; S3, fuse different blocks into a target cell image based on the screened nucleus, complete the first fusion of the target cell; S4, based on the volume and contact area of the irregular cell, the target cell is fused for the second time, so that the cutting gap inside the single cell is fused and the cutting gap outside the single cell is retained, and the segmentation of the target cell is completed.

2. The three-dimensional cell image segmentation method according to claim 1, wherein, In step S2, the adherent cell semantic segmentation result is cut into different blocks, comprising the following steps: S21, cut the connected domain into a catchment basin according to the three-dimensional watershed algorithm; S22, fuse the catchment basin into a block, if the volume of the block is greater than the set maximum volume threshold of the cell synapse, stop the fusion of the catchment basin, otherwise continue the fusion operation until the volume of the fused block approaches the maximum volume threshold of the cell synapse; S23, use an undirected graph to represent the spatial relationship between the blocks, use the three-dimensional center coordinates of each block as the vertices of the undirected graph, and assign a number to each block, use the watershed ridge line to obtain the contact relationship between the blocks as the edges of the undirected graph, and use the contact area between the watershed ridge line and the block to obtain the contact area between the blocks as the weight of the edge of the undirected graph.

3. The three-dimensional cell image segmentation method according to claim 2, wherein, Before step S21 is executed, there is a pre-step: According to the number of the cell nuclei in the connected domain, if the number of the cell nuclei is zero, the connected domain is deleted, if the number of the cell nuclei is one, block step S21 is not executed, if the number of the cell nuclei is greater than or equal to two, step S21 is executed.

4. The three-dimensional cell image segmentation method according to claim 2, wherein, Step S3 specifically comprises the following steps: S31, each cell nucleus screened out is overlapped with the blocks in the three-dimensional space, the block with the maximum overlap with each cell nucleus is found, defined as the seed block of the undirected graph, as the position of the cell nucleus; meanwhile, the vertex corresponding to the seed block in the graph structure is defined as the seed vertex; S32, the fusion operation of the blocks is performed according to the undirected graph with the seed vertex defined.

5. The three-dimensional cell image segmentation method according to claim 4, wherein, When the undirected graph has two seed vertices, step S32 specifically comprises the following steps: S321, all the connected paths between the two seed vertices are calculated according to the path search algorithm; S322, the edge with the minimum weight in each connected path is found and deleted; S323, steps S321 and S322 are repeated until there is no connected path between the two seed vertices, thereby disconnecting the undirected graph to form two subgraphs, each of which contains one seed vertex; S324, the blocks belonging to the same subgraph are fused to form the target cell image according to the block number.

6. The three-dimensional cell image segmentation method according to claim 4, wherein, When the number of the seed vertices in the undirected graph is greater than three, step S32 specifically comprises the following steps: S321`, the directly connected seed vertices are retained and the non-directly connected seed vertices are removed; S322`, all the connected paths between the directly connected seed vertices are calculated according to the path search algorithm; S323`, the edge with the minimum weight in each connected path is found and deleted; S324`, steps S322` and S323` are repeated until there is no connected path between the directly connected seed vertices, thereby disconnecting the undirected graph to form two subgraphs, each of which contains one seed vertex; S325`, the blocks belonging to the same subgraph are fused to form the target cell image according to the block number.

7. The three-dimensional cell image segmentation method of claim 1, wherein, Step S4 specifically comprises the following steps: S41, the average volume of a single cell and the fusion rule based on the volume and the contact area are set; wherein, a 26-neighborhood filter is used to find the boundary connected with the target cell in the first fusion, the number of the pixel points on the boundary is counted as the contact area between the target cells; S42, when the contact area between the synapses of the two target cells is less than the contact area between the synapses and the cell body, the volume of the synapses is less than the volume of the cell body, and the volume of the two target cells is less than the average volume of a single cell set, it is considered that the fusion rule is satisfied, the two target cells are fused, otherwise, the two target cells are not fused.

Images