Pathological image automatic segmentation and classification system

Abstract

The invention relates to the field of pathological image processing, in particular to a pathological image automatic segmentation and classification system, which specifically comprises a data enhancement module, a model training module, a data collection module and a cell nucleus segmentation and classification module, and aims to solve the problem that the number of different types of cell nucleuses in a data set is unbalanced. A method of combining the cell nucleuses of the main class and the minority class is adopted in the data enhancement module so as to expand the data of the cell nucleuses of the minority class in the data set; in order to solve the problem that the boundary of a cell nucleus in a pathological image is blurred, an SAM-ResNet-UNet model is provided, the pathological image is enhanced by using the SAM model, then the enhanced image and the pathological image are learned through the two ResNet-UNet models respectively, segmentation results are output, finally classification results of the two images are obtained through a classifier and weighted averaging is carried out, and the classification result of the two images is obtained. And an accurate cell nucleus classification result is obtained.

Description

Technical Field

[0001] The present invention relates to the field of pathological image processing, and in particular to an automatic pathological image segmentation and classification system. Background Art

[0002] Automated pathological image segmentation and classification systems play a vital role in improving the quality of medical diagnosis and have far-reaching practical significance and development potential. Such systems can convert complex pathological data into analyzable image features, provide doctors with valuable diagnostic basis, and thus improve the efficiency and accuracy of clinical diagnosis. For personalized treatment strategies for diseases such as cancer, these image features provide key references, which help optimize patients' treatment plans and improve prognosis. In most cases, the morphology and type of cell nuclei, as key structures containing rich genetic information, occupy a central position in pathological images. This information is crucial for identifying the types of cells, tissues, and cancers. Therefore, accurate segmentation and classification of cell nuclei in pathological images is not only of great value in scientific research, but also provides strong support for clinical diagnosis. However, existing automatic pathological image segmentation and classification systems have the problem of significantly reduced performance due to the imbalance in the number of different types of cell nuclei in the data set, and the problem of poor segmentation and classification effects due to the unclear boundaries of cell nuclei in pathological images. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an automatic segmentation and classification system for pathological images. In view of the problem that the performance is greatly reduced due to the imbalance in the number of different types of cell nuclei in the data set, the present invention proposes a data enhancement module, and adopts a method of combining the cell nuclei of the main class and the minority class in the data enhancement module. A new minority class cell nucleus pathological image is formed by combining the main class and the minority class cell nuclei to expand the data of the minority class cell nuclei in the data set; in view of the problem that the segmentation and classification effects are poor due to the unclear boundaries of the cell nuclei in the pathological images, the present invention proposes a SAM-ResNet-UNet model: the model first uses the SAM model to enhance the pathological image, and then uses two ResNet-UNet models to perform feature learning on the enhanced image and the original pathological image, and outputs the segmentation result; finally, the classification results of the two images are obtained by the classifier and weighted average processing is performed to obtain an accurate cell nucleus classification result.

[0004] The technical solution adopted by the present invention is as follows: The present invention provides a pathological image automatic segmentation and classification system, which specifically includes: a data enhancement module, a model training module, a data collection module, and a cell nucleus segmentation and classification module, wherein the output end of the data enhancement module is connected to the input end of the model training module, the output end of the model training module is connected to the input end of the data collection module, and the output end of the data collection module is connected to the input end of the cell nucleus segmentation and classification module;

[0005] The data enhancement module obtains a data set, and expands the data of the cell nuclei of the minority class in the data set by combining the cell nuclei of the main class with the cell nuclei of the minority class;

[0006] The model training module constructs a SAM-ResNet-UNet model and performs training based on the data set;

[0007] The data collection module obtains the pathological sections and performs digital scanning to obtain pathological images;

[0008] The cell nucleus segmentation and classification module segments and classifies the cell nuclei in the pathological image according to the SAM-ResNet-UNet model;

[0009] The specific steps of the data enhancement module to expand the data of the cell nuclei of the minority class in the data set by combining the cell nuclei of the main class with the cell nuclei of the minority class are as follows:

[0010] D1: Calculate the proportion of cell nuclei of different categories in the data set, define the threshold, take the category with the largest proportion as the largest category, take the category whose proportion difference with the largest category is within the threshold as the main category, and the rest as the minority category;

