Method, device and storage medium for recognizing pathological section pathological numbers

Abstract

The application discloses a pathological section pathological number recognition method, device and storage medium, relates to the technical field of section management, and comprises the following steps: in response to a pathological number recognition instruction, traversing a candidate area of suspected pathological number text in a pathological section image, determining a preliminary recognition result of the candidate area; based on a recognition evaluation rule, judging the fitness degree of the preliminary recognition result and the evaluation index, and calculating the confidence score of the preliminary recognition result according to the fitness degree; according to the confidence score of the preliminary recognition result, matching the execution strategy corresponding to the confidence score from the strategy chain library; executing the corresponding execution strategy on the preliminary recognition result to determine the target recognition result of the pathological number in the pathological section image. The application solves the problem of poor recognition accuracy of the pathological number by the pathological section candidate area recognition, confidence score calculation, strategy chain library matching and corresponding strategy execution scheme, and improves the recognition accuracy and automatic processing efficiency.

Description

Technical Field

[0001] This application relates to the field of slide management technology, and in particular to a method, device and storage medium for identifying pathology slide pathology numbers. Background Technology

[0002] In digital pathology intelligent diagnosis and treatment scenarios, the reliability and process continuity of optical character recognition of pathology numbers directly affect the level of digitization of pathology information and the efficiency of business flow.

[0003] In the process of automatic identification of pathology numbers on pathological slides, the relevant technologies generally use a general model to perform single text recognition. This method is prone to misidentifying pathological tissue texture as text when the background of the pathological slide is complex and the text form is varied, which leads to poor accuracy in pathology number recognition.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a method, device and storage medium for identifying pathology numbers on pathology slides, aiming to solve the technical problem of poor accuracy in identifying pathology numbers.

[0006] To achieve the above objectives, this application proposes a method for identifying pathology slide pathology numbers, the method comprising:

[0007] In response to the pathology number recognition command, the candidate regions of suspected pathology number text in the pathology slide image are traversed to determine the preliminary recognition results of the candidate regions;

[0008] Based on the identification and evaluation rules, the degree of fit between the preliminary identification results and the evaluation indicators is determined, and the confidence score of the preliminary identification results is calculated based on the degree of fit.

[0009] Based on the confidence score of the preliminary identification results, the execution strategy corresponding to the confidence score is matched from the strategy chain library;

[0010] The corresponding execution strategy is executed on the preliminary identification results to determine the target identification result of the pathology number in the pathological slide image.

[0011] In one embodiment, in response to the pathology number identification instruction, the pathology slide image is preprocessed to obtain a standardized pathology slide image;

[0012] By traversing the pathological slide images through the initial screening operation, the candidate regions of all suspected pathology number texts are located, and invalid background regions without text features are identified to obtain candidate text regions.

[0013] Local contrast enhancement and texture denoising are performed on the candidate text regions respectively to remove interference features that are not text background within the candidate text regions, thus obtaining text-standardized regions.

[0014] By performing high-precision text recognition, the text content of each standardized text region is recognized one by one, generating preliminary recognition results corresponding to the candidate regions.

[0015] In one embodiment, based on the preliminary recognition results corresponding to the candidate region, the original image of the candidate region, and the recognition process data, the recognition output probability, the pathology number rule compliance degree, and the text region image quality are extracted.

[0016] The recognition output probability, the pathology number rule compliance, and the text region image quality are matched with the evaluation indicators to determine the degree of fit for each indicator.

[0017] Based on the evaluation weights corresponding to the evaluation indicators, the fit of all individual indicators is weighted and fused to obtain the confidence score corresponding to the preliminary identification result.

[0018] In one embodiment, the preliminary identification results are divided into confidence levels according to multiple confidence level thresholds to determine abnormal areas with confidence levels lower than the qualified level.

[0019] Based on the confidence level label corresponding to the abnormal region, match the abnormal feature verification rule specific to the confidence level label;

[0020] Based on the aforementioned anomaly feature verification rules, the compliance of the pathology number rules and the validity of the text region in the anomaly region are specifically verified, and the anomaly feature distribution results of the candidate region are generated.

[0021] Based on the abnormal feature distribution results of the candidate region, the problematic link that causes the abnormality in the pathology number identification process is located, and the abnormality judgment result of the abnormal region is generated.

[0022] Based on the anomaly determination result, the pathology number is matched to identify the anomaly classification rules to determine the failure mode corresponding to the anomaly area.

[0023] In one embodiment, based on the strategy chain library and the triggering conditions corresponding to the execution strategy, a combined conditional verification is performed on the confidence score of the preliminary identification result to determine at least one candidate strategy;

[0024] Based on the strategy priority and pathological business requirements, the candidate strategies are prioritized to obtain the strategy sequence of the preliminary identification results;

[0025] The candidate strategy ranked first in the strategy sequence is taken as the execution strategy corresponding to the preliminary identification result.

[0026] In one embodiment, based on the execution strategy corresponding to the candidate region, the preliminary recognition result, associated image data and recognition process metadata of the candidate region are parsed to determine the execution action, verification rules and execution parameters corresponding to the execution strategy;

[0027] Based on the execution action and the execution parameters, a targeted processing operation is performed to match the candidate region, and the initial strategy output result of the candidate region is obtained;

[0028] According to the verification rules, the initial strategy output results of the candidate regions are subjected to confidence verification and format validity verification to generate valid recognition results of the candidate regions.

[0029] Based on the regional order of the candidate regions, the valid identification results are merged to obtain the target identification result of the pathology number.

[0030] In one embodiment, based on the high-confidence target recognition result and the auxiliary coding rules of the same slide pathology number, the business attribute reasoning and matching of the short character recognition result with the intermediate confidence level is performed to determine the tissue block number attribute corresponding to the short character recognition result;

[0031] According to the consistency rules of pathological information on the same slide, the association compliance between the high-confidence target identification result and the tissue block number attribute is verified to obtain the pathology number result;

[0032] Based on the pathology number result, the verification and recognition result of the pathology slide image is determined, and the sample data and rule execution data corresponding to this reasoning and matching process are archived and stored as difficult case data.

[0033] In one embodiment, the difficult example data in the difficult example library is screened, cleaned, deduplicated, and standardized to obtain a difficult example training dataset that meets the training requirements of the model.

[0034] The difficult example training dataset is split according to the engine parameter rules, and targeted incremental training is performed on the high-speed initial screening engine, the high-precision main recognition engine, and the dedicated difficult example engine to generate an optimized recognition engine model.

[0035] Based on the optimized recognition engine model, and combined with the strategy execution effect data corresponding to difficult examples, the trigger threshold, priority weight, and execution parameters of the execution strategies in the strategy chain library are adjusted for adaptability to obtain the linkage execution rules.

[0036] Based on the optimized recognition engine model and the linkage execution rules, the model and rules of the pathology number recognition process are updated to achieve iterative optimization of the pathology number recognition system.

[0037] In addition, to achieve the above objectives, this application also proposes a pathological slide pathological number identification device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the pathological slide pathological number identification method described above.

[0038] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the pathology slide pathology number identification method described above.

[0039] This application provides a method for identifying pathology numbers on pathological slides. The method includes: traversing the pathology slide image in response to a pathology number identification command to locate candidate regions of suspected pathology number text and determine corresponding preliminary identification results; judging the degree of fit between the preliminary identification results and preset evaluation indicators based on preset identification evaluation rules and calculating the corresponding confidence score; matching the corresponding execution strategy from a strategy chain library based on the confidence score of the preliminary identification results and the matched identification failure pattern; and executing the matched corresponding execution strategy on the preliminary identification results to determine the target identification result of the pathology number in the pathology slide image. This hierarchical, progressive, and closed-loop driven intelligent identification processing scheme solves the problems of existing methods. The optical character recognition technology for pathology slide pathology numbers only focuses on optimizing the accuracy of single recognition and lacks a systematic recovery mechanism for recognition failures. In scenarios with complex backgrounds, varied text forms, and inconsistent imaging conditions, it is prone to technical problems such as direct interruption of business processes after recognition failures, insufficient system robustness, and poor stability of automated operation. The solution improves the overall accuracy and scenario adaptability of pathology number recognition, significantly reduces the interruption rate of the pathology information digitization process, realizes intelligent hierarchical recovery of recognition failures and continuous closed-loop optimization of the system, and improves the efficiency of human-machine collaborative processing and the level of automated operation of the entire pathology business process.

[0040] In summary, this application solves the technical problem of low overall efficiency in pathology number identification by using a scheme of identifying candidate regions of pathology slides, calculating confidence scores, and matching and executing corresponding strategies with a strategy chain library. This improves the identification accuracy and automation efficiency, and achieves stable closed-loop operation of the identification process. Attached Figure Description

[0041] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating the first embodiment of the method for identifying pathology slide pathology numbers in this application;

[0044] Figure 2 Here is a flowchart for identifying the pathology number in this application;

[0045] Figure 3 This is a flowchart illustrating the fourth embodiment of the method for identifying pathology slide pathology numbers in this application;

[0046] Figure 4 This is a flowchart illustrating the eighth embodiment of the method for identifying pathology slide pathology numbers in this application;

[0047] Figure 5 This is a system framework diagram for this application;

[0048] Figure 6 This is a schematic diagram of the device for identifying the pathology number of the pathology slides in this application.

[0049] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0050] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0051] The relevant technologies generally use a general model to perform single text recognition. This method is prone to misidentifying pathological tissue texture as text when the background of the pathological slide is complex and the text form is varied, which leads to poor accuracy in pathology number recognition.

[0052] This application provides a solution: First, in response to a pathology number recognition instruction, candidate regions of suspected pathology number text in a pathology slide image are traversed to determine the preliminary recognition result of the candidate regions. Then, based on recognition evaluation rules, the degree of fit between the preliminary recognition result and the evaluation index is judged, and a confidence score of the preliminary recognition result is calculated based on the degree of fit. Next, based on the confidence score of the preliminary recognition result, an execution strategy corresponding to the confidence score is matched from a strategy chain library. Finally, the corresponding execution strategy is executed on the preliminary recognition result to determine the target recognition result of the pathology number in the pathology slide image.

[0053] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as a pathology slide pathology number recognition device. The following description uses a pathology slide pathology number recognition device as an example to illustrate this embodiment and the subsequent embodiments.

[0054] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0055] This application provides a method for identifying the pathology number of a pathology slide, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the method for identifying pathology slide pathology numbers in this application.

[0056] In this embodiment, the method for identifying the pathology number of the pathology slide includes steps S10 to S40:

[0057] Step S10: In response to the pathology number recognition instruction, traverse the candidate regions of suspected pathology number text in the pathology slide image and determine the preliminary recognition result of the candidate regions.

[0058] The pathology number recognition command is the core control command that triggers the pathology number recognition task. The candidate region for suspected pathology number text is a local image region in the pathology slide image that, after initial screening, is suspected of containing pathology number text. The preliminary recognition result is the initial text recognition output information generated after performing fine-grained recognition on the text content within the candidate region.

[0059] In this embodiment, the pathology number recognition command can be triggered in six ways. First, a timed task trigger: the system automatically generates and issues a pathology number recognition command at a specified time point according to a preset periodic scheduling rule, used to perform routine historical pathology slide archiving and recognition tasks. Second, slide scanning completion trigger: after the entire field-of-view digital image of a single pathology slide is written to the system storage medium, the system automatically generates and issues a pathology number recognition command, used to perform immediate pathology number recognition processing on newly scanned pathology slides. Third, user-initiated trigger: the user initiates a recognition request through the pathology information system interface or inputs a recognition execution command through a batch processing tool; the system receives the request and generates and issues a pathology number recognition command. Fourth, batch archiving threshold trigger: when the cumulative number of pathology slides to be processed reaches a preset batch threshold, the system automatically generates and issues a pathology number recognition command, used to control the progress of the digital archiving process for pathology slides. Fifth, system resource idle triggering: When the system detects that the real-time load of both the central processing unit and the graphics processing unit is lower than the preset idle threshold, it automatically generates and issues a background pathology number recognition instruction, utilizing the system's idle computing power to execute non-urgent historical slide recognition tasks. Sixth, pathology quality control pre-triggering: When the system receives a pathology slide quality control review request, it automatically generates and issues a pathology number recognition instruction, completing pathology number recognition and matching before initiating the quality control process, ensuring accurate binding between quality control data and pathology numbers.

[0060] Once the pathology number recognition device receives the pathology number recognition instruction, it begins loading the pathology slide image to be processed and performs candidate region localization and preliminary recognition processing.

[0061] For example, there are two methods for generating candidate region localization and preliminary identification results. The first is a global dual-engine serial identification. Before the identification task officially starts, the complete metadata of the pathological slide image is read to complete the global image standardization preprocessing. Then, the high-speed initial screening engine is called to traverse the entire pathological slide image, locating all candidate regions of suspected pathology number text in the entire region, while removing invalid background regions without text features, resulting in a set of text candidate regions after initial screening. Subsequently, the high-precision main identification engine is called to perform fine identification on each candidate region in the candidate region set, generating a preliminary identification result corresponding to each candidate region. After all regions are identified, the results are uniformly output to the evaluation module. This method has a simple and stable identification logic, is less prone to missed regions or identification errors, and can provide comprehensive and reliable basic data for subsequent confidence assessment.

[0062] The second method is a streaming, segmented, dual-engine parallel recognition approach. During the streaming reading and segmentation loading of pathological slide images, incremental initial screening and localization are performed simultaneously with image segmentation. The text candidate regions for suspected pathology numbers within the current segment are located slice by slice, and a cumulative set of text candidate regions for the entire slice is obtained. After the entire slice image is read, global deduplication and normalization of candidate regions are performed synchronously. Simultaneously, in a pipelined parallel manner, a high-precision main recognition engine is invoked to perform fine recognition on the located candidate regions, recognizing them as they are located. After the candidate regions for the entire slice are located, preliminary recognition results for all candidate regions are generated synchronously, and abnormal regions are marked during the recognition process to provide risk references for subsequent assessment. This method is executed synchronously with the image reading and segmentation loading process, without requiring additional pre-processing time. It can fully utilize idle periods in the pipeline to complete the recognition process, significantly reducing the total time of the entire process.

[0063] In an exemplary scheme for determining candidate regions and preliminary recognition results, the complete image of the pathological slide to be processed is first loaded. Global illumination and color normalization preprocessing is then performed on the image to eliminate image quality differences caused by different imaging conditions, resulting in a standardized pathological slide image. Subsequently, the standardized pathological slide image is input into a high-speed initial screening engine. Aiming for a high recall rate with no missed detections across the entire region, the engine traverses the entire standardized pathological slide image, locating all candidate regions suspected of containing pathology number text, while simultaneously removing invalid background regions without textual features, resulting in a set of text candidate regions after initial screening. Next, for each candidate region in the text candidate region set, local image normalization preprocessing is performed to weaken interference features of non-textual background within the region, resulting in a single-region standardized image adapted for high-precision recognition. Finally, all single-region standardized images are sequentially input into the high-precision main recognition engine, where character-level fine recognition is performed on the text content within each candidate region, generating a preliminary recognition result corresponding to each candidate region. Then, the validity of the preliminary identification results is verified. This verifies whether the identification result of each candidate region is empty or contains invalid or garbled characters. Completely invalid identification results are removed. Duplicate regions are then merged. If there are multiple spatially overlapping candidate regions, the one with the most complete identification result is retained, and the remaining redundant regions are removed. Finally, all candidate regions that pass the verification are summarized with their corresponding preliminary identification results to generate the final set of candidate regions and preliminary identification results, which are used for subsequent confidence assessment and strategy matching.

