An auxiliary diagnosis method and system based on pathological report text

Abstract

The application provides an auxiliary diagnosis method and system based on pathological report text, comprising: inputting pathological report text and task questions, extracting key field information in the pathological report text to form a report evidence item set; constructing a guideline knowledge base containing candidate guideline knowledge units according to a pathological guideline, recalling candidate guideline knowledge units based on the report evidence item set and the task questions, and performing pathological evidence constraint scoring and diagnosis output gating on the candidate guideline knowledge units based on a candidate knowledge scoring mechanism and a diagnosis output gating mechanism; inputting the report evidence item set, the task questions, the candidate guideline knowledge units, the scoring results and the gating results into an answer generation intelligent agent to generate auxiliary diagnosis reference information; and outputting auxiliary diagnosis result information by correcting and checking the auxiliary diagnosis reference information through the constructed reflection intelligent agent. The application improves the pathological report text auxiliary analysis efficiency and reduces the manual guideline searching and information sorting time.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and in particular relates to an auxiliary diagnostic method and system based on pathology report text. Background Technology

[0002] Current AI-assisted pathology diagnostic technologies primarily focus on the identification, classification, segmentation, and visual question answering of pathological slide images or full-field images. These technologies mainly utilize image features to assist in judging tissue structure and tumor morphology. However, in actual pathological diagnosis, pathologists also need to comprehensively consider various textual evidence from the pathology report, including the sampling site, histological morphology, immunohistochemical results, molecular detection results, depth of invasion, lymph node status, resection margin status, and clinical history, and compare this with the World Health Organization (WHO) tumor classification, the Tumor-Node-Metastasis staging system (TNM) staging criteria, and relevant pathology guidelines. Therefore, relying solely on image analysis is insufficient to meet the diagnostic assistance needs at the textual level of pathology reports.

[0003] Current natural language processing technologies for pathology reports typically focus on extracting disease names, locations, immunohistochemical markers, molecular test results, or staging fields from the report. They primarily address the problem of "information extraction" rather than "diagnostic reasoning." In other words, existing methods can identify what is written in the report, but they struggle to further determine whether a candidate diagnosis meets the necessary criteria in the guidelines, whether there is missing or conflicting evidence, and they also find it difficult to provide reliable differential diagnoses and recommendations for subsequent examinations.

[0004] When general-purpose large language models or ordinary retrieval-augmented generation methods are directly used for pathological diagnosis, their reliability is also insufficient. On the one hand, large language models may make over-inferences based on general medical knowledge, giving overly definitive diagnoses when evidence is insufficient. On the other hand, although ordinary retrieval-augmented generation (RAG) methods can recall guidelines or literature fragments, if pathological evidence matching, necessary criteria for judgment, and conflict evidence control are lacking, the model may still misuse the search content and fail to distinguish between defining evidence, supporting evidence, missing evidence, and exclusionary evidence.

[0005] Furthermore, existing systems often only output diagnostic names or natural language explanations, lacking a correspondence between reported evidence, guideline basis, and output conclusions. For pathologists, diagnostic support results should not only provide "what might be," but also explain "the basis," "what evidence is missing," "which evidence does not support it," and "what further examinations are needed." Without this traceable structure, system outputs are difficult to verify and cannot be safely used in real pathological diagnostic processes. Summary of the Invention

[0006] In view of this, this application aims to propose an auxiliary diagnostic method and system based on pathology report text to solve at least one of the above-mentioned problems.

[0007] To achieve the above objectives, the technical solution of this application is implemented as follows: Firstly, this application provides an auxiliary diagnostic method based on pathology report text, including: Input the anonymized pathology report text and the task question, and extract key field information from the pathology report text in a structured manner to form a set of report evidence items; A guideline knowledge base containing candidate guideline knowledge units is constructed based on pathology guidelines. Candidate guideline knowledge units are recalled based on the set of report evidence items and the task question. The candidate guideline knowledge units are then subjected to pathology evidence constraint scoring and diagnostic output gating based on a candidate knowledge scoring mechanism and a diagnostic output gating mechanism. The report evidence set, task question, candidate guide knowledge units, scoring results and gating results are input into the constructed answer generation agent to generate auxiliary diagnostic reference information; The constructive reflective agent corrects and verifies the auxiliary diagnostic reference information to output auxiliary diagnostic results.

