Self-supervised cancer detection system based on subspace clustering

Abstract

The invention provides a self-supervised cancer detection system based on subspace clustering. The system comprises a model training module, an image processing module and a result output module. According to the invention, a classification model for cancer discrimination is established and is applied to intelligent diagnosis of clinical pathology; when a to-be-detected pathological image is received, preprocessing operation is firstly carried out, and the preprocessing operation aims at extracting key features in the pathological image so as to better carry out cancer discrimination; inputting the preprocessed pathological image into a trained cancer discrimination classification model for discrimination to obtain a prediction result of the to-be-detected pathological image; the prediction result is displayed on a diagnosis and treatment interface of a pathologist, meanwhile, a diagnosis report is generated, and the pathologist makes more accurate cancer diagnosis and treatment decisions according to the prediction result and the diagnosis report, so that the accuracy and efficiency of pathological diagnosis are improved; therefore, a more accurate cancer judgment result is provided, and doctors are assisted in diagnosis and treatment decision making.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a self-supervised cancer detection system based on subspace clustering. Background Art

[0002] Pathological images are images obtained by observing stained tissue sections under a microscope. With the continuous advancement of science and technology, computers have played a good auxiliary role in the analysis of pathological images. Smart pathology refers to the use of advanced computer technology and algorithms to efficiently and accurately analyze and diagnose pathological images. Staining diversity has a great impact on pathological analysis, because different staining methods and experimental conditions may lead to color and texture differences between images. Self-supervised learning is a training method that does not require manual annotation of data. It completes the task by learning the intrinsic structure and representation of the input data. Subspace clustering is a clustering technology that organizes and analyzes high-dimensional data by discovering low-dimensional subspaces in the data. Self-supervised learning can help the smart pathology system learn to obtain features unrelated to staining without the need for a large amount of labeled data, thereby improving the accuracy of detection and diagnosis.

[0003] Prior art 1, application number: CN201510290950.7 discloses a cancer detection device, including a camera device and a processing device. The camera device is used to shoot a detection container at a first moment and a detection container at a second moment, respectively obtaining a first moment image and a second moment image, wherein the detection container includes a urine sample area and a detection reagent area, the urine sample area and the detection reagent area are respectively added with a urine sample to be detected and a detection reagent, and the first nematode contained in the detection reagent is stained with a first color and has movement activity; the processing device is used to analyze the movement trend of the first color, relative to the first color in the first image, if the first color in the second image has a movement trend from the detection reagent area to the urine sample area, then the detection result is judged to be cancer positive, and if the first color in the second image has a movement trend from the detection reagent area to the urine sample area, then the detection result is judged to be cancer negative. Although the detection device can be used as a low-cost, fast and effective early cancer screening device, the processing process of its pathological image is relatively simple, resulting in a low accuracy rate of the obtained cancer diagnosis result.

[0004] Prior art 2, application number: CN202110773502.8 discloses a cancer detection system, including a ventilating tube slot, an exhalation ventilating tube, a limiting sealing mechanism, a supporting card seat, a fixed connection slide slot, a positioning connection mechanism, a threaded guide rod, a breathing sleeve connector, a limiting support mechanism, a micro motor, a toothed connecting pulley and a driving pulley; the limiting guide plate is adapted to the limiting clamping slot, and the limiting support mechanism is adapted to the limiting clamping slot through the limiting guide plate and then rotated and clamped inside the positioning connection mechanism. Although the exhalation ventilating tube is set, the inner end face of the exhalation ventilating tube is hollow, so the external temperature can be effectively insulated, the thermal insulation performance of the device is improved, so that the properties of the exhaled air remain unchanged, and the detection efficiency is improved. The hollow setting inside can also improve the sound insulation and noise reduction performance of the device; but its structure is relatively simple and cannot effectively improve the intelligent control level of cancer diagnosis.

[0005] Prior art three, application number: CN202110144113.9 discloses a cancer detection kit, and also discloses the use of the antibody in the preparation of tumor treatment and / or tumor diagnosis reagents. Although its application provides an effective tool for early screening diagnosis, efficacy detection and prognosis judgment of ovarian cancer; it can only realize the diagnosis of ovarian cancer, and cannot perform detailed analysis of pathological images, resulting in high cost of use.

[0006] Currently, the existing technologies 1, 2 and 3 have the problem that the processing systems of cancer pathology images are simple in function and the processing efficiency of pathology images is low, resulting in poor accuracy of cancer identification results. Therefore, the present invention provides a self-supervised cancer detection system based on subspace clustering, which can achieve effective cancer detection in pathology images without the need for a large amount of labeled data; it integrates multiple links such as pathology section collection, labeling, algorithm processing, prediction visualization and diagnosis report, and provides a comprehensive and intelligent auxiliary diagnosis tool for pathologists. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a self-supervised cancer detection system based on subspace clustering, comprising:

[0008] The model training module is responsible for obtaining pathological tissue sections and staining them to obtain a pathological image set. 70% of the pathological images in the pathological image set are manually annotated to obtain a training set to train the cancer discrimination and classification model.

