Cellular immunohistochemical data cleaning method and system based on auto-encoder
The cellular immunohistochemistry data is cleaned by a self-encoder-based method, which solves the low-quality problems in the data collection process, realizes efficient and automated data processing, improves data quality and processing efficiency, and provides reliable data support for medical research and clinical practice.
Patent Information
- Application Number
- CN202411640626.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-18
- Publication Date
- 2025-04-25
AI Technical Summary
During the collection process of cellular immunohistochemistry data, problems such as blur, edge data or stains are prone to problems such as blur, edge data or stains, resulting in low data quality and affecting the credibility of subsequent analysis.
Using the data cleaning method based on the autoencoder, the cell immunohistochemistry data is characterized and learned by constructing the autoencoder network, a standard encoding library and a sample encoding library are generated, the cosine distance is calculated and the threshold is set, and abnormal data is eliminated to obtain high-quality data.
It realizes automation of data processing, reduces manual intervention, improves data quality and processing efficiency, and provides a reliable data foundation to support medical research and clinical practice.
Smart Images
Figure CN119888723A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of biomedical technology, and in particular to a cell immunohistochemistry data cleaning method and system based on an autoencoder. Background Art
[0002] The cell immunohistochemistry process is a series of experimental steps used to detect and locate specific proteins in tissues or cells. Cell immunohistochemistry data images are significantly different from ordinary images. They have specific staining patterns, such as ki67 staining, her2 staining, etc., which can clearly present cell morphology and rich quantitative information. In terms of tissue composition, normal breast tissue cells are arranged in an orderly manner, while abnormal cell proliferation and disordered tissue structure may occur in lesions. It is necessary to focus on these characteristics for accurate analysis.
[0003] As an innovative study, intelligent identification of cell immunohistochemistry combines cell immunohistochemistry and artificial intelligence technology. By training deep learning models, cell types can be identified and classified, protein expression levels can be evaluated, and even patients' treatment responses can be predicted. This intelligent identification technology is expected to play an important role in cancer diagnosis, immunotherapy, and drug development. In cell immunohistochemistry research, the quality of data directly affects the subsequent analysis and accuracy of results. Due to the acquisition process of cell immunohistochemistry data, the scanner is required to scan and enlarge the slices. There may be a large number of low-quality images such as blurred images due to improper operation of the scanner or some edge data or stains. If all these images are used as sample data for analysis, it will undoubtedly reduce the credibility of the data analysis. Therefore, cell immunohistochemistry data cleaning has become a crucial link in cell immunohistochemistry intelligent identification. In view of this, it is urgent to study a method that can be used to efficiently obtain high-quality cell immunohistochemistry data. Summary of the invention
[0004] The technical problem to be solved by the present invention is to provide a cell immunohistochemistry data cleaning method and system based on an autoencoder, which is suitable for cleaning noise data of a large amount of sample data to obtain high-quality and effective cell immunohistochemistry data.
[0005] In a first aspect, the present invention provides a method for cleaning cell immunohistochemistry data based on an autoencoder, comprising the following steps:
[0006] Step S1, constructing a cell immunohistochemistry sample data set, and selecting a number of data from the data as a cell immunohistochemistry standard set;
[0007] Step S2, building an autoencoder network according to the characteristics of the cell immunohistochemistry sample data set;
[0008] Step S3, inputting the cell immunohistochemistry sample data set into the autoencoder network for representation learning to obtain a trained autoencoder;
[0009] Step S4, inputting the cell immunohistochemistry standard set and the cell immunohistochemistry sample data set into the autoencoder for encoding respectively, and generating the corresponding standard encoding library and sample encoding library;
[0010] Step S5: Calculate the cosine distance between the sample encoding library and the standard encoding library respectively, and remove the data with a cosine distance greater than a preset threshold to obtain a new cell immunohistochemistry sample data set.