[0008] Secondly, based on the same inventive concept, this application also provides an auxiliary diagnostic system based on pathology report text, comprising: The report evidence structuring module is configured to take as input the anonymized pathology report text and the task question, and to extract key field information from the pathology report text in a structured manner to form a set of report evidence items; The candidate knowledge processing module is configured to construct a guideline knowledge base containing candidate guideline knowledge units based on pathology guidelines, recall candidate guideline knowledge units based on the set of report evidence items and the task question, and perform pathology evidence constraint scoring and diagnostic output gating on the candidate guideline knowledge units based on the candidate knowledge scoring mechanism and the diagnostic output gating mechanism. The reference information generation module is configured to input the report evidence item set, task question, candidate guide knowledge unit, scoring result and gating result into the constructed answer generation agent to generate auxiliary diagnostic reference information; The reflection and verification module is configured to correct and verify the auxiliary diagnostic reference information through a constructed reflection agent, so as to output auxiliary diagnostic result information.

[0009] Thirdly, based on the same inventive concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the first aspect.

[0010] Fourthly, based on the same inventive concept, this application also provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method as described in the first aspect.

[0011] Compared with existing technologies, the auxiliary diagnostic method and system based on pathology report text described in this application has the following beneficial effects: This application improves the efficiency of pathology report text-assisted analysis and the accuracy of pathology report text-assisted diagnosis by matching pathology report evidence with authoritative pathology guidelines and combining candidate knowledge scoring and diagnostic output gating. It can serve as an auxiliary reasoning tool for pathologists, reducing the time spent on manual guideline retrieval and information organization, helping them to locate key evidence more quickly, reduce omissions, and form more standardized diagnostic and follow-up examination recommendations. Attached Figure Description

[0012] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of an auxiliary diagnostic method based on pathology report text, as described in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of an auxiliary diagnostic system based on pathology report text, as described in an embodiment of this application. Figure 3 This is a schematic diagram of the hardware structure of the electronic device described in an embodiment of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0014] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0015] This application proposes an auxiliary diagnostic method based on pathology report text. This method constructs an authoritative pathology diagnostic guideline into a searchable knowledge base and combines the pathology report text, task question, retrieved diagnostic knowledge, and model self-reflection feedback into an iterative reasoning process. This scheme can provide evidence-constrained diagnostic assistance based on pathology report-level text without using pathology images, outputting the final auxiliary diagnostic result, diagnostic rationale, differential diagnosis, evidence basis, and suggestions for subsequent auxiliary examinations.

[0016] The embodiments of this application are described in detail below with reference to the accompanying drawings.

[0017] Please see Figure 1 As shown, this embodiment provides an auxiliary diagnostic method based on pathology report text, specifically including the following steps: Step S10: Input the anonymized pathology report text and the task question. The key field information in the pathology report text is extracted in a structured manner to form a set of report evidence items.

[0018] Specifically, in this embodiment, the purpose of this step is mainly to convert the pathology report text into a set of report evidence items. For information such as immunohistochemical markers, genes, mutations, fusions, rearrangements, amplifications, deletions, sampling sites, number of lymph nodes, resection margins, and depth of invasion, a rule dictionary and regular expression templates can be used for extraction. For free text such as histological morphology, diagnostic tendencies, post-treatment responses, and clinical background, language models or medical text extraction models can be used to assist in extraction. The model-assisted extraction results must be converted into preset fields and able to locate the original report text; otherwise, they will enter the low-confidence set and will not participate in determining the diagnostic gating. Each report evidence item is represented as follows: ; in, This is a category label for the evidence item, used to indicate which type of pathological information the evidence belongs to, such as the location (e.g., location). ),form( ), immunohistochemistry ( Molecular detection ), clinical information ( ) or installment information ( )wait; Standardized evidence names; The original value; This refers to the location of a segment or field in the original report. For the polarity of evidence, take , , Positive, present, detected, mutation, fusion, rearrangement, amplification, etc., are denoted as... Not detected, not provided, not mentioned, or cannot be determined is recorded as: Negative, not detected, no mutation, no fusion, no rearrangement, clearly excluded or not supported are recorded as The system distinguishes between "not detected" and "negative test result"; the former does not cause a conflict, while the latter may.