[0009] The image processing module is responsible for receiving the pathological image to be detected, performing preprocessing operations, and inputting the pathological image after the preprocessing operation into the cancer discrimination classification model for discrimination to obtain the prediction result of the pathological image to be detected;

[0010] The result output module is responsible for receiving the prediction results, displaying them on the pathologist's diagnosis and treatment interface, and generating a diagnosis report.

[0011] Optionally, the preprocessing operations of the image processing module include: scaling, cutting, subspace partitioning and feature extraction.

[0012] Optional, model training module, including:

[0013] The image set division submodule is responsible for using the classification model to divide the pathological image set into a training set and a test set, and manually annotating the training set to obtain the manually annotated training set;

[0014] The training set correction judgment submodule is responsible for inputting the training set into the cancer discrimination classification model for training to obtain the initial cancer discrimination classification model, and using the test set to determine whether the initial cancer discrimination classification model needs to correct the training set.

[0015] Optionally, the ratio of the training set to the test set of the image set partitioning submodule is 7:3.

[0016] Optionally, the model training module further includes: a training set correction submodule, which is responsible for using the current training set to train the cancer discrimination classification model if the training set does not need to be corrected; if the training set needs to be corrected, the classification model is used to search for manually annotated pathological images to be corrected in the training set.

[0017] Optionally, the training set correction submodule further includes: re-labeling the manually annotated pathological images to be corrected, correcting the training set, and returning the input to the cancer discrimination classification model for training.

[0018] Optional, image processing module, including:

[0019] The clustering processing submodule is responsible for scaling and cutting the pathological images to be detected, and using subspace clustering to divide the processed pathological images into different subspaces, where each subspace corresponds to a different color distribution;

[0020] The feature extraction submodule is responsible for extracting staining-independent features from the original pathological images using a convolutional neural network as a feature extractor, where the features are used to distinguish different subspaces;

[0021] The color transformation submodule is responsible for predicting which subspace each pathological image belongs to through the cancer discrimination classification model, so that the classification model can realize color transformation and obtain the original pathological image with color distribution, that is, obtain the prediction result of the pathological image to be detected.

[0022] Optionally, features in the feature extraction submodule are used to distinguish different subspaces.

[0023] Optional feature extraction submodule, including:

[0024] The convolution operation unit is responsible for inputting the original pathological images before and after staining into the feature extractor based on circular convolution and residual operation, and extracting the local features related to staining through convolution operation;

[0025] A position relationship determination unit is responsible for determining the features related to and unrelated to staining in the original pathological image based on the local features, and determining the position relationship of the features unrelated to staining based on the features related to staining;

[0026] The subspace discrimination unit is responsible for using the staining-independent features as the criteria for distinguishing different subspaces and inputting the staining-independent features into the cancer discrimination classification model.

[0027] Optional, model training module, including:

[0028] The instruction issuing submodule is responsible for issuing instructions to display the prediction results on the diagnosis and treatment interface, parsing the instructions and calling the prediction results;

[0029] The interface component submodule is responsible for forming the diagnosis and treatment interface. The terminal of the diagnosis and treatment interface forms the diagnosis and treatment interface display data according to the received prediction results and the diagnosis and treatment interface request set by the user;

[0030] The result processing submodule is responsible for reading the prediction results in the diagnosis and treatment interface, visualizing the prediction results using commercial-grade data chart plug-ins, and generating a diagnosis report containing commercial-grade data charts.

[0031] The model training module of the present invention obtains pathological tissue sections and performs staining processing to obtain a pathological image set, selects 70% of the pathological images from the pathological image set for manual annotation, obtains a training set, and trains the cancer discrimination and classification model; the image processing module receives the pathological image to be detected, performs a preprocessing operation, and inputs the pathological image after the preprocessing operation into the cancer discrimination and classification model for discrimination to obtain the prediction result of the pathological image to be detected; wherein the preprocessing operation includes: scaling, cutting, subspace division and feature extraction, etc.; the result output module receives the prediction result and displays it on the diagnosis and treatment interface of the pathologist, and generates a diagnosis report at the same time; the above scheme establishes a classification model for cancer discrimination and applies it to the intelligent diagnosis of clinical pathology; by obtaining pathological tissue sections and performing staining processing, a pathological image set is obtained, Data for training and testing are obtained, and 70% of the pathological images are selected for manual annotation to obtain a training set. The training set will contain samples that have been labeled whether there is cancer, which are used to train the cancer discrimination classification model; when the pathological image to be detected is received, a preprocessing operation is first performed, such as scaling, cutting, subspace division and feature extraction, etc. The preprocessing operation is intended to extract key features in the pathological image in order to better discriminate cancer; the preprocessed pathological image is input into the trained cancer discrimination classification model for discrimination to obtain the prediction result of the pathological image to be detected; the prediction result will be displayed on the diagnosis and treatment interface of the pathologist, and a diagnosis report will be generated at the same time. The pathologist will make more accurate cancer diagnosis and treatment decisions based on the prediction results and the diagnosis report, thereby improving the accuracy and efficiency of pathological diagnosis. This embodiment uses manually annotated pathological images to train a cancer discrimination classification model, and applies it to clinical pathological diagnosis to provide more accurate cancer discrimination results and assist doctors in making diagnosis and treatment decisions.