[0011] D2: Traverse each minority class, take the traversed minority class as the current minority class, calculate the difference in the number of cell nuclei between the current minority class and the largest class, and randomly select cell nuclei with the same number as the difference from the cell nuclei of the main class to form a replacement cell nucleus set;

[0012] D3: Traverse the replacement cell nucleus set, take the traversed cell nucleus as the replacement cell nucleus, randomly select one of the cell nuclei of the current minority class whose cell nucleus size is smaller than the replacement cell nucleus as the current minority cell nucleus, perform morphological expansion on the replacement cell nucleus to obtain the expanded replacement cell nucleus image, and let the pixel set of the expanded replacement cell nucleus image be ;

[0013] D4: The current minority cell nucleus is translated to coincide with the center of mass of the replacement cell nucleus, and the translated current minority cell nucleus is used as the translated minority cell nucleus image. The pixel set of the translated minority cell nucleus image is , the calculation formula for replacing the center of mass of the cell nucleus is as follows:

[0014] ;

[0015] ;

[0016] in, , Respectively represent the horizontal and vertical coordinates of the center of mass, , Respectively represent the number of rows and columns of the image that replaces the cell nucleus, , Represents the first Line The horizontal and vertical coordinate values ​​of the column, Indicates replacement of the cell nucleus Line Gray value of the column;

[0017] D5: Define three mutually exclusive pixel sets, including the external set , Internal Collection Sum and Difference , where the external set is the pixel set outside the cell nucleus image after expansion, that is, the external set , the internal set is the pixel set of a few cell nucleus images after translation, that is, the internal set , the difference set is the pixel set of the replaced nucleus image after expansion minus the internal set, that is, the difference set ;

[0018] D6: Define the mask matrix composed of masks , set the mask corresponding to the elements of the external set to 0, set the mask corresponding to the elements of the internal set to 1, subtract the normalized minimum Euclidean distance from each element of the difference set to the element of the internal set from 1, and use the calculation result as the mask corresponding to the difference set. The calculation formula for the mask corresponding to the difference set is as follows:

[0019] , ;

[0020] in, and All represent elements in the difference set, Represents the elements in the internal collection, express The corresponding mask, represents the Euclidean distance, express arrive The minimum Euclidean distance of express arrive The minimum Euclidean distance of

[0021] D7: Use the image restoration algorithm to remove the color and texture of the replaced cell nucleus, obtain the repaired cell nucleus, and convert the mask matrix The product operation is performed with the difference between the replaced nucleus and the repaired nucleus, and then added to the repaired nucleus to obtain a new minority class nucleus and add it to the data set. The specific calculation formula is as follows:

[0022] ;

[0023] in, represents the mask matrix, represents a new minority class of nuclei, represents the image inpainting algorithm, Indicates replacement of the cell nucleus.

[0024] The SAM-ResNet-UNet model specifically includes a SAM model, a first ResNet-UNet model, a second ResNet-UNet model, a first classifier and a second classifier.

[0025] The model training module constructs the SAM-ResNet-UNet model and performs training according to the data set in the following specific steps:

[0026] S1: Generate segmentation prior map and boundary prior map of pathological image using SAM model;

[0027] S2: Use the weighted average grayscale algorithm to grayscale the pathological image to obtain a grayscale image. The calculation formula for grayscale processing is as follows:

[0028] ;

[0029] in, Indicates the image Line List, Represents the gray value, , , Respectively represent the pixel values ​​of R, G, and B channels of the pathological image;

[0030] S3: Create a three-channel image with the same size as the pathological image as an enhanced image, the first channel of the enhanced image consists of a grayscale image, the second channel consists of a segmentation prior map, and the third channel consists of a boundary prior map;

[0031] S4: Generate a binary segmentation image of the enhanced image using the first ResNet-UNet model: Use the ResNet50 model as an encoder to downsample the enhanced image to extract features, use the U-Net model as a decoder to perform upsampling operations, and use the jump connection of the U-Net model between the encoder and the decoder to fuse features of different depths, and finally output a binary segmentation image of the enhanced image;

[0032] S5: Generate a binary segmentation image of the pathological image using the second ResNet-UNet model: Use the ResNet50 model as an encoder to downsample the pathological image to extract features, use the U-Net model as a decoder to perform upsampling operations, and use the jump connection of the U-Net model between the encoder and the decoder to fuse features of different depths, and finally output a binary segmentation image of the pathological image;

[0033] S6: Input the binary segmentation image of the enhanced image into the first classifier to obtain a first classification result, input the binary segmentation image of the pathological image into the second classifier to obtain a second classification result, set a first weight and a second weight, and perform a weighted average operation on the first classification result and the second classification result to obtain a final classification result. The specific formula is as follows:

[0034] ;

[0035] in, represents the final classification result, represents the first classification result, represents the second classification result, represents the first weight, Represents the second weight.

