A cancer cell nucleus segmentation method based on a two-stage codec segmentation network

Abstract

The application provides a cancer cell nucleus segmentation method based on a two-stage codec segmentation network, introduces image standardization, random cutting and splicing and affine transformation, eliminates the difference between data, and enhances the complexity of training samples; a feature extraction algorithm, a deep joint structure and a feature selection module are proposed, the feature extraction algorithm is used for extracting low-dimensional feature information, the deep joint structure is used for extracting high-dimensional feature information, and the feature selection module is used for enriching the expression ability of a feature map, finally, the inner layer decoding encoder network is used for carrying out second times of down-sampling on the feature information extracted by the outer layer encoder and feature fusion, a multi-stage skip path is introduced, different scale feature information is forwarded to the outer layer decoder, the outer layer decoder aggregates various scale feature information, fully utilizes all feature information, and can accurately segment the cell nucleus in a complex cell environment, thereby providing accurate segmentation results for subsequent quantitative analysis of the cell nucleus morphology.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation, specifically to a method for segmenting cancer cell nuclei based on a two-stage encoder-decoder segmentation network. Background Technology

[0002] Cancer, a group of diseases associated with abnormal cell growth, poses a serious public health threat in modern times, and biopsy is considered the most reliable method for diagnosing cancer. A biopsy involves collecting cell or tissue samples fixed on a glass microscope slide for subsequent hematoxylin and eosin staining and microscopic examination. Pathologists visually examine the histological sample under a microscope to distinguish cell nuclei and mark cancerous areas. However, manual nuclear segmentation is very time-consuming and suffers from high inter-observer variability, potentially impacted by low sensitivity, specificity, and reproducibility. Secondly, the appearance of manual segmentation is significantly affected by the cell nuclei, resulting in slight defocusing, staining imbalance, and blurred boundaries. Furthermore, in histopathological images, cell nuclei are very small, typically only a few pixels in radius, and vary in size, often overlapping, further complicating segmentation. Finally, the density of malignant cells in tumor tissue is usually much higher than in normal tissue, tending towards disordered and unstable morphological changes. This leads to a complex environment of glandular structural degeneration, posing a significant obstacle to manual nuclear segmentation. Therefore, manual nuclear segmentation can no longer meet the needs of practical applications.

[0003] In addition, existing cell nucleus segmentation methods include two-stage and one-stage methods. Although these methods have made some progress in cell nucleus segmentation, they still have many shortcomings. For example, two-stage methods can only segment instances within bounding boxes at very low resolution. If the predicted bounding box is smaller than the actual instance, it may lead to poor segmentation results. One-stage methods use complex post-processing rules, which limit their performance to post-processing hyperparameters. Therefore, it is necessary to further explore efficient methods for cancer cell nucleus segmentation. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a cancer cell nuclear segmentation method based on a two-stage encoder-decoder segmentation network, thereby solving the problems mentioned in the background section. The present invention overcomes the deficiencies of existing cell nuclear segmentation methods and provides accurate segmentation results for subsequent quantitative analysis of cell nuclear morphology.

[0005] To achieve the above objectives, the present invention provides a method for segmenting cancer cell nuclei based on a two-stage encoder-decoder segmentation network, comprising the following steps:

[0006] Step 1: Preprocess the histopathological image dataset and divide it into two parts: a training set and a test set. The training set is used for model training, and the test set is used to test the network model.

[0007] Step 2: Construct a two-stage encoder-decoder segmentation network;

[0008] Step 3: Construct the training function for the segmentation network;

[0009] Step 4: Set the parameters of the training function in Step 3, train the processed histopathological images, and save the optimal weights of the segmentation network;

[0010] Step 5: Input the original histopathological images of the validation set into the segmentation network trained in Step 4 to segment the sample images and obtain the final histopathological image cell nucleus segmentation map.