[0032] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0033] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0035] Figure 1 Schematic diagram of a self-supervised cancer detection system based on subspace clustering in Example 1 of the present invention;

[0036] Figure 2 This is a block diagram of a model training module in Embodiment 2 of the present invention;

[0037] Figure 3 This is a block diagram of an image processing module in Embodiment 3 of the present invention;

[0038] Figure 4 This is a block diagram of a cluster processing submodule in Embodiment 4 of the present invention;

[0039] Figure 5 The area division unit frame in Embodiment 5 of the present invention Figure 1 ;

[0040] Figure 6 The area division unit frame in Example 6 of the present invention Figure 2 ;

[0041] Figure 7 This is a block diagram of a feature extraction submodule in Embodiment 7 of the present invention;

[0042] Figure 8 This is a block diagram of the model training module in Example 8 of the present invention. DETAILED DESCRIPTION

[0043] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0044] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the embodiments of the present application. The singular forms of "a", "said" and "the" used in the embodiments of the present application are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0045] When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as described in detail. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and do not have to be used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0046] Example 1: Figure 1As shown, an embodiment of the present invention provides a self-supervised cancer detection system based on subspace clustering, comprising:

[0047] The model training module is responsible for obtaining pathological tissue sections and staining them to obtain a pathological image set. 70% of the pathological images in the pathological image set are manually annotated to obtain a training set to train the cancer discrimination and classification model.

[0048] The image processing module is responsible for receiving the pathological image to be detected, performing preprocessing operations, and inputting the pathological image after the preprocessing operation into the cancer discrimination classification model for discrimination to obtain the prediction result of the pathological image to be detected; the preprocessing operations include: scaling, cutting, subspace division and feature extraction, etc.;

[0049] The result output module is responsible for receiving the prediction results, displaying them on the pathologist's diagnosis and treatment interface, and generating a diagnosis report;

[0050] The working principle and beneficial effects of the above technical solution are as follows: the model training module of this embodiment obtains pathological tissue sections and performs staining processing to obtain a pathological image set, selects 70% of the pathological images from the pathological image set for manual annotation, obtains a training set, and trains the cancer discrimination and classification model; the image processing module receives the pathological image to be detected, performs a preprocessing operation, and inputs the pathological image after the preprocessing operation into the cancer discrimination and classification model for discrimination to obtain the prediction result of the pathological image to be detected; wherein the preprocessing operation includes: scaling, cutting, subspace division and feature extraction, etc.; the result output module receives the prediction result and displays it on the diagnosis and treatment interface of the pathologist, and generates a diagnosis report at the same time; the above scheme establishes a classification model for cancer discrimination and applies it to the intelligent diagnosis of clinical pathology; by obtaining pathological tissue sections and performing staining processing Processing, obtaining a pathological image set, obtaining data for training and testing, selecting 70% of the pathological images for manual annotation, thereby obtaining a training set, which will contain samples that have been labeled to determine whether there is cancer, for training a cancer discrimination classification model; when receiving a pathological image to be detected, first perform a preprocessing operation, such as scaling, cutting, subspace partitioning, and feature extraction, etc. The preprocessing operation is intended to extract key features in the pathological image in order to better discriminate cancer; input the preprocessed pathological image into the trained cancer discrimination classification model for discrimination, and obtain the prediction result of the pathological image to be detected; the prediction result will be displayed on the diagnosis and treatment interface of the pathologist, and a diagnosis report will be generated at the same time. The pathologist will make more accurate cancer diagnosis and treatment decisions based on the prediction results and the diagnosis report, thereby improving the accuracy and efficiency of pathological diagnosis. This embodiment uses manually annotated pathological images to train a cancer discrimination classification model, and applies it to clinical pathological diagnosis to provide more accurate cancer discrimination results and assist doctors in making diagnosis and treatment decisions.