[0019] It should be noted that the report text in this embodiment can be Chinese, English or other languages; if the knowledge base is in a different language than the report, translation retrieval, cross-language embedding or multilingual knowledge base can be used.

[0020] Step S20: Construct a guideline knowledge base containing candidate guideline knowledge units based on pathology guidelines, recall candidate guideline knowledge units based on the report evidence item set and task questions, and perform pathology evidence constraint scoring and diagnostic output gating on the candidate guideline knowledge units based on the candidate knowledge scoring mechanism and diagnostic output gating mechanism.

[0021] Specifically, in this embodiment, each knowledge unit in the guide knowledge base includes at least a disease entity. Tumor types Subtype Location of occurrence Original Chapter and the original guide The system retains the original chapters. Used for tracing the source, and at the same time generating standardized chapter fields. Key diagnostic elements and diagnostic constraint roles Computable fields, etc.

[0022] Used to unify the wording of different guideline chapters, such as diagnostic criteria Mapped to Histopathological features Mapped to Mapping Immunohistochemistry to Mapping molecular genetics to molecular genetics will enable differential diagnosis. Mapped to It will be paid in installments. Mapped to This field is used for weighting subsequent task chapters and does not replace the original. .

[0023] This indicates key diagnostic elements that can be directly compared with the pathology report, each At least include categories ,name Expected state of the guidelines and the role of evidence .in, Necessary evidence can be obtained Supporting evidence Exclusionary evidence or background evidence .For example, A positive result can usually be represented as ( , 1. Positive , TFE3 rearrangement, as a diagnostic defining event, can be represented as ( , Rearrangement ,Appear , If the reported evidence is consistent with... If consistent, it is marked as support; if not detected or mentioned, it is marked as missing; if explicitly contradictory, it is marked as conflict.

[0024] Used to indicate the strength of constraint that a knowledge unit or key piece of evidence has on a diagnostic conclusion. If the text contains key evidence... Necessary evidence Definitional evidence Diagnostic evidence Expressions such as "necessary" and "must" are denoted as... ; including exclusion ,exclude Expressions such as exclusion, non-support, etc., are denoted as ; characterized by… ,Frequent ,generally ,support Common, supportive, and other similar expressions are denoted as The rest are recorded as .

[0025] This application structures authoritative guidelines such as the WHO Classification of Cancer into knowledge units based on disease entity, tumor type, subtype, location, and chapter (including the original chapter and the original guideline text). Each knowledge unit retains the original chapter. And the original text and further generate , and Computable fields, etc. Used for weighting task chapters Used to record key diagnostic elements that can be directly compared with pathology reports. This is used to indicate the strength of the constraint of the knowledge unit or evidence item on the diagnostic conclusion. This mechanism enables the originally static guideline text to participate in the reporting of evidence matching, determination of necessary criteria, conflict identification, and generation of subsequent examination recommendations.

[0026] In this embodiment, the candidate guide knowledge retrieval is not limited to a specific retrieval tool. A preset fusion retrieval strategy can be used to simultaneously calculate term relevance and semantic relevance, and the two types of relevance are normalized to obtain a fusion retrieval score. The specific fusion retrieval strategy is as follows: Setting a task problem and pathology report text The set of report evidence items extracted from the data is as follows: The candidate guide knowledge unit is Then, the term relevance can be expressed as: ; In the formula, Indicates evidence item In knowledge unit Frequency of occurrence in Indicates inverse document frequency. Indicates the length of the knowledge unit. This represents the average length of a knowledge unit. and This is an adjustable parameter.

