Gastric biopsy pathological risk degree hierarchical identification processing method and system

Abstract

The embodiment of the invention discloses a gastric biopsy pathological risk degree hierarchical identification processing method and system, and relates to the technical field of medical artificial intelligence, and the method comprises the steps: obtaining a digital pathological image of a to-be-diagnosed gastric biopsy pathological section; analyzing the digital pathological image based on an artificial intelligence classification model to obtain a risk level corresponding to the digital pathological image; distributing the digital pathological image to a preset diagnosis path matched with the risk level; wherein the preset diagnosis paths corresponding to different risk levels are different in detail degrees of auxiliary diagnosis information provided by the system or triggered automatic auditing processes; receiving a diagnosis result of manual examination and verification for the digital pathological image, and using the diagnosis result as feedback data for updating the artificial intelligence classification model; through automatic risk layering and differentiated path distribution, diagnosis resources can be optimized, diagnosis efficiency and accuracy can be improved, and continuous evolution of the model can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical artificial intelligence, in particular to a gastric biopsy pathological risk degree stratification identification processing method and system. BACKGROUND

[0002] Digestive system diseases, especially gastric cancer, are one of the malignant tumors with high incidence and high mortality worldwide, and its early diagnosis is crucial to improve the survival rate of patients.

[0003] Pathological diagnosis is the gold standard for gastric biopsy diagnosis, however, in clinical practice, pathological diagnosis faces severe challenges. On the one hand, there is a shortage of pathologists and uneven distribution, resulting in differences in diagnostic ability and diagnostic standards between different medical institutions, and the mutual recognition of pathological reports is not high. On the other hand, pathologists need to handle a large number of gastric biopsy specimens on a daily basis, most of which are common benign lesions, which consumes a large amount of expert resources, and some rare, diagnosis boundary ambiguous difficult cases may be missed or misdiagnosed due to lack of experience or fatigue of doctors.

[0004] In order to assist doctors and improve diagnostic efficiency, there are currently schemes that use artificial intelligence to analyze digital pathology images. For example, some schemes use deep learning models to automatically screen pathological images, and classify cases into "positive" or "negative" two categories to prompt doctors to pay attention to possible lesions in the slices. However, this simple binary screening method has obvious limitations. First of all, it fails to distinguish the diagnostic difficulty of the case, and cannot optimize the allocation of diagnostic resources, that is, it cannot effectively divert benign cases with clear diagnosis, malignant cases with clear diagnosis, and difficult cases with difficult diagnosis, and the performance of the model is limited. Therefore, the current method cannot meet the urgent needs of the clinic to establish an efficient, accurate and standardized pathological diagnosis process. SUMMARY

[0005] Therefore, it is necessary to provide a gastric biopsy pathological risk degree stratification identification processing method and system to solve the technical problems of large workload of gastric biopsy pathological diagnosis, non-uniform standards, uneven allocation of diagnostic resources, and difficulty of artificial intelligence model to be effectively applied and continuously optimized in real clinical environment, and to realize automatic risk stratification of pathological cases, optimize the diagnosis workflow and continuously improve itself.

[0006] In a first aspect, a gastric biopsy pathological risk degree stratification identification processing method is provided, the method comprising:

[0007] obtaining a digital pathology image of a gastric biopsy pathological section to be diagnosed;

[0008] The digital pathological images are analyzed based on an artificial intelligence classification model to obtain the risk level corresponding to the digital pathological images;

[0009] Based on the risk level, the digital pathology image is assigned to a preset diagnostic path that matches the risk level; wherein, the preset diagnostic paths corresponding to different risk levels differ in the level of detail of the auxiliary diagnostic information provided by the system or the triggered automated review process;

[0010] The system receives diagnostic results that have been manually reviewed and confirmed for the digital pathology images, and uses these diagnostic results as feedback data to update the artificial intelligence classification model.

[0011] Secondly, a gastric biopsy pathological risk stratification identification and processing system is provided, including:

[0012] The image acquisition module is used to acquire digital pathological images of gastric biopsy slides to be diagnosed.

[0013] The risk level analysis module is used to analyze the digital pathology image based on an artificial intelligence classification model to obtain the risk level corresponding to the digital pathology image.

[0014] The diagnostic path allocation module is used to allocate the digital pathology image to a preset diagnostic path that matches the risk level, based on the risk level; wherein, the preset diagnostic paths corresponding to different risk levels differ in the level of detail of the auxiliary diagnostic information provided by the system or the automated review process triggered.

[0015] The feedback update module is used to receive the diagnostic results confirmed by manual review of the digital pathology image, and use the diagnostic results as feedback data to update the artificial intelligence classification model.

[0016] The main objective of this application is to provide a method and system for risk stratification and identification of gastric biopsy pathology. This method involves acquiring digital pathological images of gastric biopsy slides to be diagnosed; analyzing the digital pathological images based on an artificial intelligence classification model to obtain the risk level corresponding to the digital pathological images; assigning the digital pathological images to preset diagnostic paths matching the risk level; wherein the preset diagnostic paths corresponding to different risk levels differ in the level of detail of the auxiliary diagnostic information provided by the system or the triggered automated review process; receiving the diagnostic results confirmed by manual review of the digital pathological images, and using these diagnostic results as feedback data to update the artificial intelligence classification model. This method, through automated risk stratification and differentiated path allocation, can more accurately analyze gastric biopsy pathological images, optimize diagnostic resources, improve diagnostic efficiency and accuracy, and simultaneously enable continuous model evolution, establishing an efficient, accurate, and standardized pathological diagnosis process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the background art, the accompanying drawings used in the embodiments of this application or the background art will be described below.

[0018] Figure 1 A flowchart illustrating a method for stratified identification of pathological risk in gastric biopsy, provided in an embodiment of this application;

[0019] Figure 2 A schematic diagram of a hierarchical reporting system for the GBPRGSS model provided in this application embodiment;

[0020] Figure 3 This application provides a schematic diagram of the overall model prediction system process.

[0021] Figure 4 This is a schematic diagram of a gastric biopsy pathological risk level stratification identification and processing system provided in an embodiment of this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0023] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0024] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0025] The embodiments of this application are described below with reference to the accompanying drawings.