[0011] Furthermore, the step S1 specifically includes:
[0012] The immunohistochemical sections were scanned by a scanner to obtain scanned images of fixed size;
[0013] By sliding a window, the scanned image is segmented from left to right and from top to bottom at a fixed step length to obtain a number of sample images of a specific size, white edges and stain images in the sample images are removed, and the sample images are saved to form a cell immunohistochemistry sample data set;
[0014] Then, a number of standard data are selected from the cell immunohistochemistry sample data set as a cell immunohistochemistry standard set. The selection conditions of the cell immunohistochemistry standard set are as follows: a number of sample data are selected as standard images according to each type of distribution image corresponding to each staining.
[0015] Furthermore, the autoencoder in step S2 is composed of a seven-layer network, and the number of neurons contained in each layer of the network is 1024, 512, 256, 128, 256, 512, and 1024 respectively, wherein the first three layers constitute the encoder, the last three layers constitute the decoder, the middle layer is the hidden layer, and the number of output neurons is 128. In the autoencoder, the input layer data is compressed into a potential representation by the encoder, that is, the 1024-dimensional input is reduced to 128 dimensions, and the potential representation is decoded into reconstructed data by the decoder, and the reconstructed data is made close to the original input data set by training the autoencoder.
[0016] Furthermore, the step S3 specifically includes:
[0017] The cell immunohistochemistry sample data was batch-adjusted to 32×32 size, expanded to 1024 dimensions after standardization to match the model input, and then input into the autoencoder for representation learning. MSE was selected as the loss function for training to minimize the difference between the reconstructed data and the original input data, thus obtaining the trained autoencoder.
[0018] Furthermore, each coded data in the standard coding library and the sample coding library is expressed in the following format:i ∈R N×M , 0≤i≤N, where N represents the total number of codes in the corresponding code library, a i Corresponding to the code generated by the i-th image, M represents the coding dimension of a single image after the autoencoder outputs it. Each coding output is standardized, that is, a is defined i =(b 1 ,b 2 ,…,b 128 ),but Where -1≤b j ≤1, 0≤j≤128 represents a single value in the vector.
[0019] Furthermore, the step S5 is specifically as follows: sequentially obtain a coded data in the sample coding library, calculate the cosine distance between the coded data and each coded data in the standard library, take the smallest distance as the distance to the standard coding library, and if the distance is greater than a preset threshold, eliminate the sample image corresponding to the coded data.
[0020] In a second aspect, the present invention provides a cell immunohistochemistry data cleaning device based on an autoencoder, comprising:
[0021] A data set construction module is used to construct a cell immunohistochemistry sample data set and select a number of data from it as a cell immunohistochemistry standard set;
[0022] The autoencoder creation module is used to build an autoencoder network based on the characteristics of the cell immunohistochemistry sample dataset;
[0023] An autoencoder training module is used to input the cell immunohistochemistry sample data set into the autoencoder network for representation learning to obtain a trained autoencoder;
[0024] The coding library generation module is used to input the cell immunohistochemistry standard set and the cell immunohistochemistry sample data set into the autoencoder encoding respectively to generate the corresponding standard coding library and sample coding library;
[0025] The data cleaning module is used to calculate the cosine distance between the sample coding library and the standard coding library respectively, and eliminate the data with a cosine distance greater than a preset threshold to obtain a new cell immunohistochemistry sample data set.
[0026] Furthermore, the data set construction module specifically includes:
[0027] The immunohistochemical sections were scanned by a scanner to obtain scanned images of fixed size;
[0028] By sliding a window, the scanned image is segmented from left to right and from top to bottom at a fixed step length to obtain a number of sample images of a specific size, white edges and stain images in the sample images are removed, and the sample images are saved to form a cell immunohistochemistry sample data set;
[0029] Then, a number of standard data are selected from the cell immunohistochemistry sample data set as a cell immunohistochemistry standard set. The selection conditions of the cell immunohistochemistry standard set are as follows: a number of sample data are selected as standard images according to each type of distribution image corresponding to each staining.
