Ethanol-based anca-iif image processing method, system, device and medium

Abstract

The present application relates to an ANCA-IIF image processing method, system, device and medium based on an ethanol matrix, belonging to the technical field of medical detection image processing. The method comprises: performing antibody class positive and negative screening detection on a target image; performing target detection on the positive target image to obtain a region of interest in the target image and first fluorescence mode class data thereof; obtaining second fluorescence mode class data and embedded representation of each region of interest from a first model; constructing node features and edge features of the target image; determining a final fluorescence mode class of the target image based on the node features and edge features of the target image, and the final fluorescence mode class of the target image being one of multiple preset fluorescence mode classes. The present application significantly improves the classification accuracy of the fine classification of the ANCA nuclear type of the ethanol matrix, and effectively improves the detection efficiency.

Description

ANCA-IIF Image Processing Methods, Systems, Equipment, and Media Based on Ethanol Matrix Technical Field

[0001] This invention relates to the field of medical imaging processing technology, and in particular to an ANCA-IIF image processing method, system, device and medium based on an ethanol matrix. Background Technology

[0002] Immunofluorescence is a powerful technique built upon immunology, biochemistry, and microscopy. It primarily utilizes the fluorescence pattern resulting from the binding of fluorescently labeled antibodies and antigens for qualitative disease analysis. Due to its high sensitivity, specificity, and speed, it is widely used in scientific research and clinical diagnosis. Indirect immunofluorescence (IIF) is a classic technique in medical testing for autoantibody detection. Its basic principle is as follows: A solid-phase detection matrix containing a biological matrix expressing the target antigen is prepared. The sample to be tested (such as human serum) is incubated with this matrix under suitable conditions, and then a fluorescein-labeled anti-human immunoglobulin antibody (i.e., secondary antibody) is added. If the sample contains specific antibodies against the target antigen (i.e., primary antibodies), they will bind to the antigenic components on the matrix, and subsequently to the fluorescently labeled antibody, forming a specific fluorescent characteristic.

[0003] Anti-neutrophil cytoplasmic antibodies (ANCAs) are important serological markers for various autoimmune diseases and play a crucial role in clinical diagnosis. The main target antigens of ANCAs include typical target antigens closely associated with vasculitis (AAV), such as proteinase 3 (PR3) and myeloperoxidase (MPO), as well as atypical target antigens such as endothelial elastase (HLE) and cathepsin G (CG). In addition, ANCAs also involve target antigens related to other autoimmune diseases, such as antibacterial / permeability-enhancing protein (BPI) associated with ulcerative colitis and antilactin (LF) associated with systemic lupus erythematosus. Therefore, ANCA detection based on IIF method is not only the gold standard for initial screening and monitoring of disease activity of ANCA-associated vasculitis (AAV) such as granulomatous polyangiitis (GPA) and microscopic polyangiitis (MPA), but it also shows a high positive rate in other autoimmune diseases such as inflammatory bowel disease (IBD), autoimmune liver disease, antiglomerular basement membrane disease, as well as infectious diseases, malignant tumors, and drug-induced immune responses, suggesting its broad clinical relevance in various pathological states.

[0004] Clinically, ANCA detection typically employs the IIF method (hereinafter referred to as the IIF-ANCA detection method). However, the specific fluorescence characteristics in the IIF images used in the IIF-ANCA detection method (hereinafter referred to as ANCA-IIF images) exhibit various morphological features due to the diversity of ANCA target antigens and the redistribution of target antigens generated by the fixation matrix. Specifically, the fluorescence patterns in ANCA-IIF images (hereinafter referred to as IIF-ANCA fluorescence patterns) include perinuclear ANCA (p-ANCA) and cytoplasmic ANCA (c-ANCA). The perinuclear ANCA fluorescence pattern further includes typical p-ANCA (tP) and atypical p-ANCA (aP), while the cytoplasmic ANCA fluorescence pattern includes typical c-ANCA (tC) and atypical c-ANCA (aC). Compared to the single fluorescence pattern in antibody detection of a single specific target antigen, this complexity of the IIF-ANCA fluorescence pattern undoubtedly increases the difficulty of interpreting ANCA-IIF fluorescence images. In addition, the presence of other interfering factors such as antinuclear antibodies (ANA) in the sample will further increase the difficulty of interpretation.

[0005] Currently, the methods for interpreting fluorescence patterns in ANCA-IIF images include the following:

[0006] 1. Indirect immunofluorescence images are captured by a camera connected to a microscope and then manually interpreted by laboratory personnel.

[0007] Second, the readings are directly observed and interpreted under a microscope by experienced technicians.

[0008] The two methods described above are currently the commonly used interpretation methods for IIF-ANCA fluorescence patterns. They rely on manual identification of fluorescence and morphological features, resulting in high subjectivity, insufficient standardization, and low interpretation efficiency. Furthermore, for atypical ANCA, due to insufficiently significant fluorescence characteristics, the accuracy of the detection results highly depends on the professional level and interpretation experience of the technicians. This can lead to a high false positive rate and reduce detection efficiency.

[0009] Third, traditional machine learning methods are employed, involving manual feature extraction from ANCA-IIF images followed by analysis using shallow classification models. For example, the article "Expert Consensus on the Clinical Application of Anti-neutrophil Cytoplasmic Antibody Detection" published in the September 2018 issue of the Chinese Journal of Medical Laboratory Science (Vol. 41, No. 9, p. 644) discloses a scheme for classifying ANCA ethanol-fixed matrix fluorescence images using manually designed features combined with algorithms such as logistic regression and decision trees. Currently, machine learning-based interpretation methods can only achieve coarse-grained classification of basic fluorescence patterns of c-ANCA and p-ANCA, making it difficult to accurately identify atypical ANCA widely present in clinical samples. Summary of the Invention

[0010] To address the technical problems existing in the prior art, this invention proposes an ANCA-IIF image processing method, system, device, and medium based on an ethanol matrix, which is used to accurately classify various IIF-ANCA fluorescence modes based on an ethanol matrix, thereby improving detection efficiency.

[0011] To address the aforementioned technical problems, according to one aspect of the present invention, an ANCA-IIF image processing method based on an ethanol matrix is ​​provided, comprising:

[0012] The target image is subjected to a positive or negative screening test for antibody categories to determine whether the antibody category of the test sample corresponding to the target image is positive or negative. The target image is an indirect immunofluorescence image of an ANCA test sample based on an ethanol matrix, and the indirect immunofluorescence image includes multiple cell regions.

[0013] In response to the antibody category of the detection sample corresponding to the target image being positive, target detection is performed on the target image to obtain the region of interest in the target image and its first fluorescence mode category data. Each region of interest corresponds to a cell region. The first fluorescence mode category data includes one of a variety of preset fluorescence mode categories and its confidence level.

[0014] Each region of interest is input into the trained first model to obtain the second fluorescence pattern category data and embedded representation of each region of interest. The second fluorescence pattern category data includes the probabilities of multiple preset fluorescence pattern categories.

[0015] Using each region of interest in the target image as a node, the second fluorescence mode category data and the first fluorescence mode category data of each region of interest are fused to construct node features;

[0016] Construct edge features between nodes based on the embedded representation of each region of interest; and

[0017] The final fluorescence mode category of the target image is determined based on the node features and edge features of the target image. The final fluorescence mode category of the target image is one of a variety of preset fluorescence mode categories.

[0018] Optionally, the step of performing a positive or negative antibody category screening test on the target image to determine whether the antibody category of the test sample corresponding to the target image is positive or negative includes:

[0019] The target image is input into a trained classification model, which then outputs a classification category for the target image.

[0020] If the classification model outputs a fluorescent mode, the antibody class of the sample is determined to be positive; if the classification model outputs a non-fluorescent mode, the antibody class of the sample is determined to be negative.

[0021] The training sample set of the classification model is an indirect immunofluorescence image set of ANCA-detected samples based on an ethanol matrix. The training sample set is labeled as having a fluorescent mode and not having a fluorescent mode, and a training sample is labeled as having a fluorescent mode or not having a fluorescent mode.

[0022] Optionally, the step of performing target detection on the target image to obtain the region of interest and its first fluorescence pattern category data in the target image includes:

[0023] The target image is input into the trained target detection model; and

[0024] The target detection model outputs the region of interest detected from the target image and its first fluorescence mode category data;

[0025] The training samples for the target detection model are indirect immunofluorescence images of ANCA detection samples based on an ethanol matrix with cell regions annotated. The annotation information for each cell region includes one of a variety of preset fluorescence mode categories, such as typical c-ANCA, atypical c-ANCA, typical p-ANCA, atypical p-ANCA and nonspecific ANCA.

[0026] Optionally, the first model is a neural network model. Correspondingly, when inputting each region of interest into the trained first model to obtain the second fluorescence pattern category data and embedded representation of each region of interest, the first model processes the input regions of interest as follows:

[0027] The first information is obtained by extracting features from the input region of interest through multi-scale branches, wherein the first information includes feature information at multiple scales.

[0028] The second information is obtained by dynamically weighting the feature information at each scale and the inter-scale correlation information in the first information using an attention mechanism; and

[0029] The second information is used to perform prediction output operations to obtain the second fluorescence pattern category data and embedded representation of the input region of interest;

[0030] The training sample set of the neural network model consists of cell images extracted from indirect immunofluorescence images of samples detected by ANCA on an ethanol matrix; each sample image is labeled with one of several preset fluorescence mode categories.

[0031] Optionally, when extracting features from the input region of interest through multi-scale branches to obtain the first information, a preset number of feature iterations are performed to extract the first information.

[0032] Optionally, the first multi-scale feature extraction step on the input region of interest includes:

[0033] Initial feature extraction is performed on the input region of interest to transform the region of interest information in the spatial domain into region of interest information in the channel domain;

[0034] The target information is the region of interest information converted to the channel domain;

[0035] Feature extraction of target information is performed using multiple convolutional kernels of different scales to obtain feature information at multiple scales;

[0036] The feature information at multiple scales is concatenated to obtain feature concatenation information; and

[0037] Average pooling is performed on the concatenated feature information to obtain average pooling information;

[0038] The average pooling information is the target information for the next iteration.