[0011] Furthermore, step 1 includes the following steps:

[0012] Step 1.1: Standardize the histopathological images: Divide all pixels in the target histopathological image dataset by 255, then calculate the mean and standard deviation of the dataset. Finally, subtract the mean from all pixel values ​​and divide by the standard deviation to process the dataset into a standard distribution with a mean of 0 and a standard deviation of 1. Specifically, use formulas (1) and (2) for standardization:

[0013] x i =x i / 255 (1)

[0014]

[0015] Where x i For a single histopathological image in the target histopathological image dataset, These are standardized histopathological image data;

[0016] Step 1.2: Randomly crop and stitch the target tissue pathology image dataset: Randomly select four sample images from different tissues in the target tissue pathology image dataset, and randomly crop and stitch these four images; When cropping and stitching the images, the four randomly selected images are cropped sequentially with a crosshair at a random position, and the corresponding parts are stitched together. At the same time, the target box corresponding to each original image is also limited by the crosshair cropping and will not exceed the cropping range of the original image. Finally, a complete sample image of size 256*256 is stitched together.

[0017] Step 1.3: Perform affine transformation on the target tissue pathology image dataset: Construct a 256*256 black background image, and randomly move, rotate, and scale the target image samples along the center point of the black background image with a 50% probability, discarding the parts that exceed the black background image and retaining the remaining image parts as the result of the affine transformation; perform affine transformation on all target tissue pathology image datasets with a 50% probability.

[0018] Furthermore, in step 1.2, based on the number of target tissue pathology image datasets, 1000 random cropping and stitching operations are performed, and the randomly cropped and stitched sample images are added to the target tissue pathology image dataset as training samples.

[0019] Furthermore, step 2 includes the following steps:

[0020] Step 2.1: Constructing the feature extraction algorithm: The feature extraction algorithm is mainly used to extract low-dimensional feature information of the network. For the input feature vector, it will go through two branches: one branch includes a 1*1 convolutional layer, and the other branch consists of a 1*1 convolutional layer and two 3*3 convolutional layers. The second branch will add the input features and the output features, and input the feature information of the two convolutional layers of the two branches into a 1*1 convolutional layer to obtain the final feature information. Finally, the extracted feature information is activated by a nonlinear function using formula (3):

[0021]

[0022] Where z represents the input feature information. The activated feature information;

[0023] Step 2.2: Constructing a deep joint structure: The deep joint structure is mainly used to extract high-dimensional feature information of the network. The input feature vector will go through two branches: one branch includes a 1*1 convolutional layer, and the other branch consists of a 1*1 convolutional layer and four 3*3 convolutional layers. The output of the previous convolutional layer in the second branch will be used as the input of the next convolutional layer, and feature extraction will be performed in a step-like manner. Each convolutional layer in this branch will output feature information. The feature information of the five convolutional layers in the two branches will be combined and input into a 1*1 convolutional layer to obtain the final feature information. Finally, the extracted feature information will be activated by a nonlinear function using formula (4):

[0024]

[0025] Where z represents the input feature information. The activated feature information;

[0026] Step 2.3: Construct the feature selection module: The feature selection module is used for high-dimensional feature selection. It mainly consists of a 1*1 convolutional layer and three 5*5 max pooling layers with a step count of 1. The input features are passed through the convolutional layer and the output of the max pooling layer are used as the input of the max pooling layer. The output of the max pooling layer is used as the input of the max pooling layer. This process is repeated for three max pooling layers. Finally, the output features of the convolutional layer and the output features of the three max pooling layers are combined to obtain the final selected features.

[0027] Step 2.4: Construct a two-stage encoder-decoder neural network: The two-stage encoder-decoder network mainly consists of two layers: an outer encoder-decoder network and an inner decoder-decoder network. The outer encoder-decoder network mainly consists of a series of 3*3 convolutional layers, 2*2 downsampling layers with a stride of 2, 2*2 deconvolutional layers with a stride of 2, skip connection layers, and a feature extraction algorithm. The inner decoder-decoder network mainly consists of skip connection layers, a deep joint structure, and a feature selection module. The parameters of the convolutional layers are initialized in the form of a random decimal matrix.