[0036] The beneficial results achieved by the present invention using the above scheme are as follows:

[0037] (1) To address the problem of significantly reduced performance due to the imbalance in the number of different types of cell nuclei in the data set, the present invention proposes a data enhancement module and adopts a method of combining the cell nuclei of the main class with the cell nuclei of the minority class in the data enhancement module. By combining the main class with the cell nuclei of the minority class to form a new minority class cell nucleus pathology image, the data of the cell nuclei of the minority class in the data set can be expanded.

[0038] (2) To address the problem of poor segmentation and classification effects caused by unclear boundaries of cell nuclei in pathological images, the present invention proposes a SAM-ResNet-UNet model: the model first uses the SAM model to enhance the pathological image to highlight the boundary features of the cell nucleus. Subsequently, two ResNet-UNet models are used to perform feature learning on the enhanced image and the original pathological image, respectively, and the segmentation results are output, which significantly improves the model's ability to process pathological images. Finally, the classification results of the two images are obtained through a classifier and weighted average processing is performed to obtain accurate cell nucleus classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A module diagram of a pathological image automatic segmentation and classification system proposed by the present invention;

[0040] Figure 2 This is a framework diagram of the SAM-ResNet-UNet model proposed in the present invention.

[0041] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all 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.

[0043] Example 1, see Figure 1 , Figure 1 This is a module diagram of a pathological image automatic segmentation and classification system proposed by the present invention. In this embodiment, it specifically includes the following contents:

[0044] The present invention provides a pathological image automatic segmentation and classification system, which specifically includes: a data enhancement module, a model training module, a data collection module, and a cell nucleus segmentation and classification module, wherein the output end of the data enhancement module is connected to the input end of the model training module, the output end of the model training module is connected to the input end of the data collection module, and the output end of the data collection module is connected to the input end of the cell nucleus segmentation and classification module;

[0045] The data enhancement module obtains the data set and expands the data of the cell nuclei of the minority class in the data set by combining the cell nuclei of the main class with the cell nuclei of the minority class;

[0046] The model training module builds the SAM-ResNet-UNet model and trains it based on the dataset;

[0047] The data collection module obtains the pathological sections and performs digital scanning to obtain the pathological images;

[0048] The cell nucleus segmentation and classification module segments and classifies cell nuclei in pathological images based on the SAM-ResNet-UNet model.

[0049] Embodiment 2: Based on the above embodiment, this embodiment realizes the expansion of the data set through data enhancement. In this embodiment, the specific contents include the following:

[0050] The data set of this embodiment contains 318 512*512 pixel intestinal cancer tissue pathology slice images stained by hematoxylin-eosin staining, containing a total of 41277 cell nuclei, including 15898 tumor cell nuclei, 9493 immune cell nuclei, 6759 matrix cell nuclei, 5604 necrotic cell nuclei, and 3523 other cell nuclei. The categories and regions of all cell nuclei are saved in the SVG tag file;

[0051] The specific process of the data enhancement module to expand the data of the cell nuclei of the minority class in the data set by combining the cell nuclei of the main class with the cell nuclei of the minority class is as follows:

[0052] D1: Calculate the proportion of different categories of cell nuclei in the data set, define a threshold, take the category with the largest proportion as the largest category, take the category with a proportion difference within the threshold as the main category, and the remaining categories as minority categories. In this embodiment, the threshold is defined as 20%, and the proportions of tumor cell nuclei, immune cell nuclei, matrix cell nuclei, necrotic cell nuclei and other cell nuclei are 38.5%, 23.0%, 16.4%, 13.6% and 8.5% respectively. Then, tumor cell nuclei are the largest category, tumor cell nuclei and immune cell nuclei are the main categories, and matrix cell nuclei, necrotic cell nuclei and other cell nuclei are minority categories.