[0039] Optionally, the ANCA-IIF image processing method based on the ethanol matrix further includes: extracting features from the target information through residual branching to obtain the original feature information of the target information; correspondingly, the step of stitching together the feature information at multiple scales to obtain feature stitching information includes:

[0040] The feature information from the multiple scales is spliced ​​together to obtain intermediate splicing information; and

[0041] The feature splicing information is obtained by summing the intermediate splicing information and the original feature information extracted through residual branch.

[0042] Optionally, before extracting features from the target information using multiple convolutional kernels of different scales, the method further includes: performing normalization, activation function calculation, and max pooling operations on the target information in sequence.

[0043] Optionally, after extracting features from the target information to obtain feature information at multiple scales and before concatenating the feature information at multiple scales, the method further includes: performing normalization operations and activation function calculations on each scale feature information in sequence to obtain feature information of the same dimension; correspondingly, when concatenating the feature information at multiple scales, feature information of multiple scales of the same dimension is concatenated to obtain feature concatenation information.

[0044] Optionally, the step of dynamically weighting the feature information at each scale and the correlation information between scales in the first information using an attention mechanism to obtain the second information includes:

[0045] The batch size is determined based on the number of scales of the feature information in the first information; wherein, the batch size is determined based on the following formula: Where N is the total batch size. To measure the number of branches, The number of batches required for a single-scale branch. The interval distance between scale branches. The scale branch interval distance is The corresponding correlation weight coefficient at that time Indicates in In each scale branch, the interval distance is The number of scale branch pairs;

[0046] The first piece of information is divided into batches of a certain number of batches.

[0047] Attention is calculated for each batch of information to obtain the corresponding calculation results; and

[0048] The attention calculation results of all batch information are combined to obtain the second information.

[0049] Optionally, the step of concatenating the attention calculation results of all batch information to obtain the second information includes:

[0050] The attention calculation results of all batch information are concatenated to obtain the attention calculation result concatenation information; and

[0051] The second information is obtained by summing the attention calculation result, the spliced ​​information, and the first information.

[0052] Optionally, the step of constructing edge features between nodes based on the embedded representation of each region of interest includes:

[0053] Calculate the embedded representation distance between all pairs of nodes, and generate the adjacency matrix A of the target image based on the embedded representation distance between all pairs of nodes;

[0054] Based on the adjacency matrix A, a degree matrix D is constructed, where each diagonal element of the degree matrix D represents the connection between a node in the corresponding row and the remaining nodes in the same row; and

[0055] The edge feature matrix is ​​obtained by summing the adjacency matrix A and the degree matrix D. .

[0056] Optionally, the adjacency matrix element in the i-th row and j-th column of adjacency matrix A Calculated using the following formula:

[0057] ,

[0058] in, and These represent the embedded representations of the nodes in the i-th and j-th rows, respectively.

[0059] Diagonal elements in degree matrix D Calculated using the following formula:

[0060] ,

[0061] The distance threshold for the embedded representation of a node; t is the indicator function; t is the column number.

[0062] Optionally, the step of determining the final fluorescence mode category of the target image based on the node features and edge features of the target image includes:

[0063] The second model input data is constructed based on the node and edge features of the target image; and

[0064] The input data of the second model is fed into the trained second model to obtain the final fluorescence pattern category of the target image.

[0065] Optionally, the second model is a machine learning model or a deep learning model. Correspondingly, the step of constructing the input data of the second model based on the node features and edge features of the target image includes:

[0066] The node features and edge features of the target image are aggregated to obtain a feature vector of fixed length;

[0067] Correspondingly, the machine learning model or deep learning model processes the feature vector as input to obtain the final fluorescence pattern category of the target image.

[0068] Optionally, the second model is a graph neural network model; correspondingly, the step of constructing the input data of the second model based on the node features and edge features of the target image includes:

[0069] A node feature matrix V is constructed based on the number of nodes and node features of the target image;

[0070] Construct a side feature matrix based on the number of nodes and the edge features of the nodes in the target image. ;as well as

[0071] Using the node feature matrix V and the edge feature matrix As input to the graph neural network model;

[0072] Correspondingly, the graph neural network model considers the node feature matrix V and the edge feature matrix V. The final fluorescence pattern category of the target image is obtained through processing.

[0073] According to another aspect of the present invention, the present invention also provides an ANCA-IIF image processing system based on an ethanol matrix, comprising:

[0074] The positive and negative detection module is configured to perform positive and negative screening detection on the target image to determine whether the antibody category of the test sample corresponding to the target image is positive or negative. The target image is an indirect immunofluorescence image of an ANCA test sample based on an ethanol matrix, and the indirect immunofluorescence image includes multiple cell regions.

[0075] The target detection module is configured to perform target detection on the target image in response to the antibody category of the detection sample corresponding to the target image being positive, so as to obtain the region of interest in the target image and its first fluorescence mode category data. Each region of interest corresponds to a cell region. The first fluorescence mode category data includes one of a variety of preset fluorescence mode categories and its confidence level.

[0076] The first classification module is configured to input each region of interest into the trained first model to obtain the second fluorescence pattern category data and embedded representation of each region of interest. The second fluorescence pattern category data includes the probabilities of multiple preset fluorescence pattern categories.

[0077] The graph feature construction module is configured to use each region of interest (ROI) in the target image as a node, and fuses the second fluorescence mode category data and the first fluorescence mode category data of each ROI to construct node features; it also constructs edge features between nodes based on the embedded representation of each ROI; and

[0078] The second classification module is configured to determine the final fluorescence mode category of the target image based on the node features and edge features of the target image. The final fluorescence mode category of the target image is one of a variety of preset fluorescence mode categories.

[0079] According to another aspect of the present invention, an electronic device is also provided, comprising a processor and a memory, wherein a set of computer program instructions is stored in the memory, and the aforementioned ethanol-based ANCA-IIF image processing method or system is implemented when the processor executes the set of computer program instructions in the memory.

[0080] According to another aspect of the present invention, the present invention also provides a computer-readable storage medium storing a set of computer program instructions, which, when executed by a processor, implements the aforementioned ANCA-IIF image processing method or system based on an ethanol matrix.

[0081] According to another aspect of the present invention, the present invention also provides a computer program product comprising a computer program instruction set, which, when executed by a processor, implements the aforementioned ethanol-based ANCA-IIF image processing method or system.

[0082] This invention enhances the ability to distinguish similar fluorescent morphologies by improving fine-grained feature characterization and adapting to fluorescence features at different scales. It not only achieves fine classification of ANCA karyotypes in ethanol matrix, but also enables accurate differentiation between typical and atypical cytoplasmic and perinuclear karyotypes, effectively improving detection efficiency. Attached Figure Description

[0083] The preferred embodiments of the present invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0084] Figure 1 is a flowchart of an ANCA-IIF image processing method based on an ethanol matrix according to an embodiment of the present invention;

[0085] Figure 2 is a schematic diagram of the original ethanol-based ANCA-IIF image according to an embodiment of the present invention;

[0086] Figure 3 is a schematic diagram of the image to be processed after preprocessing the original ethanol-based ANCA-IIF image according to an embodiment of the present invention.

[0087] Figure 4 is a schematic diagram of a cell region with a typical perinuclear fluorescence pattern (tP);

[0088] Figure 5 is a schematic diagram of a cell region with an atypical perinuclear fluorescence pattern (aP).

[0089] Figure 6 is a schematic diagram of a cell region with a typical cytoplasmic fluorescence pattern (tC);

[0090] Figure 7 is a schematic diagram of a cell region with an atypical cytoplasmic fluorescence pattern (aC);

[0091] Figure 8 is a schematic diagram of a cellular region with a nonspecific ANCA fluorescence pattern (NS);

[0092] Figure 9 is a schematic diagram of the network structure of a first model according to an embodiment of the present invention;

[0093] Figure 10 is a flowchart of a multi-scale feature extraction method according to an embodiment of the present invention;

[0094] Figure 11 is a flowchart of an attention calculation method according to an embodiment of the present invention;

[0095] Figure 12 is a flowchart of a neural network model training method according to an embodiment of the present invention;

[0096] Figure 13 is a heatmap of the statistical data matrix applied when determining positive and negative samples according to an embodiment of the present invention;

[0097] Figure 14 is a flowchart of a method for constructing graph features according to an embodiment of the present invention;

[0098] Figure 15 is a schematic diagram of the structure of a graph neural network model according to an embodiment of the present invention;

[0099] Figure 16 is a schematic block diagram of an ANCA-IIF image processing system based on an ethanol matrix according to an embodiment of the present invention; and

[0100] Figure 17 is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0101] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0102] In the following detailed description, reference can be made to the accompanying drawings, which form part of this application and illustrate specific embodiments of the present application. In the drawings, similar reference numerals describe substantially similar components in different figures. Specific embodiments of the present application are described in sufficient detail below to enable those skilled in the art to implement the technical solutions of the present application. It should be understood that other embodiments may also be utilized, or structural, logical, or electrical changes may be made to the embodiments of the present application.

[0103] Because IIF-ANCA assays can simultaneously cover multiple target antigens, they offer irreplaceable advantages over antibody assays targeting single specific antigens. Clinically, to address the accuracy issues arising from the coexistence of multiple fluorescence patterns due to the diversity of ANCA target antigens, a common approach is to combine multiple substrates for detection. For example, ethanol-fixed neutrophils (ethanol substrate), formaldehyde-fixed neutrophils (formaldehyde substrate), and a mixture of ethanol-fixed HEp-2 and neutrophil cells (mixed substrate) can be used. This multi-substrate assay undoubtedly increases the complexity of the assay and presents a significant challenge to the clinical interpretation of IIF-based ANCA.

