Identification system based on appendicitis pathological section and storage medium
The convolutional neural network automatic identification system has solved the problems of misjudgment and resource scarcity in appendicitis pathological slide examination, and has achieved efficient and accurate automatic identification of appendicitis pathology, thus improving diagnostic efficiency and intelligence.
Patent Information
- Application Number
- CN202511206457.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-27
- Publication Date
- 2025-12-19
AI Technical Summary
Current appendicitis pathological examination relies on manual observation, which carries the risk of misdiagnosis and is difficult to carry out in resource-scarce areas, resulting in low diagnostic efficiency and high costs.
An automatic recognition system based on convolutional neural networks is adopted to achieve automatic recognition of appendicitis pathological slides through training image set acquisition module, convolution training module, feature extraction module and detection and recognition module, including tissue layer annotation, location of inflammatory cells and detection of hemorrhage and necrosis foci.
It improves the intelligence and collaboration of appendicitis pathological identification, reduces subjective bias, increases diagnostic efficiency, and reduces resource requirements.
Smart Images
Figure CN121169822A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of artificial intelligence, in particular to an identification system based on appendicitis pathological sections and a storage medium. BACKGROUND
[0002] With the routine development of post-appendectomy histopathological examination, the existing examination method mainly depends on professional pathologists to observe pathological sections under a microscope for judgment. This process relies on a complete pathological diagnosis laboratory service system, including professional slicing technicians, pathological section machines, microscopes and other professional equipment, as well as professional pathologists and pathological section storage facilities. This puts forward higher requirements for the conditions of the hospital to develop pathological services.
[0003] Under the background of the prior art, this examination method has the following deficiencies: 1. When the pathologist observes the section by naked eye, there may be misjudgment due to experience difference, and there is subjective bias; 2. In some low-income countries or regions, the pathological laboratory resources are scarce, and the corresponding pathological services may not be developed. Even if the pathological services are developed, more time, economic and human costs need to be invested. SUMMARY
[0004] The application aims to provide an identification system based on appendicitis pathological sections and a storage medium to solve the above technical problems and realize automatic identification of appendicitis pathological sections, effectively improving the intelligent and collaborative level of appendicitis pathological identification.
[0005] In order to solve the above technical problems, the application provides an identification system based on appendicitis pathological sections, which comprises a training image set acquisition module, a convolution training module, a to-be-detected image acquisition module, a feature extraction module and a detection and identification module, specifically:
[0006] The training image set acquisition module is used to acquire a tissue layer labeled image set, a hemorrhagic necrotic focus labeled image set and a tumor mucus labeled image set;
[0007] The convolution training module is used to establish an initial convolutional neural network model, train the initial convolutional neural network model based on the tissue layer labeled image set, the hemorrhagic necrotic focus labeled image set and the tumor mucus labeled image set, and acquire the convolutional neural network model until a preset convergence condition is reached.
[0008] The to-be-detected image acquisition module is used to acquire a to-be-detected appendicitis pathological section image;
[0009] The feature extraction module is used to acquire pathological section feature data, inflammatory cell position data and hemorrhagic necrotic focus detection data based on the convolutional neural network model and the to-be-detected appendicitis pathological section image;
[0010] The detection and recognition module is configured to acquire a pathological detection result of the slice based on the pathological slice feature data, the inflammatory cell position data and the hemorrhagic necrosis focus detection data, and to realize pathological recognition of the appendicitis pathological slice.
[0011] In the above scheme, the training image set acquisition module is configured to acquire the tissue layer annotation image set, the hemorrhagic necrosis focus annotation image set and the tumor mucus annotation image set, so as to provide labeled basic sample data for training of the initial convolutional neural network model, to serve as a reference standard for the initial convolutional neural network model to learn pathological features, and to ensure that the model can accurately capture target features. Then, the feature extraction module is configured to convert the to-be-detected appendicitis pathological slice image into intermediate feature data that can be used to judge the pathology on the slice by using the trained convolutional neural network model. Finally, the detection and recognition module is configured to convert the intermediate feature data into the pathological detection result of the slice, to realize automatic recognition of the pathology on the appendicitis pathological slice, and to effectively improve the intelligent and collaborative level of the appendicitis pathological recognition.
[0012] Further, the feature extraction module is configured to acquire the pathological slice feature data, the inflammatory cell position data and the hemorrhagic necrosis focus detection data based on the convolutional neural network model and the to-be-detected appendicitis pathological slice image, and includes:
[0013] Based on the convolutional neural network model and the to-be-detected appendicitis pathological slice image, acquire a layered annotation image, pathological slice feature data and preliminary hemorrhagic necrosis focus detection data
[0014] Based on the layered annotation image, acquire a layered annotation matrix, wherein an element in the layered annotation matrix corresponds to a tissue layer category in which each pixel in the layered annotation image is located.
[0015] Based on the layered annotation matrix and the preliminary hemorrhagic necrosis focus detection data, acquire hemorrhagic necrosis focus detection data.
[0016] For any tissue layer in the layered annotation image, perform the following steps: perform inflammatory cell recognition on the tissue layer to acquire an inflammatory cell recognition result of the tissue layer.
[0017] Based on the inflammatory cell recognition results of all the tissue layers in the layered annotation image, acquire inflammatory cell position data.