[0026] Please see Figure 1 , Figure 1 This is a flowchart illustrating a method for stratified identification of pathological risk in gastric biopsy provided in an embodiment of this application. The method may include:

[0027] 101. Obtain digital pathological images of the gastric biopsy pathological sections to be diagnosed.

[0028] The method described in this application can be applied to a gastric biopsy pathological risk stratification identification and processing system. This system can be deployed on an electronic device, such as a server in a medical institution. The server may include, but is not limited to, one or more hardware devices such as processors, memory, and network interfaces. The memory stores computer program instructions, which, when executed by the processor, implement the method described in this application.

[0029] The digital pathological images of the aforementioned gastric biopsy pathological sections can be digital images obtained by scanning the gastric biopsy pathological sections. Specifically, after performing gastroscopy and pathological biopsy on the subject, the extracted gastric mucosal tissue samples can be used to prepare the aforementioned gastric biopsy pathological sections.

[0030] In a specific application scenario, step 101 begins with a routine gastric biopsy case. After receiving the gastric mucosal biopsy specimen from the clinical department, the pathology department performs a series of standardized pathological procedures, including fixation, dehydration, clearing, paraffin infiltration, embedding, sectioning, and hematoxylin-eosin staining, ultimately producing one or more glass slides containing tissue sections. Subsequently, a digital pathology slide scanning operation is performed. This operation uses a high-resolution whole-slide scanner (e.g., a digital slide scanner with 20x or 40x objectives) to perform a full-field, high-magnification digital scan of the glass slide. The scanning process converts the morphological information of the entire glass slide into a high-resolution digital image file, i.e., a digital pathology image. This file typically uses a specific whole-slide scan image format, such as SVS or TIFF, and its data size can reach several gigabytes. The generated digital pathology image is assigned a unique case number and associated with the case's clinical information (such as patient name, age, and clinical presentation) and pathological information (such as sampling site and specimen quantity), and stored together in the hospital's image archive and communication system or a dedicated pathology information system server. This acquisition step 101 provides standardized, high-quality raw data input for all subsequent automated analyses.

[0031] 102. Analyze the above digital pathological images based on an artificial intelligence classification model to obtain the risk level corresponding to the above digital pathological images.

[0032] After acquiring digital pathology images, the system automatically triggers a risk stratification process. At its core is a pre-trained artificial intelligence classification model, which in this embodiment can be referred to as the GBPRGSS model. This GBPRGSS model can be a complex deep learning model; for example, its underlying feature extraction network can employ a convolutional neural network architecture, combined with an attention mechanism or transformer architecture, to enhance the ability to capture global and local features of the image.

[0033] Specifically, the system's processor (e.g., a server equipped with a high-performance graphics processing unit) loads the GBPRGSS model and takes the acquired digital pathology images as input. The model performs depth analysis on the images, a process that typically includes:

[0034] High-resolution digital pathology images are segmented into thousands or even tens of thousands of small image patches;

[0035] Feature extraction and classification prediction are performed on each image patch;

[0036] By aggregating the prediction results of all image patches and combining them with global contextual information, a comprehensive judgment is made on the entire case (i.e. the entire digital pathology image).

[0037] The key to this application's embodiments is that the output of the artificial intelligence classification model is not a simple "positive" or "negative," but a risk level preset with multiple levels (at least three levels), which represents the degree of lesion or the degree of tumor risk.

[0038] In one implementation, these three levels may include a low-risk level, a high-risk level, and an extremely high-risk level. For example, the classification is as follows:

[0039] Category A - Low Risk: This category corresponds to the most common, clearly diagnosed benign lesions in clinical practice. Examples include chronic non-atrophic gastritis and gastric fundic gland polyps. These cases have typical histomorphological features, low diagnostic controversy, and typically account for 80% to 90% of routine gastric biopsies.

[0040] Category B - High Risk: This category corresponds to a confirmed diagnosis of malignancy or high-grade neoplasia. Examples include gastric adenocarcinoma, signet ring cell carcinoma, and high-grade intraepithelial neoplasia. These cases require close attention from pathologists and directly impact the patient's subsequent treatment plan. These cases typically account for 5% to 10% of all cases.

[0041] Category C - Very High Risk: This category corresponds to complex cases with ambiguous diagnostic boundaries, prone to misdiagnosis, rare, or cases that the model itself cannot accurately determine. Examples include differentiating atypical hyperplasia, low-grade intraepithelial neoplasia from reactive hyperplasia, certain rare tumor types, or cases where poor image quality leads to low confidence in model analysis. Although these cases account for a small percentage (approximately 1% to 5%), they represent a significant challenge and pain point in clinical diagnosis.

[0042] 103. Based on the aforementioned risk level, the aforementioned digital pathology images are assigned to a preset diagnostic path that matches the risk level; wherein, the preset diagnostic paths corresponding to different risk levels differ in the level of detail of the auxiliary diagnostic information provided by the system or the automated review process triggered.

[0043] After obtaining the risk level, the system can perform automated allocation, pushing the case to different preset diagnostic pathways. The core difference between the diagnostic pathways corresponding to different risk levels lies in the level of detail of the auxiliary diagnostic information provided by the system or the automated review process triggered. This differentiated processing flow is key to achieving optimal allocation of diagnostic resources.

[0044] Figure 2 This is a schematic diagram of a hierarchical reporting system for the GBPRGSS model provided in an embodiment of this application. Figure 2 As shown, for routine gastric biopsy cases, gastric biopsy pathological slides are scanned to obtain corresponding digital pathological images, which are then input into the GBPRGSS model for processing to determine the risk level. Subsequently, the specific allocation logic based on the risk level can be as follows:

[0045] For cases classified as Category A - Low Risk, including Category A1 - Inflammation and Category A2 - Polyps, the system assigns them to a "routine diagnostic pathway." Under this pathway, cases are pushed to a work queue associated with junior or intermediate-level physician roles. The system can provide basic auxiliary diagnostic information, such as the refined classification results detailed in later embodiments. The review process under this pathway is relatively lenient; for example, the system may only be configured with a low percentage (e.g., 1%-5%) of random sampling quality control, meaning only a small number of cases are automatically pushed to more senior physicians for review.

