A method and device for identifying abnormal cervical cells and an electronic device

Abstract

The application discloses a cervical abnormal cell recognition method and device and electronic equipment, and relates to the technical field of cervical cancer detection. The method comprises the following steps: acquiring a cervical cell slice image; inputting the cervical cell slice image into a pre-trained channel separation module to obtain an H image and a DAB image, wherein the H image is a hematoxylin staining image, and the DAB image is an immunohistochemical staining image; and recognizing cervical abnormal cells by using the H image and the DAB image. The technical scheme of the application can obviously improve the cell detection speed, enhance the recognition accuracy and improve the algorithm execution efficiency in cervical cancer positive cell detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, and in particular to a cervical abnormal cell recognition method and device and electronic equipment.

[0002] The present application is the research result of the major project of scientific and technological innovation 2030 "innovation development and application of artificial intelligence test area" (2020AAA0105203). BACKGROUND

[0003] Cervical cancer is an important cause of female-specific cancer-related death. If patients are diagnosed at the precancerous stage or earlier, the cure rate can be as high as 98%, and the mortality rate will be significantly reduced. In recent years, medical microscopic image processing technology using computer image processing, artificial intelligence and other technologies is developing rapidly. Currently, the recognition of cervical cancer abnormal cells is generally carried out by the following method: finding the suspected abnormal cell location in the image RGB space through traditional image processing methods or artificial intelligence detection methods, and then obtaining the cell nucleus and cell membrane contour information through segmentation and other tasks, extracting color, texture, contour and other statistical features, and finally using classification algorithm to classify and judge whether the cell is abnormal.

[0004] Medical research has proved that neoplastic cervical cells all show different degrees of overexpression of P16 protein, so P16 is a biological marker of cervical neoplastic cells, and detecting P16 protein in cervical exfoliated cells has important reference value for improving the detection rate of abnormal cells. However, the existing technology has the following disadvantages for cervical cells stained with P16: first, all image processing operations are based on RGB color space, without fully considering the principle of cervical cell P16 staining, forcibly associating abnormal cell positive expression with RGB value, and cannot effectively obtain abnormal cell staining intensity; second, trying to extract features and classify each detected cell greatly restricts the system recognition speed. For example, each cervical cell slice image is about 50,000*50,000 pixels in size under 20 times of scanning, and even in a non-overlapping sliding window manner, 2,000 sub-images will be generated (each sub-image is calculated as 1000*1000 pixels in size). It cannot meet the requirements of data processing speed and real-time in clinical practice. SUMMARY

[0005] In order to solve the problems in the prior art, the present application provides the following technical scheme.

[0006] The present application first provides a cervical abnormal cell recognition method in a first aspect, comprising:

[0007] obtaining a cervical cell slice image;

[0008] Input the cervical cell slice image into a pre-trained channel separation module to obtain an H image and a DAB image, the H image is a hematoxylin staining image, and the DAB image is an immunohistochemical staining image;

[0009] Identify cervical abnormal cells by using the H image and the DAB image.

[0010] Optionally, the cervical cell slice image is obtained, and the method further comprises:

[0011] Obtain a low-resolution image of the cervical cell whole slice digital image, and perform coarse segmentation preprocessing on the low-resolution image;

[0012] Split the slice image after preprocessing to obtain the cervical cell slice image.

[0013] Optionally, the coarse segmentation preprocessing comprises:

[0014] Perform median filtering processing on the low-resolution image, and then use an adaptive binary algorithm to segment the image foreground region and the background region from the median filtered image.

[0015] Optionally, the channel separation module comprises a neural network and a spatial transformation matrix, and the training process of the channel separation matrix comprises:

[0016] Convert the slice image in the training data to an optical density space to obtain an optical density image;

[0017] Input the optical density image into the neural network to extract a high-dimensional feature vector of the optical density image;

[0018] Input the high-dimensional feature vector into the spatial transformation matrix, and use a preset loss function L to perform unsupervised training on the neural network and the spatial transformation matrix to obtain the channel separation module.