[0027] It should be noted that the term relevance calculation method in this embodiment can also be term frequency, inverse document frequency, phrase matching, synonym expansion, field weighting or other keyword relevance calculation methods. No specific calculation method is limited here.

[0028] Semantic relevance can be expressed as: ; In the formula, Vector representations of pathology report text and task questions. A vector representation of the candidate guide knowledge unit.

[0029] It should be noted that the semantic relevance calculation method in this embodiment can be generated by a medical text encoding model, a general semantic encoding model, a cross-language text encoding model, or other models that can calculate text similarity, and is not limited to a specific model name.

[0030] After normalizing the term relevance and semantic relevance, the fusion recall score is obtained: ; In the formula, Indicates hyperparameters, The system is based on the fusion recall score Select One candidate guide knowledge unit.

[0031] The recall steps described above are only used to provide candidate guidelines for the answer generation agent. The final output is still subject to subsequent pathological evidence scoring, diagnostic gating, and reflection and correction mechanisms.

[0032] It should be noted that the candidate knowledge retrieval method in this embodiment can also employ exact nearest neighbor, approximate nearest neighbor, inverted index, hybrid index, database retrieval, or other retrieval methods that can obtain candidate guide knowledge units. This embodiment is not limited to a specific indexing system.

[0033] In this embodiment, the candidate knowledge scoring mechanism constrained by pathological evidence specifically includes the following: For knowledge units The system is based on the pathology report text. and task issues Calculate the overall score : ; In the formula, Indicates text relevance. Indicates the coverage of case evidence. Indicates the degree to which necessary diagnostic criteria are met. Indicates the chapter task weight. Indicates conflict punishment. , , , and These are the weighting coefficients assigned respectively. In the initial state, this embodiment sets... , , , , . This indicates that the result will be restricted to... to Within the range.

[0034] Among them, text relevance Used to measure whether key evidence items from reports and questions appear in candidate guide knowledge units: ; In the formula, Key diagnostic elements in knowledge units , This indicates an exact match or a synonym match; otherwise, it is an undefined match. Gene, fusion, mutation, and rearrangement fields have higher weighting, followed by immunohistochemistry, location, morphology, and staging fields.

[0035] Case evidence coverage Used to measure the extent to which candidate guideline knowledge units cover the evidence in the current report: ; In the formula, Indicates the report's evidence items, Indicates the number of reported evidence items used in the evidence coverage calculation; Indicating the report evidence items and guidelines If they match and their polarities do not conflict, otherwise they are... .

[0036] Satisfaction of necessary diagnostic criteria Originating from the same disease entity Lowering the diagnostic constraint role The knowledge unit. Let the necessary standard set be... ,in, Represents disease entity The One necessary diagnostic criterion (necessary diagnostic criteria can be found in the guideline text). (Key evidence) (Evidence of necessity) (Definitive evidence) The expressions (diagnostic evidence), necessities, essentials, diagnostic basis, etc., are determined through rule templates, medical text extraction models, or manual review. Then, each standard is compared with the report to obtain its status. Support for , missing The conflict is Define the indicator function I(⋅), then the necessary criterion satisfaction rate, missing rate, and conflict rate are respectively: ; ; ; In the formula, For disease entity The corresponding number of necessary diagnostic criteria, For the necessary standard satisfaction, For the necessary standard missing degree, Conflict degree of necessary standards.

[0037] Chapter Task Weight The result is obtained by looking up the table based on the task type and the normalized section fields. For example, improving diagnostic tasks... , , and Weighting; Improved task of determining the primary site , and differential diagnosis Weighting; phased task improvement Weighting; increased subsequent inspection tasks , and Weights.

[0038] Conflict Punishment Used to reduce the ranking of candidate guideline knowledge units that contradict the reported evidence: ; In the formula, Indicates the proportion of immunohistochemical polarity conflict. Indicates the proportion of polarity conflict in molecular events. The locations indicated are clearly inconsistent. , , , These are the assigned weighting coefficients, in this embodiment, , , , It should be noted that items not detected or mentioned in the report are not included in the conflict calculation.