[0046] For cases classified as Category B - High Risk, including categories B1-B4, the system assigns them to a "High-Risk Diagnostic Pathway." Under this path, the case is pushed to a work queue associated with a senior or intermediate-level physician. The system provides more detailed auxiliary diagnostic information, such as the location of lesion hotspots, detailed in later embodiments. More importantly, the system automatically triggers a mandatory, 100% automated review process. Specifically, after the first senior physician completes the initial draft of the diagnostic report, the system automatically transfers the case and report to the work queue of another physician of equal or higher qualifications (e.g., another senior or intermediate-level physician or a senior subspecialist) and marks it as "Pending Review." Only after review and confirmation by the second physician does the system allow the final report to be issued.

[0047] For cases classified as Category C - Very High Risk, or those not in Category A / B, the system will assign them to an "Very High Risk Diagnostic Pathway." Under this path, the case is directly pushed to the work queue associated with the most experienced senior subspecialist. The system will display the case with the highest priority and may provide all available auxiliary analytical tools. Simultaneously, the system will automatically trigger a mandatory multi-specialist consultation process. For example, the system will automatically distribute the case to multiple pre-defined subspecialist experts and initiate an online consultation interface for all experts to review the images, discuss, and reach a consensus.

[0048] In this way, the system transforms human management processes (such as who to assign tasks to and who will review them) into the system's own rule-based automated technical behaviors (such as what information to provide and what processes to trigger), thereby achieving an intelligent and automated reshaping of the diagnostic workflow.

[0049] In an optional implementation, when the risk level is low, the method further includes:

[0050] Using the aforementioned artificial intelligence classification model or secondary classification model, multi-level label classification is performed on the aforementioned digital pathological images to output semi-quantitative grading results of specific pathological types or pathological features, and a preliminary draft of a structured report is generated accordingly.

[0051] Once the model classifies a case as Category A - Low Risk, the system does not stop analyzing. Instead, it continues to call one or more secondary classification models to perform deeper, multi-level label classification of the case. This method specifically includes the following steps:

[0052] First, the system performs preliminary subtyping of Category A - low-risk diseases. For example, a two-level classification model will determine whether the case belongs to the A1 - gastritis or the A2 - polyp category. This is a basic binary classification task that helps to further triage low-risk cases into different processing logics.

[0053] Secondly, the system performs more refined feature analysis and classification for different subtypes.

[0054] If the case is classified as A1-gastritis, the system invokes a specialized multidimensional pathological feature analysis model. This model aims to simulate international standards (such as the updated Sydney system) followed by pathologists in diagnosing gastritis, providing a semi-quantitative grading of several key pathological features. In this embodiment, these features include at least:

[0055] Chronic inflammation: The model outputs a grading result, such as "0" (none), "1+" (mild), "2+" (moderate), and "3+" (severe), by identifying and counting the density of lymphocytes and plasma cells in the interstitial tissue.

[0056] Active (acute inflammation): The model outputs corresponding grading results by identifying and locating the area and extent of neutrophil infiltration.

[0057] Intestinal metaplasia: The model determines the presence or absence of intestinal metaplasia by identifying the presence of goblet cells in the mucosal epithelium and assigning a grade based on the extent of metaplasia.

[0058] Atrophy: The model determines the presence of atrophy by assessing the reduction or disappearance of intrinsic glands and then classifies it.

[0059] Atypical hyperplasia (intraepithelial neoplasia): The model assesses the degree of atypia of glandular epithelial cells, determines the presence of dysplasia, and distinguishes its grade (e.g., none, low grade, high grade).

[0060] Helicobacter pylori: The model determines whether a sample is positive or negative by searching for typical Helicobacter pylori morphology in the image.

[0061] If a case is classified as an A2-polyp, the system will invoke another specialized polyp subtype classification model. The goal of this model is to identify the specific polyp type; for example, it can distinguish between fundic gland polyps, inflammatory gastritis polyps, hyperplastic polyps, and adenomatous polyps.

[0062] Finally, based on all the refined classification and grading results mentioned above, the system automatically generates a structured draft pathology report. This is a crucial technical step. The system retrieves a pre-defined structured report template for gastritis or gastric polyps from the database, which contains all the pathological features that need to be reported. Then, the system automatically populates the corresponding fields in the template with the results output from the secondary model (such as "Chronic Inflammation: 2+", "Intestinal Metaplasia: 1+", "Polyp Type: Fundic Gland Polyp").

[0063] When a junior or intermediate-level physician responsible for diagnosing this Category A case opens the report interface, they no longer see a blank report, but a draft pre-filled with key indicators by artificial intelligence. The physician's work has shifted from tedious manual observation, judgment, and data entry to efficient review and confirmation. They only need to quickly browse the digital pathology images, focusing on verifying the system's judgments, especially those items where the system indicates low confidence. After confirming everything is correct, the physician can confirm and issue the report with a single click. This process significantly reduces the time required to process a large number of common benign cases. Furthermore, because all grading and classification are based on the same model standards, it also significantly improves the standardization and consistency of diagnostic reports within the department and even across different institutions.

[0064] In an optional implementation, when the risk level is high, the method further includes:

[0065] Using the aforementioned artificial intelligence classification model or secondary model, the hot spots of lesions are marked on the aforementioned digital pathological images in the form of heat maps.

[0066] Specifically, for Category B high-risk cases, a solution is provided to accurately locate lesions and assist in generating standardized reports. Its main purpose is to help doctors quickly discover small or atypical cancer lesions, reduce the risk of missed diagnosis, and ensure the completeness and standardization of tumor reports.

[0067] Once the model classifies a case as Category B - High Risk, the system triggers a series of in-depth auxiliary diagnostic procedures for high-risk cases. These may include the following steps:

[0068] First, the system invokes a secondary model specifically for tumors. This model can be a built-in module of the GBPRGSS model or a standalone, pluggable external algorithm sub-model, such as a model specifically for gastric cancer subtype classification and localization (which could be called GBPRGSS-v2). This secondary model performs a more in-depth analysis of the digital pathology images, and its tasks are twofold:

[0069] Predicting specific tumor subtypes. For example, the model might output the probability that the case is most likely a specific type such as "gastric adenocarcinoma," "signet ring cell carcinoma," or "neuroendocrine tumor."

[0070] A heatmap is generated on the digital pathology image to mark the hot spots of the lesion.