[0019] Optionally, the channel separation module comprises a spatial transformation matrix, and the training process of the channel separation module comprises:

[0020] Convert the slice image in the training data to an optical density space to obtain an optical density image;

[0021] Input the optical density image into the spatial transformation matrix, and use a preset loss function L to perform unsupervised training on the spatial transformation matrix to obtain the channel separation module.

[0022] Optionally, the spatial transformation matrix is:

[0023] OD = [cosαsinβ, cosαcosβ, sinβ] T

[0024] where OD is one of R 3*3 Matrix, each row of the matrix represents the optical density value of each channel in the RGB space corresponding to one staining reagent respectively.

[0025] Optionally, the preset loss function L in the unsupervised training is a third channel loss L err , reconstruction loss l r , and staining intensity loss l h , that is:

[0026] l = L err + λ r L r + λ h L h

[0027] where λ r , λ h are weight parameters.

[0028] Optionally, the third channel loss L err is defined as:

[0029]

[0030] where M, N represent the width and height of the slice image respectively, represents the pixel value of the third channel image of the slice image after the transformation matrix operation at coordinates i, j;

[0031] The reconstruction loss L r is defined as:

[0032]

[0033] where s i,j represents the pixel value of the original image at coordinates i, j, r i,j represents the pixel value of the reconstructed image at coordinates i, j;

[0034] The staining intensity loss L h is defined as:

[0035] L h = L h1 + L h2

[0036]

[0037]

[0038] where the balance hyperparameter γ ∈ (0, 1), the channel mediation hyperparameter δ ∈ (0, 1); h i,j and di,j respectively represent the pixel values of the first two channel images after channel separation at coordinates i, j.

[0039] Optionally, the identifying the abnormal cervical cells by using the H image and the DAB image further comprises:

[0040] According to the contour position of the staining positive area, the nucleus area of the corresponding H image is found;

[0041] The cell statistical features of the nucleus area of the H image and the staining positive area of the DAB image are extracted;

[0042] The cell type is determined according to the extracted cell statistical features.

[0043] Another aspect of the present application provides an abnormal cervical cell identification device, comprising:

[0044] The acquisition module is configured to acquire the cervical cell slice image.

[0045] The calculation module is configured to input the cervical cell slice image into a pre-trained channel separation module to obtain an H image and a DAB image, wherein the H image is a hematoxylin staining image, and the DAB image is an immunohistochemical staining image.

[0046] The identification module is configured to identify the abnormal cervical cells by using the H image and the DAB image.

[0047] A third aspect of the present application provides an electronic device comprising a processor and a memory, wherein the memory stores a plurality of instructions, and the processor is configured to read the instructions and perform the abnormal cervical cell identification method according to the first aspect.

[0048] A third aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a plurality of instructions, and the plurality of instructions can be read and executed by a processor to perform the abnormal cervical cell identification method according to the first aspect.

[0049] The present application has the beneficial effects that: the abnormal cell P16 staining reagent intensity is derived by means of the adaptive staining channel separation algorithm, which is more in line with the characteristics of the medical staining model; the identification of positive cells in the DAB staining channel replaces the detection and identification of each cell in the original image, which greatly improves the algorithm speed; in the abnormal cell classification process, the average optical density and integral optical density features of the positive cells are integrated, which improves the abnormal cell identification effect. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 A flowchart of the abnormal cervical cell identification method according to the embodiment of the present application is shown.

[0051] Figure 2 Detailed flowchart of the method for identifying abnormal cervical cells according to an embodiment of the present invention.

[0052] Figure 3 Flowchart of unsupervised training of a channel separation module according to an embodiment of the present invention.

[0053] Figure 4 Flowchart of identifying abnormal cervical cells using the stained image H and the stained image DAB according to an embodiment of the present invention.

[0054] Figure 5 2 is a module diagram of the device for identifying abnormal cervical cells according to the present invention. DETAILED DESCRIPTION

[0055] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0056] The method provided by the present invention can be implemented in the following terminal environment, which may include one or more of the following components: a processor, a memory, and a display screen. The memory stores at least one instruction, which is loaded and executed by the processor to implement the method described in the following embodiments.