[0053] D2: Traverse each minority class, take the traversed minority class as the current minority class, calculate the difference in the number of cell nuclei between the current minority class and the largest class, and randomly select cell nuclei equal to the difference from the cell nuclei of the main class to form a replacement cell nucleus set. When the current minority class is the matrix cell nucleus, randomly select 15898-6759=9139 main class cell nucleus samples to form a replacement cell nucleus set;

[0054] D3: Traverse the replacement cell nucleus set, use the traversed cell nucleus as the replacement cell nucleus, randomly select one of the cell nuclei of the current minority class whose cell nucleus size is smaller than the replacement cell nucleus as the current minority cell nucleus, use a 3×3 convolution kernel to morphologically expand the replacement cell nucleus, and obtain the expanded replacement cell nucleus image. Let the pixel set of the expanded replacement cell nucleus image be ;

[0055] D4: The current minority cell nucleus is translated to coincide with the center of mass of the replacement cell nucleus, and the translated current minority cell nucleus is used as the translated minority cell nucleus image. The pixel set of the translated minority cell nucleus image is , the calculation formula for replacing the center of mass of the cell nucleus is as follows:

[0056] ;

[0057] ;

[0058] in, , Respectively represent the horizontal and vertical coordinates of the center of mass, , Respectively represent the number of rows and columns of the image that replaces the cell nucleus, , Represents the first Line The horizontal and vertical coordinate values ​​of the column, Indicates replacement of the cell nucleus Line Gray value of the column;

[0059] D5: Define three mutually exclusive pixel sets, including the external set , Internal Collection Sum and Difference , where the external set is the pixel set outside the cell nucleus image after expansion, that is, the external set , the internal set is the pixel set of a few cell nucleus images after translation, that is, the external set , the difference set is the pixel set of the replaced nucleus image after expansion minus the internal set, that is, the difference set ;

[0060] D6: Define the mask matrix composed of masks , set the mask corresponding to the elements of the external set to 0, set the mask corresponding to the elements of the internal set to 1, subtract the normalized minimum Euclidean distance from each element of the difference set to the element of the internal set from 1, and use the calculation result as the mask corresponding to the difference set. The calculation formula for the mask corresponding to the difference set is as follows:

[0061] , ;

[0062] in, and All represent elements in the difference set, Represents the elements in the internal collection, express The corresponding mask, represents the Euclidean distance, express arrive The minimum Euclidean distance of express arrive The minimum Euclidean distance of

[0063] D7: Use the image restoration algorithm to remove the color and texture of the replaced cell nucleus, obtain the repaired cell nucleus, and convert the mask matrix The difference between the replaced nucleus and the repaired nucleus is multiplied and added to the repaired nucleus to obtain a new minority class nucleus and add it to the data set. The image repair algorithm in this embodiment is specifically Patch Match. The specific calculation formula of this process is as follows:

[0064] ;

[0065] in, represents the mask matrix, represents a new minority class of nuclei, represents the image inpainting algorithm, Indicates replacement of the cell nucleus.

[0066] Example 3, see Figure 2 , Figure 2 This is a framework diagram of the SAM-ResNet-UNet model proposed in the present invention. Based on the above embodiments, this embodiment implements the training of the SAM-ResNet-UNet model. In this embodiment, the specific contents include the following:

[0067] The SAM-ResNet-UNet model specifically includes a SAM model, a first ResNet-UNet model, a second ResNet-UNet model, a first classifier and a second classifier;

[0068] The model training module builds the SAM-ResNet-UNet model and performs training based on the data set in the following specific steps:

[0069] S1: Generate a segmentation prior map and a boundary prior map of a pathological image using a SAM model. The SAM model includes an image encoder, a prompt encoder, and a lightweight mask decoder, wherein the image encoder receives an input image, the prompt encoder converts the prompt information of the image into an embedding vector that can be understood by the model, and the lightweight mask decoder predicts a segmentation mask according to the embedding vectors of the image encoder and the prompt encoder. In this embodiment, the pathological image is input into the image encoder, and the label information of the pathological image is input into the prompt encoder, and finally the segmentation prior map and the boundary prior map are obtained through the lightweight mask decoder;

