A method and system for cell nucleus segmentation using fluorescence in situ hybridization

Abstract

This invention discloses a method and system for fluorescence in situ hybridization (FISH) cell nucleus segmentation. This method uses a trained LinkNet network to segment FISH cell images, extracting high-level semantic features for abstraction. When the decoder restores the feature map resolution, it fuses the corresponding shallow-level information to capture detailed features, thereby better extracting the desired target cell foreground. The trained U-Net network is used to obtain the cell center point, which is then integrated into the watershed algorithm for secondary segmentation of the adhesion region, addressing the problem of cell adhesion and improving the accuracy of cell segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of cell segmentation, and in particular to a fluorescence in situ hybridization cell nucleus segmentation method and system. Background Art

[0002] Fluorescence in situ hybridization (FISH) is a molecular cytogenetic technique that can provide reliable imaging biomarkers for diagnosing cancer and genetic diseases. Segmentation of FISH cells is a prerequisite for quantitative analysis of these imaging biomarkers. However, cells often adhere to each other in these numerous cell images, making it difficult for many automated segmentation algorithms to accurately segment foreground cells from FISH cell images.

[0003] How to overcome the phenomenon of cell adhesion and accurately segment foreground cells from FISH cell images has become a technical problem that needs to be solved urgently. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for segmenting cell nuclei in fluorescence in situ hybridization to overcome the phenomenon of cell adhesion and accurately segment foreground cells from FISH cell images.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for segmenting cell nuclei using fluorescence in situ hybridization, comprising the following steps:

[0007] The trained LinkNet network is used to segment the fluorescence in situ hybridization cell images to obtain segmented fluorescence in situ hybridization cell images;

[0008] The trained U-net network is used to identify the cell center of the segmented fluorescence in situ hybridization cell images to obtain the cell center probability map;

[0009] According to the segmented fluorescence in situ hybridization cell image and the cell center point probability map, determining that an independent region in the segmented fluorescence in situ hybridization cell image in which the number of cell center points is greater than 1 is an adhesion region;

[0010] According to the number and distribution of cell center points in the cell center probability map of the adhesion area, the watershed algorithm was used to perform secondary segmentation of the adhesion area.

[0011] Optionally, the trained LinkNet network includes: an encoder and a decoder;

[0012] The encoder includes two convolutional layers and three residual modules, the two convolutional layers are respectively a first convolutional layer and a second convolutional layer, the three residual modules are respectively a first residual module, a second residual module and a third residual module, and the first convolutional layer, the second convolutional layer, the first residual module, the second residual module and the third residual module are connected in sequence;

[0013] The decoder includes three upsampling modules and three convolutional layers, the three upsampling modules are respectively a first upsampling module, a second upsampling module and a third upsampling module, and the three convolutional layers are respectively a third convolutional layer, a fourth convolutional layer and a fifth convolutional layer;

[0014] The input end of the first upsampling module is connected to the output end of the third residual module, the output end of the first upsampling module is connected to the first input end of the second upsampling module, and the first upsampling module is configured to perform a convolution operation on the output image of the third residual module; the second input end of the second upsampling module is connected to the output end of the second residual module, the output end of the second upsampling module is connected to the first input end of the third upsampling module, and the second upsampling module is configured to add the output image of the second residual module and the output image of the first upsampling module, and then perform a convolution operation on the result of the addition operation; the second input end of the third upsampling module is connected to the output end of the first residual module, the output end of the third upsampling module is connected to the input end of the third convolutional layer, and the third upsampling module is configured to add the output image of the first residual module and the output image of the second upsampling module, and then perform a convolution operation on the result of the addition operation; the output end of the third convolutional layer is connected to the input end of the fourth convolutional layer, and the output end of the fourth convolutional layer is connected to the input end of the fifth convolutional layer.

[0015] Optionally, the residual module includes a sixth convolutional layer, a seventh convolutional layer and an adder;

[0016] The output end of the sixth convolutional layer is connected to the input end of the seventh convolutional layer, the output end of the seventh convolutional layer is connected to the first input end of the adder, the input end of the sixth convolutional layer is connected to the second input end of the adder, the input end of the sixth convolutional layer serves as the input end of the residual module, and the output end of the adder serves as the output end of the residual module.

[0017] Optionally, the step of segmenting the fluorescence in situ hybridization cell image using the trained LinkNet network to obtain the segmented fluorescence in situ hybridization cell image further includes:

[0018] Acquiring a plurality of fluorescence in situ hybridization cell image samples, and manually segmenting the fluorescence in situ hybridization cell image samples to establish a first training sample set including the fluorescence in situ hybridization cell image samples and training samples of manual segmentation results;

[0019] Use Xaviers to initialize the linknet network;

[0020] The BCE loss function is used as the target loss function of the LinkNet network, and the initialized LinkNet network is trained using the first training sample set until the value of the target loss function of the LinkNet network is less than a first preset threshold, and the trained LinkNet network is output.