[0057] A processor can include one or more processing cores. It connects various components within the terminal using various interfaces and circuits. It executes instructions, programs, code sets, or instruction sets stored in memory, and accesses data stored in memory to perform various terminal functions and process data.

[0058] The memory may include random access memory (RAM) or read-only memory (ROM). The memory may be used to store instructions, programs, codes, code sets, or instructions.

[0059] The display is used to show the user interface of each application.

[0060] In addition, those skilled in the art will appreciate that the structure of the terminal described above does not limit the terminal. The terminal may include more or fewer components, or a combination of certain components, or a different arrangement of components. For example, the terminal may also include a radio frequency circuit, an input unit, a sensor, an audio circuit, a power supply, and other components, which will not be described in detail here.

[0061] Example 1

[0062] like Figure 1 As shown, the embodiment of the present invention first provides a method for identifying abnormal cervical cells, which is characterized by comprising:

[0063] S1, acquiring a cervical cell slice image;

[0064] S2, inputting the cervical cell slice image into a pre-trained channel separation module to obtain an H image and a DAB image, the H image being a hematoxylin staining image, and the DAB image being an immunohistochemical staining image;

[0065] S3, identifying cervical abnormal cells by using the H image and the DAB image.

[0066] The cervical abnormal cell identification method separates the P16 staining image in the RGB space into a hematoxylin staining image H and an immunohistochemical staining image DAB (diaminobenzidine staining) by using an adaptive cervical cell P16 staining image channel separation algorithm, wherein the H channel image mainly displays the cell nucleus staining, and the DAB channel image mainly displays the P16 positive expression. Because the separated DAB channel mostly presents a non-positive background area, the positive cell detection in the channel can obviously improve the cell detection speed, enhance the identification accuracy, and improve the algorithm execution efficiency.

[0067] Referring to Figure 2 In specific embodiments, the above step S1 can further include:

[0068] S101, performing full slice scanning on a cervical cell staining slide to obtain a full slice digital image.

[0069] First, in the cell sample collection stage, the pre-prepared cervical cell P16 staining slide is scanned under a 20x lens to obtain a slice digital scanning image (Whole slide).

[0070] S102, acquiring a low-resolution image of the full slice digital image, and performing rough segmentation preprocessing on the low-resolution image to obtain an image foreground area and a background area.

[0071] That is, a low-resolution image (e.g. 1 / 64) of the full slice digital image is acquired, and image preprocessing is performed to roughly segment the P16 cell area in the slice. Because the image is preprocessed at a low resolution, the number of pixels processed is significantly less than that of the original image, so the image segmentation speed can be accelerated, and the preliminary segmentation result can be obtained more timely. In a preferred embodiment, the specific process of rough segmentation is as follows:

[0072] The low-resolution image is subjected to median filtering processing, and the image foreground area is preliminarily segmented by using adaptive binaryzation. The remaining part other than the foreground area is taken as the background area.

[0073] S103, segment the foreground region of the whole slice digital image to segment the original P16 staining whole slice digital image into multiple cervical cell slice images.

[0074] Based on the above rough segmentation of the foreground and background regions, the foreground region of the whole slice digital image is segmented in a sliding window manner, and the segmented image size can be, for example, 1024*1024 pixels, and the overlapping region is 256 pixels. In this way, the original P16 staining large image is divided into about 400 sub-images.

[0075] According to an embodiment, the channel separation module in the above step S2 includes a neural network and a spatial transformation matrix in series, and the training process of the channel separation module optionally includes:

[0076] Converting the slice images in the training data to optical density space to obtain optical density images;

[0077] Inputting the optical density images into the neural network to extract a high-dimensional feature vector of the optical density images;

[0078] Inputting the high-dimensional feature vector into the spatial transformation matrix, and performing unsupervised training on the neural network and the spatial transformation matrix by using a preset loss function L to obtain the channel separation module.