[0030] Furthermore, the autoencoder in the autoencoder creation module is composed of a seven-layer network, and the number of neurons contained in each layer of the network is 1024, 512, 256, 128, 256, 512, and 1024 respectively, wherein the first three layers constitute the encoder, the last three layers constitute the decoder, the middle layer is the hidden layer, and the number of output neurons is 128. In the autoencoder, the input layer data is compressed into a potential representation through the encoder, that is, the 1024-dimensional input is reduced to 128 dimensions, and the potential representation is decoded into reconstructed data through the decoder, and the reconstructed data is made close to the original input data set by training the autoencoder.
[0031] Furthermore, the autoencoder training module specifically includes:
[0032] The cell immunohistochemistry sample data was batch-adjusted to 32×32 size, expanded to 1024 dimensions after standardization to match the model input, and then input into the autoencoder for representation learning. MSE was selected as the loss function for training to minimize the difference between the reconstructed data and the original input data, thus obtaining the trained autoencoder.
[0033] Furthermore, each coded data in the standard coding library and the sample coding library is expressed in the following format: i ∈R N×M , 0≤i≤N, where N represents the total number of codes in the corresponding code library, a i Corresponding to the code generated by the i-th image, M represents the coding dimension of a single image after the autoencoder outputs it. Each coding output is standardized, that is, a is defined i =(b 1 ,b 2 ,…,b 128 ),but Where -1≤b j ≤1, 0≤j≤128 represents a single value in the vector.
[0034] Furthermore, the data cleaning module specifically comprises: obtaining a coded data in the sample coding library in turn, calculating the cosine distance between the coded data and each coded data in the standard library, taking the smallest distance as the distance to the standard coding library, and if the distance is greater than a preset threshold, removing the sample image corresponding to the coded data
[0035] One or more technical solutions provided by the present invention have at least the following technical effects or advantages: it realizes the automation of data processing, reduces manual intervention, improves processing efficiency and accuracy, and performs data representation learning through an autoencoder network, which can efficiently extract data features and speed up data processing. By calculating the cosine distance and setting a threshold to judge the data cleaning status, it can more accurately identify and eliminate abnormal data and improve data quality. The cleaned data can support further analysis and decision making, and provide a reliable data basis for medical research and clinical practice. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] The present invention will be further described below in conjunction with embodiments with reference to the accompanying drawings.
[0037] Figure 1 The present invention is an execution flow chart of a cell immunohistochemistry data cleaning method based on an autoencoder.
[0038] Figure 2 This is a schematic diagram of the structure of a cell immunohistochemistry data cleaning system based on an autoencoder of the present invention.
[0039] Figure 3 Schematic diagram of the autoencoder network structure of the present invention.
[0040] Figure 4 Schematic diagram of the network layer structure of the autoencoder of the present invention.
[0041] Figure 5 Schematic diagram of the training and use process of the autoencoder of the present invention.
[0042] Figure 6 It is a sample diagram of each type of standard set under a certain staining of the cell immunohistochemistry data of the present invention in a specific embodiment.
[0043] Figure 7 This is a picture of the cleaned samples in the cell immunohistochemistry sample data set of the present invention in a specific embodiment. DETAILED DESCRIPTION
[0044] The embodiment of the present application provides a cell immunohistochemistry data cleaning method and system based on an autoencoder, thereby realizing automated cleaning of cell immunohistochemistry data, reducing unnecessary manual intervention, improving processing efficiency, and providing a reliable data foundation for medical research and clinical practice.
[0045] The technical solution in the embodiment of the present application has the following overall idea: first prepare the cut data set, and select several standard sets from it; construct an autoencoder network according to the characteristics of the data set; input the data set into the autoencoder for training until convergence, so as to realize the characterization learning of tissue cell color and morphology; input the extracted standard set into the autoencoder to obtain the standard encoding; set the threshold, input the data set into the autoencoder to obtain the encoding, calculate the cosine distance with the standard encoding, clean the data according to the distance, identify and eliminate abnormal data, improve the data quality, and the cleaned data can support further analysis and decision making, and provide a reliable data basis for medical research and clinical practice.
[0046] In order to make the present invention more clearly understood, a preferred embodiment is now described in detail with reference to the accompanying drawings.