[0039] When a disease entity lacks necessary diagnostic criteria, the satisfaction, absence, and conflict of necessary diagnostic criteria are not included in the comprehensive scoring and gating of the candidate disease entity, and the weights of the remaining scoring items are renormalized. When there are no valid evidence items in the report, the case evidence coverage is recorded as zero. When a candidate guideline knowledge unit lacks comparable key diagnostic elements, no evidence scoring is performed on that knowledge unit. It should be noted that the normalization operation in this embodiment can be based on min-max normalization of the current set of candidate knowledge units.

[0040] This application does not select knowledge fragments solely based on text similarity, but instead calculates text relevance for each candidate guide knowledge unit. Coverage of case evidence Degree of satisfaction of necessary diagnostic criteria Chapter task weight and conflict punishment And form a comprehensive score. This mechanism can reduce the weight of candidate entities that are inconsistent with the currently reported evidence, contain only non-specific background information, or lack necessary criteria, thereby reducing the interference of retrieval noise and competing entities on the model output.

[0041] In this embodiment, the diagnostic output gating mechanism specifically includes the following: The system processes each candidate disease entity Calculate output level .in, This represents the overall score of the candidate disease entity. Indicates the degree to which necessary diagnostic criteria are met. Indicates the degree of absence of necessary diagnostic criteria. Indicates the degree of conflict of necessary diagnostic criteria. This represents the conflict penalty value. All the above indicators are derived from the aforementioned pathological evidence constraint scoring process. The system is based on... , , , and Determine the output level of candidate disease entities: like or or ,but ; like and and and and ,but ; like and and and ,but ; like and ,but .

[0042] in, This indicates a definitive or highly supportive diagnostic auxiliary conclusion. Indicates a propensity-based diagnostic auxiliary conclusion. This indicates a supplementary conclusion for differential diagnosis or that further examination is needed for confirmation. This indicates that the current evidence does not support or substantially excludes the possibility. If the same entity meets multiple conditions simultaneously, the priority of the judgment is as follows: This mechanism prevents the system from making a direct diagnosis when necessary criteria are lacking, prioritizes downgrading or exclusion when there is clear conflict in the evidence, and retains it as a differential diagnosis or recommendation for further examination when the evidence is relevant but insufficient.

[0043] It should be noted that in this embodiment, for , , , and The threshold value is not specifically limited. This embodiment is only an example. The specific value should be flexibly set according to the actual situation. Further details will not be provided here.

[0044] This embodiment sets the necessary diagnostic criteria satisfaction level, missing evidence ratio, conflicting evidence ratio, and conflict penalty item. This mechanism prevents the system from directly outputting a conclusion when the necessary criteria are missing, prioritizes exclusion when there is clear conflict in the evidence, and retains uncertainty and recommends supplementary examinations when the evidence is relevant but insufficient.

[0045] Step S30: Input the set of report evidence items, task questions, candidate guide knowledge units, scoring results and gating results into the constructed answer generation agent to generate auxiliary diagnostic reference information.

[0046] Specifically, in this embodiment, the process integrates diverse information such as the report evidence set, task questions, candidate guide knowledge units, scoring results, and gating results into structured prompts, which are then input into an answer generation agent with strong reasoning capabilities (such as DeepSeek-R1). The answer generation agent is then called in the background to generate a structured output containing preliminary diagnostic auxiliary conclusions, supporting evidence, and an explanation of the degree of support. This forms traceable and verifiable auxiliary diagnostic reference information, laying the foundation for subsequent manual review and iterative optimization.

[0047] This application does not simply input guideline fragments into a large language model. Instead, it first recalls, scores, and gates the guideline knowledge units, and then uses reported evidence, candidate guideline knowledge, missing evidence, and conflicting evidence together to generate the answer. Therefore, the diagnostic reasons generated by the system are more likely to correspond with the original report and guideline basis, making it easier for pathologists to review.

[0048] In complex case tasks, this application system can suggest subsequent examinations based on insufficient current evidence, such as supplementing immunohistochemistry or fluorescence in situ hybridization. This includes sequencing or specific gene testing, etc. This demonstrates that this application can not only output candidate diagnoses, but also provide more targeted follow-up examination suggestions around diagnostic uncertainties.