[0064] It should be noted that in some special cases, if there is only one candidate region and a preliminary identification result, the subsequent multi-region parallel evaluation process is skipped, and the confidence evaluation process for a single result is directly entered.

[0065] Step S20: Based on the identification and evaluation rules, determine the degree of fit between the preliminary identification result and the evaluation index, and calculate the confidence score of the preliminary identification result based on the degree of fit.

[0066] The identification evaluation rules are pre-defined calculation rules used to quantitatively evaluate the reliability of preliminary identification results. Evaluation metrics are multi-dimensional criteria used to measure the quality of identification results. Fit is a quantitative value representing the degree of matching between the preliminary identification result and the evaluation metrics in the corresponding dimension. Confidence score is a comprehensive quantitative indicator used to measure the overall reliability of the preliminary identification results. For example, evaluation metrics include the probability of characters output by the identification engine, consistency of multi-engine identification results, compliance with pathology number encoding rules, image quality of text regions, and character integrity.

[0067] In this embodiment, the preliminary identification results, corresponding candidate region images, and identification process data are combined with the identification evaluation rules to perform a fit judgment and confidence score calculation, thereby initiating the confidence evaluation process for each preliminary identification result.

[0068] For example, the confidence scores for each preliminary identification result are calculated in two ways. The first method is a sequential, single-result, dimension-by-dimensional calculation. After the confidence assessment process officially starts, the preliminary identification results and associated basic data corresponding to each candidate region are extracted one by one according to the priority order of the candidate regions. Then, based on the identification assessment rules, the individual fit degree of each preliminary identification result with the corresponding preset standard is calculated in each assessment indicator dimension. After completing the full-dimensional fit degree calculation for a single result, a weighted fusion is performed to obtain the confidence score corresponding to the result. After completing the calculation of one result, the calculation of the next candidate result begins. After all results are calculated, they are uniformly output to the strategy matching module. This method has a simple and stable calculation logic, and the calculation process of each result is independent and controllable, making it less prone to calculation deviations and providing a reliable judgment basis for subsequent strategy matching.

[0069] The second approach involves dimensional splitting and batch parallel computation of all results. First, the basic data associated with the initial recognition results is dimensionally split, extracting feature subsets for engine output dimension, rule compliance dimension, image quality dimension, and character integrity dimension. Simultaneously, all initial recognition results are categorized and aligned according to these dimensions, and parallel computation across all dimensions is initiated concurrently. Within each dimension, the individual fit of all initial recognition results is calculated synchronously. After the individual fit calculations for all dimensions are completed, a weighted fusion is performed on all initial recognition results to obtain batch confidence scores for all initial recognition results. Abnormal results during the calculation process are also marked to provide risk references for subsequent strategy matching. This method executes synchronously with the multi-dimensional parallel processing flow, eliminating the need for additional serial computation time. It fully utilizes idle system computing power to complete batch computations, significantly improving the overall efficiency of confidence assessment.

[0070] Before calculating the confidence score, the corresponding identification and evaluation rules can be determined based on the following methods.

[0071] In one optional approach, a pre-set set of fixed recognition and evaluation rules is retrieved. This set pre-stores standardized evaluation rules and weight parameters adapted to different recognition scenarios and pathology number coding rules. Based on the hospital and pathology number coding rule type of the pathology slide to be processed, the corresponding pre-set recognition and evaluation rule is retrieved for subsequent fit assessment and confidence score calculation. This method is logically simple and stable, with unified and controllable rules, requiring no additional adaptation calculations. It can quickly determine the evaluation rules and is suitable for routine pathology slide recognition scenarios with high requirements for recognition response speed.

[0072] In another alternative approach, the core features of the preliminary identification results and the image quality features of the corresponding candidate regions are extracted first. Simultaneously, the hospital pathology number coding rules corresponding to the pathology slides to be processed are extracted. Then, evaluation rules optimized based on historical identification tasks are retrieved from the rule base. Based on the pathology number coding rules, image quality features, and identification result features, identification evaluation rules and corresponding weight parameters adapted to this identification task are dynamically generated. The feasibility of the generated identification evaluation rules is verified, confirming that the input and output range of the rules meets the computational requirements of the current system. After verification, these rules are used as the identification evaluation rules for this confidence assessment. This method can flexibly generate customized evaluation rules that fit the current identification task and slide features, without relying on fixed preset templates. It has higher adaptation accuracy, can fully explore the features of different pathology number coding rules, improve the accuracy of confidence scoring, and adapt to complex identification scenarios such as special coding rules and low-quality slides.

[0073] In an exemplary scheme for determining the confidence score, the set of recognition evaluation rules corresponding to all preliminary recognition results is first loaded. This set includes two categories of basic rules: evaluation rules for regular pathology numbers and evaluation rules for specially coded pathology numbers. Each rule is pre-configured with corresponding weight parameters and a single-item fit calculation rule. Then, input parameter normalization is performed, standardizing the multi-dimensional basic features associated with the collected preliminary recognition results to eliminate differences in the dimensions of different parameters. Next, for each preliminary recognition result, the normalized feature parameters are input into the corresponding recognition evaluation rule. First, the single-item fit of the engine output dimension of the preliminary recognition result is calculated; then, the single-item fit of the rule compliance dimension is calculated; then, the single-item fit of the image quality dimension is calculated; and finally, the single-item fit of the character integrity dimension is calculated. After completing the single-item fit calculation, weighted fusion is performed based on the weight parameters to obtain the preliminary confidence score of the preliminary recognition result. Next, a confidence score validity check is performed to verify whether the confidence score of each result is within a reasonable range. Abnormal scores are removed, and redundant results are then eliminated. If there are duplicate recognition results in the same area with confidence score differences less than a preset threshold, the one with the most balanced single-item fit is retained, and the remaining redundant results are removed. Finally, all the verified confidence scores are summarized to generate the final confidence score set corresponding to each preliminary recognition result. At the same time, based on the abnormal distribution characteristics of single-item fit in each dimension, a preset recognition anomaly classification rule is matched to determine the failure mode corresponding to each preliminary recognition result, which is used for subsequent execution strategy matching.

[0074] Step S30: Based on the confidence score of the preliminary identification result, match the execution strategy corresponding to the confidence score from the strategy chain library.

[0075] The strategy chain library is a pre-stored collection of execution strategies for processing recognition results with different confidence levels and failure modes. An execution strategy is a pre-defined set of standardized processing procedures and actions for recognition results at a corresponding confidence level. For example, execution strategies include strategies for directly accepting high-confidence results, directly rejecting false positives, secondary recognition of low-confidence text regions, rule-based error correction for mid-confidence results, and human-computer interaction correction for results that cannot be automatically recovered.

[0076] In this embodiment, the corresponding execution strategies can be selected in two ways. First, a serial hierarchical threshold matching method is used. All preliminary identification results are first divided into three levels: high confidence, medium confidence, and low confidence, based on a preset level threshold corresponding to the confidence score. Then, in descending order of level, the execution strategies corresponding to the confidence level in the strategy chain library are matched one by one. Simultaneously, the failure mode corresponding to the result is considered to complete the accurate matching of the strategy. Once the matching is complete, the corresponding execution strategy is directly locked without verifying other alternative strategies, thus quickly completing the matching process. This method is logically simple and stable, and the matching process progresses step by step, enabling rapid locking of the suitable execution strategy, which is suitable for conventional pathological slide identification scenarios.

[0077] Secondly, the system employs a two-dimensional parallel batch matching approach. First, it simultaneously verifies the confidence level and failure mode of all preliminary identification results, batch-marking all results with suitable conditions. Then, based on the two-dimensional combination of confidence level and failure mode, it simultaneously matches all matching execution strategies from the strategy library. Finally, based on preset strategy priority rules, it locks the highest priority target execution strategy for each identification result. This method fully utilizes parallel computing power to complete batch matching, avoiding the time-consuming nature of serial matching. Simultaneously, it comprehensively evaluates the suitability of all alternative strategies, avoiding the omission of the optimal strategy, and adapts to the batch processing needs of large-scale slices and complex identification scenarios.

[0078] In an exemplary scheme for determining an execution strategy, a pre-set strategy library is first loaded. This library contains three main categories of basic strategies: high-confidence processing strategies, mid-level confidence processing strategies, and low-confidence processing strategies. Each strategy is pre-configured with a corresponding two-dimensional trigger condition matrix, priority weights, and execution link rules. Then, a preliminary screening based on basic conditions is performed. The confidence score corresponding to the initial identification result is compared with a pre-set confidence level threshold range, and the corresponding failure mode is compared with a pre-set strategy-adapted failure mode range. Candidate execution strategies that simultaneously meet both the confidence level requirement and the failure mode adaptation requirement are selected. Next, for the candidate execution strategies that pass the preliminary screening, it is verified whether the execution conditions of the strategy conform to the business rule requirements of the current pathological slide, and whether the system resources required by the strategy are within the current system's available resource range. Candidate strategies that do not meet the execution conditions are eliminated. Finally, strategy priority ranking is performed. Based on pre-set priority rules of automation priority and accuracy priority, the remaining candidate execution strategies are ranked from highest to lowest priority. The strategy with the highest priority is selected as the target execution strategy, and the corresponding execution parameters and execution link rules are extracted. Finally, all the target execution strategies that match the preliminary identification results are summarized to generate the final execution strategy matching result set, which is used for subsequent strategy execution and result output.

[0079] It should be noted that in some special cases, if there is only one candidate execution strategy that matches the confidence score and failure mode of the identification result, the priority ranking process is skipped and the strategy is directly used as the target execution strategy corresponding to the identification result.

[0080] Step S40: Execute the corresponding execution strategy on the preliminary identification result to determine the target identification result of the pathology number in the pathological slide image.

[0081] The target identification result, after strategy execution, verification, and merging, generates the final pathology number identification output information, which can be directly connected to the pathology information management system. For example, the execution actions include directly adopting the identification result, directly removing invalid results, secondary identification by a dedicated difficult case engine, rule-based error correction, guided human-computer interaction correction, and logical fusion of multiple results.

[0082] In this embodiment, the execution rules corresponding to the execution strategy can be invoked in two ways. First, a fixed preset execution rule can be invoked directly. This method retrieves the set of execution rules pre-stored in the system, and directly matches and retrieves the corresponding standardized execution rule based on the type of the target execution strategy. This rule is then directly used for the current strategy execution. This method has simple and stable invocation logic, unified and controllable rules, fast invocation speed, and is suitable for conventional pathological slide recognition scenarios.

[0083] Secondly, the dynamic business adaptation rule generation and invocation process first extracts the core parameters of the target execution strategy, and simultaneously extracts the hospital pathology number coding rules corresponding to the pathology slides to be processed and the recognition result features of slides in the same batch. Then, it calls the execution rules optimized based on historical tasks in the rule base to dynamically generate customized execution rules adapted to this recognition task. After completing the feasibility verification of the rules, the execution rules are called for the strategy execution processing in this case. This method can flexibly generate customized rules that fit the current business scenario, with higher adaptation accuracy. It can fully combine the prior rules of pathology business to improve the accuracy of recognition results and business adaptability, and adapt to complex recognition scenarios such as special coding rules and multi-tissue block slides.

[0084] Once the execution rules are configured, the corresponding processing operations are performed on the preliminary identification results according to the target execution strategy to generate the target identification result of the pathology number corresponding to the pathology slide image.

[0085] In an exemplary scheme for executing a strategy and determining target recognition results, the execution rule set corresponding to the target execution strategy matching all preliminary recognition results is first loaded. This set includes two basic rule categories: automated processing rules and human-computer interaction processing rules. Each rule is pre-configured with corresponding execution parameters and processing flow. Then, pre-execution preparations are performed. For secondary recognition strategies, a dedicated difficult example engine is loaded and initialized. For rule correction strategies, the pathology number encoding rule library of the corresponding hospital is loaded. For human-computer interaction strategies, the highlighted areas of the interactive interface are configured. Next, for each preliminary recognition result, matching targeted processing operations are performed: for high-confidence results, a direct adoption operation is performed to generate valid recognition results; for false positives with no text, a direct rejection operation is performed to generate invalid result markers. For low-confidence valid text results, the dedicated difficult example engine is invoked to perform secondary recognition, generating secondary recognition results. For intermediate-confidence results, rule-based error correction processing is performed to generate corrected results. For results that cannot be automatically recovered, guided human-computer interaction is triggered to obtain manual correction results. After executing the strategy for all candidate regions, the validity of the execution results is verified. Based on the pathology number encoding rules, compliance review is performed on all valid recognition results. Simultaneously, the logical association verification and fusion processing of recognition results for multiple regions on the same slice are completed. The main pathology number and the auxiliary tissue block number are decomposed and labeled to generate a standardized structured pathology number result. Subsequently, the execution results are standardized and encapsulated to generate and output the target recognition result of the pathology number corresponding to the pathology slice image. At the same time, according to the preset difficult case sample screening rules, low confidence samples, strategy recovery processing samples, manually corrected samples, and false positive / false negative samples are screened from the entire process data of this recognition, generating a standardized difficult case sample set and archiving it to the difficult case sample library. The strategy execution effect data of this time is synchronized to the strategy chain library, completing the closed-loop processing of the recognition process.

[0086] Further, please refer to Figure 2 , Figure 2This is a flowchart of the pathology number recognition process for this application. In the pathology department's digital pathology slide pathology number recognition scenario, the pathology scanner processing terminal is pre-configured with a high confidence threshold of 0.90, a medium confidence range of 0.60-0.89, and a low confidence threshold of 0.60. The high-speed initial screening engine is a detection model fine-tuned for pathology text, the high-precision main engine is a character recognition model pre-trained for pathology numbers, and the dedicated difficult case engine is a model specifically fine-tuned for fuzzy pathology text. The strategy chain library contains five types of execution strategies: direct adoption, direct discard, secondary recognition, rule correction, and guided human-computer interaction. The preset hospital pathology number encoding rule is: letter prefix + 4-digit year + hyphen + 3-digit serial number + 2-digit optional tissue block number. In response to the pathology number recognition command, the system acquires a full-field image of a 40×HE-stained pathology slide. A high-speed initial screening engine traverses and locates three candidate text regions (R1, R2, and R3). A high-precision main engine then identifies these regions, yielding preliminary results: R1 is "A2023-001", R2 is meaningless gibberish, and R3 is "03". The system then performs a unified confidence assessment, calculating R1 confidence at 0.96 (no anomalies), R2 confidence at 0.22 (no false positives), and R3 confidence at 0.58 (low confidence, recognition failed). The system then applies corresponding matching strategies to each region: if R1 confidence meets the criteria, the matching strategy directly adopts the result "A2023-001". If R2 has no false positives, the matching strategy directly discards the invalid result. R3 represents a low-confidence result for valid text. A secondary recognition strategy is applied, and the dedicated difficult example engine is used for re-recognition, yielding the result "05" with a confidence level of 0.72. This result falls within the mid-confidence range and conforms to the tissue block number rules. The rule-based error correction strategy is applied, and the correction is verified as valid, so the result "05" is adopted. If the secondary recognition result still fails to meet the requirements, a guided human-computer interaction is triggered, allowing the operator to manually correct the error before adopting the result. Finally, the system merges all valid results and outputs the final pathology number recognition result "A2023-001-05". Simultaneously, the low-confidence sample and the manually corrected sample are stored in the difficult example database, completing the recognition loop.