[0104] The main reasons for the difficulty in distinguishing different IIF-ANCA fluorescence modes for ethanol-based matrices include:

[0105] 1) p-ANCA morphology is highly variable, and some p-ANCA morphologies are easily confused with c-ANCA. For example, in ethanol-fixed neutrophils with MPO as the target antigen, ethanol disrupts the phospholipid bilayer of the cell membrane, driving MPO migration with positive and negative charges. Positively charged MPO migrates towards the negatively charged nuclear region, forming a pseudo-perinuclear pattern, i.e., pseudo-p-ANCA. Furthermore, the morphology of neutrophils is complex, ranging from immature cells (1-2 nuclei) to transitionally mature cells (>5 nuclei), and different cell activation states all affect fluorescence characteristics in indirect immunofluorescence detection, such as interlobular distance and cytoplasmic granules.

[0106] 2) Significant differences exist in the expression levels and distribution locations of different antigens within neutrophils, leading to a large number of atypical fluorescence patterns (such as aC and aP). These atypical fluorescence pattern categories lack clear interpretation boundaries and are easily confused with typical pattern categories. However, these two atypical patterns can indicate different target antigens, which is of practical significance for the clinical diagnosis and disease monitoring of patients.

[0107] 3) The presence of ANA antibodies in the sample can easily lead to atypical ANCA fluorescence morphology, resulting in a large number of typical and atypical patterns being mixed together, which greatly interferes with and confuses the image interpretation task.

[0108] Based on the above analysis, it is evident that for a positive ethanol-based ANCA-IIF image, there are three main categories of fluorescence patterns: p-ANCA, c-ANCA, and non-specific ANCA (NS). Further subdivisions include tC, aC, tP, aP, and NS. Traditional processing methods struggle to accurately distinguish these subcategories of fluorescence patterns, relying primarily on manual identification. This invention proposes an IIF image processing method and system capable of accurately classifying the various IIF-ANCA fluorescence patterns based on an ethanol matrix, thereby improving detection efficiency. Referring to Figure 1, which is a flowchart of an ethanol-based ANCA-IIF image processing method according to an embodiment of the present invention, the ethanol-based ANCA-IIF image processing method includes the following steps:

[0109] Step S1: Perform positive and negative screening detection of antibody categories on the target image.

[0110] Step S2: Determine whether the positive or negative screening test result is positive. If the positive or negative screening test result is positive, proceed to step S3. If it is negative instead of positive, end the processing flow.

[0111] Step S3: Perform target detection on the target image to obtain the region of interest (ROI) and its first fluorescence pattern category data. Each ROI corresponds to a cell region, and the first fluorescence pattern category data includes one of several preset fluorescence pattern categories and their confidence levels. Preset fluorescence pattern categories include, for example, tC, aC, tP, aP, and NS.

[0112] Step S4: Input each region of interest into the trained first model to obtain the second fluorescence pattern category data and embedded representation of each region of interest. The second fluorescence pattern category data includes the probability of each preset fluorescence pattern category, such as the probability of tC, the probability of aC, the probability of tP, the probability of aP, and the probability of NS.

[0113] Step S5: Construct node features and edge features.

[0114] Step S6: Determine the final fluorescence mode category of the target image based on the node features and edge features of the target image. The fluorescence mode category of the target image is one of several preset fluorescence mode categories, that is, the final determination of which of the aforementioned five modes the fluorescence mode category of the target image is.

[0115] In step S1, before performing antibody class positive / negative screening on the target image, the original ethanol-based ANCA-IIF image is preprocessed to improve image quality stability. See Figure 2, which is a schematic diagram of the original ethanol-based ANCA-IIF image according to an embodiment of the present invention; it is a clinically collected 20X objective lens fluorescence image of a cell matrix microscope. Figure 3 is a schematic diagram of the image to be processed after preprocessing the original image according to an embodiment of the present invention. The preprocessing includes, for example, cropping the image to a specified size, normalizing the pixel values ​​of the cropped image, and then storing it as the image to be processed. The target image in step S1 is the preprocessed image to be processed. The normalization process in this invention can employ any algorithm, such as Min-Max normalization, z-score normalization, logarithmic function normalization, arctangent function normalization, etc. Those skilled in the art can choose any algorithm based on application habits, and will not be elaborated further here.

[0116] In one embodiment, step S1 employs a classification model to screen and detect the positive or negative results of the target images. This classification model can be, for example, a machine learning model or a neural network model. To train the classification model, a sufficient number of indirect immunofluorescence images taken from clinical ANCA test samples are first collected to form an initial image set. The initial images are collected, for example, by photographing fluorescent slides of the samples under a fluorescence microscope. Specifically, the sample slide is placed under a fluorescence microscope connected to a camera, and the camera randomly selects 1-8 fields of view to photograph the fluorescent slides, thus obtaining 1-8 fluorescence images of the same sample. Alternatively, the initial fluorescence images can be collected from other data sources. The initial images are then cropped and normalized to a specified size. The images are then labeled. For example, clinical laboratory experts assign fluorescence mode labels to the preprocessed fluorescence images; for example, 0 represents no fluorescence mode, indicating a negative result, and 1 represents fluorescence mode, indicating a positive result. Furthermore, the fluorescence image labeled 1 should contain a sufficient number of specific fluorescent cells. The labeled images are then stored as samples. Therefore, the label types of the training sample set of the classification model used to screen and detect the positive and negative values ​​of the target image are fluorescent mode and non-fluorescent mode, and the label of a training sample is fluorescent mode or non-fluorescent mode.

[0117] The sample set is then divided into a training set and a validation set. For example, 80% of the samples in the sample set are assigned to the training set, and 20% are assigned to the validation set. The neural network is trained based on the samples in the training set, and the model performance is evaluated on the validation set. When the model loss converges and meets the performance evaluation requirements, a classification model for positive and negative screening detection of antibody categories is obtained.

[0118] In another embodiment, for subsequent model training needs, when constructing the sample set for the classification model and labeling the samples, specific fluorescence pattern categories can also be labeled. For example, six different categories of fluorescence patterns, namely no fluorescence, tC, aC, tP, aP, and NS, can be labeled with different numbers. Among the six different categories of fluorescence patterns, except for the no fluorescence pattern, the other fluorescence patterns represent positive results. When manually labeling cells, the cell region is manually outlined with a rectangular frame, and the fluorescence pattern category of the region is labeled according to fluorescence characteristics, morphological characteristics, etc. See Figures 4 to 8. Figure 4 is a schematic diagram of a cell region with a typical perinuclear fluorescence pattern (tP); Figure 5 is a schematic diagram of a cell region with an atypical perinuclear fluorescence pattern (aP); Figure 6 is a schematic diagram of a cell region with a typical cytoplasmic fluorescence pattern (tC); Figure 7 is a schematic diagram of a cell region with an atypical cytoplasmic fluorescence pattern (aC); and Figure 8 is a schematic diagram of a cell region with a nonspecific fluorescence pattern (NS). When annotating samples, this invention combines fluorescence features and morphological features to label the fluorescence pattern category of each cell region in detail. This not only enables positive and negative screening and classification, but also serves as a tool for subsequent target detection models and the training of the first model.

[0119] In step S3, target detection is performed on the target image using a target detection model. The target detection model can be a deep learning-based neural network model. For example, a one-stage detection network or a two-stage detection network can be used. Algorithms for one-stage detection networks include YOLO (You Only Look Once) and SSD (Single Shot MultiBox Detector), while algorithms for two-stage detection networks include Faster R-CNN (Faster Region-based Convolutional Neural Network), Mask R-CNN, and Cascade R-CNN. During the training of the target detection model, cell regions are manually outlined with rectangular boxes in the sample set, and the fluorescence pattern category of each region is labeled based on fluorescence characteristics. The fluorescence pattern categories are tC, aC, tP, aP, and NS. Therefore, after image processing using the aforementioned algorithm, rectangular regions containing cell regions, i.e., regions of interest (ROIs), can be detected from the image, and corresponding descriptive information is output. The descriptive information includes region coordinates, fluorescence pattern category, and confidence level. The region coordinate value consists of four values: the coordinates of the top-left corner of the region, the region width, and the height. In this embodiment, since it is necessary to distinguish between five fluorescence pattern categories, the fluorescence pattern category data output by the target detection model is set as a data structure consisting of 5 binary digits. Each data bit represents a fluorescence pattern category, and the binary digits 0 and 1 represent the model's judgment result, respectively. 0 represents that it is not a fluorescence pattern category, and 1 represents that it is a fluorescence pattern category. The fluorescence pattern category confidence level is, for example, a number between 0 and 1. The larger the value, the greater the probability that the ROI belongs to that category, and the higher the confidence level. Therefore, the fluorescence pattern category data output by the target detection model includes one of the five fluorescence pattern categories and its confidence level.

[0120] By performing target detection on the target image in step S3, multiple regions of interest (ROIs) are obtained from the target image. Then, based on the order of information intensity from high to low, a predetermined number of m ROIs are selected from all ROIs for subsequent processing.

[0121] Referring to Figure 9, which is a schematic diagram of the network structure of a first model according to an embodiment of the present invention, the first model in step S4 is a neural network model with a specific structure, including an input layer 41, a multi-scale feature extraction layer 42, an attention calculation layer 43, and a prediction layer 44. The input layer 41 is used to receive ROI image data. In one embodiment, an ROI image data is a 256×256 image data containing 3 color channels. The multi-scale feature extraction layer 42 performs multi-scale feature extraction on the input ROI image data to obtain first information D1.

[0122] Referring to Figure 10, which is a flowchart of a multi-scale feature extraction method according to an embodiment of the present invention, this embodiment specifically includes the following steps:

[0123] Step S421: Perform initial feature extraction on the input ROI image data. For example, use a 7×7 convolution kernel to perform convolution operation on the input ROI image using a sliding window method to extract low-level feature information of the image to obtain an initial feature map.

