Multi-agent based medical image reporting automatic quality control method and system

Abstract

This application relates to the field of medical information processing and discloses an automatic quality control method and system for medical image reports based on multi-agent technology. The method involves an agent unit calling a rule engine to perform preliminary screening of clauses within a sentence segment set, identifying clauses with logical conflicts, terminological errors, or missing descriptions to form a candidate abnormal sentence segment set. Entity data of the candidate abnormal clauses is extracted and mapped to a medical knowledge graph, and the semantic confidence of the clauses is calculated. Clauses with confidence scores below a preset first threshold are selected to form a low-confidence sentence segment set. A retrieval enhancement generation unit accesses an external medical knowledge base to obtain reference segments, driving a finely tuned large language model unit to compare the low-confidence clauses with the reference segments, calculating the semantic reasonableness confidence rate. Clauses with confidence scores below a preset second threshold are identified as erroneous clauses, and a structured quality control result containing the error type, location, correction suggestions, and explanation of the cause is output.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, specifically to an automatic quality control method and system for medical image reports based on multi-agent technology. Background Technology

[0002] Medical imaging reports are the core output of radiological examinations, providing crucial information for clinical diagnosis and treatment planning. They typically include descriptions of anatomical structures, records of lesion signs, and final diagnostic conclusions. These reports are the primary tool for communication between radiologists and clinicians, and their accuracy directly impacts medical quality and patient safety.

[0003] In current clinical practice, quality control of medical imaging reports primarily relies on manual review. Meanwhile, some digital medical information systems integrate basic logic verification tools that assist physicians in checking report content through pre-defined, fixed logic. These tools are typically embedded in the report writing process to identify common typographical errors or omissions of key fields, thereby improving the report's standardization.

[0004] However, existing medical imaging report quality control technologies often rely on single-dimensional verification mechanisms, such as simple template rule matching or basic language models that focus on shallow text features. This singular technical architecture makes it difficult for the quality control process to simultaneously consider the collaborative verification of the syntactic layer, the medical professional knowledge layer, and the deep semantic layer. Due to the lack of multi-layered technical integration, existing technologies cannot effectively identify multi-level error patterns existing at different logical depths in the report, resulting in quality control being limited to surface textual logic and failing to provide comprehensive automated assurance of the medical rationality of the report content. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an automatic quality control method and system for medical image reports based on multi-agent systems, which solves the problems of multi-dimensional logical error correction and semantic quality control in medical image reports.

[0006] The first aspect of this invention provides an automated quality control method for medical image reports based on multi-agent systems, comprising the following steps: The system receives the text of a medical image report to be processed and performs segmentation based on punctuation marks within the text, generating a set of sentence segments containing multiple clauses. A built-in intelligent agent unit invokes a rule engine to perform preliminary screening of each clause within the sentence segment set, identifying clauses with logical conflicts, terminological errors, or missing descriptions. These identified clauses are output as candidate anomalous clauses, forming a candidate anomalous sentence segment set. Entity data for each candidate anomalous clause in the candidate anomalous sentence segment set is extracted and input into a medical knowledge graph for mapping and matching. The semantic confidence of each candidate anomalous clause with entities in the medical knowledge graph is calculated. Candidate abnormal clauses with semantic confidence scores below a preset first threshold are filtered and combined into a low-confidence sentence segment set, which contains multiple low-confidence clauses. An external medical knowledge base is accessed through a retrieval enhancement generation unit to retrieve and extract reference report fragments related to the low-confidence clauses in the low-confidence sentence segment set. A large language model unit, finely tuned by instructions, compares the low-confidence clauses with the reference report fragments, calculates the semantic reasonableness confidence rate of each low-confidence clause, and determines the corresponding low-confidence clause as an erroneous clause when the semantic reasonableness confidence rate is less than a preset second threshold, and outputs the corresponding structured quality control results.

[0007] The innovative principle of this invention is explained as follows: This invention constructs a progressive filtering architecture based on multi-layer intelligent fusion. By processing step by step from the basic syntax layer, medical knowledge layer to the semantic layer, it solves the technical defects of insufficient logical rigor and semantic understanding depth in the quality control of medical image reports.

[0008] At the data processing level, this invention employs a punctuation-driven segmentation mechanism for medical image report text, refining the entire report text into clause processing units, thereby improving the accuracy of error localization. In the initial screening stage, this invention introduces an intelligent agent unit with adaptive update capabilities. This intelligent agent unit dynamically supplements the rule base by learning and extracting historical error features. The rule engine intercepts surface-level logical defects by simultaneously verifying the correspondence between organs and locations, the mutual exclusion logic consistency between gender and anatomical structures, and the correspondence between descriptions and diagnostic conclusions.

[0009] During the review phase, this invention utilizes a pre-constructed medical knowledge graph to provide professional knowledge constraints. Specifically, the named entities in candidate anomaly clauses are mapped to nodes in the medical knowledge graph, and a vector similarity function is used to calculate the similarity score between the text entity vector and the medical entity node vector, which is then used as the semantic confidence score. This confidence calculation mechanism based on a medical knowledge graph provides a medical factual basis for semantic evaluation, solving the problem of uninterpretable results from purely deep learning models.