[0051] Example 2: Figure 2 As shown, based on Example 1, the model training module provided by the embodiment of the present invention includes:

[0052] The image set division submodule is responsible for using the classification model to divide the pathological image set into a training set and a test set, and manually annotating the training set to obtain the manually annotated training set, where the ratio of the training set to the test set is 7:3;

[0053] The training set correction judgment submodule is responsible for inputting the training set into the cancer discrimination classification model for training to obtain the initial cancer discrimination classification model, and using the test set to judge whether the initial cancer discrimination classification model needs to correct the training set;

[0054] The training set correction submodule is responsible for using the current training set to train the cancer discrimination classification model if the training set does not need to be corrected; if the training set needs to be corrected, the classification model is used to search for manually annotated pathological images to be corrected from the training set; the manually annotated pathological images to be corrected are re-annotated, the training set is corrected, and the images are returned to the cancer discrimination classification model for training;

[0055] The working principle and beneficial effects of the above technical scheme are as follows: the image set division submodule of this embodiment adopts a classification model to divide the pathological image set into a training set and a test set, and manually annotates the training set to obtain a manually annotated training set, wherein the ratio of the training set to the test set is 7:3; the training set correction judgment submodule inputs the training set into the cancer discrimination classification model for training to obtain an initial cancer discrimination classification model, and uses the test set to determine whether the initial cancer discrimination classification model needs to be corrected for the training set; if the training set correction submodule does not need to correct the training set, the current training set is used to train the cancer discrimination classification model; if the training set needs to be corrected, the classification model is used to search the training set for the manually annotated pathological images to be corrected; the manually annotated pathological images to be corrected are re-annotated, the training set is corrected, and returned to the cancer discrimination classification model for training; the above scheme continuously improves and optimizes the training process of the cancer discrimination classification model in an iterative manner. process to improve its performance and accuracy; by manually annotating the training set, known labels can be provided for the training set to train the cancer discrimination classification model. The model can learn to extract features from the input pathological images and classify them according to these features to determine whether there is cancer; by using the test set to evaluate the performance of the initial cancer discrimination classification model on the new data, if the model performs well on the test set, that is, it can accurately predict the category of the pathological image, then there is no need to correct the training set. At this time, the model is trained using the current training set to further improve its performance; if the initial model performs poorly on the test set, that is, the prediction error rate is high, then the training set needs to be corrected. In this case, the initial model is used to predict the training set to find samples with incorrect predictions, that is, the manually annotated pathological images to be corrected. For these samples, manual annotation can be performed again, the training set can be corrected, and the corrected training set can be input into the model for training. Through such an iterative process, the training set can be continuously corrected and improved, and the accuracy and performance of the model can be improved. This embodiment is to continuously optimize and improve the training process of the cancer discrimination classification model in an iterative manner, thereby improving the performance and accuracy of the model, so that it can better judge whether there is cancer in the pathological image.

[0056] Example 3: Figure 3 As shown, based on Example 1, the image processing module provided by the embodiment of the present invention includes:

[0057] The clustering processing submodule is responsible for scaling and cutting the pathological images to be detected, and using subspace clustering to divide the processed pathological images into different subspaces, where each subspace corresponds to a different color distribution;

[0058] The feature extraction submodule is responsible for extracting staining-independent features from the original pathological images using a convolutional neural network as a feature extractor, where the features are used to distinguish different subspaces;

[0059] The color transformation submodule is responsible for predicting which subspace each pathological image belongs to through the cancer discrimination classification model, so that the classification model can realize color transformation and obtain the original pathological image with color distribution, that is, the prediction result of the pathological image to be detected;

[0060] The working principle and beneficial effects of the above technical solution are as follows: the clustering processing submodule of this embodiment scales and cuts the pathological image to be detected, and uses subspace clustering to divide the processed pathological image into different subspaces, where each subspace corresponds to a different color distribution; the feature extraction submodule uses a convolutional neural network as a feature extractor to extract staining-independent features from the original pathological image, where the features are used to distinguish different subspaces; the color transformation submodule predicts which subspace each pathological image belongs to through a cancer discrimination classification model, so that the classification model realizes color transformation and obtains the original pathological image with color distribution, that is, obtains the prediction result of the pathological image to be detected; the above solution uses image processing and machine learning technology to realize automatic prediction and classification of pathological images, and its specific significance includes: improving the analysis efficiency of pathological images: traditional pathological image analysis needs to rely on professional doctors for visual observation and analysis , time-consuming and easily affected by subjective factors. Using automated methods to process and classify pathological images can greatly improve the efficiency of analysis and reduce manual intervention; deeply explore the characteristics of pathological images: use convolutional neural networks as feature extractors to learn features in pathological images that are not related to staining. These features may be related to the development and progression of the disease, which helps to better understand pathological characteristics and provide clinicians with more accurate diagnosis and treatment recommendations; achieve color transformation and visualization: by predicting which subspace each pathological image belongs to, the color of the original image is transformed, and the areas of different color distribution in the pathological image are more intuitively displayed, helping doctors observe and analyze pathological characteristics; provide support for pathological research: provide data analysis tools and methods for pathological research. By processing and classifying a large number of pathological images, the development laws and characteristics of the disease can be explored, providing a scientific basis for the early diagnosis and treatment of the disease.