[0049] The generative model in this embodiment is implemented using DeepSeek-R1, but it can also be replaced with GPT, Claude, Gemini, Qwen, Llama, or a medical-specific large language model. Alternatively, a local deployment model may be used, but no further restrictions are placed on the specific model implementation here.

[0050] Step S40: Correct and verify the auxiliary diagnostic reference information through the constructed reflective agent to output auxiliary diagnostic result information.

[0051] Specifically, in this embodiment, this step employs a correction mechanism based on LLM reflection results, which includes the following: After the auxiliary diagnostic reference information is generated, the reflective agent is invoked in the background to verify the previous round's output. Based on the set of reported evidence items, candidate guideline knowledge units, missing evidence, conflicting evidence, and the conclusions of the previous round, the reflective agent outputs a fixed-format result, including the status. and reasons . Pass Or needs to be improved ; This is used to explain why the validation failed.

[0052] when When the system retains the results from the previous round. At that time, the system according to The problem is mapped to insufficient evidence, inadequate consideration of conflicting evidence, omission of differential diagnosis, or over-determinism in the output, and limited corrective actions are triggered.

[0053] If the evidence is insufficient, the system will increase the required standard satisfaction level. weight ,For example, Increase , and Each reduction and will Tighten the gating conditions to , , , .like The highest output of the candidate diagnosis is ;like The highest output is .

[0054] If conflicting evidence is not adequately considered, the system will increase the conflict penalty. weight ,For example Increase , and Each reduction .like Indicating immunohistochemistry, molecular events, or site conflicts will correspondingly increase [the risk / progress]. , or If after correction or Then output ;like The highest output is ;like The highest output is .

[0055] If a differential diagnosis is missed, the system will improve... Chapter weight For example, increase And the maximum shall not exceed and the number of candidate outputs Depend on Increase to But not relaxed Gating conditions.

[0056] If the output is overly definitive, the system will not reinvent diagnostic criteria, but will instead add a conservative penalty term. Or directly reduce the output level. If corrected... , or Then no output is allowed. ;like or Then no output is allowed. The expressions "confirmed diagnosis" and "clear diagnosis" have been changed to "prone", "considered", or "requires further testing for confirmation".

[0057] The system's default maximum number of reflection rounds is If it remains the same after reaching the maximum number of rounds. If the condition is not met, the most conservative result will be output, and a manual review will be prompted. This mechanism retains LLM's ability to identify complex pathological semantics, implicit conflicts, and omissions in identification. At the same time, by using fixed states, causal mapping, and limited corrective actions, it avoids uncontrolled changes in diagnostic aid conclusions caused by open reflection, making diagnostic aid results more traceable, auditable, and controllable.

[0058] The final output includes candidate disease entities and output levels. Overall score Supporting evidence, missing evidence, conflicting evidence, corresponding guideline knowledge units, main differential diagnoses, recommendations for follow-up examinations, and reflective agent verification records (output format can be natural language report, structured JavaScript object representation). (This includes tables, pathology diagnosis recommendation forms, interface data, or fields that can be received by the electronic medical record system). It should be noted that this output is only intended to assist pathologists in their diagnosis and does not replace their final diagnosis.

[0059] This method can improve the stability and verifiability of the output results. Since purely reflective reasoning is prone to introducing new errors in subsequent revisions, this application limits the reflective process to the scope of reported evidence, guideline knowledge units, missing evidence, and conflicting evidence. It also adjusts the output level, differential diagnosis range, or statement certainty through limited revisions, thereby reducing error corrections and making it easier for pathologists to judge whether the system's conclusions are reliable.

[0060] This application improves the accuracy of diagnostic assistance based on pathology report text by matching pathology report evidence with authoritative pathology guidelines and combining candidate knowledge scoring and diagnostic output gating. In structured tasks, such as TNM staging, the implementation method of this application achieves an overall accuracy of 92.5%, higher than the 83.8% achieved by directly using large language model inference. In tumor regression grading tasks, the implementation method of this application achieves an accuracy of 99.6%.