[0021] Optionally, the method further includes: using a trained U-net network to identify cell centers in the segmented fluorescence in situ hybridization cell image to obtain a cell center probability map;

[0022] Acquire multiple fluorescence in situ hybridization cell image samples, and manually identify and mark the cell center points of the fluorescence in situ hybridization cell image samples to establish a second training sample set including the fluorescence in situ hybridization cell image samples and training samples of manual identification results;

[0023] The Dice loss function is used as the target loss function of the U-net network, and the U-net network is trained using the second training sample set until the value of the target loss function of the U-net network is less than the second preset threshold, and the trained U-net network is output.

[0024] A fluorescence in situ hybridization cell nucleus segmentation method system, the segmentation system comprising:

[0025] An image segmentation module is used to segment the fluorescence in situ hybridization cell image using the trained LinkNet network to obtain a segmented fluorescence in situ hybridization cell image;

[0026] The center point recognition module is used to identify the cell center points of the segmented fluorescence in situ hybridization cell images using the trained U-net network to obtain the cell center point probability map;

[0027] According to the segmented fluorescence in situ hybridization cell image and the cell center point probability map, determining that an independent region in the segmented fluorescence in situ hybridization cell image in which the number of cell center points is greater than 1 is an adhesion region;

[0028] The secondary segmentation module is used to perform secondary segmentation of the adhesion area using a watershed algorithm according to the number and distribution of cell center points in the cell center probability map of the adhesion area.

[0029] Optionally, the trained LinkNet network includes: an encoder and a decoder;

[0030] The encoder includes two convolutional layers and three residual modules, the two convolutional layers are respectively a first convolutional layer and a second convolutional layer, the three residual modules are respectively a first residual module, a second residual module and a third residual module, and the first convolutional layer, the second convolutional layer, the first residual module, the second residual module and the third residual module are connected in sequence;

[0031] The decoder includes three upsampling modules and three convolutional layers, the three upsampling modules are respectively a first upsampling module, a second upsampling module and a third upsampling module, and the three convolutional layers are respectively a third convolutional layer, a fourth convolutional layer and a fifth convolutional layer;

[0032] The input end of the first upsampling module is connected to the output end of the third residual module, the output end of the first upsampling module is connected to the first input end of the second upsampling module, and the first upsampling module is configured to perform a convolution operation on the output image of the third residual module; the second input end of the second upsampling module is connected to the output end of the second residual module, the output end of the second upsampling module is connected to the first input end of the third upsampling module, and the second upsampling module is configured to add the output image of the second residual module and the output image of the first upsampling module, and then perform a convolution operation on the result of the addition operation; the second input end of the third upsampling module is connected to the output end of the first residual module, the output end of the third upsampling module is connected to the input end of the third convolutional layer, and the third upsampling module is configured to add the output image of the first residual module and the output image of the second upsampling module, and then perform a convolution operation on the result of the addition operation; the output end of the third convolutional layer is connected to the input end of the fourth convolutional layer, and the output end of the fourth convolutional layer is connected to the input end of the fifth convolutional layer.

[0033] Optionally, the residual module includes a sixth convolutional layer, a seventh convolutional layer and an adder;

[0034] The output end of the sixth convolutional layer is connected to the input end of the seventh convolutional layer, the output end of the seventh convolutional layer is connected to the first input end of the adder, the input end of the sixth convolutional layer is connected to the second input end of the adder, the input end of the sixth convolutional layer serves as the input end of the residual module, and the output end of the adder serves as the output end of the residual module.

[0035] Optionally, the segmentation system further includes:

[0036] a first training sample set establishment module, configured to acquire a plurality of fluorescence in situ hybridization cell image samples, and manually segment the fluorescence in situ hybridization cell image samples to establish a first training sample set including the fluorescence in situ hybridization cell image samples and training samples of manual segmentation results;

[0037] Linknet network initialization module, used to initialize the Linknet network using Xaviers;

[0038] The linknet network training module is configured to use the BCE loss function as the target loss function of the linknet network, train the initialized linknet network using the first training sample set until the value of the target loss function of the linknet network is less than a first preset threshold, and output the trained linknet network.

[0039] Optionally, the segmentation system further includes:

[0040] a second training sample set establishment module, configured to obtain a plurality of fluorescence in situ hybridization cell image samples, and manually identify and label the cell center points of the fluorescence in situ hybridization cell image samples, thereby establishing a second training sample set including the fluorescence in situ hybridization cell image samples and training samples of the manual identification results;

[0041] A U-net network training module is used to use the Dice loss function as the target loss function of the U-net network, train the U-net network using the second training sample set until the value of the target loss function of the U-net network is less than the second preset threshold, and output the trained U-net network.