[0061] Example 4: Figure 4 As shown, based on Example 3, the clustering processing submodule provided in this embodiment of the present invention includes:

[0062] The region extraction unit is responsible for receiving the pathological image to be detected, performing a horizontal scaling operation on the pathological image to be detected, and then performing a vertical scaling operation on the pathological image to be detected until the size of the scaled pathological image is equal to the size of the diagnosis and treatment interface, cutting the scaled pathological image, extracting the region of interest in the pathological image, and determining the boundary and position of the region of interest;

[0063] The region division unit is responsible for obtaining the color histogram, texture features and shape features of the region of interest from the pathological image after cutting, extracting the feature vector, and using the clustering algorithm to divide the region of interest into different subspaces according to the similarity of the feature vectors;

[0064] The color distribution unit is responsible for selecting the color distribution according to the color histogram of the feature vector in each subspace and assigning different color distributions to each subspace;

[0065] The working principle and beneficial effects of the above technical scheme are as follows: the region extraction unit of this embodiment receives the pathological image to be detected, performs a horizontal scaling operation on the pathological image to be detected, and then performs a vertical scaling operation until the size of the scaled pathological image is equal to the size of the diagnosis and treatment interface, and cuts the scaled pathological image to extract the region of interest in the pathological image, and determines the boundary and position of the region of interest; the region division unit obtains the color histogram, texture features and shape features of the region of interest from the cut pathological image, and extracts feature vectors, and uses a clustering algorithm to divide the region of interest into different subspaces according to the similarity of the feature vectors; the color distribution unit selects a color distribution according to the color histogram of the feature vector in each subspace, and assigns different color distributions to each subspace; the above scheme analyzes and classifies the region of interest in the pathological image through the steps of scaling, cutting and feature extraction; the specific meanings are as follows: scaling and cutting operations: by scaling The zoom operation adjusts the image size to make it equal to the size of the diagnosis and treatment interface so as to better display and present it in the subsequent analysis and processing; the cutting operation can extract the region of interest, that is, the region related to a specific disease or lesion, thereby reducing the processing scope and complexity; feature extraction: extracting feature vectors such as color histogram, texture features and shape features from the pathological image after cutting, these feature vectors can reflect the color distribution, texture characteristics and shape characteristics of the region of interest, which is helpful for further analysis and diagnosis; clustering algorithm: using the clustering algorithm to divide the region of interest into different subspaces according to the similarity of the feature vectors, similar regions can be clustered, which is helpful for further analysis and classification; color distribution allocation: selecting the color distribution according to the color histogram of the feature vector in each subspace, assigning different color distributions to each subspace, and different regions of interest can be marked with different color distributions, which is convenient for doctors and pathologists to observe and analyze. This embodiment processes and analyzes the region of interest in the pathological image through the steps of scaling, cutting and feature extraction, thereby extracting useful feature information to assist doctors and pathologists in disease diagnosis and treatment decisions.

[0066] Example 5: Figure 5 As shown, based on Example 4, the area division unit provided in this embodiment of the present invention includes:

[0067] The threshold segmentation subunit is responsible for converting the pathological image after cutting into a grayscale image, using the edge detection algorithm to perform edge detection on the grayscale image to obtain the edge image; performing threshold segmentation on the edge image, setting the edge pixels as the foreground and the non-edge pixels as the background; performing morphological operations on the threshold segmentation results to fill the holes inside the edges, and performing connected region analysis on the results after morphological operations to obtain the location and shape information of the region of interest;

[0068] The feature acquisition subunit is responsible for obtaining the filtered images of the grayscale image in multiple directions, and fusing the filtered images in different directions to obtain the first target filtered image, and obtaining the curvature information of each pixel in the grayscale image in multiple directions; obtaining the second target filtered image according to the curvature information of each pixel in different directions, fusing the first target filtered image with the second target filtered image to obtain the texture features of the region of interest; calculating the aspect ratio of the region of interest, and determining the shape features according to the aspect ratio;

[0069] The vector synthesis subunit is responsible for dividing the pathological image of the region of interest into color space, quantizing the values ​​of the color space, obtaining the bins of the histograms of several sub-intervals, counting the number of pixels containing colors in each bin to obtain a color histogram, and quantizing the values ​​of the V channel, H channel and S channel in the color histogram; the values ​​in the bins are calculated from the color data of the pathological image to obtain feature statistics; and the weight values ​​of the color histogram, texture features and shape features are combined into feature vectors;