[0071] The generation of the heatmap is one of the core technologies of this embodiment. Specifically, the secondary model uses a sliding window to segment the entire digital pathology image into numerous overlapping or non-overlapping image patches. For each patch, the model calculates the probability of it containing cancer cells or high-grade lesions and outputs a probability value (e.g., from 0 to 1). Then, the system maps these probability values ​​to a color gradient; for example, areas with a probability close to 0 are displayed as blue (cool tones), areas with a probability close to 1 are displayed as red (warm tones), and intermediate probabilities are represented by transitional colors such as yellow and orange. Finally, the system overlays this semi-transparent probability map composed of colors onto the original digital pathology image, forming an intuitive heatmap. The "hottest" (e.g., reddest) areas in the image are the lesion hotspots identified by the model.

[0072] When further identifying cases using the GBPRGSS-v2 model, the focus is on predicting the probability of four types of cases: gastric adenocarcinoma, signet ring cell carcinoma, neuroendocrine tumor, and gastrointestinal stromal tumor. For example, B1 - gastric adenocarcinoma, B2 - signet ring cell carcinoma, B3 - neuroendocrine tumor, and B4 - gastrointestinal stromal tumor. Type B cases account for 5%-10% of routine gastric biopsies and differ significantly from Type A gastritis and polyps in histological and cellular morphology. Type B cases account for >90% of neoplastic lesions in the stomach. Compared to Type A, in a gastric biopsy pathological image, for example, if it is divided into 100 1024×1024 pixel image blocks, and any one or more image blocks contain a clearly defined Type B tumor area, the entire slice is classified as Type B. For Type B, the highest predicted value is taken as the final result. For Type A, the average predicted value is taken as the final result. B1-B4 show significant inter-group differences in pathological morphology and require separate prediction.

[0073] For example, in a gastric biopsy pathology report with pathology number XXX, the GBPRGSS model classifies it as B. Then, using the GBPRGSS-v2 model, the predicted values ​​are 0.31 for B1, 0.91 for B2, 0.11 for B3, and 0.01 for B4. The final prediction for this case is most likely B2-signet ring cell carcinoma. The predicted image patches are then overlaid on the original pathology image as a heatmap, allowing pathologists to quickly identify the carcinoma and avoid missing high-risk B-class cases.

[0074] When a senior physician responsible for diagnosing this Category B case opens the case file, the system loads not only the digital pathology image but also this heatmap. The physician can immediately see the highlighted hotspots of the lesion on the image. This is extremely helpful in detecting tiny lesions or finding atypical nests of cancer cells against a background of widespread inflammation. Instead of spending considerable time performing a thorough search of the entire slide, the physician can directly zoom in and focus on the area indicated by the heatmap for careful observation and confirmation, thus greatly improving diagnostic efficiency and accuracy.

[0075] Secondly, the system automatically provides structured reporting assistance based on the tumor subtype predicted by the secondary model. Once the model predicts the tumor subtype (e.g., a 95% probability of predicting "gastric adenocarcinoma"), the system retrieves a standardized report template corresponding to that tumor type from its built-in knowledge base. This template is designed in accordance with international or domestic tumor reporting guidelines (such as the CAP guidelines of the American College of Pathologists) and includes all necessary reporting items.

[0076] Furthermore, the system not only provides templates but also automatically fills in suggested auxiliary examinations in the report based on the tumor subtype. For example, if the model predicts gastric adenocarcinoma, the system may automatically select or fill in tests related to prognosis and targeted therapy, such as "HER2," "MLH1," "PMS2," "MSH2," and "MSH6," in the "Suggested Immunohistochemistry" section of the report. If it predicts signet ring cell carcinoma, it may suggest testing for "E-cadherin."

[0077] After confirming the diagnosis of the hotspot area, the doctor switches to the report interface, where a structured report has been prepared by the system, along with professional immunohistochemical testing suggestions. The doctor can adopt, modify, or supplement these suggestions based on their own judgment, and then complete the entire report. This feature ensures that critical prognostic and treatment guidance information is not missed in tumor case reports, improving the overall quality and clinical value of tumor pathology reports.

[0078] In one optional implementation, step 102 includes:

[0079] Cases that are predicted as low-risk or high-risk by the above artificial intelligence classification model with a confidence level below the first threshold, and / or cases that are predicted as extremely high-risk by the above artificial intelligence classification model with a confidence level above the second threshold, are classified as extremely high-risk.

[0080] Cases classified as Category C account for 1%-3% of routine gastric biopsies. These cases often involve rare diseases, poor diagnostic consistency, high risk of misdiagnosis, and complex or similar disease types. Currently, computer algorithm models for common Category A and B diseases face challenges in terms of difficulty and efficiency. Using a bidirectional prediction method, the GBPRGSS model classifies cases with a predicted probability of Category A or B <0.05 and / or a predicted probability of Category C >0.80 as Category C.

[0081] Further optionally, the method also includes:

[0082] The cases with the extremely high risk level mentioned above will be provided to subspecialist physicians for consultation and reporting, and the target risk level of the disease category will be determined.

[0083] The above-mentioned determination of the target risk level for this disease category specifically includes:

[0084] The target risk level for this disease category is determined to remain at the extremely high risk level mentioned above, or the target risk level for this disease category is modified to the low risk level or the high risk level mentioned above.

[0085] For cases classified as extremely high-risk (Category C), consultations and reports can be made by senior physicians or subspecialists. After the final report is issued, the subspecialist classifies the cases. Data still classified as Category C is later accumulated into a training dataset with Category C labels. Alternatively, the labels can be modified to Category A or B. Through repeated accumulation of typical Category C cases and correction of misclassified cases, the final prediction results of the GBPRGSS model will become more accurate and reliable.

[0086] 104. Receive the diagnostic results confirmed by manual review of the above-mentioned digital pathology images, and use the diagnostic results as feedback data to update the above-mentioned artificial intelligence classification model.

[0087] This step constitutes the core of the closed-loop optimization of the proposed solution. Please refer to [link / reference]. Figure 2 Once a case has undergone any of the aforementioned diagnostic pathways and has been ultimately reviewed, confirmed, and issued a diagnostic report by a pathologist with the appropriate authority, this final diagnostic result (e.g., "chronic atrophic gastritis with intestinal metaplasia," "gastric signet ring cell carcinoma," etc.) is considered the target label for the case. The system can record this final diagnostic result and bind it to the unique identifier of the corresponding digital pathology image.