[0047] Embodiment 1
[0048] This embodiment provides a cell immunohistochemistry data cleaning method based on an autoencoder. Figure 1 and Figure 5 As shown, the following steps are included:
[0049] Step S1, constructing a cell immunohistochemistry sample data set, and selecting a number of data from the data as a cell immunohistochemistry standard set;
[0050] Step S2, building an autoencoder network according to the characteristics of the cell immunohistochemistry sample data set;
[0051] Step S3, inputting the cell immunohistochemistry sample data set into the autoencoder network for representation learning to obtain a trained autoencoder;
[0052] Step S4, inputting the cell immunohistochemistry standard set and the cell immunohistochemistry sample data set into the autoencoder for encoding respectively, and generating the corresponding standard encoding library and sample encoding library;
[0053] Step S5: Calculate the cosine distance between the sample encoding library and the standard encoding library respectively, and remove the data with a cosine distance greater than a preset threshold to obtain a new cell immunohistochemistry sample data set.
[0054] Preferably, the step S1 specifically includes:
[0055] The immunohistochemical sections were scanned by a scanner to obtain scanned images of fixed size;
[0056] By sliding a window, the scanned image is segmented from left to right and from top to bottom at a fixed step length to obtain a number of sample images of a specific size, white edges and stain images in the sample images are removed, and the sample images are saved to form a cell immunohistochemistry sample data set;
[0057] Then, a number of standard data are selected from the cell immunohistochemistry sample data set as the cell immunohistochemistry standard set. The selection conditions of the cell immunohistochemistry standard set are as follows: a number of sample data are selected as standard images according to each type of distribution image corresponding to each staining. For example, Figure 6 As shown, for HER2 staining, four types of correctly stained images need to be selected as standard sets, including four types of correctly stained images: Class A (i.e., normal cell images with negative 0), Class B (i.e., negative 1+), Class C (i.e., negative 2+ uncertain), and Class D (i.e., positive 3+), and several images of each type are selected and put into the standard set. Similarly, for all types of staining in the sample data (in this embodiment, for the four stainings of PR, ki67, her2, er, and pr), several images are selected as standard sets.
[0058] Preferably, the autoencoder in step S2 is composed of a seven-layer network, such as Figure 3 and Figure 4 As shown, the number of neurons contained in each layer of the network is 1024, 512, 256, 128, 256, 512, and 1024, respectively. The first three layers constitute an encoder, the last three layers constitute a decoder, and the middle layer is a hidden layer, and the number of its output neurons is 128. In the autoencoder, the input layer data X is compressed into a potential representation c by the encoder, that is, the 1024-dimensional input is reduced to 128 dimensions. The encoder can learn the potential feature representation of the data through training. In this embodiment, the hidden layer captures the color distribution and shape distribution of the data, and the decoder decodes the potential representation into reconstructed data X'. By training the autoencoder, that is, minimizing X-X', the reconstructed data can be made as close as possible to the original input data set.
[0059] Preferably, the step S3 specifically includes:
[0060] The cell immunohistochemistry sample data is batch-adjusted to a size of 32×32, and after standardization (using (image value-mean) / variance processing or normalization processing), it is expanded to 1024 dimensions to match the input of the model, and then input into the autoencoder for representation learning. MSE is selected as the loss function for training to obtain the trained autoencoder. The training process of the autoencoder is achieved by minimizing the reconstruction error, that is, minimizing the difference between the reconstructed data and the original input data. In this way, the autoencoder can learn the effective features in the data and reconstruct the original immunohistochemistry data during decoding.
[0061] Preferably, each coded data in the standard coding library and the sample coding library is expressed in the following format: i ∈R N×M , 0≤i≤N, where N represents the total number of codes in the corresponding code library, ai Corresponding to the code generated by the i-th image, M represents the coding dimension of a single image after the autoencoder outputs it. Each coding output is standardized, that is, a is defined i =(b 1 ,b 2 ,…,b 128 ),but Where -1≤b j ≤1, 0≤j≤128 represents a single value in the vector.