[0042] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0043] The present invention discloses a method and system for segmenting fluorescence in situ hybridization cell nuclei. The segmentation method comprises the following steps: firstly, using a trained LinkNet network to segment a fluorescence in situ hybridization cell image to obtain a segmented fluorescence in situ hybridization cell image; then, using a trained U-Net network to identify cell centers in the segmented fluorescence in situ hybridization cell image to obtain a cell center probability map; then, based on the segmented fluorescence in situ hybridization cell image and the cell center probability map, determining that an independent region in the segmented fluorescence in situ hybridization cell image in which the number of cell centers is greater than 1 is an adhesion region; and finally, based on the number and distribution of cell centers in the adhesion region in the cell center probability map, using a watershed algorithm to perform secondary segmentation on the adhesion region. The present invention uses a trained LinkNet network to segment fluorescence in situ hybridization cell images to extract high-level semantic features for abstraction, and fuses the corresponding shallow information to capture detailed features when the decoder restores the feature map resolution, so as to better extract the required target cell foreground. The cell center point is obtained through the trained U-Net network, and the cell center point is integrated into the watershed algorithm to perform secondary segmentation on the adhesion area, solving the problem of mutual adhesion between cells and improving the accuracy of cell segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 A flow chart of a fluorescence in situ hybridization cell nucleus segmentation method provided by the present invention;

[0046] Figure 2 A schematic diagram of a fluorescence in situ hybridization cell nucleus segmentation method provided by the present invention;

[0047] Figure 3 The linknet network structure diagram provided by the present invention;

[0048] Figure 4 The schematic diagram of the residual module provided by the present invention, wherein: Figure 4 (a) is the schematic diagram of the first residual module, Figure 4 (b) is the schematic diagram of the second residual module. Figure 4 (c) is the schematic diagram of the third residual module;

[0049] Figure 5 The network structure diagram of U-net provided by the present invention;

[0050] Figure 6 A schematic diagram of the convolution operation principle provided by the present invention;

[0051] Figure 7 Schematic diagram of the deconvolution operation principle provided by the present invention;

[0052] Figure 8 This is a comparison diagram of the segmentation effect provided by the present invention, wherein: Figure 8 (a) is a fluorescence in situ hybridization image of cells to be segmented using the present invention, Figure 8 (b) Schematic diagram of the watershed segmentation effect based on prior information of the present invention, Figure 8 (c) is a schematic diagram of the watershed segmentation effect using no prior information. DETAILED DESCRIPTION

[0053] The purpose of the present invention is to provide a method and system for segmenting cell nuclei in fluorescence in situ hybridization to overcome the phenomenon of cell adhesion and accurately segment foreground cells from FISH cell images.

[0054] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] To effectively separate foreground cells and cell adhesion, the present invention proposes the LinkNet algorithm. This algorithm's network structure is similar to that of the U-Net, employing a symmetrical encoder-decoder topology. However, unlike the original network, the encoder uses a ResNet network to abstract high-level semantic features. The decoder then integrates the corresponding shallow-level information to capture detailed features when restoring the feature map resolution, thereby better extracting the desired target cell foreground. Furthermore, to address the issue of cell adhesion, the present invention trains a deep learning model to obtain cell center points to isolate individual cells. Cell center point identification primarily identifies the centroid of each cell. In addition to using different loss functions, the cell center point network utilizes network-learnable deconvolution to enhance the density of feature responses in the output feature map. Finally, integrating the cell centroids into the watershed algorithm significantly improves cell segmentation accuracy. Experimental results demonstrate that the proposed method can effectively separate adherent cells and achieve a moderate improvement in segmentation accuracy.

[0056] To achieve the above object, the present invention provides the following solutions:

[0057] like Figure 1 and 2The present invention provides a method for segmenting cell nuclei by fluorescence in situ hybridization, the segmentation method comprising the following steps:

[0058] Step 101: Use the trained LinkNet network to segment the fluorescence in situ hybridization cell image to obtain a segmented fluorescence in situ hybridization cell image.

[0059] Inspired by the U-Net and ResNet network structures, the present invention uses a network topology called LinkNet, which can better segment the FISH cell images of the present invention. The network diagram is shown in FIG. Figure 3 As shown:

[0060] Figure 3 The convolution in the figure represents a 3×3 convolution operation. By using upsampling operations and combining the addition operations indicated by the gray arrows, more underlying information can be effectively incorporated when restoring the resolution of the corresponding feature map. This allows the network to output a more detailed result map. Module 1 is the first residual module, module 2 is the second residual module, and module 3 is the third residual module. In addition, Figure 3 As shown, in the encoding part, module 1, module 2, and module 3 represent three residual modules. The structures of the first residual module, the second residual module, and the third residual module are respectively as follows: Figure 4 (a), 4(b) and 4(c).