[0070] S2: Use the weighted average grayscale algorithm to grayscale the pathological image to obtain a grayscale image. The calculation formula for grayscale processing is as follows:

[0071] ;

[0072] in, Indicates the image Line List, Represents the gray value, , , Respectively represent the pixel values ​​of R, G, and B channels of the pathological image;

[0073] S3: Create a three-channel image with the same size as the pathological image as an enhanced image, the first channel of the enhanced image consists of a grayscale image, the second channel consists of a segmentation prior map, and the third channel consists of a boundary prior map;

[0074] S4: Generate a binary segmentation image of the enhanced image using the first ResNet-UNet model: Use the ResNet50 model as an encoder to downsample the enhanced image to extract features, use the U-Net model as a decoder to perform upsampling operations, and use the jump connection of the U-Net model between the encoder and the decoder to fuse features of different depths, and finally output a binary segmentation image of the enhanced image;

[0075] S5: Generate a binary segmentation image of the pathological image using the second ResNet-UNet model: Use the ResNet50 model as an encoder to downsample the pathological image to extract features, use the U-Net model as a decoder to perform upsampling operations, and use the jump connection of the U-Net model between the encoder and the decoder to fuse features of different depths, and finally output a binary segmentation image of the pathological image;

[0076] The first ResNet-UNet model and the second ResNet-UNet model have the same network structure, and perform five downsampling operations through the ResNet50 model. First, it passes through a 7×7 convolution layer with an output channel of 64, then passes through a 3×3 maximum pooling layer, and then passes through four convolution groups, performing 3, 4, 6, and 3 convolution operations respectively. The downsampling of the first convolution group is realized by the maximum pooling layer, and the downsampling of the other three convolution groups is realized by the residual block of the previous convolution group.

[0077] After the downsampling operation, the U-Net model performs five upsampling operations. Each upsampling operation undergoes a 3×3 convolution operation and a ReLU activation function operation, and finally outputs a binary segmentation image of the same size as the original image.

[0078] S6: Input the binary segmentation image of the enhanced image into the first classifier to obtain a first classification result, input the binary segmentation image of the pathological image into the second classifier to obtain a second classification result, set a first weight and a second weight, and perform a weighted average operation on the first classification result and the second classification result to obtain a final classification result. The specific formula is as follows:

[0079] ;

[0080] in, represents the final classification result, represents the first classification result, represents the second classification result, represents the first weight, represents the second weight;

[0081] In this embodiment, the first classifier and the second classifier are both classified by the softmax function and the cross entropy loss function, and the first weight and the second weight are set to 0.6 and 0.4 respectively;

[0082] The model in this embodiment is built and trained by PyTorch, using Adam as the optimization function, with the initial learning rate set to 0.0001, the maximum number of iterations set to 200, and the batch size set to 4.

[0083] Embodiment 4: Based on the above embodiments, this embodiment realizes the collection and classification of pathological images. In this embodiment, the following contents are specifically included:

[0084] The data collection module obtains pathological sections and performs digital scanning to obtain pathological images. In this embodiment, the intestinal tissue of the patient is collected, and then the intestinal tissue is dehydrated, embedded with heated paraffin, cut into 5-micron thick sections using a slicer, stained with hematoxylin-eosin, and sealed with a cover glass to obtain tissue sections. The tissue sections are passed through a digital scanner to obtain a full-field digital pathological image, and the full-field digital pathological image is cut into an image of 512×512 size;

[0085] The cell nucleus segmentation and classification module segments and classifies the cell nuclei in the pathological image according to the SAM-ResNet-UNet model, specifically, the cut full-view digital pathological image is put into the SAM-ResNet-UNet model to obtain the segmented image and classification results.

Claims

1. A pathological image automatic segmentation and classification system, characterized by: Specifically include: A data enhancement module, a model training module, a data collection module, and a cell nucleus segmentation and classification module, wherein the output end of the data enhancement module is connected to the input end of the model training module, the output end of the model training module is connected to the input end of the data collection module, and the output end of the data collection module is connected to the input end of the cell nucleus segmentation and classification module; The data enhancement module obtains a data set, and expands the data of the cell nuclei of the minority class in the data set by combining the cell nuclei of the main class with the cell nuclei of the minority class; The model training module constructs a SAM-ResNet-UNet model and performs training based on the data set; The data collection module obtains the pathological sections and performs digital scanning to obtain pathological images; The cell nucleus segmentation and classification module segments and classifies cell nuclei in pathological images according to the SAM-ResNet-UNet model.