[0124] Step S422 involves feature enhancement processing of the initial features. For example, the initial feature map is input into the BRM module, which sequentially performs normalization, activation function calculation, and max pooling. Normalization eliminates dimensional differences, stabilizes the data distribution, mitigates gradient vanishing / exploding, accelerates training, and allows for larger learning rates. Activation function calculation, such as ReLU, increases non-linear expressiveness, enabling the network to fit complex functions. Finally, max pooling achieves dimensionality reduction and feature enhancement.

[0125] Step S423 involves sending the feature-enhanced information to different scale branches for feature extraction. For example, in the first convolutional scale branch (kernel branch), a 1×1 convolution kernel is used to extract combined features between channels; in the second convolutional scale branch, a 3×3 convolution kernel is used to extract mid-scale spatial features; and in the third convolutional scale branch, a 5×5 convolution kernel is used to extract large-scale spatial features. The feature extraction from each convolutional scale branch yields a feature map of the corresponding scale.

[0126] Step S424 involves performing feature fusion processing on the feature maps of multiple convolutional scale branches. This feature fusion process includes, for example, concatenating the feature maps of each convolutional scale branch along the channel dimension to obtain a fused multi-scale feature map. Furthermore, within each convolutional scale branch, after the convolution operation, normalization and activation function calculations are performed sequentially through the BR module to eliminate differences in feature distribution and ensure the consistency of the output features of different branches in the numerical space.

[0127] Step S425: Retain original features based on the residual branch. In neural network structures, gradient vanishing may occur as the number of network layers increases. Therefore, in this invention, each feature extraction network layer also includes a residual branch to retain the original features of the original image. For example, the residual branch performs convolution operations using a 1×1 convolution kernel, and then normalizes the results using a BN module.

[0128] Step S426: The multi-scale features fused from multiple convolutional scale branches and the original features from the residual branches are accumulated.

[0129] Step S427: Perform global average pooling on the accumulated data. By averaging the spatial dimensions, a global feature representation is obtained.

[0130] The currently obtained global feature representation is fed back to the input of the BRM module through the recycling path to form a loop recycling structure. From the processing steps, starting from step S422, a second feature extraction process is performed on the currently obtained global feature representation. This invention repeatedly executes the loop recycling structure a preset number of times n (e.g., n=9), achieving multiple iterations of feature optimization and layer-by-layer enhancement without significantly increasing network parameters. The last obtained global feature representation is output as the first information D1 to the attention calculation layer 43.

[0131] This invention improves the learning ability of features at different resolutions by using convolutional kernels of different scales. Let the input features of the original image be... Then, after convolution calculation and fusion using multi-size convolution kernels, the fused features are obtained. The process is represented by the following equation (1-1):

[0132] (1-1)

[0133] in, This represents a sequential splicing operation performed within a specified dimension. The calculation process representing the convolution kernel. , , These represent the convolution operations with a 1×1 kernel, a 3×3 kernel, and a 5×5 kernel, respectively, in the aforementioned embodiments. This represents the residual branch that ensures the stability of the network's gradient learning. These are the learnable parameters in the network. After multiple iterative operations, the features are fused. This process effectively condenses the fluorescence characteristics at different scales within each branch.

[0134] Referring to Figure 11, Figure 11 is a flowchart of an attention calculation method according to an embodiment of the present invention. The method includes the following steps:

[0135] Step S431: Determine the batch size based on the number of scale branches of the feature information in the first information. The batch size is calculated based on the following formula (1-2):

[0136] (1-2)

[0137] Where N is the total batch size. To measure the number of branches, To be responsible for the batch size under a single scale branch The interval distance between scale branches. The scale branch interval distance is The corresponding association weight coefficient is used to control whether the distance is within a certain range. Establish cross-scale correlation information between scale branches. Indicates in In each scale branch, the interval distance is exactly... The number of scale branch pairs.

[0138] Regarding the three convolution scale branches in the aforementioned embodiments, it can be seen that... ,set up That is, each scale branch is assigned two batches; let If we only focus on the association of continuous scale branches, then the batch size N = 3×2 + 0×(3-0) + 1×(3-1) + 0×(3-2) = 8.

[0139] Step S432: The first information is divided into multiple batch information. Specifically, in this embodiment, the first information D1 is averaged over a specified dimension and then converted into a fusion feature. Then divide it into Batch information , In particular, when segmenting the first piece of information, the segmentation order is the same as the splicing order of the scale features.

[0140] Step S433: Perform attention calculations based on multiple attention heads to obtain the corresponding calculation results. Referring to Figure 11, attention calculations are performed for the first to eighth attention levels, respectively. Attention is calculated for each batch of information based on the following formulas (1-3):

[0141] (1-3)

[0142] in, , represents the attention calculation result of the i-th self-attention head. To process batch information using a 1×1 convolution kernel After performing the calculation, we obtained: For the channel dimension within a single patch. Each Each has an independent self-attention head to perform the above calculations to obtain an independent calculation result SA.

[0143] Step S434: Concatenate the attention calculation results of all batch information. This is done using the following equation (1-4):

[0144] (1-4)

[0145] in, The attention calculation result represents the first convolutional scale branch (1×1 convolutional kernel). The results of the correlation attention calculation for the first and second convolutional scale branches are shown. The attention calculation result represents the second convolutional scale branch (3×3 convolutional kernel). The results of the correlation attention calculation represent the second and third convolutional scale branches. The attention calculation result represents the third convolution scale branch (5×5 convolution kernel). The representative will have all of Merge the attention statistics sequentially along the specified dimensions. .

[0146] This embodiment also includes an optional step S435, generating second information D2 for prediction. Specifically, the input first information D1 is compared with the attention statistics result. The two are combined to generate the second information D2 for prediction. The formula for calculating the second information D2 is shown in equation (1-5) below.

[0147] (1-5)

[0148] The prediction layer 44 of this invention includes two modules. One module performs vector output processing using the following formulas (1-6) to generate an embedded representation of the region of interest, namely, cell embedding. One module performs classification output processing using formula (1-7) to generate fluorescence pattern categories for the region of interest. ,in, .

[0149] (1-6)

[0150] (1-7)

[0151] in This indicates fully connected layer computation, primarily used for dimensional mapping and transformation.

[0152] ReLU and are activation functions, respectively. It is an improvement on the ReLU activation function, which retains more detailed information than ReLU, and is therefore used in obtaining high-dimensional embeddings.

[0153] In this embodiment, when calculating the fluorescence mode category At that time, two fully connected layer calculations were performed through two fully connected layers, thereby mitigating the problem of detail loss caused by large dimensional transformations.

[0154] Because the feature differences between typical and atypical fluorescence patterns are small, conventional neural network model learning methods have limited ability to distinguish fine-grained feature differences. Therefore, this invention also provides a contrastive learning method to train the neural network model, enabling it to learn the subtle feature differences between typical and atypical fluorescence patterns, thereby improving the ability to classify fluorescence patterns. The accuracy of the feature description is improved by using cell embedding, an embedded representation of the region of interest, to accurately describe the features of different patterns.

[0155] Referring to Figure 12, Figure 12 is a flowchart of a neural network model training method according to an embodiment of the present invention. The neural network model training method includes the following steps:

[0156] Step S401: Construct a learning sample set. Since the difficulty in interpreting ANCA-IIF images based on an ethanol matrix lies in the easy misinterpretation of fluorescence patterns, to address this problem and improve the accuracy of fluorescence pattern interpretation, this invention employs a contrastive learning training method when training the neural network model. The learning sample set includes three sample subsets: a training sample set, a positive sample set, and a negative sample set. Samples in the training sample set are called anchor samples, samples in the positive sample set are called positive samples, and samples in the negative sample set are called negative samples. Each anchor sample has corresponding positive and negative samples.

[0157] The strategy for identifying positive samples is to identify samples in the same category that are easily misclassified as other categories as positive samples of that category, with the aim of reducing the embedding distance between easily misclassified samples. The strategy for identifying negative samples is to identify samples in different categories that are easily misclassified as anchor sample categories as negative samples of each other, with the aim of increasing the distance between easily confused samples in different categories.

[0158] The samples in this invention comprise five categories: tC, aC, tP, aP, and NS. For each of these categories, positive and negative samples are constructed. For example, each cell region is extracted from the image samples used for cell annotation, forming the original sample set. Multiple ROI images labeled as tC are obtained from the original sample set. Based on experience, ROI images easily identified as belonging to other categories are selected as positive samples for the tC category. ROI images easily identified as belonging to the tC category are selected from the other four categories as negative samples for the tC category. Therefore, positive samples in a fluorescence pattern category come from samples easily identified as belonging to other categories within that category, and negative samples come from samples easily identified as belonging to the same category within other categories. This method results in three sample sets for each fluorescence pattern category: an anchor sample set, a positive sample set, and a negative sample set. Each anchor sample, along with one positive sample and one negative sample, forms a sample triplet.

[0159] In one embodiment, the process of constructing a learning sample set includes the following steps:

[0160] First, the anchor sample set is determined. Each cell region is extracted from the image samples used for cell annotation as the original sample, thus forming the original sample set. Based on the annotation information of each cell region, the original sample set is classified into five categories of anchor sample sets, corresponding to the aforementioned five categories. This invention uses all cell regions from all image samples, thus covering rich and diverse cell data, ensuring that comprehensive feature information can be learned in subsequent learning processes.