[0061] The selection of activation function is also very important for convolutional neural networks. The nonlinear unit activation function selected in this invention is ELU, which is given by formula (1):

[0062]

[0063] Where α = 1 is a constant. The ELU activation function can effectively alleviate the gradient diffusion during back propagation.

[0064] In this linknet network structure, the objective function we choose is the BCE loss function, which is defined as shown in (2).

[0065]

[0066] Where x represents the probability map of each point in the final output belonging to a cell, y represents its corresponding label, and N is the total number of pixels in the input image in each batch. By continuously iteratively updating LinkNet using the objective function, the final output data distribution of the network tends to the distribution in the label image.

[0067] In convolutional neural networks, good weight initialization is crucial. Otherwise, some parts of the network may have excessively large weights, preventing other parts from contributing effectively. In our LinkNet model, we use Xaviers to initialize the network parameters of each layer.

[0068] The trained linknet network includes: encoder and decoder;

[0069] The encoder includes two convolutional layers and three residual modules, the two convolutional layers are respectively a first convolutional layer and a second convolutional layer, the three residual modules are respectively a first residual module, a second residual module and a third residual module, and the first convolutional layer, the second convolutional layer, the first residual module, the second residual module and the third residual module are connected in sequence;

[0070] The decoder includes three upsampling modules and three convolutional layers, the three upsampling modules are respectively a first upsampling module, a second upsampling module and a third upsampling module, and the three convolutional layers are respectively a third convolutional layer, a fourth convolutional layer and a fifth convolutional layer;

[0071] The input end of the first upsampling module is connected to the output end of the third residual module, the output end of the first upsampling module is connected to the first input end of the second upsampling module, and the first upsampling module is configured to perform a convolution operation on the output image of the third residual module; the second input end of the second upsampling module is connected to the output end of the second residual module, the output end of the second upsampling module is connected to the first input end of the third upsampling module, and the second upsampling module is configured to add the output image of the second residual module and the output image of the first upsampling module, and then perform a convolution operation on the result of the addition operation; the second input end of the third upsampling module is connected to the output end of the first residual module, the output end of the third upsampling module is connected to the input end of the third convolutional layer, and the third upsampling module is configured to add the output image of the first residual module and the output image of the second upsampling module, and then perform a convolution operation on the result of the addition operation; the output end of the third convolutional layer is connected to the input end of the fourth convolutional layer, and the output end of the fourth convolutional layer is connected to the input end of the fifth convolutional layer.

[0072] Optionally, the residual module includes a sixth convolutional layer, a seventh convolutional layer and an adder;

[0073] The output end of the sixth convolutional layer is connected to the input end of the seventh convolutional layer, the output end of the seventh convolutional layer is connected to the first input end of the adder, the input end of the sixth convolutional layer is connected to the second input end of the adder, the input end of the sixth convolutional layer serves as the input end of the residual module, and the output end of the adder serves as the output end of the residual module.

[0074] The method uses the trained LinkNet network to segment the fluorescence in situ hybridization cell image to obtain the segmented fluorescence in situ hybridization cell image, and also includes: obtaining multiple fluorescence in situ hybridization cell image samples, manually segmenting the fluorescence in situ hybridization cell image samples, and establishing a first training sample set including the fluorescence in situ hybridization cell image samples and training samples of the manual segmentation results; using Xaviers to initialize the LinkNet network; using the BCE loss function as the target loss function of the LinkNet network, using the first training sample set to train the initialized LinkNet network until the value of the target loss function of the LinkNet network is less than a first preset threshold, and outputting the trained LinkNet network.

[0075] Step 102: Use the trained U-net network to identify the cell center points of the segmented fluorescence in situ hybridization cell image to obtain a cell center point probability map.

[0076] In the cell center recognition model, the present invention uses the U-net network structure, such as Figure 5 . The network structure presents an encoding-decoding topology. During the encoding process, 3×3 convolution operations are repeatedly used, and each convolution operation will be followed by a nonlinear unit (ReLU) and batch normalization operation. When performing maximum pooling, the number of feature map channels will increase exponentially. This is because the size of the feature map is reduced, and increasing the number of convolution channels can effectively reduce the loss of information. The decoding stage of U-Net can be regarded as a shape generator, and its main function is to restore the required target cell center area. And in the decoding stage, we did not simply apply linear interpolation to increase the size of the feature map. A learnable convolution (transposed convolution) is used to upsample the feature layer map of the previous layer. Its output is an enlarged and dense activation map. The operation form is as follows: Figure 6 As shown in the figure, to better locate the cell center, the network structure combines the high-resolution feature map from the encoding phase to recover the feature map size, capturing detailed information from the underlying information. A 3×3 convolution is then used to refine the output feature results based on this information. The network's final output is a probability map of the same size as the input image, indicating the probability that each pixel belongs to the cell center.