[0018] In the above scheme, the convolutional neural network is used to label the pathological section images of the appendicitis to be detected, and the spatial position and range of each tissue layer in the pathological image are determined as the spatial reference for subsequent construction of the tissue layer graph model and hierarchical recognition of inflammatory cells. The pathological section feature data obtained mainly includes features related to low-grade mucinous tumors of the appendix, which provides a basis for the subsequent detection and recognition module to determine whether it is a tumor. Then, combined with the to-be-detected tissue layer annotation matrix and the preliminary hemorrhagic necrosis focus detection data, the deviation of the preliminary hemorrhagic necrosis focus detection data can be checked and corrected, and more accurate hemorrhagic necrosis focus detection data can be output, which provides a basis for the subsequent detection and recognition module to determine whether it is acute gangrenous appendicitis. Finally, the distribution of inflammatory cells in each tissue layer is accurately positioned by hierarchical annotation of images, which reflects the spatial distribution characteristics of inflammatory cells and is used for the subsequent detection and recognition module to distinguish whether it is acute simple appendicitis.
[0019] Further, the hemorrhagic necrosis focus detection data is obtained based on the hierarchical annotation matrix and the preliminary hemorrhagic necrosis focus detection data; including:
[0020] The tissue layer graph model is constructed based on the hierarchical annotation matrix, and the tissue layer graph model includes first nodes, second nodes, third nodes and fourth nodes, and is labeled as 1, 2, 3 and 4 respectively; wherein the first nodes, the second nodes, the third nodes and the fourth nodes correspond to the mucosa layer, the submucosa layer, the muscle layer and the serosa layer respectively;
[0021] Based on the hierarchical annotation matrix, a binary mask matrix corresponding to each node is obtained;
[0022] Based on the binary mask matrix of any two nodes, the edge weight value between the tissue layers corresponding to the two nodes is calculated;
[0023] Based on the edge weight value and the pre-trained graph neural network model, the detection result of the tissue layer graph model is obtained;
[0024] Based on the detection result and the preliminary hemorrhagic necrosis focus detection data, the hemorrhagic necrosis focus detection data is obtained.
[0025] In the above scheme, the abstract organization layers are converted into specific nodes in the organization layer graph model, and the correspondence of each organization layer is clear. Moreover, the node labels in the organization layer graph model are directly related to the labels of the hierarchical labeling matrix, ensuring that the organization layer graph model corresponds one-to-one with the organization layer position of the original image. Then, the actual position distribution of each pixel in the hierarchical labeling image is clearly displayed through the binary mask matrix, which is used to locate the boundary edge between the two organization layers for subsequent positioning, ensuring the accuracy of the edge weight calculation. Then, the edge weight of the organization layer corresponding to any two nodes is calculated, and the edge weight can directly reflect the degree of damage of inflammation to the organization layer structure. Further. By inputting the edge weight into the pre-trained graph neural network model, the detection result of the organization layer graph model is output, which provides a verification standard for subsequent acquisition of optimized bleeding and necrosis focus detection data. Finally, the detection result is associated with the preliminary bleeding and necrosis focus detection data, and the accurate bleeding and necrosis focus detection data is output, which provides a reliable basis for the detection and recognition module to determine whether it is acute gangrenous appendicitis.
[0026] Further, the binary mask matrix based on any two nodes calculates the edge weight between the organization layers corresponding to the two nodes; specifically:
[0027] Based on the binary mask matrix of any two nodes, the edge pixel point set of the organization layers corresponding to the two nodes is extracted;
[0028] Based on the edge pixel point set, the pixel intensity of the edge pixel point is calculated;
[0029] According to the pixel intensity, the edge weight between the organization layers corresponding to the two nodes is obtained.
[0030] In the above scheme, since the binary mask matrix of each node marks the pixel position of the respective organization layer, the edge pixel point set of the organization layers corresponding to the two nodes is extracted through the binary mask matrix of the nodes, which can accurately lock the specific pixel coordinates of the boundary of the organization layers corresponding to the two nodes, thereby ensuring that the subsequent calculation is only for the edge pixel points, avoiding irrelevant information interference, and ensuring the pertinence and accuracy of the edge weight. Since the edge weight can directly reflect the degree of clear boundary between organization layers, it can be used to input the pre-trained graph neural network model to obtain the detection result of the organization layer graph model, which provides a basis for the detection and recognition module to determine whether it is acute gangrenous appendicitis.
[0031] Further, the following steps are performed on any organization layer in the hierarchical labeling image: performing inflammation cell recognition on the organization layer to obtain the inflammation cell recognition result of the organization layer; including:
[0032] The following steps are performed on any organization layer in the hierarchical labeling image:
[0033] The image processing algorithm is used for identifying and marking the inflammatory cells in the tissue layer to construct an inflammatory cell group graph structure of the tissue layer;
[0034] The unsupervised learning algorithm is used for latent semantic segmentation and marking of the tissue layer to construct an inflammatory cell tissue group graph structure of the tissue layer;
[0035] Based on the inflammatory cell group graph structure and the inflammatory cell tissue group graph structure, the inflammatory cell features of the tissue layer and the inflammatory cell tissue group features of the tissue layer are extracted;
[0036] Based on the inflammatory cell group graph structure and the inflammatory cell tissue group graph structure, the inflammatory cell composite features of the tissue layer are obtained;
[0037] According to the inflammatory cell composite features, the tissue layer is identified to obtain the inflammatory cell recognition result of the tissue layer.
[0038] In the above scheme, the potential inflammatory cells are identified and marked from the tissue layer by the image processing algorithm, and the interference of non-cell regions is excluded. Then the single cell features of the inflammatory cells are extracted from the inflammatory cell group graph structure, and the inflammatory cell tissue group features are extracted from the inflammatory cell tissue group graph structure, which can describe the morphology and distribution of the cells themselves and between the cells, and convert the spatial correlation information in the graph structure into quantifiable features to provide a basic feature dimension for subsequent fusion analysis. Then by splicing the inflammatory cell group graph structure and the inflammatory cell tissue group graph structure into composite features, the individual morphology and group distribution of the inflammatory cells can be reflected, and the discrimination degree of the inflammatory cells is improved. Finally, the inflammatory cell recognition result in the tissue layer is output, which can provide a basis for subsequent judgment of the detection result of the appendicitis pathological section image.