[0028] Furthermore, in step 2.4, the input feature vector sequentially passes through the outer encoder, inner decoder, inner encoder, and outer decoder. The outer encoder extracts low-dimensional feature information from the input image and inputs it into the inner decoder. The inner decoder performs feature selection on the input feature vector through a feature selection module. It then concatenates the feature information extracted by the outer encoder's co-level coding block with the skip connection layer to obtain twice the feature information of the original channel. This information is then input into the deep joint structure to extract new feature information. After that, it is input into the inner encoder to continue extracting feature information. Finally, the outer decoder's skip connection layer combines the feature information from the inner and outer encoders to obtain the final network output.

[0029] Furthermore, step 3 includes the following steps:

[0030] Step 3.1: Constructing the learning rate adjustment function: This invention combines the current iteration number and the current learning rate to dynamically adjust the learning rate, as shown in formula (5):

[0031]

[0032] Where lr is the learning rate, base_lr is the baseline learning rate, epoch is the current iteration number, epochs is the maximum number of iterations, and power is the control coefficient;

[0033] Step 3.2: Constructing the joint loss function: In this invention, the joint loss function is Equation (6), Equation (7), and Equation (8):

[0034]

[0035]

[0036] Loss=-(α*Loss1+β*Loss2) (8)

[0037] Where p i It is a predicted value, g i These are the true values, α and β are dynamic adjustment coefficients, and Loss is the final loss function;

[0038] Step 3.3: Constructing the optimization function: The optimization function constructed in this invention incorporates historical gradient information and dynamically adjusts the current parameters; the optimization function in this invention is formula (9):

[0039] w new =w old -(γ v old +lrv new (9)

[0040] Where w new w represents the weight value of the result. old The original weight values, v new Let v be the current gradient. old γ is the original gradient, γ is the momentum coefficient, and lr is the learning rate.

[0041] Furthermore, step 4 includes the following steps:

[0042] Step 4.1: Set training parameters: In step 3.1 of this invention, α is 0.6 and β is 0.4; in step 3.2, the baseline learning rate base_lr is 0.0005, the maximum number of iterations epochs is 500, and the control coefficient power is 0.9; in step 3.3, the learning rate is the learning rate adjusted in step 3.2, and the momentum coefficient γ is set to 0.9; finally, set the optimal weight parameter best.

[0043] Step 4.2: During the training process, the training learning rate is reasonably adjusted using the learning rate adjustment function in Step 3.1, and the loss value between the output result of the segmentation network and the known labeled segmentation data is calculated using the joint loss function in Step 3.2.

[0044] Furthermore, in step 4.2, the segmentation network weights are updated using the optimization function from step 3.3 combined with the current learning rate and loss value; the above process is iterated continuously, and training is terminated if the maximum number of iterations (epochs) is reached; during the iteration process, it is determined whether the current loss function result is better than the best weight parameter; if not, the current training weights are not saved; if so, the training weights are saved as the optimal weights for prediction, and the best weight parameter is updated.

[0045] The beneficial effects of this invention are:

[0046] 1. This cancer cell nucleus segmentation method based on a two-stage encoder-decoder segmentation network eliminates data variability and reduces generalization error by introducing data preprocessing methods such as image normalization, random cropping and stitching, and affine transformation, thereby accelerating model convergence. Furthermore, it generates a sufficient number of complex training samples, increasing the number of small cell nuclei in the training samples and enhancing their complexity.

[0047] 2. This cancer cell nucleus segmentation method based on a two-stage encoder-decoder segmentation network proposes a feature extraction algorithm, a deep joint structure, and a feature selection module. The feature extraction algorithm can fully extract low-dimensional feature information and increase the model's field of vision even at lower layers, which helps to extract key features for segmentation. In addition, the deep joint structure is used to extract high-dimensional feature information from the model, fully extracting deep information and obtaining more semantic information. Finally, the feature selection module enriches the expressive power of the feature map, which is beneficial to solving the problem of segmentation difficulties due to large differences in the morphology of similar cell nuclei.