[0087] By employing a dual-engine hierarchical identification system, a dual-dimensional evaluation system of confidence level and failure mode, and a hierarchical anomaly recovery mechanism driven by a strategy chain library, the problem of poor accuracy in pathology number identification has been solved, improving the accuracy and automation efficiency of pathology number identification and achieving closed-loop iterative optimization of the identification process.

[0088] Second Embodiment

[0089] This embodiment provides an exemplary scheme for locating and initially identifying candidate regions for pathology number on pathology slides. In this example, global standardization preprocessing is first performed on the pathology slide image to eliminate imaging differences. Then, a preliminary screening operation is performed to traverse the entire image to locate candidate regions for suspected pathology number text. Subsequently, local standardization processing is performed on the candidate regions to weaken background interference. Finally, a high-precision text recognition operation is used to generate preliminary identification results for each candidate region. Step S10 includes steps A11 to A14:

[0090] Step A11: In response to the pathology number identification instruction, the pathology slide image is preprocessed to obtain a standardized pathology slide image.

[0091] Step A12: By traversing the pathological slide images through the initial screening operation, the candidate regions of all suspected pathology number texts are located, and invalid background regions without text features are identified to obtain candidate text regions.

[0092] Step A13: Perform local contrast enhancement and texture denoising on the candidate text regions respectively, remove interference features that are not text backgrounds in the candidate text regions, and obtain the text-standardized regions.

[0093] Step A14: Through high-precision text recognition, text content recognition is performed on each of the standardized text regions to generate preliminary recognition results corresponding to the candidate regions.

[0094] The initial screening operation is a fast text region localization operation based on a lightweight text detection model, with high recall as its core objective. It is used to quickly locate local regions of suspected pathology number text in large-format pathology slide images, eliminating the vast majority of invalid background regions without text. Examples include lightweight model-based text detection operations, model-based fast text localization operations, and traditional text region detection operations.

[0095] The candidate text region set is a collection of candidate regions of all suspected pathology number texts that are retained after the invalid background is removed in the initial screening operation. It is the input object for subsequent local standardization processing.

[0096] Local contrast enhancement and texture denoising are local image optimization operations performed on local images of candidate regions to strengthen text features and suppress non-textual background interference. They are core preprocessing steps to improve the accuracy of subsequent text recognition. Examples include adaptive histogram equalization, non-local mean denoising, Gaussian filtering smoothing, and text edge enhancement.

[0097] Text standardization regions are candidate regions that have undergone local image optimization processing, resulting in clear text features, reduced background interference, and suitability for high-precision text recognition requirements. They serve as the direct input for high-precision text recognition.

[0098] High-precision text recognition is a character-level text content parsing operation based on a pre-trained pathology number-specific character recognition model, with high accuracy as its core objective. It is used to convert image content of standardized text regions into editable text recognition results. Examples include model-based end-to-end character recognition, model-based scene text recognition, and sequence text recognition.

[0099] The initial recognition result is the initial text recognition content corresponding to the candidate region, output by a high-precision text recognition operation. This serves as the basic input for subsequent confidence assessment and strategy matching. Examples include complete coded pathology number text, fragmented pathology number text, meaningless garbled text, and empty recognition results.

[0100] In this example, when preprocessing the pathological slide image to obtain a standardized pathological slide image, the global illumination and color normalization preprocessing method provided in the first embodiment can be referred to. Alternatively, a streaming segmentation preprocessing method can be used, where data from each segment of the pathological slide image is read while parallel normalization preprocessing is performed on each segment. After the entire slide image is read, the data is summarized and stitched together to obtain a globally standardized pathological slide image, thus completing the preprocessing operation of the entire slide image.

[0101] After standardizing the pathology slide images, the initial screening and localization process is initiated. The standardized pathology slide images are input into a high-speed initial screening engine. The initial screening operation traverses the entire pathology slide image, locating candidate regions for all suspected pathology number texts, while eliminating invalid background regions without text features. This results in a set of candidate text regions. Subsequently, for each candidate region within the set, local contrast enhancement and texture denoising are performed sequentially to weaken non-text background interference features within the region, resulting in a standardized text region for each candidate region. Finally, all standardized text regions are sequentially input into a high-precision main recognition engine. Through high-precision text recognition, character-level text content recognition is performed on each standardized text region, generating a preliminary recognition result for each candidate region. This hierarchical preprocessing and dual-engine collaborative recognition architecture improves the accuracy of preliminary pathology number recognition and the overall processing efficiency, avoiding recognition deviations caused by differences in image quality and background interference.

[0102] For example, the candidate region localization and preliminary recognition results are generated in two ways. The first is a global dual-engine serial recognition. Before the recognition task officially starts, the complete metadata of the pathological slide image is read to complete the global standardization preprocessing of the entire image, resulting in a standardized pathological slide image. Then, the high-speed primary screening engine is called to traverse the entire standardized pathological slide image, locating all candidate regions of suspected pathology number text in the entire region, while removing invalid background regions without text features, resulting in a set of candidate text regions after primary screening. Subsequently, the high-precision main recognition engine is called to perform local standardization preprocessing and character-level fine recognition on each candidate region in the candidate text region set, generating a preliminary recognition result corresponding to each candidate region. After all regions are recognized, the results are uniformly output to the subsequent confidence assessment module. This method adopts a serial execution logic of full-image preprocessing, global primary screening, and region-by-region recognition. Through unified preprocessing of the entire image and centralized primary screening localization, it ensures that no candidate regions are missed. The recognition logic is simple and stable, and it is not prone to region omissions or recognition errors, providing comprehensive and reliable basic data for subsequent confidence assessment.

[0103] The second method is a streaming, segmented, dual-engine parallel recognition. During the streaming reading and segmentation loading of pathological slide images, parallel standardization preprocessing is performed on the current segment while reading the image segments, resulting in a standardized segmented image. Simultaneously, a preliminary screening operation is performed on the preprocessed segments to locate candidate regions of suspected pathology number text within the current segment, eliminating invalid background regions, and accumulating a set of candidate text regions for the entire slide. After the entire slide image is read, global deduplication and normalization of candidate regions are performed simultaneously. At the same time, in a pipelined parallel manner, local standardization preprocessing and high-precision text recognition operations are performed on the located candidate regions, recognizing them as they are located. After the candidate regions of the entire slide are located, preliminary recognition results corresponding to all candidate regions are generated simultaneously, and abnormal regions in the recognition process are marked to provide risk reference for subsequent assessment. This method employs parallel execution logic of streaming reading, sliced ​​parallel processing, and pipeline collaborative recognition. Through sliced ​​parallel processing synchronized with the image reading process, it eliminates the need for separate preprocessing and initial screening, making full use of the pipeline's idle time to complete the entire process. This significantly reduces the total time required for pathology number recognition and is suitable for batch recognition scenarios involving large quantities of pathology slides.

[0104] Third Embodiment

[0105] This embodiment provides an exemplary scheme for quantifying the confidence score of preliminary pathology number identification results. In this example, multi-dimensional core features associated with the preliminary identification results are first extracted. Then, each feature is matched with a preset evaluation index to calculate the individual fit. Finally, a weighted fusion calculation is performed on all individual fits using preset weights to obtain a comprehensive confidence score that accurately characterizes the reliability of the identification results. Step S20 includes steps B11-B13:

[0106] Step B11: Based on the preliminary recognition results corresponding to the candidate regions, the original images of the candidate regions, and the recognition process data, extract the recognition output probability, the pathology number rule compliance degree, and the image quality of the text region.

[0107] Step B12: Match the recognition output probability, the pathology number rule compliance, and the text region image quality with the corresponding evaluation indicators to determine the degree of fit for each indicator.

[0108] Step B13: Based on the evaluation weights corresponding to the evaluation indicators, perform a weighted fusion calculation on the fit of all individual indicators to obtain the confidence score corresponding to the preliminary identification result.

[0109] The preliminary recognition result is the initial text recognition content corresponding to the candidate region, output by a high-precision text recognition operation. This is the core evaluation object for the confidence assessment in this embodiment. Examples include complete coded pathology number text, fragmented pathology number text, meaningless garbled text, and empty recognition results.

[0110] The original images of candidate regions are unprocessed local images obtained through initial screening and corresponding to preliminary identification results. They serve as the fundamental data source for extracting image quality features from text regions. Examples include original screenshots of pathology slide label areas, original images of handwritten pathology numbers on slide edges, and original images of areas with appended numbers next to tissue blocks.

[0111] The recognition process data refers to the intermediate operational data output by the high-precision text recognition engine during text recognition, which characterizes the reliability of the recognition process and is the core source for extracting the recognition output probability. Examples include single-character recognition output probability, full-sequence recognition confidence, consistency data of multi-engine recognition results, and recognition decoding path information.

[0112] The output probability is a core quantitative indicator output by a high-precision text recognition engine, representing the inherent reliability of the text recognition result. It is a core dimension for evaluating the credibility of the recognition result itself. Examples include the average output probability of a single character, the confidence score of the entire sequence recognition, the recognition probability of key coding bits, and the recognition probability of fixed first and last characters.

[0113] Pathology number rule compliance is a quantifiable value representing the degree to which the preliminary identification results match the pre-defined standardized coding rules for pathology numbers. It is a core dimension for evaluating the operational effectiveness of the identification results. Examples include coding format compliance, character type compliance, coding length compliance, and matching degree with hospital-specific pathology number coding rules.

[0114] Text region image quality is a quantitative indicator of the imaging quality of the original image of the candidate region, and it is a fundamental supporting dimension for evaluating the reliability of the recognition results. Examples include region contrast, image sharpness, character integrity, illumination uniformity, and the degree of background noise interference.

[0115] The fit metric is a standardized quantification of the degree of matching between the core features extracted from each dimension and the corresponding evaluation indicators. Its value ranges from 0 to 1 and is the basic calculation unit for weighted fusion calculation of confidence scores. Examples include the fit metric for recognition output probability, the fit metric for pathology number rule compliance, and the fit metric for text region image quality.

[0116] The evaluation weights are preset coefficients corresponding to the importance proportion of each evaluation indicator. The sum of the evaluation weights for all dimensions is 1, used to balance the influence of each dimension on the final confidence score. For example, the output recognition probability corresponds to a weight, the pathology number rule compliance corresponds to a weight, and the image quality of the text region corresponds to a weight.

[0117] In this example, when extracting the multi-dimensional core features associated with the preliminary recognition results, it can be done using a synchronous feature acquisition method. Alternatively, a streaming parallel feature extraction method can be used, where the preliminary recognition results, original images, and recognition process data of each candidate region are received simultaneously, and the corresponding core features are extracted in parallel. After all relevant data for all candidate regions have been received, the full set of multi-dimensional core features is obtained, thus completing the feature extraction operation.

[0118] After extracting multi-dimensional core features, the single-item fit calculation process is initiated. The extracted recognition output probability, pathology number rule compliance, and text region image quality are matched and compared item by item with the corresponding preset evaluation indicators to calculate the quantified fit value for each individual indicator. Subsequently, the preset evaluation weights for each evaluation indicator are retrieved, and a weighted fusion calculation is performed on the quantified fit values ​​of all individual indicators to obtain the confidence score for each preliminary recognition result. This multi-dimensional hierarchical evaluation and weighted fusion calculation logic improves the accuracy of the confidence score, avoids the result bias caused by single-dimensional evaluation, and provides a reliable judgment basis for subsequent strategy matching.

[0119] For example, there are two methods for calculating the confidence score of the preliminary identification result. The first method is a sequential single-result, dimension-wise weighted calculation. After the confidence assessment process is officially started, the preliminary identification result and associated multi-dimensional core features corresponding to each candidate region are extracted one by one according to the priority order of the candidate regions. Then, each dimension feature is compared with the corresponding evaluation index item by item, and the single-item fit quantification value of each dimension is calculated in turn. After the full-dimensional fit calculation of a single result is completed, a weighted sum is performed based on the preset evaluation weight to obtain the confidence score corresponding to the preliminary identification result. After the calculation of one result is completed, the calculation of the next candidate result is started. After all results are calculated, they are uniformly output to the subsequent strategy matching module. This method adopts the logic of sequential calculation of single-result full dimensions. The calculation process of each result is independent and controllable, and it is not easy to have calculation deviations. It can accurately adapt to the feature differences of each candidate region, provide a reliable judgment basis for subsequent strategy matching, and is suitable for the identification scenario of a conventional single pathological slide.

[0120] The second approach involves dimensional decomposition and full-result batch parallel computation. First, the multi-dimensional core features associated with all preliminary identification results are decomposed, extracting feature subsets for the identification output probability dimension, pathology number rule compliance dimension, and text region image quality dimension. Simultaneously, all preliminary identification results are categorized and aligned according to these dimensions, and parallel computation across all dimensions is initiated concurrently. Within each dimension, the quantified value of the single-item fit for all preliminary identification results is calculated synchronously. After the single-item fit calculation for all dimensions is completed, a weighted fusion calculation is performed synchronously on all preliminary identification results to obtain the confidence scores for all preliminary identification results in batches. Abnormal results during the calculation process are also marked to provide risk references for subsequent strategy matching. This method employs dimensional decomposition and full-path parallel computation logic, executed synchronously with the multi-dimensional parallel processing flow. It eliminates the need for separate, additional serial computation time, fully utilizing idle system computing power to complete batch computations, improving the overall efficiency of confidence assessment, and making it suitable for batch identification scenarios involving large numbers of pathology slides.

[0121] Fourth embodiment

[0122] This embodiment provides an exemplary scheme for determining anomalies and accurately classifying failure modes in pathology number identification results. In this example, firstly, a multi-level confidence threshold is used to classify the initial identification results into confidence levels, filtering out abnormal areas where the confidence level does not meet the acceptable standard. Then, specific anomaly feature verification rules for each anomaly area are matched to complete a dedicated verification of the pathology number rule compliance and the validity of the text region. Based on the anomaly feature distribution results obtained from the verification, the core problem link in identifying anomalies is located. Finally, the pathology number identification anomaly classification rules are matched to determine the standardized failure mode corresponding to the anomaly area, providing a reliable basis for the accurate matching of subsequent execution strategies. Please refer to... Figure 3 , Figure 3 This is a flowchart illustrating the fourth embodiment of the method for identifying pathology slide pathology numbers in this application. Following step B13, steps C11-C14 are also included:

[0123] Step C11: According to the multi-level confidence level threshold, the preliminary identification results are divided into confidence levels to determine the abnormal areas with confidence levels lower than the qualified level.