[0088] In one alternative implementation, the aforementioned artificial intelligence classification model uses a graph convolutional network to model the co-occurrence or evolutionary relationships among pathological labels;

[0089] The above-mentioned method for updating the artificial intelligence classification model involves using the diagnostic results and corresponding digital pathological images as new training data to incrementally train the artificial intelligence classification model.

[0090] Subsequently, the system can also periodically (e.g., daily, weekly, or monthly) receive these newly generated, high-quality labeled data pairs, and use these data pairs (i.e., digital pathology images and their corresponding authoritative diagnostic results) as new training data to update the GBPRGSS model.

[0091] In this embodiment, the model can be updated using incremental training. Specifically, the system will not completely retrain the model from scratch with new data, as this is both time-consuming and could lead to "catastrophic forgetting" of old knowledge. Instead, the system uses the new data to perform small-batch, short-cycle fine-tuning based on the weights of the existing GBPRGSS model. During incremental training, a relatively small learning rate is typically used to ensure that the model can retain and consolidate its existing diagnostic capabilities while learning new knowledge.

[0092] Through this complete closed loop of "analysis-assignment-diagnosis-feedback-update," the GBPRGSS model can continuously learn from real-world clinical practice, and its performance will dynamically evolve and continuously improve as new case data accumulates. For example, the model may initially be insensitive to a certain rare lesion, but when such cases diagnosed by experts are fed back to the model for learning, the model's recognition accuracy when encountering similar lesions in the future will significantly improve.

[0093] Furthermore, the system architecture of this embodiment has good scalability. Figure 3 This is a schematic diagram of an overall model prediction system provided in an embodiment of this application. Figure 3 As shown, the system includes a serial or parallel overall model prediction system for the GBPRGSS model and other algorithm models, and describes the process of dynamically supplementing the training set and algorithm iteration. Specifically, Figure 3 The flow chart from left to right is as follows:

[0094] Routine gastric biopsy cases: We first receive routine gastric biopsy case samples.

[0095] Digital pathological slide scanning: These case samples are digitally scanned and converted into digital images.

[0096] GBPRGSS Model: Scanned digital pathology images are input into the GBPRGSS model for preliminary risk stratification identification.

[0097] Self-developed tumor classification and region recognition model: The GBPRGSS model feeds images into the self-developed tumor classification and region recognition model for analysis.

[0098] External tumor classification and region identification model: At the same time, the GBPRGSS model can also send images to external tumor classification and region identification models for analysis.

[0099] Risk assessment: Based on the prediction results of self-developed and external models, the risk level of the tumor is assessed.

[0100] If the predicted probability p > 0.95, it is judged to be a malignant tumor and classified as Category B - High Risk;

[0101] If the predicted probability p is between 0.05 and 0.95, it is classified as Category C - Extremely High Risk;

[0102] If the predicted probability p < 0.05, the probability of it being a malignant tumor is low, and it is classified as Category A - low risk.

[0103] Algorithm model results as a reference: The prediction results of the above model are used as a reference to assist pathologists in making diagnoses.

[0104] The pathologist issues the report: The pathology report is ultimately issued by the pathologist based on the model results and their own professional judgment.

[0105] Dynamically supplementing the training set: The final results signed by pathologists are dynamically added to the training set of the GBPRGSS model as new labels for model iteration and optimization. Through continuous iterative updates, the performance of the GBPRGSS model is optimized, and prediction accuracy is improved.

[0106] The entire process also demonstrates the collaborative working mechanism of the GBPRGSS model with other algorithm models, as well as its adaptive learning ability to continuously optimize the model by dynamically updating the training set.

[0107] The core GBPRGSS model in this embodiment can be configured as a basic framework capable of integrating one or more external algorithm sub-models in a serial or parallel manner. For example, it can integrate an external model specifically for Helicobacter pylori detection, or a classification model developed by other research institutions for a specific tumor subtype. The GBPRGSS model can act as the overall scheduler, fusing the analysis results of its own and external sub-models before making the final risk level determination, which further enhances the functionality and adaptability of the entire system.

[0108] Based on the foregoing description, a technical solution for structurally optimizing the aforementioned artificial intelligence classification model (i.e., the GBPRGSS model) is provided below. This solution introduces a graph convolutional network into the model architecture to explicitly learn and utilize the intrinsic correlations between different pathological labels, thereby improving the diagnostic accuracy and logical consistency of the model when handling complex cases. The model architecture described in this embodiment can be applied to any artificial intelligence model in the foregoing embodiments.

[0109] Traditional classification models typically input features extracted from images directly into a fully connected layer or a softmax layer to independently predict the probability of each pathological label. This approach ignores the strong correlations between pathological labels. For example, in gastric pathology, "atrophy" and "intestinal metaplasia" often coexist, and "high-grade intraepithelial neoplasia" is an evolutionary result of "low-grade intraepithelial neoplasia." This embodiment aims to incorporate such prior pathological knowledge into the model design.

[0110] The specific implementation of this solution includes the following aspects:

[0111] Overall Model Architecture Design: The core backbone of the AI ​​classification model can be a high-efficiency convolutional neural network, such as EfficientNet-B4, used to extract high-dimensional visual feature vectors from the input digital pathology image patches. The innovation of this embodiment lies in the design of the classification head. A Graph Convolutional Network (GCN) module is introduced between the feature vectors extracted by the backbone network and the final classification output layer.

[0112] Construction of the label dependency graph: Before model training, a label dependency graph needs to be constructed. This graph can be constructed based on two types of information:

[0113] a) Prior pathological knowledge: The relationships between labels are defined by pathology experts. For example, experts can define a strong co-occurrence relationship between "atrophy" and "intestinal metaplasia," and an evolutionary relationship between "intestinal metaplasia" and "intraepithelial neoplasia."

[0114] b) Large-scale data statistics: By statistically analyzing massive amounts of pathological data with existing diagnostic results, the co-occurrence probability or conditional probability between different pathological labels is calculated. In this graph, each node represents a pathological label (e.g., "chronic inflammation," "atrophy," "intestinal metaplasia," "low-grade intraepithelial neoplasia," "high-grade intraepithelial neoplasia," "cancer," etc.). If two labels are considered pathologically related, or their co-occurrence probability exceeds a preset threshold, an edge is established between their corresponding nodes. The weight of the edge can be used to represent the strength of this association. For example, the edge weight between "atrophy" and "intestinal metaplasia" can be set relatively high.