[0048] 3. This cancer cell nucleus segmentation method based on a two-stage encoder-decoder segmentation network proposes a two-stage encoder-decoder segmentation network, which includes an outer encoder-decoder network and an inner decoder network. Its structure is similar to a decoder network nested inside an encoder-decoder network. The inner decoder network performs a second downsampling and feature fusion on the feature information extracted by the outer encoder to fully extract deep information, introduces multi-level skip paths, and forwards information of different scales to the outer decoder. The outer decoder aggregates feature information of various scales, making full use of all feature information and improving the network segmentation accuracy. Attached Figure Description

[0049] Figure 1 This is a flowchart of a cancer cell nucleus segmentation method based on a two-stage encoder-decoder segmentation network according to the present invention;

[0050] Figure 2 This is a diagram of the segmentation network for the two-stage encoder-decoder of this invention;

[0051] Figure 3 This is a graph showing the learning rate results of this invention;

[0052] Figure 4 This is a graph showing the MIOU and Dice coefficient results of this invention. Detailed Implementation

[0053] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0054] Please see Figures 1 to 4 This invention provides a technical solution: a method for segmenting cancer cell nuclei based on a two-stage encoder-decoder segmentation network, mainly comprising the following steps:

[0055] Step 1: Preprocess the histopathological image dataset and divide it into two parts: a training set and a test set. The training set is used for model training, and the test set is used to test the network model.

[0056] Step 1.1: Standardize the histopathological images: Divide all pixels in the target histopathological image dataset by 255, then calculate the mean and standard deviation of the dataset. Finally, subtract the mean from all pixel values ​​and divide by the standard deviation to process the dataset into a standard distribution with a mean of 0 and a standard deviation of 1. Specifically, use formulas (1) and (2) for standardization:

[0057] x i =x i / 255 (1)

[0058]

[0059] In this embodiment, x i For a single histopathological image in the target histopathological image dataset, These are standardized histopathological image data;

[0060] Step 1.2: Randomly crop and stitch the target tissue pathology image dataset: Randomly select four sample images from different tissues in the target tissue pathology image dataset, and randomly crop and stitch these four images. During the cropping and stitching process, the four randomly selected images are cropped sequentially with a crosshair at a random position, and the corresponding parts are stitched together. At the same time, the target bounding box corresponding to each original image is also limited by the crosshair cropping and will not exceed the cropping range of the original image. Finally, a complete sample image of size 256*256 is stitched together. Based on the number of target tissue pathology image datasets, 1000 random cropping and stitching operations are performed, and the randomly cropped and stitched sample images are added to the target tissue pathology image dataset as training samples.

[0061] Step 1.3: Perform affine transformation on the target tissue pathology image dataset: Construct a 256*256 black background image, and randomly move, rotate, and scale the target image samples along the center point of the black background image with a 50% probability, discarding the parts that exceed the black background image and retaining the remaining image parts as the result of the affine transformation; perform affine transformation on all target tissue pathology image datasets with a 50% probability.

[0062] Step 2: Construct a two-stage encoder-decoder segmentation network;

[0063] Step 2.1: Constructing the feature extraction algorithm: The feature extraction algorithm is mainly used to extract low-dimensional feature information of the network. For the input feature vector, it will go through two branches: one branch includes a 1*1 convolutional layer, and the other branch consists of a 1*1 convolutional layer and two 3*3 convolutional layers. The second branch will add the input features and the output features, and input the feature information of the two convolutional layers of the two branches into a 1*1 convolutional layer to obtain the final feature information. Finally, the extracted feature information is activated by a nonlinear function using formula (3):

[0064]

[0065] Where z represents the input feature information. The activated feature information;

[0066] Step 2.2: Constructing a deep joint structure: The deep joint structure is mainly used to extract high-dimensional feature information of the network. The input feature vector will go through two branches: one branch includes a 1*1 convolutional layer, and the other branch consists of a 1*1 convolutional layer and four 3*3 convolutional layers. The output of the previous convolutional layer in the second branch will be used as the input of the next convolutional layer, and feature extraction will be performed in a step-like manner. Each convolutional layer in this branch will output feature information. The feature information of the five convolutional layers in the two branches will be combined and input into a 1*1 convolutional layer to obtain the final feature information. Finally, the extracted feature information will be activated by a nonlinear function using formula (4):

[0067]

[0068] In this embodiment, z represents the input feature information. The activated feature information;

[0069] Step 2.3: Construct the feature selection module: The feature selection module is used for high-dimensional feature selection. It mainly consists of a 1*1 convolutional layer and three 5*5 max pooling layers with a step count of 1. The input features are passed through the convolutional layer and the output of the max pooling layer are used as the input of the max pooling layer. The output of the max pooling layer is used as the input of the max pooling layer. This process is repeated for three max pooling layers. Finally, the output features of the convolutional layer and the output features of the three max pooling layers are combined to obtain the final selected features.