[0124] The multi-level confidence level thresholds are preset numerical boundary thresholds used to divide confidence scores into multiple standardized levels. They are the core judgment criteria for achieving accurate confidence level classification. For example, the high confidence level threshold is 0.90, the lower limit threshold of the medium confidence level interval is 0.60, the low confidence level threshold is 0.60, and the acceptable level threshold is 0.90.

[0125] The confidence level classification is an operation that categorizes the confidence score corresponding to each preliminary identification result based on multiple confidence level thresholds, in order to distinguish identification results with different levels of reliability. For example, high confidence level, medium confidence level, low confidence level, qualified level, and unqualified level.

[0126] In this embodiment, the aforementioned confidence level classification and abnormal region identification can be accomplished in four ways. First, a fixed three-level classification: the system presets three fixed confidence level thresholds (high, medium, and low). The confidence scores of all preliminary identification results are compared one by one with these fixed thresholds, directly classifying them to the corresponding confidence level. Simultaneously, all abnormal regions with levels below the acceptable level are filtered out. This method is logically simple and stable, with uniform and controllable thresholds, and is suitable for conventional standardized pathology number identification scenarios. Second, a hospital-specific rule-based dynamic classification: based on the pathology number identification quality control requirements of the hospital to which the pathology slide belongs, the system dynamically adjusts multiple confidence level thresholds and acceptable level standards, and then performs classification and abnormal region filtering on the preliminary identification results. This method can flexibly adapt to the differentiated quality control requirements of different hospitals and is suitable for pathology identification scenarios involving multiple hospital campuses. Third, adaptive grading of slide quality: Based on the overall imaging quality of the pathological slide images, the confidence grading thresholds are dynamically adjusted at multiple levels. For low-quality slides, the acceptable grade threshold is tightened, while for high-quality slides, the acceptable grade threshold is relaxed. Then, confidence grading and abnormal region screening are performed. This method can adapt to pathological slides of different imaging qualities, avoiding misjudgments of abnormal regions due to differences in slide quality. Fourth, priority grading of batch tasks: For large-scale pathological slide identification tasks, the grading thresholds are dynamically adjusted based on the urgency of the task. For urgent archiving tasks, the acceptable grade threshold is relaxed to reduce anomaly processing time, while for routine quality control tasks, the acceptable grade threshold is tightened to improve recognition accuracy. Then, grading and abnormal region screening are performed. This method balances the processing efficiency and recognition accuracy of batch tasks.

[0127] After the pathology number recognition system completes the confidence score calculation for all preliminary recognition results, it begins to load the preset multi-level confidence grading thresholds and initiates the confidence level classification and abnormal area screening process.

[0128] For example, there are two implementation methods for confidence level classification and abnormal region determination. The first method is serial single-result successive classification: After the classification process is officially started, the confidence score corresponding to each preliminary identification result is extracted according to the priority order of the candidate regions. This score is then compared with the multiple confidence level thresholds to determine the confidence level of the result. Simultaneously, it is determined whether the level is lower than the qualified level. If it is lower, it is marked as an abnormal region. After the classification of one result is completed, the processing of the next result begins. After all results are processed, all marked abnormal regions are summarized and uniformly output to the subsequent anomaly verification stage. This method adopts a serial logic of successive judgment of single results. The classification process of each result is independent and controllable, which can accurately adapt to the confidence differences of each candidate region, reduce the likelihood of misjudgment, and provide an accurate list of abnormal regions for subsequent anomaly verification. It is suitable for the identification scenario of a conventional single pathological slide.

[0129] The second approach involves batch parallel classification of all results. First, the confidence scores corresponding to all preliminary identification results are read and normalized in batches. Simultaneously, the multiple confidence level thresholds are broken down into multiple parallel judgment conditions. Parallel judgment of the confidence levels of all preliminary identification results is initiated concurrently, completing the batch classification of all results. At the same time, all abnormal regions with confidence levels below the acceptable level are batch-filtered out, and the confidence level label corresponding to each abnormal region is simultaneously marked, providing basic label data for subsequent anomaly verification. This method uses the logic of parallel judgment of all data, which can fully utilize the system's parallel computing power to complete batch classification, significantly shortening the time required for classifying large batches of identification results. It does not require additional serial processing time and is suitable for batch identification scenarios involving large batches of pathological slides.

[0130] After completing the confidence level classification and identifying of all preliminary identification results and anomaly regions, the marking and output of anomaly regions can be completed in the following manner.

[0131] In one alternative approach, the system first retrieves preset fixed multi-level confidence thresholds and pass / fail standards. The confidence scores of all preliminary identification results are compared item by item with these thresholds. After completing the confidence level classification, all candidate regions with levels below the pass / fail standard are directly filtered out and marked as abnormal regions. Simultaneously, a corresponding confidence level label is assigned to each abnormal region. The results are then aggregated and directly output to the subsequent anomaly verification stage. This method relies on preset fixed thresholds to complete the level classification and anomaly screening. It is logically intuitive, computationally efficient, and fast, making it suitable for conventional standardized pathology number identification scenarios.

[0132] In another alternative approach, the hospital information, image quality information, and priority information of the current identification task for the pathological slide to be processed are first extracted. Simultaneously, the system's preset dynamic threshold adjustment rules are retrieved. Based on this information, multiple confidence level thresholds and pass / fail standards are dynamically generated to suit the current identification task. After comparing the confidence scores of all preliminary identification results with the dynamically generated thresholds item by item to complete the confidence level classification, all candidate regions with levels below the pass / fail level are selected, marked as abnormal regions, and bound with corresponding confidence level labels before being output to the subsequent anomaly verification stage. This method can flexibly generate personalized grading thresholds that fit the current identification task, with higher accuracy in level classification. It can adapt to differentiated identification scenarios of different hospitals, different slide qualities, and different task priorities, improving the accuracy of abnormal region selection.

[0133] In an exemplary scheme for identifying anomaly regions, a pre-set set of multiple confidence level thresholds is first loaded. This set includes four basic thresholds: a high confidence threshold of 0.90, a medium confidence range of 0.60-0.89, a low confidence threshold of 0.60, and a pass / fail threshold of 0.90. The pass / fail level is pre-configured as a high confidence level. Then, a full confidence score is read, summarizing the confidence scores and associated candidate region information for all preliminary identification results. Next, a result-by-result classification is performed, comparing the confidence score of each preliminary identification result with the classification thresholds one by one. Confidence scores ≥ 0.90 are classified as high confidence pass / fail, scores 0.60 ≤ confidence score < 0.90 are classified as medium confidence fail / fail, and scores < 0.60 are classified as low confidence fail / fail. Next, anomaly region filtering is performed, marking all candidate regions classified as intermediate or low-confidence unqualified as anomalies. Each anomaly region is then associated with a corresponding confidence level label, confidence score, preliminary identification result, and the original image information of the candidate region. Next, anomaly region validity verification is performed, removing overlapping anomaly regions with duplicate labels, retaining the anomaly region with the lowest confidence level, and removing other redundant labels. Finally, all valid anomaly regions and their corresponding label information are summarized to generate a final anomaly region list for subsequent anomaly feature verification and rule matching.

[0134] It should be noted that in some special cases, if the confidence level of all preliminary identification results reaches the qualified level and there are no abnormal areas with a confidence level lower than the qualified level, then the subsequent abnormal feature verification, abnormal link location and failure mode determination processes are skipped, and the process directly enters the execution strategy matching stage.

[0135] Step C12: Match the anomaly feature verification rule specific to the confidence level label corresponding to the anomaly region.

[0136] Anomaly feature verification rules are pre-defined, standardized execution rules corresponding to different confidence levels, used to verify anomaly features in abnormal regions. Different confidence level labels correspond to specific verification dimensions and standards, which are the core basis for achieving accurate anomaly feature verification. For example, there are verification rules specific to medium confidence levels, verification rules specific to low confidence levels, verification rules specific to invalid regions, verification rules for compliance with pathology number rules, and verification rules for the validity of text regions.

[0137] In this embodiment, the matching of the above-mentioned anomaly feature verification rules can be completed in four ways. First, fixed label and rule one-to-one matching: the system pre-configures a one-to-one mapping relationship between each confidence level label and its corresponding exclusive anomaly feature verification rule. Based on the confidence level label of the anomaly region, the corresponding exclusive verification rule is directly retrieved. This method has simple and stable matching logic, fast rule retrieval speed, and is suitable for conventional standardized pathology number recognition scenarios. Second, multi-label combination rule matching: for the same anomaly region, confidence level labels, slide quality labels, and hospital rule labels are simultaneously bound. Based on the multi-label combination conditions, the corresponding composite anomaly feature verification rule is matched. This method can achieve more refined rule matching, improve the accuracy of anomaly verification, and is suitable for complex pathology number recognition scenarios. Third, dynamic rule generation matching: based on the confidence level label, confidence score value, and preliminary identification result features of the anomaly region, exclusive anomaly feature verification rules adapted to the anomaly region are dynamically generated. After completing the rule feasibility verification, they are used for subsequent anomaly verification. This method can flexibly adapt to the personalized features of each anomaly region and meet the verification needs of special anomaly scenarios. Fourth, historical data optimization rule matching involves retrieving the optimal verification rule corresponding to historical samples with the same confidence level and anomaly type from the difficult case library and matching it to the current anomaly area. This method can continuously improve the accuracy of rule matching based on historical optimization data and adapt to the continuously iterating pathology number recognition system.

[0138] After the pathology number recognition system completes the identification of abnormal areas and binds them with confidence level labels, it begins to load the system's preset abnormal feature verification rule library and starts the rule matching process.

[0139] For example, there are two matching methods for anomaly feature verification rules. The first is serial single-region label matching. After the rule matching process is officially started, the anomaly regions are sorted from high to low confidence level, and the confidence level labels corresponding to each anomaly region are extracted one by one. Based on the preset label and rule mapping relationship, the exclusive anomaly feature verification rule corresponding to the label is retrieved. After the rule matching of one anomaly region is completed, the processing of the next anomaly region begins. After all anomaly regions are matched, each anomaly region is bound to the matched exclusive verification rule and uniformly output to the subsequent special verification stage. This method adopts a serial logic of sequential matching of single regions. The rule matching process of each anomaly region is independent and controllable, which can accurately adapt to the differentiated verification requirements of different confidence levels, is not prone to rule mismatch, and can provide accurate rule basis for subsequent special verification, adapting to the recognition scenario of conventional single pathological slides.

[0140] The second method is batch parallel matching of all regions with the same label. First, all abnormal regions are grouped and categorized according to their confidence level labels. Abnormal regions with the same confidence level label are grouped into the same processing group. Simultaneously, anomaly feature verification rules specific to the corresponding label are matched to each group. Then, the matched verification rules are synchronously bound to all abnormal regions within the same group, completing the rule matching for all abnormal regions in batches. At the same time, anomaly labels are marked during the matching process to provide risk references for subsequent verification. This method, employing label grouping and parallel matching logic across all groups, can significantly reduce repetitive rule retrieval and matching operations, fully utilize the system's parallel computing power to complete batch rule matching, shorten the rule matching time for large batches of abnormal regions, and is suitable for batch identification scenarios involving large numbers of pathological slides.

[0141] After reading the confidence level labels of all abnormal regions, the matching of the corresponding abnormal feature verification rules can be completed in the following way.

[0142] In one alternative approach, the system first retrieves a pre-defined fixed mapping relationship library of tags and rules, along with an anomaly feature verification rule library. This mapping relationship library pre-stores a one-to-one mapping relationship between each confidence level tag and its corresponding exclusive verification rule. Based on the confidence level tag corresponding to the anomaly region, the corresponding exclusive anomaly feature verification rule is directly retrieved from the rule library, bound to the corresponding anomaly region, and then output to the subsequent verification stage. This method relies on the pre-defined fixed mapping relationship to complete rule matching, is logically intuitive, has a fast calling speed, is less prone to matching errors, and is suitable for conventional standardized pathology number recognition scenarios.

[0143] In another alternative approach, the confidence level label, confidence score, preliminary identification result features, pathology number rule features of the affiliated hospital, and slide imaging quality features corresponding to the abnormal region are first extracted. Then, the rule generation logic optimized based on historical difficult case samples is retrieved from the rule base. Based on the above multi-dimensional features, a unique abnormal feature verification rule adapted to the abnormal region is dynamically generated. Simultaneously, the feasibility of the generated verification rule is verified, confirming that the verification dimensions and judgment criteria of the rule meet the processing requirements of the current system. After verification, it is bound to the corresponding abnormal region and output to the subsequent verification stage. This method can flexibly generate customized verification rules that fit the individual characteristics of each abnormal region, without relying on fixed preset mapping relationships. It has higher matching accuracy, can adapt to differentiated verification needs of different confidence intervals and different abnormal types, improves the accuracy of subsequent abnormal feature verification, and is suitable for complex pathology number recognition scenarios.

[0144] In an exemplary scheme for determining the matching of anomaly feature verification rules, the system's pre-set anomaly feature verification rule library is first loaded. This rule library contains two core rule categories: mid-confidence level exclusive verification rules and low-confidence level exclusive verification rules. The mid-confidence level exclusive verification rules are pre-configured with a two-dimensional verification standard: in-depth verification of pathology number rule compliance and basic verification of text region validity. The low-confidence level exclusive verification rules are also pre-configured with a two-dimensional verification standard: in-depth verification of text region validity and basic verification of pathology number rule compliance. A one-to-one mapping relationship between confidence level labels and verification rules is also pre-configured. Then, anomaly region grouping is performed. Based on the confidence level labels corresponding to all anomaly regions, the anomaly regions are divided into mid-confidence level anomaly groups and low-confidence level anomaly groups. Next, group rule matching is performed. For all anomaly regions within the mid-confidence level anomaly group, mid-confidence level exclusive anomaly feature verification rules are retrieved; for all anomaly regions within the low-confidence level anomaly group, low-confidence level exclusive anomaly feature verification rules are retrieved. Next, the rules are bound to regions. The matched specific verification rules are assigned to each abnormal region within the corresponding group, and the judgment criteria and verification process corresponding to the verification rule are loaded for each abnormal region. Then, rule matching validity is verified to check whether the verification rule bound to each abnormal region matches the confidence level label. Mismatched rules are removed, and the matching is re-completed. Finally, all abnormal regions with completed rule bindings and their corresponding verification rules are summarized to generate the final rule matching results, which are used for subsequent specific anomaly feature verification.

[0145] It should be noted that in some special cases, if the same abnormal area is bound to multiple confidence level labels at the same time, the corresponding verification rule will be matched based on the label with the highest priority, where the low confidence level label has a higher priority than the medium confidence level label.

[0146] Step C13: Based on the abnormal feature verification rules, perform a special verification on the compliance of the pathology number rules and the validity of the text region in the abnormal region, and generate the abnormal feature distribution results of the candidate region.

[0147] Pathology number rule compliance refers to the degree of matching between the preliminary identification result of the abnormal region and the preset standardized coding rules for pathology numbers. It is a core dimension for verifying the business validity of the identification result and is used to determine whether the identification result conforms to the coding specifications of pathology numbers. For example, compliance with coding format, character type, coding length, and hospital-specific pathology number coding rules. Text region validity refers to the degree of conformity of the original image of the abnormal region to whether it contains valid pathology number text content. It is a core dimension for verifying whether the candidate region is a valid text region and is used to determine whether the candidate region is a false positive or invalid background region. For example, text feature completeness, character edge clarity, text structure rationality, proportion of non-text background, and probability of valid text existence.