[0062] Preferably, the step S5 is specifically as follows: sequentially obtain a coded data in the sample coding library, calculate the cosine distance between the coded data and each coded data in the standard library, take the smallest distance as the distance to the standard coding library, and if the distance is greater than a preset threshold, remove the sample image corresponding to the coded data. Figure 7 As shown, sample images that are too blurred, not successfully stained, long cells, white blood cells, etc. will be cleaned up when no images with high similarity can be found in the standard set, leaving only usable high-quality sample images.
[0063] Based on the same inventive concept, the present application also provides a device corresponding to the method in Example 1, see Example 2 for details.
[0064] Embodiment 2
[0065] This embodiment provides a cell immunohistochemistry data cleaning device based on an autoencoder, such as Figure 2 and Figure 5 As shown, including:
[0066] A data set construction module is used to construct a cell immunohistochemistry sample data set and select a number of data from it as a cell immunohistochemistry standard set;
[0067] The autoencoder creation module is used to build an autoencoder network based on the characteristics of the cell immunohistochemistry sample dataset;
[0068] An autoencoder training module is used to input the cell immunohistochemistry sample data set into the autoencoder network for representation learning to obtain a trained autoencoder;
[0069] The coding library generation module is used to input the cell immunohistochemistry standard set and the cell immunohistochemistry sample data set into the autoencoder encoding respectively to generate the corresponding standard coding library and sample coding library;
[0070] The data cleaning module is used to calculate the cosine distance between the sample coding library and the standard coding library respectively, and eliminate the data with a cosine distance greater than a preset threshold to obtain a new cell immunohistochemistry sample data set.
[0071] Preferably, the data set construction module specifically includes:
[0072] The immunohistochemical sections were scanned by a scanner to obtain scanned images of fixed size;
[0073] By sliding a window, the scanned image is segmented from left to right and from top to bottom at a fixed step length to obtain a number of sample images of a specific size, white edges and stain images in the sample images are removed, and the sample images are saved to form a cell immunohistochemistry sample data set;
[0074] Then, a number of standard data are selected from the cell immunohistochemistry sample data set as the cell immunohistochemistry standard set. The selection conditions of the cell immunohistochemistry standard set are as follows: a number of sample data are selected as standard images according to each type of distribution image corresponding to each staining. For example, Figure 6 As shown, for HER2 staining, four types of correctly stained images need to be selected as standard sets, including four types of correctly stained images: Class A (i.e., normal cell images with negative 0), Class B (i.e., negative 1+), Class C (i.e., negative 2+ uncertain), and Class D (i.e., positive 3+), and several images of each type are selected and put into the standard set. Similarly, for all types of staining in the sample data (in this embodiment, for the four stainings of PR, ki67, her2, er, and pr), several images are selected as standard sets.
[0075] Preferably, the autoencoder in the autoencoder creation module is composed of a seven-layer network, such as Figure 3 and Figure 4 As shown, the number of neurons contained in each layer of the network is 1024, 512, 256, 128, 256, 512, and 1024, respectively. The first three layers constitute an encoder, the last three layers constitute a decoder, and the middle layer is a hidden layer, and the number of its output neurons is 128. In the autoencoder, the input layer data X is compressed into a potential representation c by the encoder, that is, the 1024-dimensional input is reduced to 128 dimensions. The encoder can learn the potential feature representation of the data through training. In this embodiment, the hidden layer captures the color distribution and shape distribution of the data, and the decoder decodes the potential representation into reconstructed data X'. By training the autoencoder, that is, minimizing X-X', the reconstructed data can be made as close as possible to the original input data set.
[0076] Preferably, the autoencoder training module specifically includes:
[0077] The cell immunohistochemistry sample data is batch-adjusted to a size of 32×32, and after standardization (using (image value-mean) / variance processing or normalization processing), it is expanded to 1024 dimensions to match the input of the model, and then input into the autoencoder for representation learning. MSE is selected as the loss function for training to obtain the trained autoencoder. The training process of the autoencoder is achieved by minimizing the reconstruction error, that is, minimizing the difference between the reconstructed data and the original input data. In this way, the autoencoder can learn the effective features in the data and reconstruct the original immunohistochemistry data during decoding.