[0077] Figure 5 U-net architecture. Figure 5 In the figure, each grayscale box corresponds to a multi-channel feature map, and the number of channels is indicated at the top of these boxes. The size of the feature map is provided at the lower left edge of the box. The box without grayscale represents the feature map information copied from the encoding part, and the arrows represent different operations. The schematic diagram of the convolution operation principle is shown in the figure. Figure 6 As shown, the deconvolution operation diagram is as follows Figure 7 shown.

[0078] Furthermore, deep convolutional neural networks are often difficult to optimize due to the problem of internal covariate shift; the input distribution in each layer changes iteratively during training as the parameters of its previous layers are updated. Optimizing deeper networks is often problematic because changes in distribution are amplified step by step as they propagate across layers. Therefore, we perform batch normalization after each activation function to reduce internal covariate shift by normalizing the input distribution of each layer to a standard Gaussian distribution. We observe that batch normalization plays an important role in optimizing our network structure.

[0079] In the center point region recognition model, the present invention uses the Dice loss function to optimize our model. The loss function can be defined as (3):

[0080]

[0081] Where N is the number of images fed into each batch, Y j is the true label of each sample, P j Represents the final predicted probability map.

[0082] The training process is as follows: obtaining multiple fluorescence in situ hybridization cell image samples, and manually identifying and labeling the cell center points of the fluorescence in situ hybridization cell image samples, establishing a second training sample set including the fluorescence in situ hybridization cell image samples and the training samples of the manual identification results; using the Dice loss function as the target loss function of the U-net network, using the second training sample set to train the U-net network until the value of the target loss function of the U-net network is less than a second preset threshold, and outputting the trained U-net network.

[0083] Step 103 : determining, based on the segmented FISH cell image and the cell center probability map, that an independent region in which the number of cell center points in the segmented FISH cell image is greater than 1 is an adhesion region.

[0084] Step 104 : Based on the number and distribution of cell center points in the cell center probability map of the adhesion area, a watershed algorithm is used to perform secondary segmentation on the adhesion area.

[0085] Watershed algorithms are often used to separate different regions that are adhered to each other. However, general watershed algorithms are easily affected by small noises in the image and may cause over-segmentation due to the presence of multiple extreme points in a connected region. The practicality of these results is usually very low. To address the above problems, the present invention first uses the cell center point regions obtained from the center point recognition model and uses these regions as prior information to guide the marker-based watershed algorithm to separate the cell regions that are adhered to each other in the cell segmentation results. The final effect is shown in the figure below. Figure 8 (a). In Figure 8(b), the post-processing algorithm proposed by the present invention can accurately separate the mutually adhered cell regions. In contrast, in Figure 8(c), due to the lack of prior information, the ordinary watershed algorithm itself cannot find a stable water injection point, resulting in many over-segmentation phenomena.

[0086] The present invention also provides a fluorescence in situ hybridization cell nucleus segmentation method system, the segmentation system comprising:

[0087] The image segmentation module is used to segment the fluorescence in situ hybridization cell image using the trained LinkNet network to obtain the segmented fluorescence in situ hybridization cell image.

[0088] The trained LinkNet network includes: an encoder and a decoder; the encoder includes two convolutional layers and three residual modules, the two convolutional layers are respectively the first convolutional layer and the second convolutional layer, the three residual modules are respectively the first residual module, the second residual module and the third residual module, the first convolutional layer, the second convolutional layer, the first residual module, the second residual module and the third residual module are connected in sequence; the decoder includes three upsampling modules and three convolutional layers, the three upsampling modules are respectively the first upsampling module, the second upsampling module and the third upsampling module, the three convolutional layers include respectively the third convolutional layer, the fourth convolutional layer and the fifth convolutional layer; the input end of the first upsampling module is connected to the output end of the third residual module, the output end of the first upsampling module is connected to the first input end of the second upsampling module, and the first upsampling module is used to upsample the third residual module. The output image of the module is convolved with the output image of the second residual module; the second input end of the second upsampling module is connected to the output end of the second residual module, the output end of the second upsampling module is connected to the first input end of the third upsampling module, and the second upsampling module is used to add the output image of the second residual module and the output image of the first upsampling module, and then perform a convolution operation on the addition operation result; the second input end of the third upsampling module is connected to the output end of the first residual module, the output end of the third upsampling module is connected to the input end of the third convolutional layer, and the third upsampling module is used to add the output image of the first residual module and the output image of the second upsampling module, and then perform a convolution operation on the addition operation result; the output end of the third convolutional layer is connected to the input end of the fourth convolutional layer, and the output end of the fourth convolutional layer is connected to the input end of the fifth convolutional layer.