[0070] Step 2.4: Constructing a Two-Stage Encoder-Decoder Neural Network: The two-stage encoder-decoder segmentation network mainly consists of two layers: an outer encoder-decoder network and an inner decoder-decoder network. The outer encoder-decoder network primarily comprises a series of 3x3 convolutional layers, 2x2 downsampling layers with a stride of 2, 2x2 deconvolutional layers with a stride of 2, skip connection layers, and a feature extraction algorithm. The inner decoder-decoder network mainly consists of skip connection layers, a deep joint structure, and a feature selection module. The convolutional layer parameters are initialized using a random fractional matrix. The input feature vector sequentially passes through the outer encoder and inner decoder. The network consists of an inner encoder, an outer encoder, and a layer decoder. The outer encoder extracts low-dimensional feature information from the input image and inputs it into the inner decoder. The inner decoder performs feature selection on the input feature vector through a feature selection module. It then concatenates the feature information extracted by the outer encoder's corresponding layer through a skip connection layer to obtain twice the original channel's feature information, which is then input into the deep joint structure to extract new feature information. This new feature information is then input into the inner encoder to continue extracting feature information. Finally, the outer decoder's skip connection layer combines the feature information from the inner and outer encoders to obtain the final network output.

[0071] Step 3: Construct the training function for the segmentation network;

[0072] Step 3.1: Constructing the learning rate adjustment function: This invention combines the current iteration number and the current learning rate to dynamically adjust the learning rate, as shown in formula (5):

[0073]

[0074] Where lr is the learning rate, base_lr is the baseline learning rate, epoch is the current iteration number, epochs is the maximum number of iterations, and power is the control coefficient;

[0075] Step 3.2: Constructing the joint loss function: In this invention, the joint loss function is Equation (6), Equation (7), and Equation (8):

[0076]

[0077]

[0078] Loss=-(α*Loss1+β*Loss2) (8)

[0079] Where p i It is a predicted value, g i These are the true values, α and β are dynamic adjustment coefficients, and Loss is the final loss function;

[0080] Step 3.3: Constructing the optimization function: The optimization function constructed in this invention incorporates historical gradient information and dynamically adjusts the current parameters; the optimization function in this invention is formula (9):

[0081] w new =w old -(γ v old +lrv new (9)

[0082] In this embodiment, w new w represents the weight value of the result. old The original weight values, v new Let v be the current gradient. old γ is the original gradient, γ is the momentum coefficient, and lr is the learning rate.

[0083] Step 4: Set the parameters of the training function in Step 3, train the processed histopathological images, and save the optimal weights of the segmentation network;

[0084] Step 4.1: Set training parameters: In step 3.1 of this invention, α is 0.6 and β is 0.4; in step 3.2, the baseline learning rate base_lr is 0.0005, the maximum number of iterations epochs is 500, and the control coefficient power is 0.9; in step 3.3, the learning rate is the learning rate adjusted in step 3.2, and the momentum coefficient γ is set to 0.9; finally, set the optimal weight parameter best.

[0085] Step 4.2: During training, the training learning rate is reasonably adjusted using the learning rate adjustment function in Step 3.1. The loss value between the output result of the segmentation network and the known labeled segmentation data is calculated using the joint loss function in Step 3.2. The segmentation network weights are updated using the optimization function in Step 3.3 combined with the current learning rate and loss value. The above process is iterated continuously. If the maximum number of iterations (epochs) is reached, training is terminated. During the iteration, it is determined whether the current loss function result is better than the best weight parameter. If not, the current training weights are not saved. If so, the training weights are saved as the optimal weights for prediction, and the best weight parameter is updated.