[0161] Then, the positive and negative sample sets are determined. To determine the corresponding positive and negative sample sets for each category of anchor sample set, each anchor sample of each category is first input into the neural network model trained without using a contrastive learning strategy to obtain the corresponding predicted category. Then, a statistical data matrix heatmap is constructed based on the actual labeled category and the predicted category, as shown in Figure 13. Figure 13 is a statistical data matrix heatmap applied when determining positive and negative samples according to an embodiment of the present invention. Different colors in the figure represent different sample numbers, as shown in the legend on the right side of the matrix diagram. The vertical axis of the matrix heatmap diagram represents the actual labeled category of the sample, and the horizontal axis represents the predicted category obtained by the neural network model trained conventionally (without using a contrastive learning strategy). As can be seen from the figure, the boxes in the diagonal represent samples whose predicted results are the same as the actual labels, that is, samples accurately predicted by the currently conventionally trained neural network model, and also samples that are easy to predict. When determining the positive and negative sample sets for a category, taking the tP category as an example, based on the principle that samples easily identified as belonging to other categories are considered positive samples, the samples in the first to fourth columns of the bottom row can all be considered positive samples. Similarly, based on the principle that samples easily identified as belonging to the tP category from the other four categories are considered negative samples, the samples in the second row from the bottom of the fifth column to the top row can all be considered negative samples. Furthermore, since different major categories, such as perinuclear and cytoplasmic categories, are relatively easy to distinguish, while different subcategories within the same major category, such as the typical and atypical subcategories within the perinuclear category, are difficult to distinguish, this embodiment primarily selects different subcategories within the same major category as positive and negative samples, respectively. For example, the samples in the red squares of the bottom row, second column (true label is category tP, predicted result is category aP) are taken as positive samples of category tP and negative samples of category aP, respectively. Similarly, the samples in the blue squares of the second row, fifth column (true label is category aP, predicted result is category tP) are taken as negative samples of category tP and positive samples of category aP, respectively. The positive and negative sample sets for other categories are obtained in the same way.

[0162] Step S402: Construct the contrastive learning loss function. Specifically, the present invention calculates the contrastive learning loss using the following formula (2-1). :

[0163] (2-1)

[0164] in, The cross-entropy loss is calculated based on the sample fluorescence pattern categories. Used to represent the sample distance between typical and atypical fluorescence patterns within the same class. and They are respectively and The weight parameters. Among them, Calculate using the following formula (2-2), The following formula (2-3) is used for calculation.

[0165] (2-2)

[0166] in, For the first one sample Category labels corresponding to each fluorescence pattern category For the neural network model to the first The prediction result obtained from predicting the nth sample belongs to the nth sample. The probability of each fluorescent pattern category.

[0167] (2-3)

[0168] in, This represents the process by which a neural network model generates sample feature vectors (i.e., cell embeddings). The dimension of the sample image. Represents the input number anchor samples, Represents the input number One positive sample, Represents the input number One negative sample, , It is the set of all triples in the learning sample set, with a cardinality of N, which is the total number of anchors. This is the edge constant, representing the boundary between positive and negative samples, used to prevent gradient vanishing. This is a distance metric function used to calculate the shape difference between two vectors X and Y. The calculation formula is shown in equation (2-4) below:

[0169] (2-4)

[0170] in, Let X represent the root mean square error, and Y represent two vectors. It is a constant used to prevent gradient vanishing during training. and The value is calculated using the center-normalized operation. The singular value decomposition (SVD) process decomposes a matrix into a left singular matrix, a right singular matrix, and a singular value matrix. The left and right singular matrices represent the optimal alignment angles between two vectors in different directions, while the singular value matrix represents the degree of matching in different alignment directions. In this embodiment, the SVD operation rotates vector Y to the direction of optimal alignment with X. After normalization and rotation alignment, the difference between the two vectors is calculated to obtain the final measure of the distance between the feature vectors of the two samples (cell regions).

[0171] in, Represents a vector The calculated value after normalization operation at the center is shown in equation (2-5) below.

[0172] (2-5)

[0173] Represents a vector The calculated value is obtained by normalizing the center, as shown in equation (2-6) below.

[0174] (2-6)

[0175] In the formula This represents the Frobenius norm, used to eliminate brightness differences among fluorescent cells in different samples; and They are vectors sum vector The average value.

[0176] Combining formulas (2-4) and (2-5) (2-6), it can be seen that the present invention first eliminates the brightness difference of cells through normalization, then aligns the cell morphology, and finally compares the differences in cell morphology by measuring vector distance.

[0177] As can be seen from the aforementioned process of constructing the loss function, through and The effective combination of these factors ensures that the neural network model can accurately classify samples while quantifying the similarity between them.

[0178] Step S403: Divide the learning samples into multiple batches. Each batch includes a certain number of anchor samples, positive samples, and negative samples.

[0179] Step S404: Initialize the trainable parameters in the neural network model. The initialization method can be random initialization or initialization based on a preset distribution.

[0180] Step S405, forward propagation computation. Specifically, a batch of training samples is input into the neural network model. The neural network model performs a forward propagation operation layer by layer to obtain the prediction output for each sample. The prediction output includes the cell embedding of the training sample and the probabilities of five fluorescence pattern categories. .

[0181] Step S406: Calculate the contrastive learning loss based on the loss function.

[0182] For each learning sample in the training batch, a triplet sample is constructed for each anchor sample, and the contrastive learning loss is calculated according to formulas 2-1 to 2-6. .

[0183] Step S407, backpropagation and parameter update. Based on the contrastive learning loss... The backpropagation algorithm is used to calculate the gradient information of each trainable parameter in the neural network model relative to the loss value; based on the gradient information, the trainable parameters are updated using a preset parameter update strategy to reduce the loss value.

[0184] Step S408: Determine whether a preset stopping condition is met. If met, the model training process ends. The stopping condition includes one of the following: loss value convergence, training iterations reaching a preset threshold, or model performance meeting expected requirements. If the preset stopping condition is not met, return to step S405 and continue iterative training.

[0185] After training is completed, the trained neural network model is output, which is used to perform inference or prediction tasks on the data to be processed.

[0186] This invention increases the distance between samples of different categories that are easily classified as belonging to the same category, and decreases the distance between samples of the same category that are easily classified as belonging to different categories. This enables the first model to learn easily confused fluorescence pattern features, thereby effectively improving the prediction accuracy of typical and atypical fluorescence patterns within the same fluorescence pattern.

[0187] After processing in step S4, multiple ROIs, their corresponding fluorescence pattern category probabilities, and embedded representations expressing fluorescence pattern information are obtained from an ethanol-based ANCA-IIF image. Since ethanol-based ANCA-IIF images typically contain both typical and atypical fluorescence patterns, this potential cell coexistence relationship can significantly impact the final classification results. Considering the powerful role of graph structures in relationship mining, this invention constructs node and edge features in step S5 to facilitate the determination of fluorescence pattern coexistence relationships and dominant fluorescence patterns in an image through graph structures.

[0188] Figure 14 is a flowchart of a method for constructing graph features according to an embodiment of the present invention. Specifically, it includes the following steps:

[0189] Step S51: Select the top m ROIs in the target image based on their confidence ranking as nodes in the graph structure.

[0190] Step S52: Construct node features. Specifically, fuse the second fluorescence pattern category data and the first fluorescence pattern category data for each ROI to construct node features. Since there are a total of 5 positive fluorescence pattern categories in this invention, the first category One-hot code of the first fluorescence pattern category data for each ROI is obtained based on the output of the target detection model. , Class confidence in the first fluorescence pattern category data , Based on the output of the first model, the second fluorescence mode category data for each ROI is obtained as the second category probability distribution. , The fluorescence pattern category data structure in this embodiment includes five data bits, each representing a fluorescence pattern category. For example, the five data bits, from front to back, represent tC, aC, tP, aP, and NS. The target detection model uses one-hot encoding as the category output label. When a specific fluorescence pattern category of an ROI is determined, the data bits corresponding to that fluorescence pattern category in the first fluorescence pattern category data structure are set to 1, and the remaining data bits are set to 0. For example, when the category output label of an ROI1 output by the target detection model is (1,0,0,0,0;0.87), the first five bits (1,0,0,0,0) are the first category one-hot encoding of the first fluorescence pattern category data, indicating that the fluorescence pattern category of ROI1 is tC, and the last bit is the confidence level, indicating that the confidence level of determining the fluorescence pattern category of the ROI as tC is 0.87. The second fluorescence pattern category data output by the first model consists of the probabilities of the five fluorescence pattern categories. For example, the second fluorescence mode category data of ROI1 output by the first model is (0.83, 0.08, 0.04, 0.05, 0), which means that the probability of the current output ROI being tC is 0.83, the probability of aC is 0.08, the probability of tP is 0.04, the probability of aP is 0.05, and the probability of NS is 0, respectively.

[0191] Then, the probability distributions of the second category are merged according to formula (3-1). Category 1 one-hot encoding and confidence level This allows us to obtain the node features of a node. .

[0192] (3-1)

[0193] For example, the node features of ROI1 mentioned above. As shown in equation (3-2):

[0194] (3-2)

[0195] Furthermore, in each node feature Add 1 complement bit , ,when =1 indicates that the total number of cells in the image is less than 1. ,when =0 indicates that the total number of cells in the image is not less than 0. One. That is, modify formula (3-1) to formula (3-3):

[0196] (3-3)

[0197] If there are at least m ROIs in the image containing ROI1, the node features of ROI1 are represented by the following formula (3-4):

[0198] (3-4)

[0199] If there are fewer than m ROIs in the image containing ROI1, for example, when m=50, and the node containing ROI1 contains 45 ROIs, then the node features starting from the 46th node are represented by the following formula (3-5):

[0200] (3-5)

[0201] -1 represents the confidence level.

[0202] Therefore, the node characteristics of each node The m node features of the target image constitute a node feature matrix. , .

[0203] Step S53: Calculate the adjacency matrix A of the target image based on the ROI-based embedded representation. Wherein, The row index order and column index order of the adjacency matrix A are the same, and the same row index and column index represent the same ROI in the target image. Matrix elements The distance between the embedded representation of the ROI in the i-th row and the embedded representation of the ROI in the j-th column is shown in formula (4-1).

[0204] (4-1)

[0205] and These represent the embedded representations of the nodes in the i-th and j-th rows, respectively.

[0206] Step S54: Construct a degree matrix D based on the adjacency matrix A. The degree matrix D reflects the distribution pattern between nodes. Each diagonal element in the degree matrix D... This represents the connection between the nodes in the corresponding row and other nodes. The specific calculation formula is shown in equation (4-2) below.