[0039] Further, the inflammatory cell composite features of the tissue layer are obtained based on the inflammatory cell group graph structure and the inflammatory cell tissue group graph structure; comprising:
[0040] Based on the inflammatory cell group graph structure, the inflammatory cell tissue group graph structure and the graph feature evolution algorithm, the inflammatory cell peripheral cell features of the tissue layer and the inflammatory cell tissue peripheral tissue features of the tissue layer are extracted;
[0041] The inflammatory cell features, the inflammatory cell peripheral cell features and the inflammatory cell tissue peripheral tissue features are spliced to obtain the inflammatory cell composite features of the tissue layer.
[0042] In the above scheme, by capturing the correlation characteristics of inflammatory cells and the surrounding environment, the limitations of single cell characteristics are compensated for, the feature dimension is enriched, the characteristics of the surrounding cells and the surrounding tissue region of the inflammatory cells can be reflected, the synergistic characteristics of the cell population are captured, and the limitations of relying only on single cell characteristics are broken. Then, the inflammatory cell characteristics, the inflammatory cell peripheral cell characteristics and the inflammatory cell tissue peripheral tissue characteristics are spliced, multi-dimensional features are fused, and a complete feature vector including "self-peripheral cell-peripheral tissue" is formed, providing a comprehensive basis for subsequent identification of inflammatory cells. The inflammatory cell complex feature formed after splicing integrates multi-dimensional information, can more accurately distinguish inflammatory cells from non-inflammatory cells, greatly improves the accuracy of subsequent full connection neural network classification, and lays a foundation for accurate identification of inflammatory cells.
[0043] Further, the identification of the tissue layer according to the inflammatory cell complex feature obtains an inflammatory cell identification result of the tissue layer; comprising:
[0044] Obtain an inflammatory cell labeled image set;
[0045] Establish an initial full connection neural network model, train the initial full connection neural network model based on the inflammatory cell labeled image set until a preset convergence condition is reached, and obtain a full connection neural network model;
[0046] According to the inflammatory cell complex feature and the full connection neural network model, the tissue layer is identified, and an inflammatory cell identification result of the tissue layer is obtained.
[0047] In the above scheme, by learning the inflammatory cell labeled image set by the initial full connection neural network, the parameters of the model can be continuously adjusted until the model converges. The trained full connection neural network model is used to identify the inflammatory cells of the tissue layer, and the identification result of the inflammatory cells in the tissue layer is output, which provides basic data for subsequent judgment of the distribution range of inflammatory cells and directly supports the subsequent identification of the appendicitis pathological section.
[0048] Further, the detection and identification module is configured to obtain a pathological detection result of the section based on the pathological section feature data, the inflammatory cell position data and the hemorrhagic necrosis focus detection data, and to realize identification of the appendicitis pathological section; specifically:
[0049] If the pathological section feature data exceeds the preset threshold, it is determined that the detection result of the to-be-detected appendicitis pathological section image is a low-grade mucinous tumor of the appendix;
[0050] If the pathological section feature data does not exceed the preset threshold and the inflammatory cell position data is in the preset area, it is determined that the detection result of the to-be-detected appendicitis pathological section image is acute simple appendicitis;
[0051] If the pathological section feature data does not exceed the preset threshold value, the inflammatory cell position data exceeds the preset region, and the hemorrhagic necrotic focus detection data meets the preset hemorrhagic necrotic standard, it is judged that the detection result of the appendicitis pathological section image to be detected is acute gangrenous appendicitis.
[0052] If the pathological section feature data does not exceed the preset threshold value, the inflammatory cell position data exceeds the preset region, and the hemorrhagic necrotic focus detection data does not meet the preset hemorrhagic necrotic standard, it is judged that the detection result of the appendicitis pathological section image to be detected is acute suppurative appendicitis.
[0053] In the above scheme, by judging the tumor first and then distinguishing the type of appendicitis according to the inflammatory cell position data and the hemorrhagic necrotic focus detection data in a progressive rule, the quantitative features of the pathological section are converted into a structured recognition result, realizing the automatic and standardized recognition of the three types of acute appendicitis and low-grade mucinous tumor of the appendix, reducing subjective bias and improving diagnosis efficiency.
[0054] Further, if the pathological section feature data exceeds the preset threshold value, it is judged that the detection result of the appendicitis pathological section image to be detected is low-grade mucinous tumor of the appendix; specifically:
[0055] The pathological section feature data includes tumor cell mass data and mucous area data;
[0056] If the tumor cell mass data exceeds the first preset threshold value or the mucous area data exceeds the second preset threshold value, it is judged that the detection result of the appendicitis pathological section image to be detected is low-grade mucinous tumor of the appendix.
[0057] In the above scheme, by locking the core specificity indicators of low-grade mucinous tumor of the appendix, the tumor cell mass data and the mucous area data can exclude features unrelated to tumors. Through the double threshold value logic, different forms of tumors are covered, ensuring that no typical cases are missed, and the accuracy of tumor recognition is improved.
[0058] The present application provides a non-transitory computer readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the functions of the system as described above. BRIEF DESCRIPTION OF DRAWINGS
[0059] Figure 1 An architecture diagram of an identification system based on appendicitis pathological sections is provided for an embodiment of the present application;
[0060] Figure 2 A flowchart of an identification system based on appendicitis pathological sections is provided for an embodiment of the present application. DETAILED DESCRIPTION
[0061] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by those ordinarily skilled in the art without creative effort belong to the scope of the present application.