[0070] The working principle and beneficial effects of the above technical solution are as follows: the threshold segmentation subunit of this embodiment converts the pathological image after cutting processing into a grayscale image, uses an edge detection algorithm to perform edge detection on the grayscale image to obtain an edge image; performs threshold segmentation on the edge image, sets edge pixels as the foreground, and sets non-edge pixels as the background; performs morphological operations on the threshold segmentation results to fill the holes inside the edges, performs connected region analysis on the results after the morphological operations, and obtains the position and shape information of the region of interest; the feature acquisition subunit obtains filtered images of the grayscale image in multiple directions, and performs fusion processing on the filtered images in different directions to obtain a first target filtered image, and obtains the information of each pixel point in the grayscale image in multiple directions. curvature information; according to the curvature information of each pixel in different directions, a second target filtered image is obtained, and the first target filtered image is merged with the second target filtered image to obtain the texture features of the region of interest; the aspect ratio of the region of interest is calculated, and the shape features are determined according to the aspect ratio; the vector synthesis subunit divides the pathological image of the region of interest into color space, quantizes the values ​​of the color space, obtains the group distances of the histograms of several sub-intervals, counts the number of pixels containing colors in each group distance, obtains a color histogram, and quantifies the values ​​of the V channel, H channel and S channel in the color histogram; the values ​​in the group distance are calculated from the color data of the pathological image. The characteristic statistics; according to the color histogram, The weight values ​​of texture features and shape features are combined into feature vectors; the above scheme converts pathological images into feature vectors, so that pathological images can be quantitatively described and analyzed; specifically, the significance includes the following aspects: Feature extraction: Through edge detection, threshold segmentation, morphological operations and filtering, multiple features such as color histogram, texture features and shape features are extracted from pathological images. These features can reflect the color distribution, texture characteristics and shape information of pathological images, which helps to deeply understand the characteristics of pathological images; Feature fusion: Different features are fused together, and the color, texture and shape characteristics of pathological images are comprehensively considered to improve the description ability and discrimination of pathological images. Feature fusion can be achieved through weighted averaging. and feature connection, and make reasonable combinations according to the importance and relevance of features; feature quantification: quantify the color histogram, texture features and shape features and convert them into numerical form, which can facilitate subsequent calculations and analysis, making the feature vectors comparable and computable. The quantification process can be achieved through methods such as statistics and distribution modeling; feature vector representation: combine each feature into a feature vector according to certain rules, and be able to represent multiple features of the pathological image in a unified form. The feature vector can be used as input data for machine learning, pattern recognition and other tasks, such as pathological image classification, target detection, etc. Through the comparison and analysis of feature vectors, automatic processing and judgment of pathological images can be achieved.This embodiment converts the pathological image into a feature vector to achieve quantitative description and analysis of the pathological image, providing a basis for subsequent computer-aided diagnosis, disease prediction, etc. Through feature extraction, fusion and quantification, it can better capture the characteristic information of the pathological image and improve the understanding and analysis capabilities of the pathological image.

[0071] Example 6: Figure 6 As shown, based on Example 4, the area division unit provided in this embodiment of the present invention further includes:

[0072] The vector dot product subunit is responsible for extracting feature vectors, calculating the center vector of the region of interest using a clustering algorithm (k-means clustering algorithm), and calculating the dot product of the center vector and the feature vector to be compared, that is, multiplying the elements of the corresponding positions of the two vectors and then summing them;

[0073] The vector similarity calculation subunit is responsible for calculating the Euclidean length of the center vector and the feature vector to be compared, that is, summing the squares of each element in the vector and taking the square root; using cosine similarity to calculate the similarity between the center vector and the feature vector to be compared, the value range of cosine similarity is -1 to 1. The closer the similarity is to 1, the more similar the center vector is to the feature vector to be compared, and the closer it is to -1, the less similar the center vector is to the feature vector to be compared;

[0074] The subspace division subunit is responsible for dividing the feature vectors with similarity into a subspace according to the similarity comparison results between the feature vector and the center vector, so as to realize the division of different subspaces of the region of interest;

[0075] The working principle and beneficial effects of the above technical solution are as follows: the vector dot product subunit of this embodiment extracts the feature vector, calculates the center vector of the region of interest by using a clustering algorithm (k-means clustering algorithm), and calculates the dot product of the center vector and the feature vector to be compared, that is, multiplying the elements of the corresponding positions of the two vectors and then summing them; the vector similarity calculation subunit calculates the Euclidean length of the center vector and the feature vector to be compared, that is, summing the squares of each element in the vector and taking the square root; the cosine similarity is used to calculate the similarity between the center vector and the feature vector to be compared, and the value range of the cosine similarity is -1 to 1. The closer the similarity is to 1, the more similar the center vector is to the feature vector to be compared, and the closer it is to -1, the less similar the center vector is to the feature vector to be compared; the subspace The division subunit divides the feature vectors with similar similarity into a subspace according to the similarity comparison result between the feature vector and the center vector, so as to realize the division of different subspaces of the region of interest; the above scheme divides the vectors with similar features into the same subspace by calculating the similarity between the feature vector and the center vector, in order to realize the distinction and classification of the region of interest; by clustering similar feature vectors together, the characteristics and attributes of the region of interest can be better understood and analyzed, and this division can help to perform more effective data processing, pattern recognition, classification and clustering tasks; in addition, by calculating the similarity, different feature vectors can also be sorted to find the feature vector most similar to the center vector, so as to perform more accurate matching and recommendation. This embodiment can realize the division and classification of the region of interest by calculating the similarity between the feature vector and the center vector, so as to better understand and process the relevant data.