[0079] According to another embodiment, the channel separation module in the above step S2 includes a spatial transformation matrix, and the training process of the channel separation module optionally includes:

[0080] Converting the slice images in the training data to optical density space to obtain optical density images;

[0081] Inputting the optical density images into the spatial transformation matrix, and performing unsupervised training on the spatial transformation matrix by using a preset loss function L to obtain the channel separation module.

[0082] In the training process of the channel separation module, the present application adopts an adaptive P16 staining image separation method based on unsupervised training, performs unsupervised training on the channel separation module by using preset training data, learns a spatial transformation matrix of the slice images, realizes conversion of the image from RGB space to staining space (H-DAB), and obtains a nucleus staining image H and a P16 expression image DAB.

[0083] According to the Lambert-Beer imaging rule, the transmitted light intensity is related to the staining quantity, and the formula is as follows:

[0084]

[0085] wherein, I 0, is the incident light intensity, I cis the transmitted light intensity, A refers to the dosage intensity of the dye, is the absorption coefficient, and C represents the type of staining reagent. From the above formula, we can see that the intensity of transmitted light is not linearly related to the concentration of the staining reagent, that is, the grayscale value of each channel depends on the concentration of the staining reagent in a nonlinear manner. During the whole slide scanning process, each pixel value is obtained by the CCD photosensitive device. Therefore, it is inaccurate to directly perform color separation and quantitative analysis of staining intensity based solely on grayscale values. In physics, the optical density (OD) of an image is defined as:

[0086]

[0087] As can be seen from the above formula, the optical density intensity of each channel is linearly related to the dye dose. Therefore, before training, it is necessary to convert the slice images in the form of RGB images in the training data into the optical density space, and then perform unsupervised training on the spatial transformation matrix.

[0088] like Figure 3 As shown, taking the channel separation module including the series-connected neural network and spatial transformation matrix as an example, the process is as follows:

[0089] First, the slice image (i.e., P16 subimage) in the training data is converted to optical density space to obtain a first optical density image, which is then input into the neural network. The backbone of the neural network uses a classification network, such as VGG, ResNet, or DenseNet, to extract high-dimensional feature vectors from the first optical density image. Specifically, if the output layer outputs a 6-dimensional column vector, the 6-dimensional column vector is reshaped into two 1*3 vectors, represented as α = [a1, a2, a3] and β = [b1, b2, b3].

[0090] The above vectors are input into the space transformation matrix to obtain the images of channels H and DAB in the optically dense space. The space transformation matrix is:

[0091] OD=[cosαsinβ, cosαcosβ, sinβ] T

[0092] Where OD is an R 3*3 The matrix, R, is an orthogonal matrix. Each row of the matrix represents the optical density value of each staining reagent in each channel in the RGB space. Then, by calculating the product of the first optical density image and the OD matrix, the image of channels H and DAB in the optical density space can be calculated, that is, the second optical density image.

[0093] Finally, the images of the channels H and DAB in the optical density space are converted from the optical density space to the RGB space, so that the single-channel nucleus staining H and the single-channel P16 positive cell staining DAB images can be obtained. In the training process, the preset neural network loss function L is used to unsupervisedly train the neural network and the spatial transformation matrix, and the channel separation module is obtained through the training.

[0094] According to the principle of immunohistochemical staining, the neural network loss function L used in the application mainly considers the following three parts: the third channel loss L err , the reconstruction loss L r , and the staining intensity loss L h .

[0095] 1) The third channel loss L err : The P16 staining process mainly includes two coloring reagents, hematoxylin and DAB, so in the ideal case, the third channel value of the channel separation should be 0, and the formula is as follows:

[0096]

[0097] In the above formula, M and N respectively represent the width and height of the image, represents the pixel value of the third channel image of the slice image after the transformation matrix operation at coordinates i and j.