[0148] In this embodiment, the aforementioned specialized verification and anomaly feature distribution results can be generated in two ways. First, dynamic specialized verification based on dimension priority: Based on the confidence level label of the anomaly region, the priority and depth of the two verification dimensions are dynamically adjusted. Regions with medium confidence levels prioritize pathology number rule compliance verification, while regions with low confidence levels prioritize text region validity verification. The verification results are then aggregated to generate anomaly feature distribution results. This method can accurately adapt to anomaly features of different confidence levels, improving verification efficiency and accuracy, and adapting to differentiated anomaly scenarios. Second, multi-dimensional composite specialized verification: In addition to the two core dimensions of pathology number rule compliance and text region validity, two auxiliary verification dimensions—recognition engine output stability and character recognition integrity—are added. Multi-dimensional composite specialized verification is then performed, and the results of all dimensions are aggregated to generate anomaly feature distribution results. This method can extract more comprehensive anomaly features and adapt to the verification needs of complex composite anomaly scenarios.

[0149] After the pathology number recognition system completes the matching and binding of the abnormal feature verification rules for all abnormal areas, it initiates a two-dimensional special verification process to perform targeted verification of the compliance of the pathology number rules and the validity of the text area for each abnormal area.

[0150] For example, the implementation of specialized verification and abnormal feature distribution result generation includes dimensional splitting and full-region batch parallel verification. First, the verification rules bound to all abnormal regions are dimensionally split, extracting the verification standards for the pathology number rule compliance verification dimension and the text region validity verification dimension respectively. Simultaneously, all abnormal regions are categorized and aligned according to the two dimensions, and parallel specialized verification for both dimensions is initiated simultaneously. Within the pathology number rule compliance dimension, batch rule compliance verification is performed on the preliminary identification results of all abnormal regions, generating batch abnormal verification results for all abnormal regions in this dimension. Within the text region validity dimension, batch text validity verification is performed on the original images of all abnormal regions, generating batch abnormal verification results for all abnormal regions in this dimension. After all verifications in both dimensions are completed, the verification results for all abnormal regions in both dimensions are summarized and merged simultaneously, generating batch abnormal feature distribution results corresponding to all abnormal regions. At the same time, severely abnormal regions during the verification process are marked to provide risk reference for subsequent location analysis. This approach employs dimensional splitting and full-path parallel verification logic, which is executed synchronously with the multi-dimensional parallel processing flow. It can fully utilize the system's parallel computing power to complete the special verification of a large number of abnormal regions, improve the overall verification efficiency, and is suitable for batch identification scenarios of a large number of pathological slides.

[0151] After completing the preparation for the special verification of all abnormal areas, the special verification and abnormal feature distribution results can be generated in the following way.

[0152] In one alternative approach, specific anomaly feature verification rules, confidence level labels, preliminary identification result features, and original image quality features bound to the anomaly region are first extracted. Based on the verification rules, the verification depth and judgment criteria for both dimensions are dynamically adjusted. For regions with medium confidence, the verification depth for compliance with pathology number rules is increased; for regions with low confidence, the verification depth for the validity of the text region is increased. Then, based on the adjusted verification criteria, specialized verifications for both dimensions are performed. Simultaneously, an anomaly severity quantification score is added, recording the anomaly items, anomaly severity quantification values, and anomaly types for each dimension. These are then aggregated to generate a refined anomaly feature distribution result corresponding to the anomaly region. This method can dynamically adjust the verification depth based on the features of the anomaly region, without relying on a fixed verification process. It offers higher verification accuracy and can precisely extract differentiated anomaly features from different types of anomaly regions, providing a more refined basis for subsequent anomaly location and failure mode determination, and is adaptable to complex pathology number recognition scenarios.

[0153] In an exemplary scheme for generating abnormal feature distribution results, the exclusive abnormal feature verification rules bound to all abnormal regions are first loaded. Among them, the verification rules for the intermediate confidence abnormal regions stipulate that: the pathology number rule compliance verification is a deep verification, including four sub-items of verification: encoding format, character type, encoding length, and hospital-specific rules; and the text region validity verification is a basic verification, including one sub-item of verification: the probability of valid text existence. The verification rules for the low confidence abnormal regions stipulate that: the text region validity verification is a deep verification, including four sub-items of verification: text feature integrity, character edge clarity, text structure rationality, and non-text background ratio; and the pathology number rule compliance verification is a basic verification, including one sub-item of verification: the encoding format. Subsequently, multi-dimensional specific verifications are performed. For each intermediate-confidence anomaly region, a deep verification of the four sub-items of pathology number rule compliance is first performed, recording the verification results and anomalies for each sub-item. Then, a basic text region validity verification is performed, and the verification results are recorded. For each low-confidence anomaly region, a deep verification of the four sub-items of text region validity is first performed, recording the verification results and anomalies for each sub-item. Then, a basic pathology number rule compliance verification is performed, and the verification results are recorded. Next, anomaly feature summarization is performed. For each anomaly region, the verification results of all sub-items in both dimensions are summarized, and anomaly sub-items, anomaly degree, and anomaly type are marked. Sub-items that do not meet the verification criteria are marked as anomalies, multiple anomalies are marked as composite anomalies, and single anomalies are marked as single-dimensional anomalies. Then, anomaly feature standardization processing is performed, classifying all anomalies according to a preset standardization format to generate standardized anomaly feature distribution results for each anomaly region. Finally, the anomaly feature distribution results of all anomaly regions are summarized to generate the final full set of anomaly feature distribution results for subsequent anomaly localization.

[0154] It should be noted that in some special cases, if an abnormal area has no abnormal items after special verification and the abnormal feature distribution result is no abnormality, then the abnormal area is remarked as a qualified area, skipping the subsequent abnormal link location and failure mode determination process, and directly entering the result adoption stage.

[0155] Step C14: Based on the abnormal feature distribution results of the candidate region, locate the problematic link that causes the abnormality in the pathology number identification process, and generate the abnormality judgment result of the abnormal region.

[0156] In an exemplary scheme for generating anomaly determination results, a pre-set mapping library of anomaly features and problem stages is first loaded. This mapping library is pre-configured with: invalid text region anomalies corresponding to the initial screening stage of candidate regions, pathology number rule non-compliance anomalies corresponding to the high-precision text recognition stage, image quality anomalies corresponding to the image preprocessing stage, and multi-dimensional composite anomalies corresponding to multiple stages of the entire process. Then, anomaly feature analysis is performed, analyzing the anomaly items, anomaly types, anomaly severity, and anomaly dimensions for the anomaly feature distribution results of each anomaly region. Next, problem stage matching and localization are performed. For regions with single-dimensional anomalies, the corresponding core problem stage is directly matched based on the anomaly type. For regions with multi-dimensional composite anomalies, a full-process backtracking verification is performed, verifying the execution results of the four stages—initial screening, preprocessing, recognition, and evaluation—step by step to locate all problem stages causing the anomalies, while distinguishing between core and secondary problem stages. Finally, anomaly severity level determination is performed, classifying the anomaly determination results into three levels: mild, moderate, and severe anomalies based on the number and severity of anomalies. Next, the anomaly assessment results are standardized and encapsulated, generating a standardized anomaly assessment result for each anomaly region that includes the core problem steps, secondary problem steps, anomaly severity level, and a summary of anomaly characteristics. Finally, the anomaly assessment results for all anomaly regions are aggregated to generate a final full set of anomaly assessment results, which is used for subsequent failure mode determination.

[0157] It should be noted that in some special cases, if the abnormal area cannot be located as a single problematic link after backtracking and verification, and is a complex anomaly involving multiple links in the entire process, then all related problematic links will be marked in the anomaly judgment result, and it will also be marked as a severe complex anomaly, providing a comprehensive basis for subsequent failure mode determination.

[0158] Step C15: Based on the anomaly determination result, match the pathology number to identify the anomaly classification rules and determine the failure mode corresponding to the abnormal area.

[0159] The pathology number identification anomaly classification rules are pre-defined rules used to categorize anomaly judgment results into standardized failure modes. These rules are the core basis for achieving accurate failure mode classification, and pre-configure the mapping relationship between anomaly judgment results and standardized failure modes. Examples include single-stage anomaly classification rules, multi-stage composite anomaly classification rules, and anomaly severity grading classification rules. Failure modes are standardized pathology number identification anomaly type labels generated after matching with the anomaly classification rules. These labels accurately characterize the final type of identified anomaly and are the core basis for subsequent strategy matching. Examples include false detection (tissue texture misjudged as text), detected region but recognition failure, character recognition error, encoding format mismatch, and multi-stage composite recognition failure.

[0160] In this embodiment, the matching and determination of the aforementioned failure modes can be accomplished in two ways. First, multi-dimensional composite classification matching: based on the core problem links, secondary problem links, severity level of the anomaly, and the distribution results of anomaly features in the anomaly judgment results, multi-dimensional composite classification judgment is performed to match the corresponding composite failure modes. This method can achieve more refined failure mode classification and adapt to complex composite anomaly scenarios. Second, dynamic classification rule matching: based on the features of the anomaly judgment results and the current hospital's pathology number quality control rules, suitable anomaly classification rules are dynamically generated. After rule verification, the corresponding failure mode is matched and determined. This method can flexibly adapt to the differentiated classification requirements of different hospitals and is suitable for pathology identification scenarios involving multiple hospital campuses.

[0161] After the pathology number recognition system generates the anomaly determination results for all abnormal areas, it loads the system's preset pathology number recognition anomaly classification rule library and starts the failure mode matching and determination process.

[0162] For example, the implementation of failure mode matching and determination includes batch parallel matching of all results from anomaly type clustering. First, clustering is performed on the anomaly judgment results of all anomaly regions, grouping anomaly regions with the same core problem steps and the same anomaly type into the same cluster group. Simultaneously, corresponding anomaly classification rules are matched for each cluster group. Then, parallel matching is initiated simultaneously for all cluster groups. Within each cluster group, based on the unified anomaly judgment results and classification rules within the group, failure mode matching and determination of all anomaly regions within the group are completed in batches. Simultaneously, a corresponding standardized failure mode label is bound to each anomaly region. After all parallel matching of all cluster groups is completed, the failure mode labels of all anomaly regions are summarized, and special failure modes of composite anomalies are marked to provide risk reference for subsequent strategy matching. This approach, employing anomaly type clustering and full-group parallel matching logic, reduces repetitive rule matching operations, fully utilizes the system's parallel computing power to complete the failure mode classification of a large number of anomaly regions, significantly shortens the overall time consumption of the classification process, and is suitable for batch identification scenarios of large batches of pathological slides.

[0163] After reading the anomaly determination results for all abnormal regions, the failure mode can be matched and determined in the following way.

[0164] In one alternative approach, the system first retrieves a pre-defined rule base for classifying abnormal pathology numbers. This rule base stores a fixed mapping relationship between abnormal judgment results and standardized failure modes. Based on the core problem steps and abnormal types in the abnormal judgment results corresponding to the abnormal regions, the system directly matches the corresponding standardized failure modes from the rule base, binds them to the corresponding abnormal regions, and outputs them to the subsequent execution strategy matching stage. This method relies on pre-defined fixed classification rules to complete the matching, is logically intuitive, has a fast matching speed, and provides a unified classification standard, making it suitable for conventional standardized pathology number recognition scenarios.

[0165] In another alternative approach, the anomaly judgment results, anomaly feature distribution results, pathology number quality control rules of the affiliated hospital, and preliminary identification result features corresponding to the abnormal region are first extracted. Then, multi-dimensional composite classification rules optimized based on historical difficult case samples are retrieved from the rule base. Based on the above multi-dimensional features, multi-dimensional composite classification judgment is performed. First, the core failure mode category is determined, and then the subdivided failure modes are determined based on secondary anomaly features. At the same time, the matching degree between the determined failure modes and the anomaly features is verified. After verification, they are bound to the corresponding abnormal region and output to the subsequent execution strategy matching stage. This method adopts the logic of multi-dimensional composite classification, which does not rely on a fixed one-to-one mapping relationship. It has higher classification accuracy and can accurately adapt to the differentiated features of complex composite anomalies, generating more refined standardized failure modes. This provides a more reliable basis for the accurate matching of subsequent execution strategies and is suitable for complex pathology number identification scenarios.

[0166] Fifth Embodiment

[0167] This embodiment provides an exemplary scheme for accurate matching and optimal strategy locking of pathology number identification execution strategies. In this example, based on a preset strategy chain library and the triggering conditions corresponding to each execution strategy, a combined conditional verification is performed on the confidence score of the preliminary identification result, the association failure mode, and the pathology business attribute features to screen out all suitable candidate strategies. Then, according to the preset strategy priority rules and pathology business process requirements, the candidate strategies are sorted by hierarchical priority to generate an ordered strategy sequence. Finally, the optimal candidate strategy ranked first in the strategy sequence is determined as the target execution strategy corresponding to the preliminary identification result. Step S30 includes steps D11~D13:

[0168] Step D11: Based on the strategy chain library and the triggering conditions corresponding to the execution strategy, perform a combined conditional verification on the confidence score of the preliminary identification result to determine at least one candidate strategy.

[0169] Step D12: Based on the strategy priority and pathological business requirements, prioritize the candidate strategies to obtain the strategy sequence of the preliminary identification results.

[0170] Step D13: Select the candidate strategy ranked first in the strategy sequence as the execution strategy corresponding to the preliminary identification result.

[0171] Triggering conditions are multi-dimensional combination rules pre-configured on each execution strategy to determine whether the strategy is suitable for the current recognition result. Only when the recognition result fully meets all triggering conditions can it be matched as a candidate strategy. For example, confidence level threshold conditions, failure mode matching conditions, pathology number rule compliance conditions, text region validity conditions, and business scenario adaptation conditions.

[0172] Combined conditional validation is a multi-condition synchronous validation operation performed on the multi-dimensional features of the recognition result based on the trigger conditions of the policy chain library. It is used to filter out all execution policies that are suitable for the current recognition result and is the core operation of candidate policy selection. Examples include dual-dimensional combined validation of confidence and failure mode, composite validation of confidence and business scenario, and multi-condition synchronous validation across all dimensions.

[0173] Strategy priority rules are pre-defined core rules used to prioritize candidate strategies. These rules clearly define the priority weights, ranking dimensions, and judgment criteria for different strategies, distinguishing the suitability of different candidate strategies. Examples include automation priority rules, accuracy priority rules, processing efficiency priority rules, and business compliance priority rules.

[0174] In this example, when performing combined condition checks to determine suitable candidate strategies, the matching method of confidence level and failure mode provided in the fourth embodiment can be used as a reference. Alternatively, a multi-dimensional feature synchronous check can be used, simultaneously reading the confidence score, associated failure mode, and pathological business attribute features corresponding to the preliminary identification result, while simultaneously disassembling the trigger conditions of each execution strategy in the strategy chain library. A full match check is then performed on each of the multi-dimensional features and trigger conditions, accumulating and filtering out all candidate strategies that meet the conditions, thus completing the full screening of candidate strategies. After the candidate strategy screening is completed, a priority sorting process is initiated. Based on preset strategy priority rules and the pathological business process requirements of the current task, all candidate strategies are graded and prioritized, generating an ordered strategy sequence arranged from high to low suitability. Finally, the candidate strategy ranked first in the strategy sequence is identified as the target execution strategy corresponding to the preliminary identification result. This multi-dimensional combined check and the graded sorting logic for business suitability improve the matching accuracy between the execution strategy and the identification result, avoiding strategy mismatch caused by single threshold matching, and providing accurate and compliant execution basis for subsequent strategy execution.