[0062] The present embodiment provides an identification system based on appendicitis pathological sections, the architecture diagram of which is shown in Figure 1 , the flowchart of which is shown in Figure 2 , which comprises a training image set acquisition module, a convolution training module, a to-be-detected image acquisition module, a feature extraction module and a detection and identification module, specifically:
[0063] The training image set acquisition module is configured to acquire a tissue layer annotated image set, a hemorrhagic necrotic focus annotated image set and a tumor mucous annotated image set.
[0064] The convolution training module is configured to establish an initial convolutional neural network model, train the initial convolutional neural network model based on the tissue layer annotated image set, the hemorrhagic necrotic focus annotated image set and the tumor mucous annotated image set, until a preset convergence condition is reached, and acquire a convolutional neural network model.
[0065] The to-be-detected image acquisition module is configured to acquire a to-be-detected appendicitis pathological section image.
[0066] The feature extraction module is configured to acquire pathological section feature data, inflammatory cell position data and hemorrhagic necrotic focus detection data based on the convolutional neural network model and the to-be-detected appendicitis pathological section image.
[0067] The detection and identification module is configured to acquire a pathological detection result of the section based on the pathological section feature data, the inflammatory cell position data and the hemorrhagic necrotic focus detection data, and realize pathological identification of the appendicitis pathological section.
[0068] In the embodiment, the training image set acquisition module is used to acquire the tissue layer annotation image set, the hemorrhagic necrosis focus annotation image set and the tumor mucous annotation image set (tumor cell cluster data and mucous area data are respectively annotated as 1 and 2, and the background is marked as 0), which can provide annotated basic sample data for the training of the initial convolutional neural network model, serve as a reference standard for the initial convolutional neural network model to learn pathological features, and ensure that the model can accurately capture target features. The initial convolutional neural network model adjusts its weights and feature detectors by learning the marked layers in the annotation image set until a preset convergence condition is reached, and obtains the convolutional neural network model. Then, the feature extraction module converts the appendicitis pathological section image to be detected into intermediate feature data that can be used to judge the pathology on the section by using the trained convolutional neural network model. Finally, the detection and recognition module converts the intermediate feature data into the pathological detection result of the section, realizes the automatic recognition of the pathology on the appendicitis pathological section, and effectively improves the intelligent and collaborative level of the appendicitis pathology recognition.
[0069] Further, the feature extraction module is configured to acquire pathological section feature data, inflammatory cell position data and hemorrhagic necrosis focus detection data based on the convolutional neural network model and the appendicitis pathological section image to be detected, and includes:
[0070] acquire hierarchical annotation images, pathological section feature data and preliminary hemorrhagic necrosis focus detection data based on the convolutional neural network model and the appendicitis pathological section image to be detected
[0071] acquire a hierarchical annotation matrix based on the hierarchical annotation images, wherein elements in the hierarchical annotation matrix correspond to tissue layer categories in which each pixel in the hierarchical annotation images is located;
[0072] acquire hemorrhagic necrosis focus detection data based on the hierarchical annotation matrix and the preliminary hemorrhagic necrosis focus detection data;
[0073] perform the following steps on any tissue layer in the hierarchical annotation images: perform inflammatory cell recognition on the tissue layer to obtain inflammatory cell recognition results of the tissue layer;
[0074] acquire inflammatory cell position data based on the inflammatory cell recognition results of all tissue layers in the hierarchical annotation images.
[0075] In the embodiment, the convolutional neural network is used to stratify and label the appendicitis pathological section images to be detected, and the spatial positions and ranges of each tissue layer in the pathological images are determined as the spatial reference for subsequent construction of tissue layer graph model and stratified recognition of inflammatory cells. The element values in the stratification and labeling matrix are 0, 1, 2, 3 or 4, which correspond to background, mucosa layer, submucosa layer, muscle layer or serosa layer respectively. The obtained pathological section feature data mainly includes features related to low-grade mucinous tumor of the appendix, which provides a basis for the subsequent detection and recognition module to determine whether it is a tumor. The trained convolutional neural network model is applied to the appendicitis pathological section images to be detected, and the preliminary hemorrhagic necrosis focus region detection data of the appendicitis pathological section images to be detected can be obtained, wherein 0 represents background and 1 represents a cell region with hemorrhagic necrosis. Then, combined with the to-be-detected tissue layer labeling matrix and the preliminary hemorrhagic necrosis focus detection data, the deviation of the preliminary hemorrhagic necrosis focus detection data can be checked and corrected, and more accurate hemorrhagic necrosis focus detection data can be output, which provides a basis for the subsequent detection and recognition module to determine whether it is acute gangrenous appendicitis. Finally, the distribution of inflammatory cells in each tissue layer is accurately positioned by stratified labeling images, which reflects the spatial distribution characteristics of inflammatory cells and is used for the subsequent detection and recognition module to distinguish whether it is acute simple appendicitis.
[0076] Further, the hemorrhagic necrosis focus detection data is obtained based on the stratification and labeling matrix and the preliminary hemorrhagic necrosis focus detection data; including:
[0077] The tissue layer graph model is constructed based on the stratification and labeling matrix, and the tissue layer graph model includes first node, second node, third node and fourth node, and is labeled as 1, 2, 3 and 4 respectively; wherein the first node, the second node, the third node and the fourth node correspond to mucosa layer, submucosa layer, muscle layer and serosa layer respectively;
[0078] Based on the stratification and labeling matrix, a binary mask matrix corresponding to each node is obtained;
[0079] Based on the binary mask matrix of any two nodes, the edge weight value between the tissue layers corresponding to the two nodes is calculated;
[0080] Based on the edge weight value and the pre-trained graph neural network model, the detection result of the tissue layer graph model is obtained;
[0081] Based on the detection result and the preliminary hemorrhagic necrosis focus detection data, the hemorrhagic necrosis focus detection data is obtained.