[0098] 2) The reconstruction loss L r : In order to ensure that the original image loses as little information as possible after the spatial transformation matrix operation, the difference between the image generated by the inverse transformation of the H-DAB image and the original image is calculated, and the mathematical expression is as follows:

[0099]

[0100] Where s i,j represents the pixel value of the original image at coordinates i and j, and r i,j represents the pixel value of the reconstructed image at coordinates i and j.

[0101] 3) The staining intensity loss L h : The two staining reagents are equally important in channel separation, so in order to avoid being too biased towards a certain staining, a hyperparameter γ is set to balance the two staining reagents.

[0102]

[0103] In the above formula, h i,j and d i,j respectively represent the pixel values of the first two channel images at coordinates i and j after channel separation, and γ is the balance hyperparameter, which is 0.6 in the experiment.

[0104] In addition, in order to avoid the single-channel image dyeing concentration being too high, the application designs a channel mediation super parameter δ ∈ (0, 1) to control the dyeing intensity, and δ = 0.8 is used in the experiment;

[0105]

[0106] Therefore, the dyeing intensity loss L h Can be expressed as:

[0107] L h = L h1 + L h2

[0108] The neural network loss function L is

[0109] L = L err + λ r L r + λ h L h

[0110] Where λ r , λ h are weight parameters.

[0111] As an optional embodiment, in order to further improve the algorithm execution speed, in the transformation matrix training process, the spatial transformation matrix is directly trained without passing through the neural network, and a transformation matrix (only 6 training parameters) is learned. Using the trained transformation matrix as the channel separation matrix can also achieve the conversion of the slice image from the RGB space to the H-DAB space. Compared with the above method, the image separation quality of this embodiment is reduced, but the algorithm execution speed can be improved by 60%.

[0112] In a preferred embodiment, the above step S2 further comprises:

[0113] The cervical cell slice image obtained in the input step S1 is converted from the RGB space to the optical density space, the channel separation is performed using the trained channel separation module, the single-channel H and DAB images in the optical density space are obtained, and then the single-channel image is converted from the optical density space to the RGB space. The output single-channel nucleus dyeing image H and positive cell dyeing image DAB.

[0114] In a preferred embodiment, the above step S3 further comprises:

[0115] The DAB image is preprocessed to extract the dyeing positive area contour;

[0116] According to the position of the dyeing positive area contour, the nucleus area of the corresponding H image is found;

[0117] extracting cell statistical features of the nucleus region of the H image and the staining positive region of the DAB image;

[0118] According to the extracted cell statistical features, the cell type is determined.

[0119] That is, on the basis of separating the staining channel, the DAB image is detected for abnormal cells, and it is identified whether the positive region is an abnormal cell. See Figure 4 In the abnormal cell detection stage, the DAB image needs to be preprocessed first. The preprocessing process may include, for example, image gamma correction, selecting the top 20% of the color histogram as the binarization threshold, performing a binarization operation, then performing an opening operation with a kernel of nuclear size 15, filtering small block regions, and then extracting the staining positive region contour.

[0120] Since manual operation in the process of cervical cell preparation may cause misstaining, the extracted contour needs to be preliminarily screened to filter out some obviously impure positive regions.

[0121] Specifically, according to the position of the positive contour in the DAB channel, the corresponding H channel nucleus contour is searched. If the nucleus cannot be found, it is also considered as a misstaining case. If the nucleus is found, the H channel nucleus statistical features are extracted, including the nucleus area, perimeter, and elliptical approximation degree, and the DAB channel positive region area, perimeter, and other features are also extracted. In addition, the average optical density value and integral optical density value of the positive cell are also calculated as nucleus and cytoplasm features for subsequent discrimination.

[0122] Finally, according to the nucleus and cytoplasm feature information extracted by the above process, the abnormal cell recognition is performed. The cell recognition step can be considered as a traditional classification task. Here, the cell recognition result is divided into three types: abnormal cell, normal cell, and impurity. In this classification task, any traditional classification algorithm can be used, for example, a three-layer fully connected neural network is used for classification in the present application, the input layer is an 18-dimensional vector, the hidden layer parameters are 256, 256 respectively, the activation function uses ReLu, the output layer is 3, the activation function is sigmoid, and the loss uses cross-entropy loss. The training samples are 5500, including 1900 abnormal cells, 1500 normal cells, and 2100 impurities.