[0061] In complex entity-level diagnostic tasks, this application also improves diagnostic accuracy. For example, in rare tumor diagnosis tasks, the accuracy rate of this application's system reaches 94.6%; in metastatic tumor primary site inference tasks, the accuracy rate reaches 94.5%, reducing the number of misclassifications by approximately 39% compared to the baseline of direct large language model inference. These results demonstrate that this application can better utilize morphological, immunohistochemical, molecular detection, and clinical background evidence in the report, reducing misclassifications arising from relying solely on internal model knowledge.

[0062] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0063] Based on the same inventive concept, and corresponding to the methods of any of the above embodiments, the embodiments of this application also provide an auxiliary diagnostic system based on pathology report text.

[0064] like Figure 2 As shown, the pathology report-based assisted diagnostic system includes: The report evidence structuring module 11 is configured to take the anonymized pathology report text and task questions as input, and to extract key field information from the pathology report text in a structured manner to form a set of report evidence items; The candidate knowledge processing module 12 is configured to construct a guideline knowledge base containing candidate guideline knowledge units based on pathology guidelines, recall candidate guideline knowledge units based on the report evidence item set and task questions, and perform pathology evidence constraint scoring and diagnostic output gating on the candidate guideline knowledge units based on the candidate knowledge scoring mechanism and diagnostic output gating mechanism. The reference information generation module 13 is configured to input the report evidence item set, task question, candidate guide knowledge unit, scoring result and gating result into the constructed answer generation agent to generate auxiliary diagnostic reference information; The reflection and verification module 14 is configured to correct and verify the auxiliary diagnostic reference information through the constructed reflection agent in order to output auxiliary diagnostic result information.

[0065] For ease of description, the above system is described by dividing it into various modules based on their functions. Of course, in implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware.

[0066] The system can be deployed on the hospital intranet, pathology department server, cloud-based private environment, or local workstation. Input can be anonymized pathology report text, or structured or semi-structured report content exported from the pathology information system.

[0067] In human-computer collaborative experiments, the system described in this application can assist pathologists in organizing report evidence, highlighting missing information, controlling diagnostic uncertainty, and improving the quality of diagnosis and subsequent recommendations. In tasks such as TNM staging, diagnosis of rare tumors, inference of the primary site of metastatic tumors, and analysis of complex cases, pathologists assisted by this application system outperformed those without assistance or with ordinary search assistance, and were able to shorten the completion time for some tasks.

[0068] Therefore, this application can not only improve the diagnostic assistance performance of the model itself, but also serve as an auxiliary reasoning tool for pathologists, helping them to locate key evidence more quickly, reduce omissions, and form more standardized diagnostic and follow-up examination recommendations.

[0069] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the methods described in any of the above embodiments.

[0070] Figure 3This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0071] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0072] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0073] The input / output interface 1030 is used to connect input / output modules to realize information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0074] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0075] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0076] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0077] The electronic devices described above are used to implement the corresponding methods in any of the foregoing embodiments and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0078] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium that stores computer instructions for causing the computer to perform the methods described in any of the above embodiments.

[0079] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0080] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to perform the methods described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0081] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0082] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the claims of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A diagnostic aid method based on pathology report text, characterized in that, include: Input the anonymized pathology report text and the task question, and extract key field information from the pathology report text in a structured manner to form a set of report evidence items; A guideline knowledge base containing candidate guideline knowledge units is constructed based on pathology guidelines. Candidate guideline knowledge units are recalled based on the set of report evidence items and the task question. The candidate guideline knowledge units are then subjected to pathology evidence constraint scoring and diagnostic output gating based on a candidate knowledge scoring mechanism and a diagnostic output gating mechanism. The report evidence set, task question, candidate guide knowledge units, scoring results and gating results are input into the constructed answer generation agent to generate auxiliary diagnostic reference information; The constructive reflective agent corrects and verifies the auxiliary diagnostic reference information to output auxiliary diagnostic results.