[0086] Step 5: Input the original histopathological images of the validation set into the segmentation network trained in Step 4 to segment the sample images and obtain the final histopathological image cell nucleus segmentation map.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0088] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for segmenting cancer cell nuclei based on a two-stage encoder-decoder segmentation network, characterized in that, Includes the following steps: Step 1: Preprocess the histopathological image dataset and divide it into two parts: a training set and a test set. The training set is used for model training, and the test set is used to test the segmentation performance of the network model. Step 2: Construct a two-stage encoder-decoder segmentation network; Step 3: Construct the training function for the segmentation network; Step 4: Set the parameters of the training function in Step 3, train the processed histopathological images, and save the optimal weights of the segmentation network; Step 5: Input the original histopathological images of the validation set into the segmentation network trained in Step 4 to segment the sample images and obtain the final histopathological image cell nucleus segmentation map; Step 2 includes the following steps: Step 2.1: Building the Feature Extraction Algorithm: The feature extraction algorithm is mainly used to extract low-dimensional feature information from the network. For the input feature vector, it will go through two branches: one branch includes a 1 One convolutional layer has one branch, and the other branch consists of one convolutional layer. A 1-layer convolutional layer and two 3-layer convolutional layers The system consists of 3 convolutional layers, where the second branch adds the input and output features, combining the feature information from both convolutional layers into a single 1-layer convolutional layer. In the convolutional layer of 1, the final feature information is obtained; Finally, the extracted feature information is activated by a nonlinear function using formula (3): ； in For the input feature information, The activated feature information; Step 2.2: Constructing a Deep Joint Structure: The deep joint structure is mainly used to extract high-dimensional feature information from the network. For the input feature vector, it will go through two branches: one branch includes a 1 One convolutional layer has one branch, and the other branch consists of one convolutional layer. A 1-layer convolutional layer and four 3-layer convolutional layers A 3-convolutional layer, where the output of the preceding convolutional layer in the second branch serves as the input to the following convolutional layer, extracting features in a step-like manner. Each convolutional layer in this branch outputs feature information separately. The feature information from the five convolutional layers in both branches is combined and input into a 1-convolutional layer. In the convolutional layer of 1, the final feature information is obtained; Finally, the extracted feature information is activated by a nonlinear function using formula (4): ； in For the input feature information, The activated feature information; Step 2.3: Building the Feature Selection Module: The feature selection module is used for high-dimensional feature selection, mainly consisting of a 1 A convolutional layer of 1 and three 5-step convolutional layers of 1 step each. The system consists of 5 max pooling layers, where the input features are passed through the output of the convolutional layers and then used as the input of the max pooling layers. The output of the max pooling layers is then used as the input of the next max pooling layer. This process continues through three max pooling layers. Finally, the output features of the convolutional layers and the output features of the three max pooling layers are combined to obtain the final selected features. Step 2.4: Constructing a Two-Stage Encoder-Decoder Neural Network: The two-stage encoder-decoder network mainly consists of two layers: an outer encoder-decoder network and an inner decoder-decoder network. The outer encoder-decoder network primarily consists of a series of 3... 3 convolutional layers, 2 2 downsampling layers with a step size of 2, 2 The system consists of two deconvolutional layers with a stride of 2, skip connection layers, and a feature extraction algorithm. The inner decoder network mainly consists of skip connection layers, a deep joint structure, and a feature selection module. The parameters of the convolutional layers are initialized in the form of a random fractional matrix.

2. The cancer cell nucleus segmentation method based on a two-stage encoder-decoder segmentation network according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1: Standardize the histopathological images: Divide all pixels in the target histopathological image dataset by 255, then calculate the mean and standard deviation of the dataset. Finally, subtract the mean from all pixel values ​​and divide by the standard deviation to process the dataset into a standard distribution with a mean of 0 and a standard deviation of 1. Specifically, use formulas (1) and (2) for standardization: ； in For a single histopathological image in the target histopathological image dataset, These are standardized histopathological image data; Step 1.2: Random cropping and stitching of the target tissue pathology image dataset: Four sample images from different tissues are randomly selected from the target tissue pathology image dataset, and these four images are randomly cropped and stitched together. During the cropping and stitching process, the four randomly selected images are sequentially cropped using a crosshair at a random position, and the corresponding portions are stitched together. Simultaneously, the target bounding box corresponding to each original image is also constrained by the crosshair cropping and will not exceed the cropping range of the original image. Finally, a complete 256-pixel image is stitched together. A sample image of size 256; Step 1.3: Perform affine transformation on the target tissue pathology image dataset: construct a 256 Given a black background image of size 256, the target image sample is randomly moved, rotated, and scaled along the center point of the black background image with a 50% probability. The portion exceeding the black background image is discarded, and the remaining image portion is retained as the result of the affine transformation. The affine transformation is performed on all target tissue pathology image datasets with a 50% probability.