[0207] (4-2)

[0208] Where i and j are the row and column indices of the adjacency matrix A, respectively, and in the degree matrix D, the diagonal elements... The row index and column index are the same, i.e., i=j; Ⅱ(∙) is an indicator function, which is 1 when the condition is met and 0 otherwise; Let be the matrix element in the i-th row and t-th column of the adjacency matrix A, and τ be the node activation threshold. According to formula (4-2), when the cosine similarity between a node (cell) and another node exceeds τ, the node connection between them will be activated, and the accumulated connection information between the corresponding node and all other nodes will be obtained.

[0209] Step S55: Calculate the sum of the adjacency matrix A and the degree matrix D to obtain the edge feature matrix. Specifically, the edge feature matrix is ​​calculated based on the following formula (4-3). .

[0210] (4-3)

[0211] Since the ROI embedded representation cell embedding generated by the neural network model through contrastive learning characterizes the differences between different fluorescence patterns, especially the differences between typical and atypical fluorescence patterns, this invention calculates the values ​​of all ROIs in the target image. The cosine similarity is used as the edge feature of the node. Furthermore, when the diagonal of the adjacency matrix A calculated according to formula (4-1) is all 0, this introduces some invalid learning information. To obtain richer edge relationships, this invention utilizes the Laplacian matrix transformation principle to calculate the degree matrix D. The value of each element in the degree matrix D reflects whether the fluorescence pattern of the cell corresponding to the current node is the dominant pattern of the entire image. Therefore, the edge features in this invention can represent the correlation between cells. Based on the correlation between all cells in an image and combined with the fluorescence pattern category corresponding to the node, the coexistence relationship of different types of cells in the image can be depicted.

[0212] As seen in step S5, this invention constructs graph features composed of node features and edge features. , graph features It can be represented as .

[0213] Based on the graph features obtained in step S5 Various methods can be used to obtain the final specific fluorescence pattern category of the target image.

[0214] Example 1

[0215] In this embodiment, the final specific fluorescence pattern category of the target image is obtained based on a statistical method. This includes the following steps:

[0216] Get the maximum value of the diagonal elements in the degree matrix .

[0217] The row number that determines the maximum value of the diagonal elements. .

[0218] Row number based on the maximum value of the diagonal elements Determine the node features of the corresponding region of interest. .

[0219] Determine node features The model checks whether the determined category in the first fluorescence pattern category data output by the target detection model is the same as the category with the highest probability in the second fluorescence pattern category data output by the first model. If they are the same, the category determined in the first fluorescence pattern category data is directly output as the final fluorescence pattern category of the target image. If they are different, the category with the highest probability in the second fluorescence pattern category data is determined. Confidence level in the first fluorescence mode category data Based on the size relationship, the fluorescence pattern category with the larger value is taken as the final fluorescence pattern category of the target image.

[0220] For example, when determining the maximum value relative to the diagonal elements The node features of the region of interest corresponding to the row number for At that time, the probability of the first element corresponding to the tC category in the second fluorescence pattern category data is 0.83, and the probability of the first element corresponding to the tC category in the first fluorescence pattern category data is 0.87. Since the two determinations are the same, the tC category is taken as the final fluorescence pattern category of the target image. When the maximum value of the diagonal elements is determined... The node features of the region of interest corresponding to the row number for When the probability of the first position in the second fluorescence pattern category data corresponding to the tC category is 0.83, and the probability of the second position in the first fluorescence pattern category data corresponding to the aC category is 0.87, the two judgments are different. Therefore, the aC category corresponding to the second position with a confidence level of 0.87 is taken as the final fluorescence pattern category of the target image.

[0221] Since the graph features generated in step S5 are a rich representation, such as node features that can represent cell morphology, fluorescence intensity, and fluorescence category, and edge features that can represent the morphological similarity between cells, the present invention can also use machine learning models, deep learning models, or graph neural network models (hereinafter referred to as the second model) based on these graph features to obtain the final specific fluorescence pattern category of the target image.

[0222] Example 2

[0223] In this second embodiment, a machine learning model is used as the second model. The second model is a classifier composed of support vector machines, random forests, gradient boosting decision trees, etc. The machine learning model obtains the final specific fluorescence pattern category of the target image.

[0224] First, a learning dataset is constructed. Specifically, using the graph features of multiple images obtained by the aforementioned method, the graph features of each image are vectorized. For example, the node feature matrix in the original graph features is... Sum of edge feature matrices A fixed-length feature vector is formed by concatenating the features along a specified dimension. In this embodiment, the number of dimensions of the node features is 12.

[0225] In another embodiment, before feature concatenation, the edge features are first globally aggregated, for example, by aggregating the original edge feature matrix along the node dimension. Calculate one or more global statistics, such as the mean. Standard deviation The maximum, minimum, and median values ​​are then calculated, and these global statistics are added to the edge feature matrix. In this way, the original edge feature matrix is ​​obtained. Corrected to a new edge feature matrix The new edge feature matrix , This increases the number of global statistics. Then, the original node features and the new edge features are concatenated to obtain the feature vector. .

[0226] Then, a supervised learning dataset is constructed based on the corresponding image category labels. , }

[0227] Next, the classifier is trained and evaluated based on the learning dataset until it meets the requirements. Then, the optimal model weights obtained from the training are used in the actual prediction task. Those skilled in the art can refer to traditional machine learning model training methods to complete the training of the second model in this embodiment, which will not be described in detail here.

[0228] In application, the graph features of the target image are used to construct a feature vector. The data is then fed into a classifier, and the classifier outputs the final specific fluorescence pattern category of the target image.

[0229] Example 3

[0230] In this third embodiment, a deep learning model, such as a Transformer-based classification model, is used to obtain the final specific fluorescence pattern category of the target image.

[0231] First, a sample set is constructed. Specifically, the graph features obtained using the aforementioned method are used for unordered fusion of graph features. Specifically, to ensure compatibility with standard deep learning frameworks, the graph structure data is converted into a fixed-size tensor. In this embodiment, node features and defined edge features are fused and flattened into patches, forming a fixed-length feature vector F. As a classification feature, the node feature in this embodiment has 12 dimensions.

[0232] Then, a supervised learning dataset is constructed based on the corresponding image category labels. , }

[0233] Next, the classification model is trained and validated. The Transformer-based classification model includes an input embedding layer, a multi-head self-attention computation layer, and a classification layer.

[0234] During the forward propagation, the input embedding layer adds a learnable standard sinusoidal positional encoding to the feature vector F in the samples. Since the order of ROIs in the input feature vector F is random or artificially defined, this positional encoding introduces positional information that is independent of real biological space.

[0235] Next, in the multi-head self-attention computation layer, dense feature extraction is achieved through the computation of multiple attention heads, thereby learning the global cell relationships contained in the feature vector F.

[0236] The classification layer takes the vector output by the last self-attention head, performs global average pooling on it to obtain a global feature representation, and finally outputs the class probability through a multilayer perceptron classifier.

[0237] Then, the cross-entropy method is used to calculate the classification loss, and the loss value obtained from the classification head is used for back gradient propagation and parameter update. When the optimal loss is obtained, the weights are saved as the optimal model.

[0238] In application, the graph features of the target image are used to construct a feature vector. The data is then input into the classification model, and the category with the highest probability output by the classification model is determined as the final specific fluorescence pattern category of the target image.

[0239] Example 4

[0240] In this fourth embodiment, a graph neural network model is used to determine the final specific fluorescence pattern category of the target image.

[0241] In this embodiment, the structure of the graph neural network model is shown in Figure 15, which is a schematic diagram of the structure of a graph neural network model according to an embodiment of the present invention. The graph neural network model in this embodiment includes a two-level graph convolutional structure and a classification head.

[0242] When the map features of a target image When input is given to a graph neural network model, the computation process of the graph neural network model is as follows:

[0243] First, the first-level graph convolution calculation is performed according to the following formula (5-1) to obtain the first-level graph convolution calculation result. .

[0244] (5-1)

[0245] in, These are the learnable parameters in the network, where This indicates the number of channels in the hidden layer. , is the node feature matrix. , where is the edge feature matrix.

[0246] The second-level graph convolution calculation result is obtained by performing the second-level graph convolution calculation as shown in equation (5-2). .

[0247] (5-2)

[0248] in, The parameters are learnable in the network. The second-level graph convolutional layer uses the same propagation mechanism as the first-level graph convolutional layer, but no longer uses an activation function to preserve the linear separability of the original feature space.

[0249] Next, in the classification layer, the node features are aggregated into a graph-level representation using formula (5-3). .

[0250] (5-3)

[0251] These are the learnable parameters in the network. Indicates the index of the subgraph.

[0252] Finally, the graph-level representation is transformed using the Softmax operation shown in formula (5-4). The fluorescence pattern is converted into a probability distribution, and the category corresponding to the position index of the maximum probability value is determined as the final fluorescence pattern category.

[0253] (5-4)

[0254] The training process of the graph neural network model in this embodiment is as follows.

[0255] First, a training sample set is constructed. For example, an initial sample set is obtained, which is a set of ANCA-IIF images on an ethanol matrix. Then, the dominant fluorescence mode of each ANCA-IIF image is labeled and used as the category label for the ANCA-IIF image.

[0256] Each ANCA-IIF image is then input into the target detection model to obtain cell regions within the ANCA-IIF images, and to determine the fluorescence pattern category and confidence level of each cell region. The sample set includes multiple subsets of cell region images, each subset comprising a cell region image from one ANCA-IIF image. The fluorescence pattern category and confidence level of each cell region image serve as the initial label information. For example, the initial label information data structure is set to 6 fields, with the first 5 fields corresponding to the 5 fluorescence pattern categories. The field corresponding to the fluorescence pattern category of each cell region image is set to 1, the remaining fields are set to 0, and the last field is the confidence level. For example, the initial label information for the i-th cell region image... i =[1,0,0,0,0,0.87].

[0257] Then, the image of each cell region in the sample set is output to the neural network model to obtain the corresponding second fluorescence mode category data and embedded representation. The second fluorescence mode category data includes the probability of each preset fluorescence mode category. The data structure of the second fluorescence mode category data is set to 5 fields, which correspond to the 5 fluorescence mode categories respectively. The data of each field is the probability of the corresponding category. For example, a second fluorescence mode category data is represented as [0.83, 0.05, 0.07, 0.05, 0].