[0175] For example, there are three ways to implement strategy matching and optimal strategy locking. These three methods are logically independent, have significantly different adaptation scenarios, and can cover the strategy matching needs of all business scenarios:

[0176] The first method is a two-dimensional lockstep progressive matching method based on confidence level and failure mode. After the strategy matching process officially starts, a first-level lockstep verification is performed. This method compares the confidence level of the initial identification results with the confidence trigger thresholds of each execution strategy in the strategy chain library, locking all execution strategies that meet the confidence level conditions and eliminating all strategies that do not match the confidence level conditions, thus completing the first-level candidate pool selection. A second-level lockstep verification is then performed. This method precisely compares the failure modes associated with the initial identification results with the failure mode trigger conditions of each strategy in the first-level candidate pool, further locking all execution strategies that simultaneously meet both the confidence level and failure mode conditions, eliminating all strategies that do not match the failure mode conditions, thus completing the second-level candidate pool selection. Finally, all strategies in the second-level candidate pool are the suitable candidate strategies. After the candidate strategy selection is completed, the candidate strategies are sorted based on preset fixed priority rules to generate a strategy sequence, and the first strategy in the sequence is locked as the target execution strategy. This method employs a two-level lockstep progressive verification logic, which gradually narrows the policy matching range through two levels of condition locking. Each step of the verification is irreversible, completely avoiding cross-dimensional and cross-level policy mismatch problems. The matching logic is rigorous and stable, and the accuracy of the fit verification is extremely high, making it suitable for the recognition scenario of standard pathology slide archiving in hospitals.

[0177] The second method is the dynamic weight adaptation and matching method based on business priority. After the strategy matching process is officially launched, the business type tags of the current identification task are first read, including four core tags: emergency archiving, routine quality control, scientific research data archiving, and batch historical slice processing. Based on the business type tags, corresponding dynamic weight configuration rules are matched. Specifically, emergency archiving tasks are matched with weight rules that prioritize processing efficiency, routine quality control tasks with weight rules that prioritize accuracy, scientific research archiving tasks with weight rules that prioritize compliance, and batch processing tasks with weight rules that prioritize automation. Then, a combined condition verification is performed, comparing the confidence score and failure mode of the preliminary identification results with the trigger conditions of each execution strategy in the strategy chain library to filter out all candidate strategies that meet the basic conditions. Subsequently, based on the matched dynamic weight configuration rules, dynamic weights are assigned to multiple adaptation dimensions for each candidate strategy, and the comprehensive adaptation score of each candidate strategy is calculated. The candidate strategies are sorted from high to low according to the comprehensive adaptation score, generating a strategy sequence that dynamically adapts to business needs, and locking the first strategy in the sequence as the target execution strategy. This method adopts a business-driven dynamic weight adaptation logic, which breaks the limitations of fixed priority sorting. It can flexibly adjust the strategy sorting logic according to the core needs of different business scenarios, and achieves deep adaptation between the execution strategy and business scenarios. It can simultaneously balance recognition accuracy, processing efficiency and business compliance, and adapt to the complex pathology number recognition scenario of multiple business lines in hospitals.

[0178] The third method is the difficult example sample benchmarking adaptive matching method. After the strategy matching process is officially started, a basic combined condition check is first performed to screen out candidate strategies that meet the basic conditions of confidence and failure mode, and an initial candidate strategy pool is constructed. Then, the confidence score, failure mode, pathology number coding features, and candidate region image quality features of the current preliminary identification results are extracted to generate a multi-dimensional feature vector of the current sample. Based on this feature vector, a preset number of historical similar difficult example samples are retrieved from the difficult example library. Similar difficult example samples are historical samples whose feature similarity with the current sample is higher than a preset threshold and have completed final processing and effect verification. Subsequently, the final processing success rate, accuracy, and processing time of each candidate strategy in all similar difficult example samples are statistically analyzed. Taking the best historical processing effect as the core standard, the candidate strategies in the initial candidate pool are prioritized and scored. The candidate strategies are sorted from high to low according to the priority of the best historical effect, generating an adaptive optimization strategy sequence. The first strategy in the sequence is locked as the target execution strategy. At the same time, the feature data of this strategy matching is synchronized to the difficult example library to complete the continuous iterative optimization of the rules. This method employs an adaptive benchmarking logic driven by historical difficult case data. It does not rely on fixed preset rules and can continuously optimize the policy matching accuracy based on historical successful cases. It becomes more accurate with use and can effectively solve the policy mismatch problem in rare and abnormal scenarios and low-quality slide scenarios. It significantly improves the policy matching accuracy of difficult samples and is suitable for complex recognition scenarios containing a large number of low-quality and specially coded pathological slides.

[0179] In an exemplary scheme for determining the target execution strategy, the system first loads the pathology number identification strategy chain library, which contains five core execution strategies: direct discard strategy (strategy 0), direct adoption strategy (strategy 1), rule correction strategy (strategy 2), secondary identification strategy (strategy 3), and human-computer interaction correction strategy. Each strategy is pre-configured with a dual-dimensional trigger condition matrix of confidence level and failure mode, priority weight, and hospital pathology business adaptation rules. Then, basic features are read: A single digital pathology slide scan image is input, summarizing the preliminary identification results, confidence levels, and failure modes of the three candidate regions R1, R2, and R3 identified in the initial screening. Simultaneously, the hospital's pathology number coding rules (letter + year + serial number + optional tissue block number) and routine archiving business tags are loaded. Specifically, R1 result is "A2023-001", confidence level 0.92, no identification anomalies; R2 result is "Gland", confidence level 0.15, failure mode is "false detection (tissue texture misjudged as text)"; R3 result is "fuzzy and unrecognizable", confidence level 0.05, failure mode is "region detected but recognition failed". Next, a combined conditional validation is performed, comparing the confidence level and failure mode of each region with the strategy triggering conditions item by item, constructing an initial candidate strategy pool for each region: R1 matches the direct adoption strategy, R2 matches the direct discard strategy, and R3 matches the secondary identification strategy and the human-computer interaction fallback strategy. Subsequently, based on the business rules prioritizing automation and accuracy, candidate strategies are prioritized to generate an ordered strategy sequence. The compliance of the execution conditions of the first strategy in the sequence is then verified, identifying R1 as the direct adoption strategy, R2 as the direct discard strategy, and R3 as the secondary identification strategy. For the "05" result (confidence 0.70, median confidence, not conforming to the complete pathology number rule) output by the dedicated difficult case engine for R3, a re-verification and sorting process is performed, locking the rule correction strategy as the target execution strategy. Finally, the strategy matching results of the entire slice are summarized to generate a complete strategy matching set for subsequent strategy execution, result output, and closed-loop archiving of difficult case samples.

[0180] Sixth Embodiment

[0181] This embodiment provides an exemplary scheme for targeted execution of a pathology number recognition strategy and accurate generation of target results. In this example, the core elements of the execution strategy corresponding to the candidate region are first analyzed to determine the matching execution action, verification rules, and execution parameters. Then, based on the execution elements, a specific targeted processing operation is performed on the candidate region to obtain the initial strategy output result. Subsequently, the confidence level and format validity of the result are verified through preset verification rules to generate a standardized and valid recognition result. Finally, based on the spatial distribution order of the candidate region and the pathology number encoding logic, all valid recognition results are merged to obtain the target recognition result of the pathology number corresponding to the pathology slide image. Step S40 includes steps E11~E14:

[0182] Step E11: Based on the execution strategy corresponding to the candidate region, parse the preliminary recognition results, associated image data and recognition process metadata of the candidate region to determine the execution action, verification rules and execution parameters corresponding to the execution strategy.

[0183] Step E12: Based on the execution action and the execution parameters, perform a targeted processing operation to match the candidate region, and obtain the initial strategy output result of the candidate region.

[0184] Step E13: According to the verification rules, perform confidence verification and format validity verification on the initial strategy output results of the candidate regions to generate valid recognition results for the candidate regions.

[0185] Step E14: Based on the regional order of the candidate regions, merge the valid identification results to obtain the target identification result of the pathology number.

[0186] An execution action is the core processing operation specified in the execution strategy for the current candidate region. It is the core execution unit of the strategy, and each execution strategy corresponds to a unique core execution action. Examples include directly adopting the identification result, directly discarding invalid results, secondary identification by a dedicated difficult case engine, error correction processing based on pathology number rules, and guided manual correction result entry.

[0187] Targeted processing operations are specific processing operations performed on the features of the current candidate region based on the execution actions and parameters. They are the core of the execution strategy and are designed to perfectly match the recognition results and anomaly types of the current candidate region. Examples include direct encapsulation of high-confidence results, removal and marking of false detection results, secondary recognition of ambiguous regions, and rule-based reasoning and error correction for short character results.

[0188] In this example, when parsing the execution actions, verification rules, and execution parameters corresponding to the execution strategy, the method of reading the strategy matching results provided in the fifth embodiment can be referred to. Alternatively, a streaming incremental parsing method can be used, receiving the execution strategy, preliminary recognition results, associated image data, and recognition process metadata of each candidate region while simultaneously parsing the core elements of the corresponding execution strategy. The extraction and adaptation of execution actions, verification rules, and execution parameters are completed region by region. After all candidate regions are parsed, the strategy execution element set of the entire slice is obtained, thus completing the preparatory work for strategy execution. After completing the parsing of the strategy execution elements of all candidate regions, the targeted processing flow is started. Based on the execution actions and execution parameters corresponding to each candidate region, the matching targeted processing operation is performed to generate the initial strategy output result for each candidate region. Subsequently, according to the corresponding verification rules, the confidence level and format validity of each initial strategy output result are checked, and valid recognition results that pass the verification are selected. Finally, based on the spatial region order of the candidate regions and the pathology number encoding logic, all valid recognition results are structurally merged to generate the pathology number target recognition result corresponding to the pathology slice image.

[0189] For example, there are two ways to implement the execution strategy and generate target recognition results:

[0190] The first method is a multi-region hierarchical pipeline linkage execution method. After the strategy execution process is officially launched, all candidate regions are first classified into three levels according to the processing time and priority of the execution strategy: immediate processing level, routine processing level, and fallback processing level. Direct adoption and direct discard strategies are classified as immediate processing level, secondary identification and rule correction strategies as routine processing level, and human-computer interaction correction strategies as fallback processing level. Then, a three-level pipeline execution channel is established, and the processing processes of the three levels are launched in parallel. The immediate processing level channel simultaneously completes the strategy execution and valid result generation for all high-priority candidate regions; the routine processing level channel simultaneously completes the targeted processing and verification for all medium-priority candidate regions; and the fallback processing level channel simultaneously completes the abnormal result processing for low-priority candidate regions. During pipeline execution, valid identification results generated by previous channels can be synchronized to subsequent channels, providing prior evidence for rule correction and logical reasoning in subsequent channels. After all candidate regions in the three channels have been processed, all valid identification results are summarized and merged according to spatial order and encoding logic to generate the final pathology number target identification result. This method employs a hierarchical pipeline-linked execution logic, breaking through the time bottleneck of traditional serial execution. It maximizes the utilization of system computing power through hierarchical parallel processing, while realizing the cross-channel linkage reuse of prior results. This shortens the total execution time of the full slice strategy and ensures the accuracy of the results, making it suitable for high-throughput recognition scenarios involving large-volume pathology slice archiving.

[0191] The second method is a closed-loop execution approach involving the linkage of primary and secondary pathology numbers. After the strategy execution process officially starts, the execution strategies for all candidate regions are first categorized into primary pathology number processing strategies and secondary number processing strategies. The primary pathology number processing process is initiated first, executing the entire strategy for candidate regions matching the direct adoption or secondary identification strategies. This generates valid primary pathology number identification results that have passed verification, locking in the core primary number of the pathology slide. Subsequently, using the confirmed primary pathology number result as a priori benchmark, the secondary number processing process is initiated. For candidate regions with short characters and intermediate confidence levels from the matching rule correction and secondary identification strategies, targeted processing operations and rule reasoning verification are performed, combined with the hospital coding rules corresponding to the primary pathology number. For example, based on the coding rules of primary pathology number A2023-001, the short character 05 is inferred to be a compliant tissue block subsidiary number, generating the initial strategy output result for the secondary number. Then, based on the verification rules of the primary pathology number, confidence and format validity checks are performed on the secondary number results. After passing the checks, valid identification results for the secondary number are generated. Finally, the main pathology number result and the auxiliary number result are structurally merged to generate a complete target recognition result with the main number and the auxiliary block number. At the same time, the sample data of this linkage reasoning is stored in the hard case library to complete the continuous optimization of the rules. This method adopts the closed-loop execution logic of linkage reasoning between main and auxiliary pathology numbers, which breaks through the limitations of traditional single-region independent processing. It realizes accurate reasoning and error correction of auxiliary numbers through the prior benchmark of the main pathology number, solves the industry pain points of low recognition accuracy of short tissue block numbers in pathological slides and difficulty in rule verification, improves the recognition accuracy of complex pathology number scenarios, and is adapted to complex pathological slide recognition scenarios with multiple tissue blocks and multiple auxiliary numbers.

[0192] In an exemplary scheme for determining the target identification result of pathology numbers, the closed-loop resilient architecture for pathology number identification, evaluation, decision-making, recovery, and learning, as described in this application, is used to execute the entire process through a multi-level heterogeneous collaborative mechanism consisting of a high-speed initial screening engine, a high-precision main identification engine, and a dedicated difficult case engine. First, a pre-configured extensible strategy chain library is loaded. This library contains five core execution strategies: direct discard (strategy 0), direct adoption (strategy 1), rule correction (strategy 2), secondary identification (strategy 3), and guided human-computer interaction (strategy 4). Each strategy is pre-configured with confidence levels and two-dimensional trigger conditions for failure modes, execution actions, verification rules, and parameters adapted to hospital pathology business. Subsequently, the entire process of basic data is read: A high-speed initial screening engine locates three text candidate regions (R1, R2, and R3) on a single digital pathology slide. These regions are then identified by a high-precision main engine to obtain preliminary recognition results. A multi-dimensional unified evaluation system is then used to calculate the comprehensive confidence score of each region and automatically classify failure modes. Specifically, R1 is "A2023-001" with a comprehensive confidence score of 0.92 and no recognition anomalies; R2 is "Gland" with a comprehensive confidence score of 0.15 and a failure mode of "false detection (tissue texture misjudged as text)"; and R3 is "fuzzy and unrecognizable" with a comprehensive confidence score of 0.05 and a failure mode of "region detected but recognition failed." Next, based on the two-dimensional criteria of "confidence level and failure mode," corresponding execution strategies are matched for each candidate region: R1 is matched with strategy 1, R2 with strategy 0, and R3 with strategy 3. Subsequently, targeted strategy processing and compliance verification were performed: For R1, the parsed action was to directly adopt the result, with the verification rule being a confidence level ≥ 0.90 and compliance with the hospital's pathology number coding rule of "letter + 4-digit year + hyphen + 3-digit serial number". After passing the confidence level review and format validity verification, a valid recognition result "A2023-001" was generated; For R2, the parsed action was to directly discard the invalid result, and after passing the false positive validity verification, an invalid region marker was generated; For R3, the parsed action was to call the special difficult example index for fuzzy text. The engine performs secondary recognition and outputs the initial result "05". Combined with the confirmed primary pathology number prior rules, Strategy 2 is triggered to perform error correction processing based on the pathology number coding rules. After confidence verification and format validation, it is confirmed that it is a compliant tissue block number, and a valid recognition result "Block Number: 05" is generated. If the dedicated engine's secondary recognition still cannot obtain a valid result, Strategy 4's non-blocking guided human-computer interaction is triggered. The uncertain area of ​​the target is highlighted to guide the user to perform the minimum necessary operations of clicking to confirm, dragging to select, or manually entering, while providing a skip option to ensure that the main process is not blocked.Finally, based on the spatial order of candidate regions and the pathology number encoding logic, all valid recognition results are merged to generate the final pathology number target recognition result of the pathology slide, "A2023-001, block number: 05". At the same time, the low confidence samples, processing logs and manually corrected samples in the R2 and R3 regions are automatically labeled and stored in the difficult sample library, which is used regularly for incremental training and fine-tuning of the recognition engines at all levels.