[0123] In an optional embodiment, the classification step can use a traditional classification model other than a neural network to recognize abnormal cells, such as SVM, random forest, etc., which can also achieve good classification results.

[0124] It can be seen that the above cervical abnormal cell recognition method of the present application effectively recognizes the cervical slice cells by means of the adaptive staining channel separation algorithm. Compared with the cell recognition method of the prior art, the following advantages are possessed:

[0125] Firstly, by means of the adaptive staining channel separation method, the abnormal cell P16 staining reagent intensity is derived, which is more in line with the staining model principle in medicine.

[0126] Secondly, by recognizing the positive cells in the DAB staining channel, the direct detection and recognition of each cell in the original image are replaced, and the speed and efficiency of the algorithm are greatly improved.

[0127] Thirdly, in the abnormal cell classification process, the multi-dimensional discriminant features such as the average optical density and the integral optical density of the positive cells are fused, and the recognition effect and accuracy of the abnormal cells are improved.

[0128] Embodiment two

[0129] As Figure 5 shown, another aspect of the present application also includes a functional module architecture completely corresponding to the above method process, that is, the present application embodiment also provides a cervical abnormal cell recognition device, which comprises:

[0130] The acquisition module 201 is configured to acquire a cervical cell slice image.

[0131] The calculation module 202 is configured to input the cervical cell slice image into a pre-trained channel separation module to obtain an H image and a DAB image, the H image being a hematoxylin staining image, and the DAB image being an immunohistochemical staining image.

[0132] The recognition module 203 is configured to recognize cervical abnormal cells by using the H image and the DAB image.

[0133] The device can be realized by the cervical abnormal cell recognition method provided in the above embodiment one, and the specific implementation method can be referred to the description in embodiment one, which will not be repeated here.

[0134] The present application also provides a memory which stores a plurality of instructions for implementing the cervical abnormal cell recognition method as described in embodiment one.

[0135] The present application also provides an electronic device comprising a processor and a memory connected to the processor, wherein the memory stores a plurality of instructions which can be loaded and executed by the processor to enable the processor to perform the method as described in embodiment one.

[0136] While the preferred embodiments of the application have been described, additional variations and modifications can be made to these embodiments by those skilled in the art once they have the benefit of the foregoing description without departing from the spirit and scope of the application. Accordingly, it is intended that the appended claims be interpreted as including all such variations and modifications as fall within the spirit and scope of the application. It is further intended that the disclosure of all such modifications and variations be included within the scope of the application, the terms used herein being defined solely for purposes of the description being applied thereto unless otherwise indicated.

Claims

1. A method for identifying abnormal cervical cells, characterized in that: include: Acquire cervical cell section images; Inputting the cervical cell section image into a pre-trained channel separation module to obtain an H image and a DAB image, wherein the H image is a hematoxylin staining image and the DAB image is an immunohistochemical staining image; identifying abnormal cervical cells using the H image and the DAB image; The channel separation module includes a spatial transformation matrix, and the training process of the channel separation module includes: Convert the slice images in the training data into optical density space to obtain optical density images; Inputting the optically dense image into the spatial transformation matrix, and performing unsupervised training on the spatial transformation matrix using a preset loss function L to obtain the channel separation module; The spatial transformation matrix is: OD=[cosαsinβ,cosαcosβ,sinβ] T Where OD is an R 3*3 Matrix, each row of the matrix represents the optical density value of each channel in the RGB space corresponding to a dye reagent; α = [a1, a2, a3] and β = [b1, b2, b3] are the eigenvectors of the image respectively; The loss function L preset in the unsupervised training is the third channel loss L err , reconstruction loss L r and the loss of staining intensity L h The sum is: L=L err +λ r L r +λ h L h where λ r ,λ h is the weight parameter; The third channel loss L err Defined as: Wherein, M and N represent the width and height of the slice image respectively. represents the pixel value of the third channel image of the slice image at coordinates i, j after the transformation matrix operation; The reconstruction loss L is defined as: Among them, s i，j Represents the pixel value of the original image at coordinates i, j, r i，j Represents the pixel value of the reconstructed image at coordinates i, j; the reconstruction loss Lr is used to calculate the difference between the image generated after the H-DAB image is inversely transformed and the original image; The staining intensity loss L h Defined as: L h =L h1 +L h2 Wherein the balance hyperparameter γ∈(0,1), the channel adjustment hyperparameter δ∈(0,1); h i，j and d i，j Represent the pixel values ​​of the H image and DAB image at coordinates i and j after channel separation respectively.