[0258] Then, the second fluorescence pattern category data of the same cell region image are merged into the initial annotation information label, and a complement bit is added to obtain the first annotation information label1. For example, the first annotation information label1 of the i-th cell region image. i =[0.83,0.05,0.07,0.05,0,1,0,0,0,0,0.87,0]. The embedded representation of this cell region image is used as the second annotation information label2.

[0259] After the aforementioned processing, the first annotation information label1 and the second annotation information label2 of each cell region image were obtained.

[0260] Based on the first annotation information (label1) and the second annotation information (label2) of each cell region image in a subset of cell regions from an ANCA-IIF image, a node feature matrix V and an edge feature matrix are constructed. Thus, a graphical feature sample of an ANCA-IIF image was obtained, and the graphical features are represented as follows: The category labels of the ANCA-IIF images are used as the category labels of the graph feature samples. Specifically, each cell region image is modeled as a node, and the first annotation information of the cell region image is used as the category label. As node features All cell region image node features Construct a node feature matrix V, and construct an edge feature matrix based on the second annotation information label2 of all cell region images according to the aforementioned formulas (4-1) to (4-3).

[0261] The graph neural network model is then trained and evaluated until it meets the requirements.

[0262] This embodiment abstracts cellular regions in an image as graph nodes, cell morphology and fluorescence characteristics as node features, and the similarity relationships between cells as edges. Through learning from sample data and using message passing and aggregation in a graph neural network, nodes with similar features (cell populations belonging to the same fluorescence pattern) reinforce each other. The model can automatically identify and amplify signals from the largest and most coherent cell populations based on graph topology (such as node degree and subgraph density) or attention weights, effectively suppressing interference from sparse and scattered abnormally stained cells. Through learning from samples, this embodiment can identify patterns represented by cell populations with more consistent morphology and more distinct features. For example, even with only a few typical cytoplasmic fluorescence patterns (tC) and a large number of atypical cytoplasmic fluorescence patterns (aC), this embodiment, through the transmission of edge relationships in the graph neural network model, can still infer that the former is more consistent and representative of the pattern, thus correctly outputting the typical cytoplasmic fluorescence pattern (tC) as the dominant category.

[0263] As described above, this invention significantly improves the automatic recognition and classification performance of complex fluorescence patterns in IIF images through a multi-level collaborative modeling mechanism. On one hand, by enhancing fine-grained feature representation and discriminative constraints, it effectively reduces the misclassification rate between typical and atypical fluorescence patterns, improves the ability to distinguish similar fluorescence and morphology, achieves fine classification of ANCA karyotypes in ethanol-based matrices, and also accurately distinguishes between typical and atypical karyotypes within cytoplasmic and perinuclear karyotypes. On the other hand, by introducing a graph structure, the complex relative relationships in multi-cell, multi-fluorescence patterns are transformed into a learnable structured representation, thereby significantly improving the classification accuracy of samples with coexisting multi-fluorescence patterns. Simultaneously, addressing the issue of large variations in fluorescence feature scale, this invention employs a multi-branch and adaptive weight adjustment mechanism to enhance the model's robustness and generalization ability to fluorescence features at different scales. Overall, this invention achieves higher classification accuracy, stronger adaptability, and more stable recognition results in complex fluorescence pattern recognition scenarios.

[0264] On the other hand, the present invention also provides an ANCA-IIF image processing system based on an ethanol matrix. Referring to Figure 16, Figure 16 is a schematic block diagram of an ANCA-IIF image processing system based on an ethanol matrix according to an embodiment of the present invention. The system includes a positive / negative detection module 1, a target detection module 2, a first classification module 3, an image feature construction module 4, and a second classification module 5. The positive / negative detection module 1 is used to perform positive / negative screening detection of antibody categories on the target image to determine whether the antibody category of the corresponding test sample is positive or negative. The target image is an indirect immunofluorescence image of an ANCA test sample based on an ethanol matrix, and the indirect immunofluorescence image includes multiple cellular regions. In one embodiment, the positive / negative detection module 1 inputs the target image into a trained classification model, which outputs a classification category for the target image. If the classification category output by the model is fluorescence mode, the antibody category of the sample is determined to be positive; if the classification category output by the model is non-fluorescence mode, the antibody category of the sample is determined to be negative. The training sample set of the classification model is a set of indirect immunofluorescence images of ANCA-detected samples based on an ethanol matrix. The training sample set is labeled with fluorescence mode and non-fluorescence mode, and a training sample is labeled with either fluorescence mode or non-fluorescence mode. When the positive / negative detection module 1 determines that the antibody category of the sample corresponding to the target image is positive, it sends a notification to the target detection module 2.

[0265] The target detection module 2 receives a notification from the positive / negative detection module 1. When the antibody category of the test sample corresponding to the target image is positive, it performs target detection on the target image to obtain the region of interest (ROI), the first fluorescence pattern category data of the ROI, and its confidence level. Each ROI corresponds to a cell region, and the first fluorescence pattern category data includes one of several preset fluorescence pattern categories. In one embodiment, the target detection module 2 inputs the target image to the trained target detection model; and the target detection model outputs the ROI detected from the target image, the first fluorescence pattern category data of the ROI, and its confidence level. The training sample for the target detection model is an indirect immunofluorescence image of an ANCA test sample based on an ethanol matrix, annotated with cell regions. The annotation information for each cell region includes one of several preset fluorescence pattern categories, including tC, aC, tP, aP, and NS. After processing, the target detection module 2 sends a notification to the first classification module 3.

[0266] After receiving the notification from the target detection module 2, the first classification module 3 inputs each region of interest into the trained first model to obtain the second fluorescence pattern category data and embedded representation of each region of interest. The second fluorescence pattern category data includes the probability of each preset fluorescence pattern category.

[0267] The first model is a neural network model. The process of generating the second fluorescence pattern category data and embedded representation for each region of interest is described in the aforementioned method section and will not be repeated here. After processing all regions of interest, the first classification module 3 sends a notification to the graph feature construction module 4.

[0268] The graph feature construction module 4, based on the received notification, uses each region of interest (ROI) in the target image as a node, and fuses the second fluorescence mode category data and the first fluorescence mode category data of each ROI to construct node features; it also constructs edge features between nodes based on the embedded representation of each ROI. For details, please refer to the method section; further elaboration is omitted here. After constructing the node and edge features of the target image, it is sent to the second classification module 5.

[0269] The second classification module 5 determines the final fluorescence mode category of the target image based on the node and edge features of the target image. The final fluorescence mode category of the target image is one of several preset fluorescence mode categories, including tC, aC, tP, aP, and NS. The second classification module 5 can determine the final fluorescence mode category of the target image using statistical methods, or it can use machine learning models, deep learning models, or graph neural network models. For details, please refer to the description in the aforementioned method section; further elaboration is omitted here.

[0270] Figure 17 is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. The electronic device can be implemented as a server or other various terminal devices, such as desktop personal computers, tablet computers, laptop computers, etc., including a processor 601 and a memory 602. The memory 602 stores a program instruction set. When the processor 601 executes the program instruction set in the memory 602, it implements any of the aforementioned ANCA-IIF image processing methods and systems based on ethanol matrix.

[0271] Specifically, the processor 601 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of the present invention.

[0272] Memory 602 may include mass storage for data or instructions. For example, and not limitingly, memory 602 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 602 may include removable or non-removable (or fixed) media. Where appropriate, memory 602 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 602 is non-volatile solid-state memory.

[0273] The memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, a memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the ethanol-based ANCA-IIF image processing method and system provided by this invention.

[0274] In one example, the electronic device may also include a communication interface 603 and a bus 610. The processor 601, memory 602, and communication interface 603 are connected via the bus 610 and communicate with each other.

[0275] The communication interface 603 is mainly used to realize communication between various modules, systems, units and / or devices in the embodiments of the present invention.

[0276] Bus 610 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 610 may include one or more buses. While specific buses are described and illustrated in embodiments of the invention, the invention contemplates any suitable bus or interconnect.

[0277] The present invention also provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement any of the ethanol-based ANCA-IIF image processing methods and systems described in the foregoing embodiments. The computer-readable storage medium can be any medium that can tangibly contain or store computer-executable instructions for use by or in connection with instruction execution systems, apparatuses, and devices. The storage medium can be a transient computer-readable storage medium or a non-transitory computer-readable storage medium. Non-transitory computer-readable storage media may include, but are not limited to, magnetic storage devices, optical storage devices, and / or semiconductor storage devices. Examples of such storage devices include, for example, magnetic disks, optical discs based on CD, DVD, or Blu-ray technology, and persistent solid-state storage such as flash memory and solid-state drives.

[0278] This invention also provides a computer program product comprising a set of computer program instructions, which, when executed by a processor, implement any of the ANCA-IIF image processing methods and systems based on an ethanol matrix as described in the foregoing embodiments. The computer program product includes, but is not limited to, application installation packages, application plugins, and mini-programs that can run within certain applications, all published on websites or in app stores.

[0279] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0280] The above embodiments are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the scope of the invention. Therefore, all equivalent technical solutions should also fall within the scope of the invention.

Claims

1. An ANCA-IIF image processing method based on an ethanol matrix, characterized in that, The method includes: performing antibody class positive / negative screening on a target image to determine whether the antibody class of the corresponding test sample is positive or negative, wherein the target image is an indirect immunofluorescence image of an ANCA test sample based on an ethanol matrix, and the indirect immunofluorescence image includes multiple cell regions; in response to the target image corresponding to a positive antibody class, performing target detection on the target image to obtain regions of interest (ROIs) and their first fluorescence pattern class data, each ROI corresponding to a cell region, the first fluorescence pattern class data including one of multiple preset fluorescence pattern classes and their confidence levels; inputting each ROI to a trained first model to obtain second fluorescence pattern class data and an embedded representation for each ROI, the second fluorescence pattern class data including the probabilities of multiple preset fluorescence pattern classes; using each ROI in the target image as a node, fusing the second fluorescence pattern class data and the first fluorescence pattern class data of each ROI to construct node features; constructing edge features between nodes based on the embedded representation of each ROI; and determining the final fluorescence pattern class of the target image based on the node features and edge features of the target image, the final fluorescence pattern class of the target image being one of multiple preset fluorescence pattern classes.