2. The automatic pathological image segmentation and classification system according to claim 1, characterized in that: The specific steps of the data enhancement module to expand the data of the cell nuclei of the minority class in the data set by combining the cell nuclei of the main class with the cell nuclei of the minority class are as follows: D1: Calculate the proportion of cell nuclei of different categories in the data set, define the threshold, take the category with the largest proportion as the largest category, take the category whose proportion difference with the largest category is within the threshold as the main category, and the rest as the minority category; D2: Traverse each minority class, take the traversed minority class as the current minority class, calculate the difference in the number of cell nuclei between the current minority class and the largest class, and randomly select cell nuclei with the same number as the difference from the cell nuclei of the main class to form a replacement cell nucleus set; D3: traverse the replacement cell nucleus set, take the traversed cell nucleus as the replacement cell nucleus, randomly select one of the cell nuclei of the current minority class whose cell nucleus size is smaller than the replacement cell nucleus as the current minority cell nucleus, perform morphological expansion on the replacement cell nucleus to obtain the expanded replacement cell nucleus image; D4: The current minority cell nucleus is translated to coincide with the center of mass of the replacement cell nucleus, and the translated current minority cell nucleus is used as the translated minority cell nucleus image; D5: Define three mutually exclusive pixel sets, including the external set, the internal set and the difference set; D6: Define a mask matrix composed of masks, set the mask corresponding to the elements of the external set to 0, set the mask corresponding to the elements of the internal set to 1, subtract the normalized minimum Euclidean distance from each element of the difference set to the element of the internal set from 1, and use the calculated result as the mask corresponding to the difference set; D7: Use the image restoration algorithm to remove the color and texture of the replaced cell nucleus to obtain the repaired cell nucleus. Multiply the mask matrix with the difference between the replaced cell nucleus and the repaired cell nucleus, and add it to the repaired cell nucleus to obtain a new minority class cell nucleus and add it to the dataset.

3. The automatic pathological image segmentation and classification system according to claim 1, characterized in that: The SAM-ResNet-UNet model specifically includes a SAM model, a first ResNet-UNet model, a second ResNet-UNet model, a first classifier and a second classifier.

4. The automatic pathological image segmentation and classification system according to claim 1, characterized in that: The model training module constructs the SAM-ResNet-UNet model and performs training according to the data set in the following specific steps: S1: Generate segmentation prior map and boundary prior map of pathological image using SAM model; S2: grayscale the pathological image using a weighted average grayscale algorithm to obtain a grayscale image; S3: Create a three-channel image with the same size as the pathological image as the enhanced image; S4: Generate a binary segmentation image of the enhanced image using the first ResNet-UNet model: Use the ResNet50 model as an encoder, the U-Net model as a decoder, and use a jump connection of the U-Net model between the encoder and the decoder, and finally output a binary segmentation image of the enhanced image; S5: Generate a binary segmentation image of the pathological image using the second ResNet-UNet model: use the ResNet50 model as an encoder, use the U-Net model as a decoder, and use a jump connection of the U-Net model between the encoder and the decoder, and finally output a binary segmentation image of the pathological image; S6: Input the binary segmentation image of the enhanced image into the first classifier to obtain a first classification result, input the binary segmentation image of the pathological image into the second classifier to obtain a second classification result, set a first weight and a second weight, and perform a weighted average operation on the first classification result and the second classification result to obtain a final classification result.

5. The automatic pathological image segmentation and classification system according to claim 2, characterized in that: The external set is the pixel set outside the expanded replaced cell nucleus image, the internal set is the pixel set of a few cell nucleus images after translation, and the difference set is the pixel set of the expanded replaced cell nucleus image minus the internal set.

6. The automatic pathological image segmentation and classification system according to claim 4, characterized in that: The first channel of the enhanced image is composed of a grayscale image, the second channel of the enhanced image is composed of a segmentation prior map, and the third channel of the enhanced image is composed of a boundary prior map.

Images