2. The method according to claim 1, characterized in that The step of obtaining a cervical cell slice image further comprises: Acquiring a low-resolution image of a full-slice digital image of cervical cells, and performing coarse segmentation preprocessing on the low-resolution image; The pre-processed slice image is segmented to obtain the cervical cell slice image.

3. The method according to claim 2, characterized in that The coarse segmentation preprocessing includes: The low-resolution image is subjected to median filtering, and then an adaptive binarization algorithm is used to segment the image foreground area and background area from the median-filtered image.

4. The method according to claim 1, wherein The method of identifying abnormal cervical cells using the H image and the DAB image further includes: Preprocess the DAB image and extract the outline of the positive staining area; According to the outline position of the positive staining area, find the cell nucleus area corresponding to the H image; Extract cell statistical features of the cell nucleus area of ​​the H image and the positive staining area of ​​the DAB image; The cell types were determined based on the extracted cell statistical features.

5. A device for identifying abnormal cervical cells, characterized in that: include: An acquisition module, used for acquiring cervical cell slice images; a calculation module, configured to input the cervical cell section image into a pre-trained channel separation module to obtain an H image and a DAB image, wherein the H image is a hematoxylin-stained image and the DAB image is an immunohistochemical staining image; an identification module, configured to identify abnormal cervical cells using the H image and the DAB image; The channel separation module includes a spatial transformation matrix, and the training process of the channel separation module includes: Convert the slice images in the training data into optical density space to obtain optical density images; Inputting the optically dense image into the spatial transformation matrix, and performing unsupervised training on the spatial transformation matrix using a preset loss function L to obtain the channel separation module; The spatial transformation matrix is: OD=[cosαsinβ,cosαcosβ,sinβ] T Where OD is an R 3*3 Matrix, each row of the matrix represents the optical density value of each channel in the RGB space corresponding to a dye reagent; α = [a1, a2, a3] and β = [b1, b2, b3] are the eigenvectors of the image respectively; The loss function L preset in the unsupervised training is the third channel loss L err , reconstruction loss L r and the loss of staining intensity L h The sum is: L=L err +λ r L r +λ h L h where λ r ,λ h is the weight parameter; The third channel loss L err Defined as: Wherein, M and N represent the width and height of the slice image respectively. represents the pixel value of the third channel image of the slice image at coordinates i, j after the transformation matrix operation; The reconstruction loss L r Defined as: Among them, s i，j Represents the pixel value of the original image at coordinates i, j, r i，j Represents the pixel value of the reconstructed image at coordinates i, j; The reconstruction loss Lr is used to calculate the difference between the image generated by the inverse transformation of the H-DAB image and the original image; The staining intensity loss L h Defined as: L h =L h1 +L h2 Wherein the balance hyperparameter γ∈(0,1), the channel adjustment hyperparameter δ∈(0,1); h i，j and d i，j Represent the pixel values ​​of the H image and DAB image at coordinates i and j after channel separation respectively.

6. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a plurality of instructions, and the processor is used to read the instructions and execute the method for identifying abnormal cervical cells according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a plurality of instructions, and the plurality of instructions can be read by a processor and executed by the method for identifying abnormal cervical cells according to any one of claims 1 to 4.

Images