2. The ANCA-IIF image processing method based on an ethanol matrix according to claim 1, characterized in that, The step of performing antibody class positive / negative screening on a target image to determine whether the antibody class of the corresponding test sample is positive or negative includes: inputting the target image into a trained classification model, and having the classification model output a classification class for the target image; and determining that the antibody class of the test sample is positive if the classification class output by the classification model is fluorescent mode; and determining that the antibody class of the test sample is negative if the classification class output by the classification model is non-fluorescent mode; wherein, the training sample set of the classification model is an indirect immunofluorescence image set of ANCA test samples based on an ethanol matrix, and the labels of the training sample set are fluorescent mode and non-fluorescent mode, and the label of a training sample is fluorescent mode or non-fluorescent mode.

3. The ANCA-IIF image processing method based on an ethanol matrix according to claim 1, characterized in that, The steps of performing target detection on a target image to obtain the region of interest and its first fluorescence pattern category data in the target image include: inputting the target image into a trained target detection model; and outputting the region of interest and its first fluorescence pattern category data detected from the target image by the target detection model; wherein, the training sample of the target detection model is an indirect immunofluorescence image of an ANCA detection sample based on an ethanol matrix with cell regions annotated, and the annotation information of each cell region includes one of a variety of preset fluorescence pattern categories, including typical c-ANCA, atypical c-ANCA, typical p-ANCA, atypical p-ANCA and nonspecific ANCA.

4. The ANCA-IIF image processing method based on an ethanol matrix according to claim 1, characterized in that, The first model is a neural network model. Correspondingly, when inputting each region of interest into the trained first model to obtain the second fluorescence pattern category data and embedded representation of each region of interest, the first model processes the input region of interest by: extracting features from the input region of interest through multi-scale branches to obtain first information, wherein the first information includes feature information at multiple scales; performing dynamic weighted calculations on the feature information at each scale and the correlation information between scales in the first information through an attention mechanism to obtain second information; and performing prediction output operations on the second information to obtain the second fluorescence pattern category data and embedded representation of the input region of interest; wherein the sample images in the training sample set of the neural network model are cell images extracted from indirect immunofluorescence images of samples detected by ANCA based on ethanol matrix; and the label of each sample image is one of multiple preset fluorescence pattern categories.

5. The ANCA-IIF image processing method based on an ethanol matrix according to claim 4, characterized in that, When extracting features from the input region of interest through multi-scale branches to obtain the first information, a preset number of feature iterations are performed to extract the first information.

6. The ANCA-IIF image processing method based on an ethanol matrix according to claim 5, characterized in that, The steps for the first multi-scale feature extraction of the input region of interest include: performing initial feature extraction on the input region of interest to transform the region of interest information in the spatial domain to the region of interest information in the channel domain; using the transformed region of interest information in the channel domain as the target information; extracting features from the target information using convolutional kernels of multiple different scales to obtain feature information at multiple scales; concatenating the feature information at multiple scales to obtain feature concatenation information; and performing average pooling operation on the feature concatenation information to obtain average pooling information; wherein, the average pooling information is the target information for the next iteration.

7. The ANCA-IIF image processing method based on an ethanol matrix according to claim 6, characterized in that, Further includes: The original feature information of the target information is obtained by extracting features from the target information through residual branching; Correspondingly, the step of splicing feature information at multiple scales to obtain feature splicing information includes: splicing feature information at multiple scales to obtain intermediate splicing information; The feature splicing information is obtained by accumulating the intermediate splicing information and the original feature information extracted through residual branch.

8. The ANCA-IIF image processing method based on an ethanol matrix according to claim 6, characterized in that, Before extracting features from the target information using multiple convolutional kernels of different scales, the process also includes: performing normalization, activation function calculation, and max pooling operations on the target information in sequence.

9. The ANCA-IIF image processing method based on an ethanol matrix according to claim 6, characterized in that, After extracting features from the target information to obtain feature information at multiple scales, and before splicing the feature information at multiple scales, the method further includes: performing normalization operations and activation function calculations on each scale feature information in turn to obtain feature information of the same dimension; correspondingly, when splicing the feature information at multiple scales, feature information of the same dimension at multiple scales is spliced ​​to obtain feature splicing information.

10. The ANCA-IIF image processing method based on an ethanol matrix according to claim 4, characterized in that, The step of dynamically weighting the feature information at each scale and the correlation information between scales in the first information using an attention mechanism to obtain the second information includes: determining the batch size based on the number of scales of the feature information in the first information; wherein, the batch size is determined based on the following formula: Where N is the total batch size. To measure the number of branches, The number of batches required for a single-scale branch. The interval distance between scale branches. The scale branch interval distance is The corresponding correlation weight coefficient, Indicates in In each scale branch, the interval distance is The number of scale branch pairs; dividing the first information into batches of a certain number; performing attention calculations on each batch of information to obtain the corresponding calculation results; and concatenating the attention calculation results of all batches of information to obtain the second information.

11. The ANCA-IIF image processing method based on an ethanol matrix according to claim 10, characterized in that, The step of concatenating the attention calculation results of all batch information to obtain the second information includes: concatenating the attention calculation results of all batch information to obtain attention calculation result concatenation information; and accumulating the attention calculation result concatenation information and the first information to obtain the second information.

12. The ANCA-IIF image processing method based on an ethanol matrix according to claim 1, characterized in that, The steps for constructing edge features between nodes based on the embedded representation of each region of interest include: calculating the embedded representation distance between every pair of nodes in the total number of nodes, and generating an adjacency matrix A of the target image based on the embedded representation distance between every pair of nodes; constructing a degree matrix D based on the adjacency matrix A, where each diagonal element of the degree matrix D represents the connection between a node in the corresponding row and the remaining nodes in the same row; and calculating the sum of the adjacency matrix A and the degree matrix D to obtain the edge feature matrix. 。 13. The ANCA-IIF image processing method based on an ethanol matrix according to claim 12, characterized in that, The adjacency matrix element in the i-th row and j-th column of adjacency matrix A Calculated using the following formula: ,in, and The embedded representations of the nodes in the i-th and j-th rows are respectively; the diagonal elements in the degree matrix D Calculated using the following formula: ， The distance threshold for the embedded representation of a node; t is the indicator function; t is the column number.

14. The ANCA-IIF image processing method based on an ethanol matrix according to claim 1, characterized in that, The steps for determining the final fluorescence mode category of the target image based on the node features and edge features of the target image include: constructing second model input data based on the node features and edge features of the target image; and inputting the second model input data into the trained second model to obtain the final fluorescence mode category of the target image.

15. The ANCA-IIF image processing method based on an ethanol matrix according to claim 14, characterized in that, The second model is a machine learning model or a deep learning model. Correspondingly, the step of constructing the input data of the second model based on the node features and edge features of the target image includes: aggregating the node features and edge features of the target image to obtain a feature vector of fixed length; correspondingly, the machine learning model or deep learning model processes the feature vector as input to obtain the final fluorescence mode category of the target image.

16. The ANCA-IIF image processing method based on an ethanol matrix according to claim 14, characterized in that, The second model is a graph neural network model; correspondingly, the step of constructing the input data of the second model based on the node features and edge features of the target image includes: constructing a node feature matrix V based on the number of nodes and node features of the target image; and constructing an edge feature matrix based on the number of nodes and edge features of the nodes in the target image. ; and the node feature matrix V and edge feature matrix As input to the graph neural network model; correspondingly, the graph neural network model uses the node feature matrix V and the edge feature matrix V as inputs. The final fluorescence pattern category of the target image is obtained through processing.

17. An ANCA-IIF image processing system based on an ethanol matrix, characterized in that, include: The positive and negative detection module is configured to perform positive and negative screening detection on the target image to determine whether the antibody category of the test sample corresponding to the target image is positive or negative. The target image is an indirect immunofluorescence image of an ANCA test sample based on an ethanol matrix, and the indirect immunofluorescence image includes multiple cell regions. The target detection module is configured to perform target detection on the target image in response to the antibody category of the detection sample corresponding to the target image being positive, so as to obtain the region of interest in the target image and its first fluorescence mode category data. Each region of interest corresponds to a cell region. The first fluorescence mode category data includes one of a variety of preset fluorescence mode categories and its confidence level. The first classification module is configured to input each region of interest into the trained first model to obtain the second fluorescence pattern category data and embedded representation of each region of interest. The second fluorescence pattern category data includes the probabilities of multiple preset fluorescence pattern categories. The graph feature construction module is configured to use each region of interest in the target image as a node, and fuse the second fluorescence mode category data and the first fluorescence mode category data of each region of interest to construct node features; Construct edge features between nodes based on the embedded representation of each region of interest; The second classification module is configured to determine the final fluorescence mode category of the target image based on the node features and edge features of the target image. The final fluorescence mode category of the target image is one of a variety of preset fluorescence mode categories.

18. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a set of computer program instructions, which, when executed by the processor, execute the ethanol-based ANCA-IIF image processing method according to any one of claims 1-16 or implement the ethanol-based ANCA-IIF image processing system according to claim 17.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a set of computer program instructions, which, when executed by a processor, perform the ethanol-based ANCA-IIF image processing method according to any one of claims 1-16 or implement the ethanol-based ANCA-IIF image processing system according to claim 17.

20. A computer program product, characterized in that, It includes a computer program instruction set, which, when executed by a processor, performs the ethanol-based ANCA-IIF image processing method according to any one of claims 1-16 or implements the ethanol-based ANCA-IIF image processing system according to claim 17.

Images