[0078] Preferably, each coded data in the standard coding library and the sample coding library is expressed in the following format: i ∈R N×M , 0≤i≤N, where N represents the total number of codes in the corresponding code library, a i Corresponding to the code generated by the i-th image, M represents the coding dimension of a single image after the autoencoder outputs it. Each coding output is standardized, that is, a is defined i =(b 1 ,b 2 ,…,b 128 ),but Where -1≤b j ≤1, 0≤j≤128 represents a single value in the vector.
[0079] Preferably, the data cleaning module specifically comprises: obtaining a coded data in the sample coding library in turn, calculating the cosine distance between the coded data and each coded data in the standard library, taking the smallest distance as the distance to the standard coding library, and if the distance is greater than a preset threshold, removing the sample image corresponding to the coded data, such as Figure 7 As shown, sample images that are too blurred, not successfully stained, long cells, white blood cells, etc. will be cleaned up when no images with high similarity can be found in the standard set, leaving only usable high-quality sample images.
[0080] The technical solution provided in the embodiment of the present application has at least the following technical effects or advantages: it realizes the automation of data processing, reduces manual intervention, improves processing efficiency and accuracy, and performs cellular immunohistochemistry data characterization learning through the autoencoder network, which can efficiently extract data features, speed up data processing, and eliminate data with a large distance from the standard library to obtain effective high-quality data, and then provide a reliable data basis for medical research and clinical practice. After training, the autoencoder of the present invention can be used for subsequent similar data or data with similar distribution. After the standard library is created, it can also be directly used in the subsequent data cleaning process, with a high reuse rate, which improves the overall data processing efficiency.
[0081] Although the specific implementation modes of the present invention are described above, those skilled in the art should understand that the specific implementation modes described are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.
Claims
1. A cell immunohistochemistry data cleaning method based on an autoencoder, characterized in that: The steps include: Step S1, constructing a cell immunohistochemistry sample data set, and selecting a number of data from the data as a cell immunohistochemistry standard set; Step S2, building an autoencoder network according to the characteristics of the cell immunohistochemistry sample data set; Step S3, inputting the cell immunohistochemistry sample data set into the autoencoder network for representation learning to obtain a trained autoencoder; Step S4, inputting the cell immunohistochemistry standard set and the cell immunohistochemistry sample data set into the autoencoder for encoding respectively, and generating the corresponding standard encoding library and sample encoding library; Step S5: Calculate the cosine distance between the sample coding library and the standard coding library respectively, and remove the sample data with a cosine distance greater than a preset threshold to obtain a new cell immunohistochemistry sample data set.
2. The method for cleaning cell immunohistochemistry data based on an autoencoder according to claim 1, characterized in that: The step S1 specifically includes: The immunohistochemical sections were scanned by a scanner to obtain scanned images of fixed size; By sliding a window, the scanned image is segmented from left to right and from top to bottom at a fixed step length to obtain a number of sample images of a specific size, white edges and stain images in the sample images are removed, and the sample images are saved to form a cell immunohistochemistry sample data set; Then, a number of standard data are selected from the cell immunohistochemistry sample data set as a cell immunohistochemistry standard set. The selection conditions of the cell immunohistochemistry standard set are as follows: a number of sample data are selected as standard images according to each type of distribution image corresponding to each staining.
3. The method for cleaning cell immunohistochemistry data based on an autoencoder according to claim 1, characterized in that: The autoencoder in step S2 is composed of a seven-layer network, and the number of neurons contained in each layer of the network is 1024, 512, 256, 128, 256, 512, and 1024 respectively, wherein the first three layers constitute an encoder, the last three layers constitute a decoder, the middle layer is a hidden layer, and the number of its output neurons is 128. In the autoencoder, the input layer data is compressed into a potential representation by the encoder, that is, the 1024-dimensional input is reduced to 128 dimensions, and the potential representation is decoded into reconstructed data by the decoder, and the reconstructed data is made close to the original input data set by training the autoencoder.