2. The method according to claim 1, characterized in that: Key field information is extracted to convert the pathology report text into a set of report evidence items. Each report evidence item contains an evidence item category label, standardized evidence name, original value, original text fragment or field position, and evidence polarity. The methods for extracting key field information include: rule dictionaries, regular expression templates, language models, or medical text extraction models.

3. The method according to claim 1, characterized in that: Each candidate guideline knowledge unit in the guideline knowledge base includes at least the disease entity, tumor type, subtype, location, original chapter, and original guideline text.

4. The method according to claim 1, characterized in that: A preset fusion recall strategy is used to recall candidate guide knowledge units based on the set of reported evidence items and the task question, providing candidate guide basis for the answer generation agent; wherein, the fusion recall strategy includes: The relevance of terms is determined based on the set of evidence items in the report, the task question, and the knowledge units of the candidate guidelines. Semantic relevance is determined based on the vector representations of pathology report text and task questions, as well as the vector representations of candidate guideline knowledge units; The term relevance and semantic relevance are normalized to obtain a fusion recall score, and candidate guide knowledge units with higher scores are selected based on the fusion recall score.

5. The method according to claim 1, characterized in that, The candidate knowledge scoring mechanism includes: Text relevance is determined based on the pathology report text, task questions, and candidate guideline knowledge units; The pathological evidence coverage, the satisfaction of necessary diagnostic criteria, and the conflict penalty are determined based on the pathology report text and the candidate guideline knowledge unit, respectively. The chapter task weights are determined based on the task questions and the candidate guide knowledge units. The comprehensive score of the candidate disease entity is obtained based on the text relevance, pathological evidence coverage, satisfaction of necessary diagnostic criteria, chapter task weight, and conflict penalty.

6. The method according to claim 5, characterized in that, The diagnostic output gating mechanism includes: The output level of the candidate disease entity is determined based on the comprehensive score of the candidate disease entity, the degree of satisfaction of necessary diagnostic criteria, the degree of absence of necessary diagnostic criteria, the degree of conflict of necessary diagnostic criteria, and the conflict penalty item, and is used as the output gating result. The output levels include definitive or highly supportive diagnostic aid conclusions, propensity diagnostic aid conclusions, differential diagnostic aid conclusions or conclusions requiring further examination for confirmation, and levels where current evidence does not support or substantially excludes the diagnosis.

7. The method according to claim 1, characterized in that: The reflective agent performs verification based on the set of reported evidence items, candidate guideline knowledge units, missing evidence, conflicting evidence, and auxiliary diagnostic reference information to output a fixed-format result, including a status and a reason; wherein, the status is either passed or needs improvement, and the reason is used to explain why the verification failed. If the status is passed, the auxiliary diagnostic reference information is retained; In response to the state that improvement is needed, the problem is mapped to insufficient evidence, insufficient consideration of conflicting evidence, omission of differential diagnosis, or overly certain output statement based on the reason, and a corrective action is triggered. When the state is passed or the maximum number of reflection rounds is reached, auxiliary diagnostic result information is output.

8. An auxiliary diagnostic system based on pathology report text, characterized in that, include: The report evidence structuring module is configured to take as input the anonymized pathology report text and the task question, and to extract key field information from the pathology report text in a structured manner to form a set of report evidence items; The candidate knowledge processing module is configured to construct a guideline knowledge base containing candidate guideline knowledge units based on pathology guidelines, recall candidate guideline knowledge units based on the set of report evidence items and the task question, and perform pathology evidence constraint scoring and diagnostic output gating on the candidate guideline knowledge units based on the candidate knowledge scoring mechanism and the diagnostic output gating mechanism. The reference information generation module is configured to input the report evidence item set, task question, candidate guide knowledge unit, scoring result and gating result into the constructed answer generation agent to generate auxiliary diagnostic reference information; The reflection and verification module is configured to correct and verify the auxiliary diagnostic reference information through a constructed reflection agent, so as to output auxiliary diagnostic result information.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium, characterized in that, in, The non-transitory computer-readable storage medium stores computer instructions for causing a computer to perform the method described in any one of claims 1-7.

Images