[0115] The training and inference process of a model may include the following steps:

[0116] 1. During training, for each input image patch, the backbone network first extracts a visual feature vector. Simultaneously, the initial state of each label node can be represented by a learnable word embedding vector. Then, the visual feature vector is used to update the state of each node in the graph, ensuring that the node state simultaneously incorporates visual information and the semantic information of the label itself.

[0117] 2. The graph convolutional network module begins operation. It propagates and aggregates information multiple times along the edges of the pre-constructed label dependency graph. In each round of propagation, each node receives information from its neighbors and updates its state by combining this information with its own current information. For example, the "gut-like" node receives information from its neighbor, the "shrinking" node. After several rounds of propagation, each node's state vector incorporates structural information from itself, its neighbors, and even the entire graph.

[0118] 3. The state vectors of each node after being updated by the graph convolutional network are fed into their respective classifiers (such as a simple linear layer) to predict whether the image patch contains the pathological features represented by the node.

[0119] In this way, the model is no longer making decisions in isolation. For example, when the model detects a very clear visual feature of "atrophy" in an image, even if the visual feature of "intestinal metaplasia" is less typical, the graph convolutional network will correspondingly increase the prediction confidence of the "intestinal metaplasia" label because the "atrophy" node will propagate a strong "belief" to the "intestinal metaplasia" node through the edges of the graph. This makes the model's final output (e.g., predicting the presence of both "atrophy" and "intestinal metaplasia") more consistent with the inherent logic of pathology.

[0120] This graph convolutional network-based model architecture significantly enhances the system's ability to handle complex cases with multiple coexisting pathological features. It reduces logically contradictory diagnostic outputs (e.g., directly diagnosing high-grade lesions in the absence of atrophy and intestinal metaplasia), making the AI ​​model's predictions more robust, interpretable, and reliable, thus providing clinicians with higher-quality auxiliary diagnostic opinions.

[0121] The training and verification of the GBPRGSS algorithm model in the embodiments of this application are further explained below.

[0122] First, in the data processing stage, a large number of gastric biopsy digital slide samples were labeled and balanced: labels that appeared frequently, had similar histological morphology or clinical prognosis were semantically merged to form an 81-dimensional multi-label vector. Then, through color normalization, 512×512 pixel uniform scaling, Otsu automatic tissue region extraction, and multiple enhancements such as flipping, rotation, Mixup, CutMix, and staining perturbation, cell-tissue details were preserved while significantly reducing domain drift caused by staining differences.

[0123] For example, the initially set label frequency distribution can be:

[0124] High-frequency tags (0001-0020): those that appear more than 1000 times (e.g., "gastric fundic gland polyps", "moderate chronic inflammation");

[0125] Mid-frequency tags (0021-0080): Occurring 300-1000 times (e.g., "mild active inflammation", "mild atrophy");

[0126] Low-frequency tags (0081-0380): Occurrence frequency <300 times (e.g., rare atypical hyperplasia, special infection characteristics).

[0127] Most of the high-frequency tags are classified as Category A – Low Risk; most of the medium-frequency tags are classified as Category A – Low Risk and Category B – High Risk, with only a few classified as Category C – Extremely High Risk. Most of the low-frequency tags are classified as Category C – Extremely High Risk.

[0128] The label categories are imbalanced; high-frequency labels (such as "moderate chronic inflammation") may account for more than 50% of the total labels, while low-frequency labels (such as "intestinal metaplasia with dysplasia") may account for only 0.1%. Direct training will cause the model to favor high-frequency labels and ignore rare pathological features. This problem can be solved by the label cleaning and category balancing described above.

[0129] The model backbone uses EfficientNet-B4 pre-trained on ImageNet, with CBAM attention embedded between its convolutional blocks: channel attention is generated by parallel feeding of global average pooling and max pooling into a shared MLP, while spatial attention is generated by 7×7 convolutions after channel compression. The two are cascaded sequentially, enabling the network to simultaneously focus on "what features" and "where features are located." The multi-scale feature pyramid (FPN) integrates shallow texture and deep semantics, and 81 independent sigmoid outputs enable multi-label prediction. Label co-occurrence relationships are modeled by a graph convolutional network (GCN) to alleviate the problem of sparse positive samples for rare labels.

[0130] The training employs a three-stage progressive fine-tuning approach: for example, first, the backbone is frozen, and only the classification head is trained (10-15 epochs); then, the last 2-3 convolutional blocks are unfrozen (15-20 epochs); finally, global low-learning-rate fine-tuning is performed (10-15 epochs). A cosine annealing scheduler, combined with an early stopping mechanism (if the F1 score does not improve after 5 consecutive epochs of validation), avoids overfitting. Dropout, Mixup, and label smoothing further enhance robustness. The validation phase includes five-fold cross-validation and independent testing with multiple cases from external hospitals. Multi-dimensional metrics such as macro / micro F1, AUC-ROC, and PPV / NPV are comprehensively analyzed, and a confusion matrix is ​​plotted to analyze misdiagnosis patterns. Daily diagnostic feedback data is continuously collected, and the training set is incrementally updated daily, weekly, and monthly to achieve online model iteration and closed-loop performance improvement.

[0131] Optionally, the above multi-dimensional evaluation indicators include:

[0132] Overall performance: Macro F1 score (balancing the contributions of each category); Micro F1 score (focusing on overall prediction accuracy). AUC-ROC curve (evaluating the model's discriminative ability).

[0133] Category-specific metrics: Calculate precision, recall, and F1 score for each of the 81 labels. Plot a confusion matrix and analyze misdiagnosis patterns.

[0134] Clinically applicable indicators: PPV (Positive Predictive Value): The probability that a predicted pathological type is actually that type. NPV (Negative Predictive Value): The probability that a predicted non-pathological type is actually not that type.