4. The method for cleaning cell immunohistochemistry data based on an autoencoder according to claim 1, characterized in that: The step S3 specifically includes: The cell immunohistochemistry sample data was batch-adjusted to 32×32 size, expanded to 1024 dimensions after standardization to match the model input, and then input into the autoencoder for representation learning. MSE was selected as the loss function for training to minimize the difference between the reconstructed data and the original input data, thus obtaining the trained autoencoder.
5. The method for cleaning cell immunohistochemistry data based on an autoencoder according to claim 1, characterized in that: Each coded data in the standard coding library and the sample coding library is expressed in the following format: i ∈R N×M , 0≤i≤N, where N represents the total number of codes in the corresponding code library, a i Corresponding to the code generated by the i-th image, M represents the encoding dimension of the output of a single image after the autoencoder. Each encoding output is standardized, that is, a is defined i =(b1,b2,…,b 128 ),but Where -1≤b j ≤1, 0≤j≤128 represents a single value in the vector.
6. The method for cleaning cell immunohistochemistry data based on an autoencoder according to claim 1, characterized in that: The step S5 is specifically as follows: sequentially obtain a coded data in the sample coding library, calculate the cosine distance between the coded data and each coded data in the standard library, take the smallest distance as the distance to the standard coding library, and if the distance is greater than a preset threshold, remove the sample image corresponding to the coded data.
7. A cell immunohistochemistry data cleaning device based on an autoencoder, characterized in that: include: A data set construction module is used to construct a cell immunohistochemistry sample data set and select a number of data from it as a cell immunohistochemistry standard set; The autoencoder creation module is used to build an autoencoder network based on the characteristics of the cell immunohistochemistry sample dataset; An autoencoder training module is used to input the cell immunohistochemistry sample data set into the autoencoder network for representation learning to obtain a trained autoencoder; The coding library generation module is used to input the cell immunohistochemistry standard set and the cell immunohistochemistry sample data set into the autoencoder encoding respectively to generate the corresponding standard coding library and sample coding library; The data cleaning module is used to calculate the cosine distance between the sample coding library and the standard coding library respectively, and eliminate the data with a cosine distance greater than a preset threshold to obtain a new cell immunohistochemistry sample data set.
8. The cell immunohistochemistry data cleaning device based on an autoencoder as claimed in claim 7, characterized in that: The data set construction module specifically includes: The immunohistochemical sections were scanned by a scanner to obtain scanned images of fixed size; By sliding a window, the scanned image is segmented from left to right and from top to bottom at a fixed step length to obtain a number of sample images of a specific size, white edges and stain images in the sample images are removed, and the sample images are saved to form a cell immunohistochemistry sample data set; Then, a number of standard data are selected from the cell immunohistochemistry sample data set as a cell immunohistochemistry standard set. The selection conditions of the cell immunohistochemistry standard set are as follows: a number of sample data are selected as standard images according to each type of distribution image corresponding to each staining.
9. The cell immunohistochemistry data cleaning device based on an autoencoder as claimed in claim 7, characterized in that: The autoencoder in the autoencoder creation module is composed of a seven-layer network, and the number of neurons contained in each layer of the network is 1024, 512, 256, 128, 256, 512, and 1024 respectively, wherein the first three layers constitute an encoder, the last three layers constitute a decoder, and the middle layer is a hidden layer, and the number of its output neurons is 128. In the autoencoder, the input layer data is compressed into a potential representation by the encoder, that is, the 1024-dimensional input is reduced to 128 dimensions, and the potential representation is decoded into reconstructed data by the decoder, and the reconstructed data is made close to the original input data set by training the autoencoder.
10. The cell immunohistochemistry data cleaning device based on an autoencoder as claimed in claim 7, characterized in that: The data cleaning module specifically comprises the following steps: sequentially obtaining a coded data in the sample coding library, calculating the cosine distance between the coded data and each coded data in the standard library, taking the smallest distance as the distance to the standard coding library, and if the distance is greater than a preset threshold, eliminating the sample image corresponding to the coded data.