[0089] The residual module includes a sixth convolutional layer, a seventh convolutional layer and an adder; the output end of the sixth convolutional layer is connected to the input end of the seventh convolutional layer, the output end of the seventh convolutional layer is connected to the first input end of the adder, the input end of the sixth convolutional layer is connected to the second input end of the adder, the input end of the sixth convolutional layer serves as the input end of the residual module, and the output end of the adder serves as the output end of the residual module.

[0090] In order to implement LinkNet network training, the segmentation system also includes: a first training sample set establishment module, used to obtain multiple fluorescence in situ hybridization cell image samples, and manually segment the fluorescence in situ hybridization cell image samples to establish a first training sample set including the fluorescence in situ hybridization cell image samples and training samples of the manual segmentation results; a LinkNet network initialization module, used to initialize the LinkNet network using Xaviers; and a LinkNet network training module, used to use the BCE loss function as the target loss function of the LinkNet network, and use the first training sample set to train the initialized LinkNet network until the value of the target loss function of the LinkNet network is less than a first preset threshold, and output the trained LinkNet network.

[0091] The center point recognition module is used to use the trained U-net network to identify the cell center points of the segmented fluorescence in situ hybridization cell images and obtain the cell center point probability map.

[0092] In order to implement the training of the U-net network, the segmentation system also includes: a second training sample set establishment module, which is used to obtain multiple fluorescence in situ hybridization cell image samples, and manually identify and mark the cell center points of the fluorescence in situ hybridization cell image samples to establish a second training sample set including the fluorescence in situ hybridization cell image samples and the training samples of the manual identification results; a U-net network training module, which is used to use the Dice loss function as the target loss function of the U-net network, train the U-net network using the second training sample set until the value of the target loss function of the U-net network is less than a second preset threshold, and output the trained U-net network; based on the segmented fluorescence in situ hybridization cell image and the cell center point probability map, determine that the independent region in the segmented fluorescence in situ hybridization cell image with a number of cell center points greater than 1 is an adhesion region;

[0093] The secondary segmentation module is used to perform secondary segmentation of the adhesion area using the watershed algorithm according to the number and distribution of cell center points in the cell center probability map of the adhesion area.

[0094] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0095] The present invention discloses a method and system for segmenting fluorescence in situ hybridization cell nuclei. The segmentation method comprises the following steps: firstly, using a trained LinkNet network to segment a fluorescence in situ hybridization cell image to obtain a segmented fluorescence in situ hybridization cell image; then, using a trained U-Net network to identify cell centers in the segmented fluorescence in situ hybridization cell image to obtain a cell center probability map; then, based on the segmented fluorescence in situ hybridization cell image and the cell center probability map, determining that an independent region in the segmented fluorescence in situ hybridization cell image in which the number of cell centers is greater than 1 is an adhesion region; and finally, based on the number and distribution of cell centers in the adhesion region in the cell center probability map, using a watershed algorithm to perform secondary segmentation on the adhesion region. The present invention uses a trained LinkNet network to segment fluorescence in situ hybridization cell images to extract high-level semantic features for abstraction, and fuses the corresponding shallow information to capture detailed features when the decoder restores the feature map resolution, so as to better extract the required target cell foreground. The cell center point is obtained through the trained U-Net network, and the cell center point is integrated into the watershed algorithm to perform secondary segmentation on the adhesion area, solving the problem of mutual adhesion between cells and improving the accuracy of cell segmentation.

[0096] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0097] This article uses specific examples to illustrate the principles and implementation methods of the invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. The described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