[0082] In the embodiment, the abstract organization layers are converted into specific nodes in the organization layer graph model, and the correspondence of the organization layers is determined. The labels of the nodes in the organization layer graph model are directly related to the labels of the hierarchical labeling matrix, ensuring that the nodes in the organization layer graph model correspond one-to-one with the organization layers of the original image. The connection between nodes is represented by edges, and the energy value (i.e., edge weight) of the edge represents the clearness of the boundary between the organization layers, thereby providing a visual representation of the hierarchical layers. Then, the actual position distribution of each pixel in the hierarchical labeling image is clearly displayed through the binary mask matrix (where 0 represents the background and 1 represents the corresponding region position), which is used to locate the boundary edge between the two organization layers in the subsequent step, ensuring the accuracy of the edge weight calculation. Next, the edge weight between the organization layers corresponding to any two nodes is calculated, and the edge weight obtained by the calculation can directly reflect the degree of damage of inflammation to the organization layer structure. Further, by inputting the edge weight into the pre-trained graph neural network model, the detection result of the organization layer graph model is outputted, which refers to the probability that the appendicitis pathological section image to be detected is acute gangrenous appendicitis or acute suppurative appendicitis, which can provide a verification standard for subsequent acquisition of optimized hemorrhagic necrosis focus detection data. Finally, the detection result is associated with the preliminary hemorrhagic necrosis focus detection data for analysis, to determine whether there is a necrosis focus and obvious structural change, thereby outputting accurate hemorrhagic necrosis focus detection data, which provides a reliable basis for the detection and recognition module to determine whether it is acute gangrenous appendicitis.
[0083] Further, the binary mask matrix based on any two nodes is used to calculate the edge weight between the organization layers corresponding to the two nodes. Specifically,
[0084] The binary mask matrix based on any two nodes is used to extract the edge pixel point set of the organization layers corresponding to the two nodes.
[0085] Based on the edge pixel point set, the pixel intensity of the edge pixel point is calculated.
[0086] According to the pixel intensity, the edge weight between the organization layers corresponding to the two nodes is obtained.
[0087] In the embodiment, since the binary mask matrix of each node marks the pixel position of the respective tissue layer, the edge pixel set of the two nodes corresponding to the tissue layers can be extracted by the binary mask matrix of the node, so as to accurately lock the specific pixel coordinates of the intersection of the two nodes corresponding to the tissue layers, thereby ensuring that the subsequent calculation is only for the edge pixel points, avoiding irrelevant information interference, and ensuring the pertinence and accuracy of the edge weight. Since the edge weight can directly reflect the clear degree of the boundary between the tissue layers, it can be used to obtain the detection result of the tissue layer graph model by inputting the pre-trained graph neural network model in the following, thereby providing a basis for the detection and recognition module to determine whether it is acute gangrenous appendicitis. In the embodiment, the edge weight can be calculated by an energy function, and the calculation formula of the edge weight between the i th node and the j th node is: wherein, E(i,j) is the edge weight (also referred to as the energy value of the edge) between the i th node and the j th node, P(i,j) represents the edge pixel set of the two tissue layer regions corresponding to the i th node and the j th node, and I represents the pixel intensity at the pixel point (x,y). The calculation of the edge weight considers the pixel intensity gradient in the image, so the higher the energy value, the clearer the boundary between the layers.
[0088] Further, the following steps are performed on any tissue layer in the layered annotation image: performing inflammation cell recognition on the tissue layer to obtain an inflammation cell recognition result of the tissue layer; comprising:
[0089] The following steps are performed on any tissue layer in the layered annotation image:
[0090] Performing inflammation cell recognition and inflammation cell labeling on the tissue layer based on an image processing algorithm to construct an inflammation cell group graph structure of the tissue layer;
[0091] Performing potential semantic segmentation and labeling on the tissue layer based on an unsupervised learning algorithm to construct an inflammation cell tissue group graph structure of the tissue layer;
[0092] Based on the inflammation cell group graph structure and the inflammation cell tissue group graph structure, extracting inflammation cell features of the tissue layer and inflammation cell tissue group features of the tissue layer;
[0093] Based on the inflammation cell group graph structure and the inflammation cell tissue group graph structure, obtaining an inflammation cell composite feature of the tissue layer;
[0094] According to the inflammation cell composite feature, recognizing the tissue layer to obtain an inflammation cell recognition result of the tissue layer.