[0076] Example 7: Figure 7 As shown, based on Example 3, the feature extraction submodule provided in this embodiment of the present invention includes:

[0077] The convolution operation unit is responsible for inputting the original pathological images before and after staining into the feature extractor based on circular convolution and residual operation, and extracting the local features related to staining through convolution operation;

[0078] A position relationship determination unit is responsible for determining the features related to and unrelated to staining in the original pathological image based on the local features, and determining the position relationship of the features unrelated to staining based on the features related to staining;

[0079] The subspace discrimination unit is responsible for using the staining-independent features as the criteria for distinguishing different subspaces and inputting the staining-independent features into the cancer discrimination classification model;

[0080] The working principle and beneficial effects of the above technical solution are as follows: the convolution operation unit of this embodiment inputs the original pathological image before and after staining into the feature extractor based on circular convolution and residual operation, and extracts the local features related to staining through convolution operation; the position relationship determination unit determines the features related to and unrelated to staining in the original pathological image according to the local features, and determines the position relationship of the features unrelated to staining according to the features related to staining; the subspace distinction unit uses the features unrelated to staining as the standard for distinguishing different subspaces, and inputs the features unrelated to staining into the cancer discrimination and classification model; the above solution uses computer vision technology and machine learning methods to extract the features related to and unrelated to staining from the pathological image, and uses these features to perform cancer discrimination and classification; the significance is: automated feature extraction: traditional Pathological diagnosis usually relies on manual observation and analysis, while the use of computer vision technology to automatically extract and analyze features from pathological images reduces the subjectivity and time cost of manual operations; improves diagnostic accuracy: by extracting features related to staining, it can better capture subtle changes and abnormal features in pathological images, thereby improving the accuracy and sensitivity of cancer diagnosis; reduces the risk of human misjudgment: since human judgment may be subject to subjective influences and experience limitations, the use of machine learning models can reduce the risk of human misjudgment, thereby improving the reliability and consistency of cancer discrimination; assists doctors in decision-making: the automated cancer discrimination classification model can be used as an auxiliary tool for doctors, providing doctors with more objective references and decision-making support, helping doctors to more accurately assess cancer risks and lesion levels in pathological images. This embodiment improves the efficiency and accuracy of pathological diagnosis, provides better decision-making support for clinicians, and thus improves the diagnosis and treatment results and survival rate of cancer patients.

[0081] Example 8: Figure 8 As shown, based on Example 1, the model training module provided by the embodiment of the present invention includes:

[0082] The instruction issuing submodule is responsible for issuing instructions to display the prediction results on the diagnosis and treatment interface, parsing the instructions and calling the prediction results;

[0083] The interface component submodule is responsible for forming the diagnosis and treatment interface. The terminal of the diagnosis and treatment interface forms the diagnosis and treatment interface display data according to the received prediction results and the diagnosis and treatment interface request set by the user;

[0084] The result processing submodule is responsible for reading the prediction results in the diagnosis and treatment interface, visualizing the prediction results using commercial-grade data chart plug-ins, and generating a diagnosis report containing commercial-grade data charts;

[0085] The working principle and beneficial effects of the above technical scheme are as follows: the instruction issuing submodule of this embodiment issues an instruction to display the prediction results on the diagnosis and treatment interface, parses the instruction and retrieves the prediction results; the interface component submodule builds the diagnosis and treatment interface, and the terminal of the diagnosis and treatment interface builds the diagnosis and treatment interface to display data according to the received prediction results and the diagnosis and treatment interface request set by the user; the result processing submodule reads the prediction results on the diagnosis and treatment interface, uses a commercial-grade data chart plug-in to visualize the prediction results, and generates a diagnostic report containing commercial-grade data charts; the above scheme combines the prediction results of cancer detection with the diagnosis and treatment interface to provide doctors and patients with more intuitive and visual information, thereby helping to make more accurate diagnosis and treatment decisions. The specific meanings are as follows: Provide real-time prediction results: By displaying the prediction results on the diagnosis and treatment interface, doctors and patients can obtain the results of cancer detection in real time without additional waiting time, which speeds up the diagnosis; Visual data display: Visualizing the prediction results using a commercial-grade data chart plug-in can more intuitively present information such as the presence, type and malignancy of cancer, making it easier for doctors and patients to understand and analyze; Personalized diagnosis and treatment interface: The diagnosis and treatment interface can be customized according to user settings and requests to meet the needs of different doctors and patients, and improve the personalization and flexibility of diagnosis; Generate diagnostic reports: By combining the charts generated by the commercial-grade data chart plug-in with other diagnostic information, the diagnosis and treatment interface can generate a diagnostic report containing charts, which is convenient for doctors and patients to conduct further analysis and record, and provide a basis for treatment decisions. This embodiment combines the prediction results of cancer detection with the diagnosis and treatment interface, which can improve diagnostic efficiency, increase information visualization and personalization, and provide doctors and patients with a better diagnosis and treatment experience and decision support.