[0193] It should be noted that in some special cases, if only one valid identification result exists after verification and there are no other valid results with subordinate numbers, the result merging process is skipped, and the valid identification result is directly used as the target identification result of the pathology number corresponding to the pathology slide image.

[0194] Seventh Embodiment

[0195] This embodiment provides an exemplary scheme for the generation of pathology number primary and secondary number linkage reasoning and verification identification results. In this example, based on the verified high-confidence primary pathology number target identification result and the auxiliary coding rules of the pathology number on the same slide, business attribute reasoning matching is performed on the short character identification result with medium confidence to determine its corresponding tissue block number attribute. Then, based on the consistency rules of pathology information on the same slide, the association compliance verification of primary and secondary numbers is completed to generate standardized pathology number results. Finally, the verification identification result is output, and the full data of this reasoning process is archived and stored as difficult case data, realizing accurate identification of pathology number auxiliary numbers and closed-loop optimization of the identification process. After step S40, steps F11~F13 are also included:

[0196] Step F11: Based on the high-confidence target recognition result and the auxiliary coding rules of the same slide pathology number, perform business attribute reasoning and matching on the short character recognition result with medium confidence to determine the tissue block number attribute corresponding to the short character recognition result.

[0197] Step F12: According to the consistency rules of pathological information on the same slide, verify the compliance of the association between the high-confidence target identification result and the tissue block number attribute to obtain the pathology number result.

[0198] Step F13: Based on the pathology number result, determine the verification and recognition result of the pathology slide image, and archive and store the sample data and rule execution data corresponding to this reasoning and matching process as difficult case data.

[0199] The auxiliary coding rules for pathology numbers on the same slide are pre-defined by the hospital, binding the master pathology number and the auxiliary tissue block number on the same pathology slide. These rules standardize the naming of multiple tissue blocks on the same slide and are the core rule basis for inferring short-character business attributes. Examples include: tissue block numbering rules with a hyphen followed by a two-digit Arabic numeral after the master pathology number; tissue block numbering rules with numbers in parentheses following the master pathology number; and rules allowing only one set of master pathology numbers to correspond to multiple sets of two-digit tissue block numbers on the same slide.

[0200] The short character recognition result with intermediate confidence is the text recognition result output by the recognition engine. It is a short character with a confidence level in the preset intermediate range and does not conform to the encoding rules of the complete pathology number. It is the processing object of this business attribute reasoning.

[0201] The tissue block number attribute is a business attribute determined by reasoning and matching the short character recognition result. It is used to identify the short character as the sequential number of the pathological tissue block corresponding to the master pathology number on the same slide, and is a core component of the structured pathology number result. For example, the "05" tissue block number attribute corresponding to the master pathology number "A2023-001" and the sequential number attribute of the attached tissue blocks on the same slide.

[0202] The consistency rules for pathology information on the same slide are pre-defined business consistency verification rules that all pathology number-related information on the same pathology slide must meet. These rules ensure the business relevance and compliance of the master pathology number and the associated tissue block numbers, and are the core basis for verifying pathology number results. Examples include rules such as allowing only one valid master pathology number on the same slide, ensuring the tissue block number matches the hospital coding rule of the master pathology number, and prohibiting duplicate tissue block numbers on the same slide.

[0203] Difficult example data consists of all sample data and execution data generated during this inference matching process, used for subsequent model optimization and rule iteration. It is the core data source of the closed-loop learning mechanism. Examples include the original images of candidate regions corresponding to short characters with intermediate confidence, rule execution logs of the inference matching process, process data of primary and secondary number association verification, and comparison data after manual correction.

[0204] In this example, when performing business attribute reasoning and matching on the short character recognition results with intermediate confidence to determine the tissue block number attribute, the primary and secondary pathology number linkage reasoning method provided in the sixth embodiment can be used. Alternatively, a linkage reasoning method based on the same batch of slide rules can be used. This involves simultaneously reading the primary pathology number and tissue block number rules of the pathology slides already identified in the same batch, and combining the high-confidence target recognition results and auxiliary coding rules of the current slide to perform batch linkage reasoning and matching on the short character recognition results. This process accumulates and completes the attribute determination for all short character results, thereby determining the tissue block number attribute.

[0205] After determining the tissue block number attribute, the associated compliance verification process is initiated. Based on the consistency rules of pathological information on the same slide, the business relevance, coding compliance, and information consistency of the high-confidence target identification result and the tissue block number attribute are verified item by item. After the verification is passed, a standardized pathology number result integrating the primary and secondary numbers is generated. Subsequently, the verification and identification result of the pathology slide image is generated based on the pathology number result. At the same time, the sample data and rule execution data corresponding to this reasoning and matching process are archived and stored in the difficult case sample library as difficult case data. In this way, through the full-process logic of business reasoning anchored by the primary pathology number, compliance verification, and closed-loop archiving, the industry pain points of low recognition accuracy of short numbers of tissue blocks on the same slide and the inability to automatically determine business attributes are solved, thereby improving the completeness and business compliance of the pathology number identification result.

[0206] For example, there are three ways to implement the linkage reasoning and review result generation of pathology number primary and secondary numbering. These three methods are logically independent, adapt to significantly different scenarios, and can cover the reasoning needs of all business scenarios:

[0207] The first method, a single-step closed-loop reasoning method uniquely anchored by the primary pathology number, first locks the verified, high-confidence primary pathology number target identification result as the sole anchoring benchmark for this reasoning after the reasoning process officially starts. It then retrieves the auxiliary coding rules for pathology numbers on the same slide from the hospital to which the primary pathology number belongs, clarifying the coding format, length range, and character type requirements for tissue block numbers. Next, it matches the short character identification result with the intermediate-confidence level with the auxiliary coding rules item by item to determine if it conforms to the coding specifications of tissue block numbers. If it fully conforms, it directly determines the tissue block number attribute corresponding to the short character, completing the single-step reasoning matching. Subsequently, based on the consistency rules of pathology information on the same slide, it performs a compliance check on the association between primary and auxiliary numbers. After the check passes, it generates the pathology number result and the verification identification result. This method uses a single-step closed-loop reasoning logic uniquely anchored by the primary number, resulting in a short reasoning chain, high execution efficiency, and rigorous rule matching. It completely avoids reasoning bias caused by multiple anchor points, is suitable for the identification scenario of routine standardized pathology slides in hospitals, and is the preferred execution solution for routine clinical scenarios.

[0208] The second method is a fuzzy adaptation reasoning method based on multi-dimensional feature linkage. After the reasoning process officially starts, feature data from four core dimensions are extracted: the coding rule features of the high-confidence master pathology number, the association features of the master and auxiliary numbers of the pathology slides identified in the same batch, the image spatial location features of the short character recognition results, and the image quality features of the candidate regions corresponding to the short characters. Then, a multi-dimensional weighted reasoning model is constructed, and adaptation weight coefficients are assigned to the four dimensions of features, with the master pathology number coding rule features having the highest weight, and the other dimensions serving as auxiliary correction criteria. Subsequently, based on the weighted reasoning model, comprehensive reasoning matching is performed on the short character recognition results of the intermediate confidence level. Even if there are slight character recognition errors or partial compliance with coding rules for the short characters, fuzzy adaptation can be completed through multi-dimensional feature linkage to accurately determine its tissue block number attribute. Finally, based on multi-dimensional consistency rules, the compliance verification of the master and auxiliary numbers is completed, generating the pathology number result and the review recognition result. This method employs a fuzzy adaptation reasoning logic based on multi-dimensional feature linkage, breaking through the limitations of traditional fixed rule matching. It can effectively solve the problem of being unable to determine attributes due to slight fuzziness of short characters and recognition errors of some characters, significantly improving the recognition accuracy of tissue block numbers in edge scenarios and adapting to low-quality pathological slide recognition scenarios with slight defects.

[0209] The third method is the difficult example library-based adaptive inference method. After the inference process officially starts, it first locks the high-confidence primary pathology number target recognition result and the corresponding hospital affiliated coding rule, and at the same time extracts the full-dimensional features of the mid-confidence short character recognition result to generate the feature vector of the current sample. Then, based on this feature vector, it retrieves a preset number of source historical inference samples from the difficult example library. Source historical samples are those from the same hospital, with the same coding rule, the same type of short characters, and have completed final verification and effect validation. Subsequently, it statistically analyzes the inference rules, matching thresholds, and final verification results of all source historical samples. Taking the highest historical processing success rate as the core standard, it adaptively generates inference matching rules and thresholds suitable for the current sample. Based on the adaptively generated rules, it completes the business attribute inference matching of short characters to determine their tissue block number attribute. Then, it completes compliance verification to generate pathology number results and review recognition results. At the same time, it synchronizes the sample data and rule execution data of this inference to the difficult example library to complete the continuous iterative optimization of the inference rules. This method employs adaptive reasoning logic driven by historical data from a difficult case library, which can effectively solve the problem of tissue block numbering reasoning under special hospitals and special coding rules. It is suitable for pathological slide recognition scenarios in specialized hospitals and research hospitals with personalized coding rules.

[0210] In an exemplary scheme for determining the verification and identification results of a pathological slide, the high-confidence target identification result "A2023-001" of the pathological slide generated by the sixth embodiment is first loaded. Simultaneously, the hospital's preset auxiliary coding rules for pathology numbers on the same slide are loaded: only one master pathology number is allowed per slide, and the tissue block number is a two-digit Arabic numeral, subordinate to the master pathology number. Then, the intermediate-confidence short character identification result "05" generated by the secondary identification of the slide is read. Its overall confidence level is 0.70, falling within the preset intermediate-confidence range. Subsequently, business attribute reasoning and matching are performed. Based on the master pathology number "A2023-001" and the auxiliary coding rules, it is determined that the short character "05" fully conforms to the coding specifications of the tissue block number, thus determining its corresponding tissue block number attribute. Next, a compliance check is performed. Based on the consistency rule of pathological information on the same slide, the check confirms that the same slide has only one master pathology number, the tissue block number conforms to the hospital's coding rules, and there are no duplicate numbers. After the check passes, a structured pathology number result is generated: master pathology number "A2023-001" and tissue block number "05". Subsequently, based on this pathology number result, a standardized verification and recognition result of the pathology slide image is generated: master pathology number: A2023-001, tissue block number: 05. Finally, the original image of the candidate region corresponding to the short character "05", the rule execution log, the inference matching process data, and the verification process data are used as difficult case data during this inference matching process. They are automatically labeled and archived in the difficult case sample library for subsequent incremental training of the recognition engine and iterative optimization of the inference rules.

[0211] Eighth embodiment

[0212] This embodiment provides an exemplary scheme for closed-loop iterative optimization driven by difficult cases in a pathology number recognition system. In this example, the difficult case data archived in the difficult case library is first filtered, cleaned, and standardized to generate a difficult case training dataset adapted to the training requirements of each level of the recognition engine. Then, the dataset is split based on engine parameter adaptation rules, and targeted incremental training is performed on the high-speed initial screening engine, the high-precision main recognition engine, and the dedicated difficult case engine to obtain optimized engine models. Next, combined with the strategy execution effect data corresponding to the difficult cases, the strategy triggering threshold, priority weight, and execution parameters of the strategy chain library are adjusted to generate linked execution rules. Finally, the model and rules of the recognition system are fully updated, achieving continuous self-evolution of the pathology number recognition system. Please refer to... Figure 4 , Figure 4 This is a flowchart illustrating the eighth embodiment of the method for identifying pathology slide pathology numbers in this application. Following step F13, steps G11-G14 are also included:

[0213] Step G11: Filter the difficult example data in the difficult example library, clean and remove duplicates and standardize the annotations to obtain a difficult example training dataset that meets the training requirements of the model.

[0214] Step G12: Split the difficult example training dataset according to the engine parameter rules, and perform targeted incremental training on the high-speed initial screening engine, the high-precision main recognition engine, and the dedicated difficult example engine respectively to generate an optimized recognition engine model.

[0215] Step G13: Based on the optimized recognition engine model and combined with the strategy execution effect data corresponding to the difficult examples, the trigger threshold, priority weight, and execution parameters of the execution strategies in the strategy chain library are adjusted for adaptability to obtain the linkage execution rules.

[0216] Step G14: Based on the optimized recognition engine model and the linkage execution rules, update the model and rules of the pathology number recognition process to achieve iterative optimization of the pathology number recognition system.

[0217] In this example, when selecting difficult example data to prepare the difficult example training dataset, the difficult example data archiving method provided in the seventh embodiment can be referred to. Alternatively, a streaming incremental dataset preparation method can be used, where newly archived difficult example data is incrementally read from the difficult example library while simultaneously performing cleaning, deduplication, standardization annotation, and format conversion, accumulating to generate a difficult example training dataset that meets the training requirements, thereby completing the entire process of training dataset preparation.

[0218] After preparing the difficult example training dataset, the engine-targeted incremental training process is initiated. Following the parameter adaptation rules for each engine level, the difficult example training dataset is split and matched to the high-speed initial screening engine, the high-precision main recognition engine, and the dedicated difficult example engine. Targeted incremental fine-tuning training is performed on the core weaknesses of each engine to generate optimized recognition engine models at each level. Subsequently, combining the strategy execution effect data corresponding to the difficult example data, the trigger thresholds, priority weights, and execution parameters of each execution strategy in the strategy chain library are optimized for adaptability, generating linkage execution rules that match the engine and strategy. Finally, based on the optimized model and linkage execution rules, a full update of the production environment recognition process is completed. This difficult example-driven model and strategy joint optimization mechanism achieves continuous closed-loop iterative optimization of the pathology number recognition system, increasing accuracy with use and reducing reliance on manual intervention in the long term.