[0095] In this embodiment, potential inflammatory cells are identified and labeled from the tissue layer by an image processing algorithm, and the cell area is labeled as 1 and the background is labeled as 0 in the output image, excluding the interference of non-cell areas. An unsupervised learning algorithm is used to perform potential semantic segmentation and labeling on the area that may contain information helpful for cell type judgment, and the tissue area is labeled as 1 and the background is labeled as 0 in the output image. For each inflammatory cell, the cells within a given distance range are extracted to form the nodes of the cell group graph structure, and the distance between the cells is judged. If the distance is less than a given distance threshold, an edge is established between the two cells, otherwise not, thereby forming the cell group graph structure of the inflammatory cell. The given distance threshold is set according to the situation of the pathological image in the actual application scenario. In the 40x pathological image in this embodiment, it is set to 20. The construction of the corresponding inflammatory cell tissue group graph structure for each inflammatory cell tissue is also a similar establishment method. Then the monomer features of the inflammatory cells are extracted from the inflammatory cell group graph structure, and the inflammatory cell tissue group features are extracted from the inflammatory cell tissue group graph structure, which can describe the morphology and distribution of the cells themselves and between the cells, and convert the spatial correlation information in the graph structure into quantifiable features, providing a basic feature dimension for subsequent fusion analysis. By calculating the inflammatory cell group graph structure and the inflammatory cell tissue group graph structure respectively, the mean value of the gray value, the variance of the gray value, the area, the perimeter, the long axis length, the short axis length, the eccentricity, and the compactness of the corresponding inflammatory cell and inflammatory tissue area are obtained. The corresponding inflammatory cell eight-dimensional vector and inflammatory cell tissue eight-dimensional vector formed are called inflammatory cell features and inflammatory cell tissue group features. Then by splicing the inflammatory cell group graph structure and the inflammatory cell tissue group graph structure into a composite feature, the individual morphology and population distribution of the inflammatory cells can be reflected, and the discrimination degree of the inflammatory cells is improved. Finally, the inflammatory cell recognition result in the tissue layer is output, which can provide a basis for subsequent detection of the detection result of the appendicitis pathological section image to be detected.
[0096] Further, the inflammatory cell composite feature of the tissue layer is obtained based on the inflammatory cell group graph structure and the inflammatory cell tissue group graph structure; comprising:
[0097] Based on the inflammatory cell group graph structure, the inflammatory cell tissue group graph structure, and the graph feature evolution algorithm, the inflammatory cell peripheral cell feature of the tissue layer and the inflammatory cell tissue peripheral tissue feature of the tissue layer are extracted;
[0098] The inflammatory cell features, the inflammatory cell peripheral cell features, and the inflammatory cell tissue peripheral tissue features are spliced to obtain the inflammatory cell composite feature of the tissue layer.
[0099] In this embodiment, the specific calculation process of the inflammatory cell peripheral cell features and the inflammatory cell tissue peripheral tissue features based on the graph feature evolution according to the inflammatory cell group graph structure and the inflammatory cell tissue group graph structure respectively is as follows: wherein, E (0) is the single-cell feature of the node, A and D are the adjacency matrix and the degree matrix of the corresponding graph structure respectively. Set k to and K=3. Thus, the node features corresponding to the inflammatory cells or the inflammatory cell tissues can be obtained from E. By capturing the associated features of the inflammatory cells and the peripheral environment, the limitations of the single-cell features are compensated for, the feature dimension is enriched, and the features of the neighboring cells and the peripheral tissue regions around the inflammatory cells can be reflected, the collaborative features of the cell groups can be captured, and the limitations of relying only on single-cell features are broken. Then, the inflammatory cell features, the inflammatory cell peripheral cell features, and the inflammatory cell tissue peripheral tissue features are spliced, multi-dimensional features are fused, and a complete feature vector containing “self-peripheral cell-peripheral tissue” is formed, which provides a comprehensive basis for the subsequent identification of inflammatory cells. The inflammatory cell composite features formed after splicing integrate multi-dimensional information, can more accurately distinguish inflammatory cells from non-inflammatory cells, greatly improve the accuracy of subsequent full-connection neural network classification, and lay a foundation for the accurate identification of inflammatory cells.
[0100] Further, the identification of the tissue layer according to the inflammatory cell composite features to obtain the inflammatory cell identification result of the tissue layer comprises:
[0101] obtaining an inflammatory cell labeled image set;
[0102] establishing an initial full-connection neural network model, training the initial full-connection neural network model based on the inflammatory cell labeled image set until a preset convergence condition is reached, and obtaining a full-connection neural network model;
[0103] identifying the tissue layer according to the inflammatory cell composite features and the full-connection neural network model to obtain the inflammatory cell identification result of the tissue layer.
[0104] In this embodiment, by enabling the initial full-connection neural network to learn the inflammatory cell labeled image set, the parameters of the model can be continuously adjusted until the model converges. The trained full-connection neural network model is used to identify the inflammatory cells of the tissue layer, and the identification result of the inflammatory cells in the tissue layer is output, which provides basic data for the subsequent judgment of the distribution range of the inflammatory cells and directly supports the subsequent identification of the appendicitis pathological section.
[0105] Further, the detection recognition module is configured to acquire a pathological detection result of the slice based on the pathological slice feature data, the inflammatory cell position data and the hemorrhagic necrosis focus detection data, so as to realize recognition of the appendicitis pathological slice; specifically,
[0106] If the pathological slice feature data exceeds the preset threshold, it is determined that the detection result of the to-be-detected appendicitis pathological slice image is low-grade mucus tumor of the appendix.
[0107] If the pathological slice feature data does not exceed the preset threshold and the inflammatory cell position data is in the preset region, it is determined that the detection result of the to-be-detected appendicitis pathological slice image is acute simple appendicitis.
[0108] If the pathological slice feature data does not exceed the preset threshold, the inflammatory cell position data exceeds the preset region, and the hemorrhagic necrosis focus detection data meets the preset hemorrhagic necrosis standard, it is determined that the detection result of the to-be-detected appendicitis pathological slice image is acute gangrenous appendicitis.
[0109] If the pathological slice feature data does not exceed the preset threshold, the inflammatory cell position data exceeds the preset region, and the hemorrhagic necrosis focus detection data does not meet the preset hemorrhagic necrosis standard, it is determined that the detection result of the to-be-detected appendicitis pathological slice image is acute purulent appendicitis.