[0086] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention belong to the scope of the present invention and its equivalent technologies, the present invention is also intended to include these changes and variations.

Claims

1. A self-supervised cancer detection system based on subspace clustering, characterized in that: Include: The model training module is responsible for obtaining pathological tissue sections and staining them to obtain a pathological image set; pathological images are selected from the pathological image set for manual annotation to train the cancer discrimination and classification model; The image processing module is responsible for receiving the pathological image to be detected, performing preprocessing operations, and inputting the pathological image after the preprocessing operation into the cancer discrimination classification model for discrimination to obtain the prediction result of the pathological image to be detected; The result output module is responsible for receiving the prediction results, displaying them on the pathologist's diagnosis and treatment interface, and generating a diagnosis report.

2. The self-supervised cancer detection system based on subspace clustering according to claim 1, characterized in that: The preprocessing operations of the image processing module include: scaling, cutting, subspace partitioning and feature extraction.

3. The self-supervised cancer detection system based on subspace clustering as claimed in claim 1, characterized in that: Model training module, including: The image set division submodule is responsible for using the classification model to divide the pathological image set into a training set and a test set, and manually annotating the training set to obtain the manually annotated training set; The training set correction judgment submodule is responsible for inputting the training set into the cancer discrimination classification model for training to obtain the initial cancer discrimination classification model, and using the test set to determine whether the initial cancer discrimination classification model needs to correct the training set.

4. The self-supervised cancer detection system based on subspace clustering as claimed in claim 3, characterized in that: The ratio of the training set to the test set of the image set partitioning submodule is 7:

3.

5. The self-supervised cancer detection system based on subspace clustering as claimed in claim 3, characterized in that: The model training module also includes: a training set correction submodule, which is responsible for using the current training set to train the cancer discrimination classification model if the training set does not need to be corrected; if the training set needs to be corrected, the classification model is used to search for the manually annotated pathological images to be corrected from the training set.

6. The self-supervised cancer detection system based on subspace clustering according to claim 5, characterized in that: The training set correction submodule also includes: re-labeling the manually annotated pathological images to be corrected, correcting the training set, and returning the input to the cancer discrimination classification model for training.

7. The self-supervised cancer detection system based on subspace clustering as claimed in claim 1, characterized in that: Image processing module, including: The clustering processing submodule is responsible for scaling and cutting the pathological images to be detected, and using subspace clustering to divide the processed pathological images into different subspaces, where each subspace corresponds to a different color distribution; The feature extraction submodule is responsible for extracting staining-independent features from the original pathological images using a convolutional neural network as a feature extractor, where the features are used to distinguish different subspaces; The color transformation submodule is responsible for predicting which subspace each pathological image belongs to through the cancer discrimination classification model, so that the classification model can realize color transformation and obtain the original pathological image with color distribution, that is, obtain the prediction result of the pathological image to be detected.

8. The self-supervised cancer detection system based on subspace clustering according to claim 7, characterized in that: The features in the feature extraction submodule are used to distinguish different subspaces.

9. The self-supervised cancer detection system based on subspace clustering according to claim 7, characterized in that: Feature extraction submodule, including: The convolution operation unit is responsible for inputting the original pathological images before and after staining into the feature extractor based on circular convolution and residual operation, and extracting the local features related to staining through convolution operation; A position relationship determination unit is responsible for determining the features related to and unrelated to staining in the original pathological image based on the local features, and determining the position relationship of the features unrelated to staining based on the features related to staining; The subspace discrimination unit is responsible for using the staining-independent features as the criteria for distinguishing different subspaces and inputting the staining-independent features into the cancer discrimination classification model.

10. The self-supervised cancer detection system based on subspace clustering according to claim 1, characterized in that: Model training module, including: The instruction issuing submodule is responsible for issuing instructions to display the prediction results on the diagnosis and treatment interface, parsing the instructions and calling the prediction results; The interface component submodule is responsible for forming the diagnosis and treatment interface. The terminal of the diagnosis and treatment interface forms the diagnosis and treatment interface display data according to the received prediction results and the diagnosis and treatment interface request set by the user; The result processing submodule is responsible for reading the prediction results in the diagnosis and treatment interface, visualizing the prediction results using commercial-grade data chart plug-ins, and generating a diagnosis report containing commercial-grade data charts.

Images