3. The cancer cell nucleus segmentation method based on a two-stage encoder-decoder segmentation network according to claim 2, characterized in that: In step 1.2, based on the number of target tissue pathology image datasets, 1000 random cropping and stitching operations are performed, and the randomly cropped and stitched sample images are added to the target tissue pathology image dataset as training samples.

4. The cancer cell nucleus segmentation method based on a two-stage encoder-decoder segmentation network according to claim 1, characterized in that: In step 2.4, the input feature vector passes sequentially through the outer encoder, inner decoder, inner encoder, and outer decoder. The outer encoder extracts low-dimensional feature information from the input image and inputs it into the inner decoder. The inner decoder performs feature selection on the input feature vector through a feature selection module. It then concatenates the feature information extracted by the outer encoder's co-level coding block with the skip connection layer to obtain twice the feature information of the original channel, which is then input into the deep joint structure to extract new feature information. This new feature information is then input into the inner encoder to continue extracting feature information. Finally, the outer decoder's skip connection layer combines the feature information from the inner and outer encoders to obtain the final network output.

5. The cancer cell nucleus segmentation method based on a two-stage encoder-decoder segmentation network according to claim 1, characterized in that: Step 3 includes the following steps: Step 3.1: Constructing the learning rate adjustment function: This invention combines the current iteration number and the current learning rate to dynamically adjust the learning rate, as shown in formula (5): ； Where lr is the learning rate, base_lr is the baseline learning rate, epoch is the current iteration number, epochs is the maximum number of iterations, and power is the control coefficient; Step 3.2: Constructing the joint loss function: In this invention, the joint loss function is Equation (6), Equation (7), and Equation (8): ； in It is a predicted value. It is the actual value. and It is the dynamic adjustment coefficient, and Loss is the final loss function; Step 3.3: Constructing the optimization function: The optimization function constructed in this invention incorporates historical gradient information and dynamically adjusts the current parameters; the optimization function in this invention is formula (9): ； in For the result weight value, These are the original weight values. For the current gradient, For the original gradient, The momentum coefficient, This is the learning rate.

6. The cancer cell nucleus segmentation method based on a two-stage encoder-decoder segmentation network according to claim 5, characterized in that: Step 4 includes the following steps: Step 4.1: Setting training parameters: In step 3.1 of this invention, It is 0.

6. It is 0.4; In step 3.2, the baseline learning rate (base_lr) is 0.0005, the maximum number of iterations (epochs) is 500, and the control coefficient (power) is 0.

9. In step 3.3, the learning rate is the learning rate adjusted in step 3.

2. The momentum coefficient was set to 0.9; finally, the optimal weighting parameter was set to best. Step 4.2: During the training process, the training learning rate is reasonably adjusted using the learning rate adjustment function in Step 3.1, and the loss value between the output result of the segmentation network and the known labeled segmentation data is calculated using the joint loss function in Step 3.

2.

7. The cancer cell nucleus segmentation method based on a two-stage encoder-decoder segmentation network according to claim 6, characterized in that: In step 4.2, the segmentation network weights are updated using the optimization function from step 3.3, combined with the current learning rate and loss value. The above process is iterated continuously. If the maximum number of iterations (epochs) is reached, the training is terminated. During the iteration, it is determined whether the current loss function result is better than the best weight parameter. If not, the current training weight is not saved. If so, the training weight is saved as the optimal weight for prediction, and the best weight parameter is updated.

Images