[0110] In this embodiment, by using the progressive rule of first judging tumor and then distinguishing the types of appendicitis according to the inflammatory cell position data and the hemorrhagic necrosis focus detection data, the quantitative features of the pathological slice are converted into a structured recognition result, so as to realize automatic and standardized recognition of the three types of acute appendicitis and low-grade mucus tumor of the appendix, reduce subjective bias and improve diagnosis efficiency. It is determined whether the inflammatory cells exist only in the preset region (the preset region) by the inflammatory cell position data. For example, if the pathological slice feature data does not exceed the preset threshold and the inflammatory cell position data is in the preset region, the to-be-detected appendicitis pathological slice image is recognized as acute simple appendicitis. In this embodiment, the preset hemorrhagic necrosis standard is that there is a necrosis focus and obvious structural change, such as destruction of the gland or disorder of the hierarchy, so that the to-be-detected appendicitis pathological slice image is determined to be acute gangrenous appendicitis.
[0111] Further, if the pathological slice feature data exceeds the preset threshold, it is determined that the detection result of the to-be-detected appendicitis pathological slice image is low-grade mucus tumor of the appendix; specifically,
[0112] The pathological slice feature data includes tumor cell cluster data and mucus region data;
[0113] If the tumor cell cluster data exceeds the first preset threshold or the mucus region data exceeds the second preset threshold, it is determined that the detection result of the to-be-detected appendicitis pathological slice image is low-grade mucus tumor of the appendix.
[0114] In the embodiment, by locking the core specificity indicators of the low-grade mucus tumor of the appendix, the tumor cell mass data and the mucus area data (both refer to the number of pixels contained therein), features unrelated to the tumor can be excluded. The first preset threshold is set for the tumor cell mass data, and the second preset threshold is set for the mucus area data. By double threshold logic, if the tumor cell mass data or the mucus area data is greater than the set threshold, it is judged that the tumor cell mass or the mucus area exists, which guarantees to cover different forms of tumors, ensures not to miss typical cases, and improves the accuracy of tumor identification. The first preset threshold and the second preset threshold in the embodiment can be set according to the specific use scene.
[0115] The embodiment provides a non-transitory computer-readable storage medium, which stores a computer program. The computer program is executed by a processor to realize the functions of the system described above.
[0116] The embodiment can reflect the good effect of the recognition system based on the appendix inflammation pathological section provided by the embodiment by calculating the precision and recall rates. Taking acute simple appendicitis as an example, the precision rate represents the proportion of the to-be-detected appendix inflammation pathological section images that are judged to be acute simple appendicitis, and the actual detection is also the to-be-detected appendix inflammation pathological section images; the recall rate represents the proportion of the actual detection that is correctly judged to be acute simple appendicitis. The recognition system based on the appendix inflammation pathological section provided by the embodiment has a recognition precision rate of 0.8475 and a recall rate of 0.8772 for acute simple appendicitis, a recognition precision rate of 0.9111 and a recall rate of 0.8506 for acute gangrenous appendicitis, and a recognition precision rate of 0.8182 and a recall rate of 0.8889 for acute purulent appendicitis. It can be seen that the recognition system based on the appendix inflammation pathological section provided by the embodiment has high accuracy and sensitivity for acute appendicitis recognition, and can accurately identify all low-grade mucus tumors of the appendix, and has good working performance and can provide important value for clinical pathology work.
[0117] The above describes the preferred embodiments of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements are also considered within the protection scope of the present application.
Claims
1. A recognition system based on appendicitis pathological slides, characterized in that, It includes a training image set acquisition module, a convolution training module, a target image acquisition module, a feature extraction module, and a detection and recognition module, specifically: The training image set acquisition module is used to acquire the tissue layer labeled image set, the hemorrhage and necrosis foci labeled image set, and the tumor mucus labeled image set; The convolutional training module is used to establish an initial convolutional neural network model. The initial convolutional neural network model is trained based on the tissue layer labeled image set, the hemorrhage and necrosis foci labeled image set, and the tumor mucus labeled image set until the preset convergence condition is reached, and the convolutional neural network model is obtained. The image acquisition module is used to acquire pathological slide images of appendicitis to be detected. The feature extraction module is used to obtain pathological slide feature data, inflammatory cell location data, and hemorrhage and necrosis foci detection data based on the convolutional neural network model and the pathological slide image of the appendicitis to be detected. The detection and identification module is used to obtain the pathological test results of the slides based on pathological slide feature data, inflammatory cell location data, and hemorrhage and necrosis foci detection data, thereby realizing the pathological identification of appendicitis pathological slides.
2. The identification system based on appendicitis pathological sections according to claim 1, characterized in that, The feature extraction module is used to acquire pathological slide feature data, inflammatory cell location data, and hemorrhage and necrosis foci detection data based on a convolutional neural network model and the pathological slide image of the appendicitis to be detected; including: Based on the convolutional neural network model and the pathological slide image of the appendicitis to be detected, hierarchical labeled images, pathological slide feature data, and preliminary hemorrhage and necrosis foci detection data are obtained. A hierarchical annotation matrix is obtained based on the hierarchical annotation image; wherein, the elements in the hierarchical annotation matrix correspond to the tissue layer category in which each pixel in the hierarchical annotation image is located; Based on the hierarchical annotation matrix and preliminary hemorrhage and necrosis foci detection data, obtain hemorrhage and necrosis foci detection data; For any tissue layer in the layered labeled image, perform the following steps: identify inflammatory cells in the tissue layer and obtain the inflammatory cell identification results for the tissue layer; Based on the identification results of inflammatory cells in all tissue layers of the layered labeled image, the location data of inflammatory cells are obtained.