Claims

1. A method for cell nucleus segmentation by fluorescence in situ hybridization, characterized in that: The segmentation method comprises the following steps: The trained LinkNet network is used to segment the fluorescence in situ hybridization cell images to obtain segmented fluorescence in situ hybridization cell images; The trained U-net network is used to identify the cell center of the segmented fluorescence in situ hybridization cell images to obtain the cell center probability map; According to the segmented fluorescence in situ hybridization cell image and the cell center point probability map, determining that an independent region in the segmented fluorescence in situ hybridization cell image in which the number of cell center points is greater than 1 is an adhesion region; According to the number and distribution of cell center points in the cell center probability map of the adhesion area, the watershed algorithm is used to perform secondary segmentation on the adhesion area. The U-net network presents an encoding-decoding topology. In the encoding process, Convolution operation, and each convolution operation will be followed by nonlinear units and batch normalization operations. When performing maximum pooling, the number of feature map channels will increase exponentially. This is because the size of the feature map is reduced, and the number of convolution channels is increased to effectively reduce the loss of information. In the decoding stage of U-Net, it is regarded as a shape generator. Its function is to recover the required target cell center area, and in the decoding stage, a learnable convolution is used to upsample the feature layer map of the previous layer. Its output is an enlarged and dense activation map. At the same time, in order to locate the position of the cell center, the U-net network obtains the detailed information in the underlying information by combining the high-resolution feature map in the encoding stage when restoring the feature map size, and then uses Convolution refines the output feature results based on this information. The final output of the network is a probability map of the same size as the input image, indicating the probability that each pixel belongs to the center point of the cell. Batch normalization is performed after each activation function to reduce internal covariate shift by normalizing the input distribution of each layer to a standard Gaussian distribution. The trained linknet network includes: encoder and decoder; The encoder includes two convolutional layers and three residual modules, the two convolutional layers are respectively a first convolutional layer and a second convolutional layer, the three residual modules are respectively a first residual module, a second residual module and a third residual module, and the first convolutional layer, the second convolutional layer, the first residual module, the second residual module and the third residual module are connected in sequence; The decoder includes three upsampling modules and three convolutional layers, the three upsampling modules are respectively a first upsampling module, a second upsampling module and a third upsampling module, and the three convolutional layers are respectively a third convolutional layer, a fourth convolutional layer and a fifth convolutional layer; An input end of the first upsampling module is connected to an output end of the third residual module, and an output end of the first upsampling module is connected to a first input end of the second upsampling module. The first upsampling module is configured to perform a convolution operation on an output image of the third residual module. A second input end of the second upsampling module is connected to an output end of the second residual module, and an output end of the second upsampling module is connected to a first input end of the third upsampling module. The second upsampling module is configured to add an output image of the second residual module to an output image of the first upsampling module, and then perform a convolution operation on the result of the addition operation. A second input end of the third upsampling module is connected to an output end of the first residual module, and an output end of the third upsampling module is connected to an input end of the third convolutional layer. The third upsampling module is configured to add an output image of the first residual module to an output image of the second upsampling module, and then perform a convolution operation on the result of the addition operation. An output end of the third convolutional layer is connected to an input end of the fourth convolutional layer, and an output end of the fourth convolutional layer is connected to an input end of the fifth convolutional layer. The residual module includes a sixth convolutional layer, a seventh convolutional layer and an adder; The output end of the sixth convolutional layer is connected to the input end of the seventh convolutional layer, the output end of the seventh convolutional layer is connected to the first input end of the adder, the input end of the sixth convolutional layer is connected to the second input end of the adder, the input end of the sixth convolutional layer serves as the input end of the residual module, and the output end of the adder serves as the output end of the residual module.

2. The fluorescence in situ hybridization cell nucleus segmentation method according to claim 1, characterized in that: The method further includes: using the trained LinkNet network to segment the fluorescence in situ hybridization cell image to obtain the segmented fluorescence in situ hybridization cell image; Acquiring a plurality of fluorescence in situ hybridization cell image samples, and manually segmenting the fluorescence in situ hybridization cell image samples to establish a first training sample set including the fluorescence in situ hybridization cell image samples and training samples of manual segmentation results; Use Xaviers to initialize the linknet network; The BCE loss function is used as the target loss function of the LinkNet network, and the initialized LinkNet network is trained using the first training sample set until the value of the target loss function of the LinkNet network is less than a first preset threshold, and the trained LinkNet network is output.

3. The method for cell nucleus segmentation by fluorescence in situ hybridization according to claim 1, wherein The method uses the trained U-net network to identify the cell center points of the segmented fluorescence in situ hybridization cell image to obtain a cell center point probability map, and also includes: Acquire multiple fluorescence in situ hybridization cell image samples, and manually identify and mark the cell center points of the fluorescence in situ hybridization cell image samples to establish a second training sample set including the fluorescence in situ hybridization cell image samples and training samples of manual identification results; The Dice loss function is used as the target loss function of the U-net network, and the U-net network is trained using the second training sample set until the value of the target loss function of the U-net network is less than the second preset threshold, and the trained U-net network is output.