[0219] For example, there are two ways to implement closed-loop iterative optimization of the pathology number recognition system:

[0220] The first approach is a precise, targeted optimization method based on failure mode decomposition. After the optimization process officially begins, all difficult example data in the difficult example library is fully decomposed and categorized according to preset failure mode classification rules, dividing them into four core difficult example clusters: false detection, recognition failure, character ambiguity, and rule non-compliance. Each cluster corresponds to a specific optimization objective. Then, for each failure mode cluster, a dedicated difficult example training subset is prepared and matched to the corresponding responsible engine: the false detection difficult example set matches the high-speed initial screening engine, the recognition failure difficult example set matches the high-precision main recognition engine, and the character ambiguity difficult example set matches the dedicated difficult example engine. For each engine, incremental fine-tuning training is performed targeting a single failure mode, precisely addressing the specific weakness of the corresponding engine. After training, based on the strategy execution performance data corresponding to each failure mode cluster, the trigger thresholds and execution parameters of the corresponding strategies are adjusted accordingly. For example, the text detection threshold of the initial screening engine is optimized for false detection difficult examples, and the trigger conditions for the direct discard strategy are adjusted simultaneously. Finally, the joint optimization of the engine model and strategy rules is completed, generating a new version of the engine model and linked execution rules, thus completing the system update. This method employs a precise targeted optimization logic based on failure mode decomposition, enabling precise iteration for a class of problems, a subset, a single optimization, and a complete solution. It completely avoids the overfitting and capability shift problems caused by full training, resulting in extremely high optimization efficiency and strong targeting, making it suitable for the periodic system optimization needs of routine clinical scenarios.

[0221] The second approach is a hospital-specific clustered closed-loop optimization method. After the optimization process officially begins, the difficult case data in the difficult case library is first clustered according to the hospital, pathology number coding rules, and slide imaging equipment, creating multiple independent hospital-specific difficult case clusters. Then, for each hospital-specific difficult case cluster, a hospital-specific difficult case training dataset is prepared. Based on this dataset, incremental fine-tuning training is performed on each level of the recognition engine to generate a hospital-specific optimization engine model adapted to the hospital's coding rules and imaging equipment characteristics. Simultaneously, combining the strategy execution effect data of the hospital cluster, the strategy trigger threshold, priority weight, and execution parameters of the corresponding hospital in the strategy chain library are customized and adjusted to generate hospital-specific linkage execution rules. Finally, a corresponding customized optimization model and linkage rules are deployed for each cooperating hospital, achieving customized iterative optimization for different hospital scenarios. This method adopts a hospital-specific clustered customized optimization logic, breaking through the limitations of traditional general models that apply a one-size-fits-all approach. It can accurately adapt to the personalized coding rules of different hospitals and the imaging characteristics of different equipment, improving the recognition accuracy in specific hospital scenarios and adapting to the application scenarios of group pathology centers with multiple campuses and multiple cooperating hospitals.

[0222] In an exemplary scheme for implementing closed-loop iterative optimization of a pathology number recognition system, the system first loads the full set of difficult case data archived in the difficult case library, including false detection samples in region R2, secondary recognition samples of fuzzy text in region R3, manually corrected pathology number samples, and short character samples for rule reasoning, totaling 1200 sets of valid difficult case data. Then, dataset preparation is performed: the full set of difficult case data is cleaned and deduplicated, removing 20 duplicate samples and 15 invalid blank samples, resulting in 1165 sets of valid samples. Next, standardized annotation is performed on the valid samples, completing the annotation of pathology number text truth values, failure mode labels, text region coordinates, and hospital coding rule attributes, generating a difficult case training dataset adapted to the model training requirements. Then, dataset splitting is performed: according to the engine parameter adaptation rules, false detection difficult case samples of text detection are split into the high-speed primary screening engine training set, complete pathology number recognition difficult case samples are split into the high-precision main recognition engine training set, and fuzzy text and short character difficult case samples are split into the dedicated difficult case engine training set. Subsequently, targeted incremental training was performed: For the high-speed initial screening engine, incremental fine-tuning was performed based on false detection samples to optimize the false detection suppression capability of text detection. For the high-precision main recognition engine, incremental fine-tuning was performed based on complete pathology number samples to improve the character recognition accuracy of low-quality images. For the dedicated difficult example engine, incremental fine-tuning was performed based on fuzzy short character samples to strengthen the recognition capability of tissue block numbers. After training, optimized recognition engine models at each level were generated. Next, strategy rule optimization was performed: Combining the strategy execution effect data corresponding to the difficult example data, the success rate of strategy execution in different confidence intervals was statistically analyzed to adapt to the optimized engine recognition capability. The confidence trigger threshold of each execution strategy in the strategy chain library was adjusted, lowering the high confidence threshold from 0.90 to 0.88, and adjusting the intermediate confidence interval to 0.58-0.87. At the same time, the execution parameters of the rule correction strategy and the engine calling rules of the secondary recognition strategy were optimized to generate linkage execution rules that match the optimized engine. Finally, the system update is performed: During off-peak business hours, the optimized recognition engine models at all levels and the linked execution rules are deployed to the pathology number recognition system in the production environment to complete the model and rule update of the entire process.

[0223] Furthermore, in an exemplary scheme of a hierarchical, decision-driven intelligent recovery and interactive processing system for identifying pathology slide pathology numbers, please refer to... Figure 5 , Figure 5This is a system framework diagram for this application. A pathology slide pathology number recognition system constructs a complete technical framework with self-recovery capabilities for recognition difficulties or failures. The system mainly comprises three core units: a multi-level recognition engine module, a unified confidence assessment module, and a policy executor module. It also incorporates a closed-loop learning mechanism. The multi-level recognition engine module includes at least two heterogeneous recognition engines: a high-speed initial screening engine aiming for high recall to quickly locate all possible text candidate regions in the image; a high-precision main recognition engine based on a deep learning OCR model for fine recognition of the initial screening candidate regions; and a dedicated hard-example engine trained for specific high-frequency failure types such as illegible handwriting and low-contrast text. The unified confidence assessment module receives the raw outputs of the multi-level engines and calculates the final confidence of the recognition result through a comprehensive scoring model, combining four dimensions: model confidence, consistency of multi-engine results, conformity to prior rules for pathology numbers, and image quality of the text region. It also analyzes and classifies failure modes such as "no text region detected at all," "region detected but character sequence invalid," and "partial character recognition blurred." The core innovation of this invention is the strategy executor module, which maintains a configurable strategy chain. It dynamically selects the execution path based on the confidence level and failure mode output by the confidence assessment module. The strategy chain includes a complete execution logic: direct output of high-confidence results; triggering a rule-based error correction algorithm when the result is of medium confidence and partially conforms to the rule; switching or parallel invoking a dedicated hard-example engine for secondary recognition when the result is of low confidence but a text region is detected; and initiating non-blocking guided human-computer interaction when the dedicated engine still fails or the confidence level is extremely low, until a reliable recognition result is obtained. Simultaneously, the system incorporates a closed-loop learning mechanism. All samples processed by the strategy executor, especially those processed through human interaction, are automatically labeled and stored in a hard-example sample library for subsequent incremental training or fine-tuning of the recognition engines at each level, enabling the system to continuously evolve.

[0224] This application provides a pathology slide pathology number identification device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the pathology slide pathology number identification method in the above embodiment 1.

[0225] The following is for reference. Figure 6 The diagram illustrates a structural schematic of a pathology slide identification device suitable for implementing embodiments of this application. The pathology slide identification device in this application may include, but is not limited to, mobile terminals such as automatic digital pathology slide scanners, pathology slide barcode scanners, and automatic pathology slide number identification instruments, as well as fixed terminals such as whole-slide imaging systems and digital pathology slide scanning systems. Figure 6The pathology slide identification device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0226] like Figure 6 As shown, the pathology slide identification device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The random access memory 1004 also stores various programs and data required for the operation of the pathology slide identification device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the pathology slide identification device to communicate wirelessly or wiredly with other devices to exchange data. Although pathology slide identification devices with various systems are shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems may be implemented alternatively.

[0227] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0228] The pathology slide identification device provided in this application, employing the pathology slide identification method described in the above embodiments, can solve the technical problem of poor accuracy in pathology slide identification. Compared with the prior art, the beneficial effects of the pathology slide identification device provided in this application are the same as those of the pathology slide identification method provided in the above embodiments, and other technical features of this pathology slide identification device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0229] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0230] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0231] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the pathology slide pathology number identification method in the above embodiments.

[0232] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, radio frequency (RF), etc., or any suitable combination thereof.

[0233] The aforementioned computer-readable storage medium may be included in a pathological slide pathological number identification device; or it may exist independently and not be assembled into a pathological slide pathological number identification device.

[0234] The aforementioned computer-readable storage medium carries one or more programs that, when executed by a pathology slide pathology number recognition device, cause the pathology slide pathology number recognition device to: respond to a pathology number recognition instruction, traverse candidate regions of suspected pathology number text in the pathology slide image, and determine a preliminary recognition result for the candidate regions; based on recognition evaluation rules, determine the degree of fit between the preliminary recognition result and the evaluation index, and calculate a confidence score for the preliminary recognition result based on the degree of fit; based on the confidence score of the preliminary recognition result, match an execution strategy corresponding to the confidence score from a strategy chain library; and execute the corresponding execution strategy on the preliminary recognition result to determine the target recognition result of the pathology number in the pathology slide image.

[0235] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0236] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation that may be implemented in systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0237] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0238] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described method for identifying pathology slide pathology numbers, thereby solving the technical problem of poor accuracy in pathology number identification. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the pathology slide pathology number identification method provided in the above embodiments, and will not be repeated here.

[0239] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for identifying pathology slide pathology numbers, characterized in that, The method includes: In response to the pathology number recognition command, the candidate regions of suspected pathology number text in the pathology slide image are traversed to determine the preliminary recognition results of the candidate regions; Based on the preliminary recognition results corresponding to the candidate regions, the original images of the candidate regions, and the recognition process data, the recognition output probability, the pathology number rule compliance degree, and the text region image quality are extracted. The recognition output probability, the pathology number rule compliance, and the text region image quality are matched with the corresponding evaluation indicators to determine the degree of fit for each indicator. Based on the evaluation weights corresponding to the evaluation indicators, the fit of all individual indicators is weighted and fused to obtain the confidence score corresponding to the preliminary identification result. Based on the confidence score of the preliminary identification result, the execution strategy corresponding to the confidence score is matched from the strategy chain library, including: performing a combined condition check on the confidence score of the preliminary identification result based on the strategy chain library and the triggering conditions corresponding to the execution strategy, and determining at least one candidate strategy; Based on the strategy priority and pathological business requirements, the candidate strategies are prioritized to obtain the strategy sequence of the preliminary identification results; The candidate strategy ranked first in the strategy sequence is taken as the execution strategy corresponding to the preliminary identification result; The execution strategy corresponding to the preliminary identification result is executed to determine the target identification result of the pathology number in the pathological slide image, including: based on the execution strategy corresponding to the candidate region, parsing the preliminary identification result of the candidate region, associated image data and identification process metadata, and determining the execution action, verification rules and execution parameters corresponding to the execution strategy; Based on the execution action and the execution parameters, a targeted processing operation is performed to match the candidate region, and the initial strategy output result of the candidate region is obtained; According to the verification rules, the initial strategy output results of the candidate regions are subjected to confidence verification and format validity verification to generate valid recognition results of the candidate regions. Based on the regional order of the candidate regions, the effective identification results are merged to obtain the target identification result of the pathology number; After the step of executing the corresponding execution strategy on the preliminary identification result to determine the target identification result of the pathology number in the pathology slide image, the method for identifying the pathology number of the pathology slide further includes: Based on the high-confidence target recognition results and the auxiliary coding rules of the same slide pathology number, business attribute reasoning and matching are performed on the short character recognition results with medium confidence to determine the tissue block number attribute corresponding to the short character recognition results; According to the consistency rules of pathological information on the same slide, the association compliance between the high-confidence target identification result and the tissue block number attribute is verified to obtain the pathology number result; Based on the pathology number result, the verification and recognition result of the pathology slide image is determined, and the sample data and rule execution data corresponding to this reasoning and matching process are archived and stored as difficult case data.

2. The method for identifying pathology slide pathology numbers as described in claim 1, characterized in that, The step of responding to the pathology number recognition command, traversing candidate regions of suspected pathology number text in the pathology slide image, and determining the preliminary recognition result of the candidate regions includes: In response to the pathology number identification command, the pathology slide image is preprocessed to obtain a standardized pathology slide image; By traversing the pathological slide images through the initial screening operation, the candidate regions of all suspected pathology number texts are located, and invalid background regions without text features are identified to obtain candidate text regions. Local contrast enhancement and texture denoising are performed on the candidate text regions respectively to remove interference features that are not text background within the candidate text regions, thus obtaining text-standardized regions. By performing high-precision text recognition, the text content of each standardized text region is recognized one by one, generating preliminary recognition results corresponding to the candidate regions.

3. The method for identifying pathology slide pathology numbers as described in claim 1, characterized in that, After the step of performing a weighted fusion calculation on the fit of all individual indicators according to the evaluation weights corresponding to the evaluation indicators to obtain the confidence score corresponding to the preliminary identification result, the method for identifying the pathology slide pathology number further includes: The preliminary identification results are divided into confidence levels according to multiple confidence level thresholds to identify abnormal areas with confidence levels lower than the qualified level. Based on the confidence level label corresponding to the abnormal region, match the abnormal feature verification rule specific to the confidence level label; Based on the aforementioned anomaly feature verification rules, the compliance of the pathology number rules and the validity of the text region in the anomaly region are specifically verified, and the anomaly feature distribution results of the candidate region are generated. Based on the abnormal feature distribution results of the candidate region, the problematic link that causes the abnormality in the pathology number identification process is located, and the abnormality judgment result of the abnormal region is generated. Based on the anomaly determination result, the pathology number is matched to identify the anomaly classification rules to determine the failure mode corresponding to the anomaly area.

4. The method for identifying pathology slide pathology numbers as described in claim 1, characterized in that, After determining the verification and recognition result of the pathological slide image based on the pathological number result, and archiving and storing the sample data and rule execution data corresponding to this inference matching process as difficult case data, the method for recognizing the pathological number of the pathological slide further includes: The difficult example data in the difficult example library is filtered, cleaned, deduplicated, and standardized to obtain a difficult example training dataset that meets the training requirements of the model. The difficult example training dataset is split according to the engine parameter rules, and targeted incremental training is performed on the high-speed initial screening engine, the high-precision main recognition engine, and the dedicated difficult example engine to generate an optimized recognition engine model. Based on the optimized recognition engine model, and combined with the strategy execution effect data corresponding to difficult examples, the trigger threshold, priority weight, and execution parameters of the execution strategies in the strategy chain library are adjusted for adaptability to obtain the linkage execution rules. Based on the optimized recognition engine model and the linkage execution rules, the model and rules of the pathology number recognition process are updated to achieve iterative optimization of the pathology number recognition system.

5. A device for identifying pathology slide pathology numbers, characterized in that, The pathology slide pathology number identification device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the pathology slide pathology number identification method as described in any one of claims 1 to 4.

6. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the method for identifying the pathology number of a pathology slide as described in any one of claims 1 to 4.

Images