3. The identification system based on appendicitis pathological sections according to claim 2, characterized in that, The process of obtaining hemorrhagic necrosis foci detection data based on a hierarchical annotation matrix and preliminary hemorrhagic necrosis foci detection data includes: A tissue layer diagram model is constructed based on a hierarchical annotation matrix. The tissue layer diagram model includes a first node, a second node, a third node, and a fourth node, which are labeled as 1, 2, 3, and 4, respectively. The first node, the second node, the third node, and the fourth node correspond to the mucosa, the submucosa, the muscularis propria, and the serosa, respectively. Based on the hierarchical labeling matrix, obtain the binary mask matrix corresponding to each node; Calculate the edge weights between the corresponding organizational layers of any two nodes based on the binarized mask matrix of any two nodes. Based on the edge weights and the pre-trained graph neural network model, the detection results of the tissue layer graph model are obtained; Based on the detection results and the preliminary hemorrhagic necrosis foci detection data, the hemorrhagic necrosis foci detection data are obtained.
4. The identification system based on appendicitis pathological sections according to claim 3, characterized in that, The edge weights between the organizational layers corresponding to these two nodes are calculated based on the binarized mask matrix of any two nodes. Specifically: Based on the binary mask matrix of any two nodes, extract the set of edge pixels of the corresponding organization layer of these two nodes; Based on the set of edge pixels, calculate the pixel intensity of the edge pixels; The edge weight between the corresponding tissue layers of these two nodes is obtained based on the pixel intensity.
5. The identification system based on appendicitis pathological sections according to claim 2, characterized in that, The steps for performing the following operations on any tissue layer in the layered labeled image are: identifying inflammatory cells in the tissue layer and obtaining the inflammatory cell identification results for the tissue layer; including: Perform the following steps on any organizational layer in a layered labeled image: Based on image processing algorithms, inflammatory cells in this tissue layer are identified and labeled, and an inflammatory cell population map structure of this tissue layer is constructed. Based on unsupervised learning algorithms, the tissue layer is segmented and labeled with latent semantics, and the inflammatory cell tissue group graph structure of the tissue layer is constructed. Based on the inflammatory cell population map structure and the inflammatory cell tissue population map structure, the inflammatory cell features and the inflammatory cell tissue population features of the tissue layer are extracted. Based on the inflammatory cell population map structure and the inflammatory cell tissue population map structure, the complex inflammatory cell features of the tissue layer are obtained. The tissue layer is identified based on the complex characteristics of the inflammatory cells, and the inflammatory cell identification results of the tissue layer are obtained.
6. The identification system based on appendicitis pathological sections according to claim 4, characterized in that, The step of obtaining the complex inflammatory cell features of the tissue layer based on the inflammatory cell population map structure and the inflammatory cell tissue population map structure includes: Based on the inflammatory cell swarm graph structure, the inflammatory cell tissue swarm graph structure, and the graph feature evolution algorithm, the cell features surrounding the inflammatory cells in the tissue layer and the tissue features surrounding the inflammatory cells in the tissue layer are extracted. The inflammatory cell features, the cell features surrounding the inflammatory cells, and the tissue features surrounding the inflammatory cell tissue are spliced together to obtain the composite inflammatory cell features of the tissue layer.
7. The identification system based on appendicitis pathological sections according to claim 4, characterized in that, The step of identifying the tissue layer based on the complex characteristics of the inflammatory cells and obtaining the inflammatory cell identification results of the tissue layer includes: Obtain a set of labeled images of inflammatory cells; An initial fully connected neural network model is established, and the initial fully connected neural network model is trained based on a set of inflammatory cell labeled images until the preset convergence condition is met, thereby obtaining the fully connected neural network model. Based on the complex features of the inflammatory cells and the fully connected neural network model, the tissue layer is identified, and the inflammatory cell identification results of the tissue layer are obtained.
8. The identification system based on appendicitis pathological sections according to claim 1, characterized in that, The detection and identification module is used to obtain the pathological examination results of the slides based on pathological slide feature data, inflammatory cell location data, and hemorrhage and necrosis foci detection data, thereby achieving pathological identification of appendicitis pathological slides; specifically: If the pathological section feature data exceeds the preset threshold, the detection result of the appendicitis pathological section image to be detected is determined to be low-grade mucinous tumor of the appendix. If the pathological section feature data does not exceed the preset threshold and the inflammatory cell location data is within the preset area, then the detection result of the appendicitis pathological section image to be detected is determined to be acute simple appendicitis. If the pathological section feature data does not exceed the preset threshold, the inflammatory cell location data exceeds the preset area, and the hemorrhage and necrosis foci detection data meet the preset hemorrhage and necrosis criteria, then the detection result of the appendicitis pathological section image to be detected is determined to be acute gangrenous appendicitis. If the pathological section feature data does not exceed the preset threshold, the inflammatory cell location data exceeds the preset area, and the hemorrhage and necrosis foci detection data does not meet the preset hemorrhage and necrosis criteria, then the detection result of the appendicitis pathological section image to be tested is determined to be acute suppurative appendicitis.
9. The identification system based on appendicitis pathological sections according to claim 7, characterized in that, If the pathological section feature data exceeds a preset threshold, the detection result of the appendicitis pathological section image to be detected is determined to be a low-grade mucinous tumor of the appendix; specifically: The pathological section feature data includes tumor cell cluster data and mucus region data; If the tumor cell cluster data exceeds the first preset threshold or the mucus region data exceeds the second preset threshold, the detection result of the appendicitis pathological section image to be detected is determined to be a low-grade mucinous tumor of the appendix.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the functions of the system as described in any one of claims 1 to 9.