[0135] Given the current state of healthcare, gastrointestinal diseases account for a large proportion of cases. This presents challenges for clinical pathology diagnosis due to the heavy workload, inconsistent standards, and differing formats, making inter-institutional recognition of results difficult. Consequently, the processing and analysis of gastric biopsy pathology images still face several issues:

[0136] The shortage of pathologists makes it impossible to meet the short reporting time for digestive pathology reports, resulting in lengthy reporting times and high costs. Current AI algorithm models are generally designed for predicting specific diseases and cannot be readily applied in clinical practice. The lack of standardized histopathological specimen classification means that subspecialists primarily handle common cases, consuming significant time. Meanwhile, junior and mid-level non-subspecialist physicians are prone to diagnostic errors in rare and complex cases, leading to decreased accuracy and consistency in pathological diagnoses. The limited resources of subspecialists, coupled with a lack of updated subspecialty pathology knowledge in primary care hospitals and non-specialist general hospitals, have resulted in significant diagnostic biases between hospitals.

[0137] The innovative aspects of the gastric biopsy pathological risk stratification identification method in this application include, but are not limited to:

[0138] 1. Based on real-world clinical settings, this system uses gastric biopsy histopathology as a starting point, covering a variety of diseases. It categorizes gastric biopsy digital pathology into three risk levels based on a comprehensive assessment of the algorithm model's prediction accuracy and the risk value diagnosed by pathologists.

[0139] 2. The computer model rapidly classifies diseases into low- and high-risk categories. The algorithm then further predicts the results and outputs a structured final pathology diagnosis report, which is subsequently signed and confirmed by a pathologist – a complete reporting and diagnostic system. This system can quickly issue pathology diagnoses for over 97% of common diseases.

[0140] 3. A dual-safety mechanism, combining algorithmic models and manual screening by doctors, avoids inconsistencies in diagnostic criteria for rare, unusual, complex, and error-prone diseases. This process filters out cases with extremely high risk values ​​from a massive database.

[0141] 4. Transform pathologists from the traditional microscope-based slide reading mode to an online computer screen-based mode. This enables rapid triage and referral to other subspecialists or simultaneous consultations and discussions by multiple doctors. It eliminates geographical limitations associated with the office microscope, allowing specialist consultations and discussions to be conducted while traveling or at home. This improves the timely consultation rate for complex pathologies and significantly reduces patient waiting times for difficult cases.

[0142] 5. Currently, most AI algorithm models for histopathology, including those for gastrointestinal diseases, are only suitable for predicting results for a single disease and a single scenario. This solution, however, can integrate multi-layered AI algorithm models sequentially or in parallel into its overall architecture. This allows AI algorithm models that are currently unsuitable for practical application to be transferred to real-world pathological diagnostic environments, improving diagnostic efficiency and accuracy, identifying rare and atypical cases, and avoiding missed or misdiagnosed cases of related diseases.

[0143] 6. After this case is applied to clinical trials, the latest test set results are continuously received and fed back. Finally, the pathologist's diagnosis results are used as the new round of training set data to dynamically supplement the upgrade and iteration of the model algorithm.

[0144] Based on the description of the foregoing method embodiments, this application also discloses a gastric biopsy pathological risk level hierarchical identification and processing system.

[0145] Figure 4 This is a schematic diagram of a gastric biopsy pathological risk stratification identification and processing system provided in an embodiment of this application. Figure 4 As shown, the system 400 includes:

[0146] Image acquisition module 410 is used to acquire digital pathological images of gastric biopsy pathological sections to be diagnosed;

[0147] The risk level analysis module 420 is used to analyze the above-mentioned digital pathological images based on an artificial intelligence classification model to obtain the risk level corresponding to the above-mentioned digital pathological images.

[0148] The diagnostic path allocation module 430 is used to allocate the digital pathology images to a preset diagnostic path that matches the risk level, based on the risk level. The preset diagnostic paths corresponding to different risk levels differ in the level of detail of the auxiliary diagnostic information provided by the system or the automated review process triggered.

[0149] The feedback update module 440 is used to receive the diagnostic results confirmed by manual review of the above-mentioned digital pathology images, and use the diagnostic results as feedback data to update the above-mentioned artificial intelligence classification model.

[0150] According to the specific implementation of the embodiments of this application, Figure 1 The relevant steps in the illustrated embodiment can be executed by various modules in the above-described device, and will not be described in detail here.

[0151] The gastric biopsy pathology risk stratification identification and processing system 400 in this embodiment can acquire digital pathological images of gastric biopsy pathological slides to be diagnosed; analyze the digital pathological images based on an artificial intelligence classification model to obtain the risk level corresponding to the digital pathological images; allocate the digital pathological images to preset diagnostic paths matching the risk level according to the risk level; wherein the preset diagnostic paths corresponding to different risk levels differ in the level of detail of the auxiliary diagnostic information provided by the system or the automated review process triggered; receive the diagnostic results confirmed by manual review of the digital pathological images, and use the diagnostic results as feedback data to update the artificial intelligence classification model; this method, through automated risk stratification and differentiated path allocation, can more accurately realize gastric biopsy pathological image analysis, optimize diagnostic resources, improve diagnostic efficiency and accuracy, and at the same time enable continuous evolution of the model, establishing an efficient, accurate and standardized pathological diagnosis process.

[0152] This application also discloses an electronic device. The electronic device includes a processor and a memory. The electronic device may further include a bus, and the processor and memory can be interconnected via the bus. The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The electronic device may also include input / output devices, which may include a display screen, such as a liquid crystal display (LCD). The memory is used to store one or more programs containing instructions; the processor is used to call the instructions stored in the memory to execute, such as... Figure 1 Some or all of the steps in the illustrated embodiments.

[0153] It should be understood that, in the embodiments of this application, the processor may be a Central Processing Unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0154] Input devices may include touchpads, microphones, etc., while output devices may include displays (LCDs, etc.), speakers, etc. The memory may include read-only memory and random access memory (RAM), and provides instructions and data to the processor. A portion of the memory may also include non-volatile RAM. For example, the memory may also store device type information.

[0155] The electronic device described in this application embodiment can acquire digital pathological images of gastric biopsy slides to be diagnosed; analyze the digital pathological images based on an artificial intelligence classification model to obtain the risk level corresponding to the digital pathological images; allocate the digital pathological images to preset diagnostic paths matching the risk level according to the risk level; wherein the preset diagnostic paths corresponding to different risk levels differ in the level of detail of the auxiliary diagnostic information provided by the system or the triggered automated review process; receive the diagnostic results confirmed by manual review of the digital pathological images, and use the diagnostic results as feedback data to update the artificial intelligence classification model; this method, through automated risk stratification and differentiated path allocation, can more accurately realize the analysis of gastric biopsy pathological images, optimize diagnostic resources, improve diagnostic efficiency and accuracy, and at the same time enable continuous evolution of the model, establishing an efficient, accurate and standardized pathological diagnosis process.