4. A fluorescence in situ hybridization cell nucleus segmentation system, characterized in that: The segmentation system comprises: An image segmentation module is used to segment the fluorescence in situ hybridization cell image using the trained LinkNet network to obtain a segmented fluorescence in situ hybridization cell image; The center point recognition module is used to identify the cell center points of the segmented fluorescence in situ hybridization cell images using the trained U-net network to obtain the cell center point probability map; According to the segmented fluorescence in situ hybridization cell image and the cell center point probability map, determining that an independent region in the segmented fluorescence in situ hybridization cell image in which the number of cell center points is greater than 1 is an adhesion region; The secondary segmentation module is used to perform secondary segmentation of the adhesion area using the watershed algorithm based on the number and distribution of cell center points in the cell center probability map; The U-net network presents an encoding-decoding topology. In the encoding process, Convolution operation, and each convolution operation will be followed by nonlinear units and batch normalization operations. When performing maximum pooling, the number of feature map channels will increase exponentially. This is because the size of the feature map is reduced, and the number of convolution channels is increased to effectively reduce the loss of information. In the decoding stage of U-Net, it is regarded as a shape generator. Its function is to recover the required target cell center area, and in the decoding stage, a learnable convolution is used to upsample the feature layer map of the previous layer. Its output is an enlarged and dense activation map. At the same time, in order to locate the position of the cell center, the U-net network obtains the detailed information in the underlying information by combining the high-resolution feature map in the encoding stage when restoring the feature map size, and then uses Convolution refines the output feature results based on this information. The final output of the network is a probability map of the same size as the input image, indicating the probability that each pixel belongs to the center point of the cell. Batch normalization is performed after each activation function to reduce internal covariate shift by normalizing the input distribution of each layer to a standard Gaussian distribution. The trained LinkNet network includes an encoder and a decoder; The encoder includes two convolutional layers and three residual modules, the two convolutional layers are respectively a first convolutional layer and a second convolutional layer, the three residual modules are respectively a first residual module, a second residual module and a third residual module, and the first convolutional layer, the second convolutional layer, the first residual module, the second residual module and the third residual module are connected in sequence; The decoder includes three upsampling modules and three convolutional layers, the three upsampling modules are respectively a first upsampling module, a second upsampling module and a third upsampling module, and the three convolutional layers are respectively a third convolutional layer, a fourth convolutional layer and a fifth convolutional layer; An input end of the first upsampling module is connected to an output end of the third residual module, and an output end of the first upsampling module is connected to a first input end of the second upsampling module. The first upsampling module is configured to perform a convolution operation on an output image of the third residual module. A second input end of the second upsampling module is connected to an output end of the second residual module, and an output end of the second upsampling module is connected to a first input end of the third upsampling module. The second upsampling module is configured to add an output image of the second residual module to an output image of the first upsampling module, and then perform a convolution operation on the result of the addition operation. A second input end of the third upsampling module is connected to an output end of the first residual module, and an output end of the third upsampling module is connected to an input end of the third convolutional layer. The third upsampling module is configured to add an output image of the first residual module to an output image of the second upsampling module, and then perform a convolution operation on the result of the addition operation. An output end of the third convolutional layer is connected to an input end of the fourth convolutional layer, and an output end of the fourth convolutional layer is connected to an input end of the fifth convolutional layer. The residual module includes a sixth convolutional layer, a seventh convolutional layer and an adder; The output end of the sixth convolutional layer is connected to the input end of the seventh convolutional layer, the output end of the seventh convolutional layer is connected to the first input end of the adder, the input end of the sixth convolutional layer is connected to the second input end of the adder, the input end of the sixth convolutional layer serves as the input end of the residual module, and the output end of the adder serves as the output end of the residual module.

5. The fluorescence in situ hybridization cell nucleus segmentation system according to claim 4, characterized in that: The segmentation system further comprises: a first training sample set establishment module, configured to acquire a plurality of fluorescence in situ hybridization cell image samples, and manually segment the fluorescence in situ hybridization cell image samples to establish a first training sample set including the fluorescence in situ hybridization cell image samples and training samples of manual segmentation results; Linknet network initialization module, used to initialize the Linknet network using Xaviers; The linknet network training module is configured to use the BCE loss function as the target loss function of the linknet network, train the initialized linknet network using the first training sample set until the value of the target loss function of the linknet network is less than a first preset threshold, and output the trained linknet network.

6. The fluorescence in situ hybridization cell nucleus segmentation system according to claim 4, characterized in that: The segmentation system further comprises: a second training sample set establishment module, configured to obtain a plurality of fluorescence in situ hybridization cell image samples, and manually identify and label the cell center points of the fluorescence in situ hybridization cell image samples, thereby establishing a second training sample set including the fluorescence in situ hybridization cell image samples and training samples of the manual identification results; A U-net network training module is used to use the Dice loss function as the target loss function of the U-net network, train the U-net network using the second training sample set until the value of the target loss function of the U-net network is less than the second preset threshold, and output the trained U-net network.

Images