[0156] This application also provides a computer storage medium storing a computer program for electronic data exchange, which causes a computer to perform some or all of the steps of any of the gastric biopsy pathological risk stratification identification processing methods described in the above method embodiments.

[0157] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0158] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0159] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the division of modules is merely a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. The coupling, direct coupling, or communication connection shown or discussed between each other may be indirect coupling or communication connection through some interfaces, apparatuses, or modules, and may be electrical, mechanical, or other forms.

[0160] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0161] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media can be read-only memory (ROM), random access memory (RAM), or magnetic media, such as floppy disks, hard disks, magnetic tapes, magnetic disks, or optical media, such as digital versatile discs (DVDs), or semiconductor media, such as solid state disks (SSDs).

Claims

1. A method for stratified identification and processing of pathological risk in gastric biopsy, characterized in that, The method includes: Obtain digital pathological images of gastric biopsy slides to be diagnosed; The digital pathological images are analyzed based on an artificial intelligence classification model to obtain the risk level corresponding to the digital pathological images; Based on the risk level, the digital pathology image is assigned to a preset diagnostic path that matches the risk level; wherein, the preset diagnostic paths corresponding to different risk levels differ in the level of detail of the auxiliary diagnostic information provided by the system or the triggered automated review process; The system receives diagnostic results that have been manually reviewed and confirmed for the digital pathology images, and uses these diagnostic results as feedback data to update the artificial intelligence classification model.

2. The method according to claim 1, characterized in that, When the risk level is low, the method further includes: The digital pathological images are classified using the artificial intelligence classification model or the secondary classification model to output semi-quantitative grading results of specific pathological types or characteristics, and a draft structured report is generated accordingly.

3. The method according to claim 1, characterized in that, When the risk level is high risk, the method further includes: The artificial intelligence classification model or secondary model is used to mark the hot spots of lesions on the digital pathological image in the form of a heat map.

4. The method according to claim 1, characterized in that, The analysis of the digital pathology image based on the artificial intelligence classification model to obtain the risk level corresponding to the digital pathology image includes: Cases predicted by the AI ​​classification model as low risk or high risk with a confidence level below the first threshold, and / or cases predicted by the AI ​​classification model as extremely high risk with a confidence level above the second threshold, are classified as extremely high risk.

5. The method according to claim 4, characterized in that, The method further includes: Cases classified as extremely high risk will be provided to subspecialist physicians for consultation and reporting, and the target risk level for that disease category will be determined. Determining the target risk level for this disease category specifically includes: The target risk level for the disease category is determined to remain at the extremely high risk level, or the target risk level for the disease category is modified to the low risk level or the high risk level.

6. The method according to claim 5, characterized in that, The artificial intelligence classification model uses graph convolutional networks to model the co-occurrence or evolution relationships between pathological labels; The method for updating the artificial intelligence classification model is to use the diagnostic results and the corresponding digital pathological images as new training data to perform incremental training on the artificial intelligence classification model.

7. The method according to claim 2, characterized in that, The secondary labeling of the low-risk cases includes gastritis and polyps; among which: The gastritis is diagnosed semi-quantitatively based on six dimensions, including: acute inflammation, chronic inflammation, intestinal metaplasia, atrophy, dysplasia, and Helicobacter pylori. The polyps are classified into fundic gland polyps, gastritis polyps, gastric hyperplasia polyps, and other types.

8. The method according to claim 7, characterized in that, The step of analyzing the digital pathological image based on an artificial intelligence classification model to obtain the risk level corresponding to the digital pathological image also includes: Compared to the low-risk level, if any one or more image blocks in the digital pathology image contain a region of the high-risk level, the entire gastric biopsy pathology slide is classified as the high-risk level. For the high-risk level, the highest predicted value is taken as the final result; For the low-risk level, the average predicted value is taken as the final result.

9. The method for stratified identification and processing of pathological risk in gastric biopsy according to claim 1, characterized in that, The architecture of the artificial intelligence classification model includes: A pre-trained EfficientNet-B4 was used as the backbone network to extract features from digital pathology images. The CBAM attention mechanism was employed to enhance the model's focus on key pathological features; The feature pyramid network is used to fuse shallow detail features with deep semantic features; A multi-label classification head is adopted, including a global average pooling layer and a fully connected layer. Each node independently predicts a pathological label, and the Sigmoid activation function is used to support multiple labels being positive at the same time. Graph Convolutional Networks (GCNs) are introduced to capture the co-occurrence relationships between labels, construct a label dependency graph, and learn the conditional probability distribution between labels; The training method for the artificial intelligence classification model includes: Transfer learning is employed, and the backbone network is initialized using ImageNet pre-trained weights, while convolutional layer parameters are retained and classification head weights are randomly initialized. Implement a three-stage fine-tuning strategy: Freeze the backbone network and train only the classification head; Unfreeze the final preset number of convolutional blocks and perform fine-tuning. Global fine-tuning involves unfreezing the entire network and adjusting it using a preset learning rate. A cosine annealing scheduler is used to periodically adjust the learning rate to avoid getting trapped in local optima; training stops if the score does not improve after 5 consecutive rounds of validation. Regularization techniques are applied, including Dropout, Mixup training, and label smoothing.

10. A gastric biopsy pathological risk stratification identification and processing system, characterized in that, The gastric biopsy pathological risk stratification identification and processing system includes: The image acquisition module is used to acquire digital pathological images of gastric biopsy slides to be diagnosed. The risk level analysis module is used to analyze the digital pathology image based on an artificial intelligence classification model to obtain the risk level corresponding to the digital pathology image. The diagnostic path allocation module is used to allocate the digital pathology image to a preset diagnostic path that matches the risk level, based on the risk level; wherein, the preset diagnostic paths corresponding to different risk levels differ in the level of detail of the auxiliary diagnostic information provided by the system or the automated review process triggered. The feedback update module is used to receive the diagnostic results confirmed by manual review of the digital pathology image, and use the diagnostic results as feedback data to update the artificial intelligence classification model.

Images