[0010] In the final decision-making stage, this invention combines instruction fine-tuning of the large language model unit (MLM) with retrieval enhancement generation technology. The retrieval enhancement generation unit extracts highly similar standard reference report texts from an external vector database as the reference context for the MLM's inference. The MLM performs deep reasoning based on the semantic consistency between the reference context and low-confidence clauses. This approach can identify semantic contradictions and generate structured quality control results containing error localization, correction plans, and explanations of causes based on the retrieved medical standards, thus improving the effectiveness of automated quality control.

[0011] A second aspect of the present invention provides an automated quality control system for medical image reports based on multi-agent technology, comprising: The report segmentation module receives the medical image report text to be processed and performs segmentation operations based on the punctuation marks in the medical image report text to generate a set of sentence segments containing multiple clauses; The rule engine initial screening module is used to call the rule engine through the built-in intelligent agent unit to perform preliminary screening of each clause in the sentence segment set, identify clauses with logical conflicts, terminological errors or missing descriptions, and output the identified clauses as candidate abnormal clauses to form a candidate abnormal sentence segment set. The knowledge graph screening module is used to extract entity data of each candidate abnormal clause in the candidate abnormal sentence segment set, input the entity data into the medical knowledge graph for mapping and matching, calculate the semantic confidence of each candidate abnormal clause with the entity of the medical knowledge graph, and filter to generate a set of sentence segments with low confidence. The final review module is used to access an external medical knowledge base through the retrieval enhancement generation unit to obtain reference context data, and drive the large language model unit, which has been fine-tuned by instructions, to compare low-confidence clauses with the reference context data, thereby calculating the semantic reasonableness confidence rate and outputting the corresponding structured quality control results.

[0012] Through the coordinated operation of various modules, this invention achieves automated processing from medical image report input to final quality control result output, reduces the labor intensity of manual review, and enhances the interpretability of error correction conclusions while ensuring the medical rationality of the quality control process.

[0013] This invention provides an automated quality control method and system for medical image reports based on multi-agent technology. It offers the following advantages: 1. This invention performs progressive screening by sequentially inputting the clauses obtained from report segmentation into a rule engine, a medical knowledge graph, and a large language model unit that has been fine-tuned by instructions. It uses rules to intercept basic logical contradictions, uses the knowledge graph to verify medical professional facts, and uses the large language model unit to evaluate the semantic consistency of complex contexts. This achieves full-dimensional automated quality control of medical imaging reports from local syntactic structure and professional knowledge logic to the semantic level of the entire text. It solves the technical deficiency that a single quality control method cannot cover multi-level error patterns and ensures the comprehensiveness and accuracy of the quality control results.

[0014] 2. This invention introduces an intelligent agent unit with adaptive update capabilities into the initial screening module of the rule engine. The intelligent agent unit collects and analyzes historical corpus features and human quality control feedback signals in real time. Based on the extracted error and omission patterns, it dynamically generates new logical judgment templates and automatically adds new rules to the rule engine. This achieves adaptive growth and continuous optimization of the quality control rule system, reduces the complexity and time cost of manually maintaining and updating the rule base, and enhances the quality control system's ability to flexibly adapt to different medical institution reporting styles and new error patterns.

[0015] 3. This invention employs a technical architecture combining a large language model unit with a retrieval enhancement generation unit in the final review stage. It utilizes a vector database to retrieve standard reference reports from an authoritative external medical knowledge base as the basis for reasoning. The logical reasoning results of the large language model unit are then transformed into structured quality control results containing error type, error location, correction suggestions, and explanations of causes. This achieves the interpretability and clinical guidance value of automated quality control conclusions, assisting doctors in quickly performing report correction operations based on an understanding of the error correction basis, and improving the efficiency of writing medical image diagnostic reports in a human-machine collaborative mode. Attached Figure Description

[0016] Figure 1 This is a structural block diagram of the automated quality control system for medical image reports of the present invention; Figure 2 This is a flowchart illustrating the automatic quality control method for medical image reports of the present invention. Figure 3 This is a working logic diagram of the report segmentation module of the present invention; Figure 4 This is a working logic diagram of the rule engine initial screening module of the present invention; Figure 5 This is a working logic diagram of the knowledge graph screening module of the present invention; Figure 6 This is a working logic diagram of the final review module of the present invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see the appendix Figure 1 -Appendix Figure 6 This invention provides an automatic quality control system for medical image reports based on multi-layer intelligent fusion. The automatic quality control system for medical image reports includes: a report segmentation module, a rule engine initial screening module, a knowledge graph screening module, and a final review module.

[0019] The report segmentation module is used to acquire the medical image report text to be processed. The module segments the medical image report text according to punctuation marks (specifically commas, pauses, and periods) to obtain corresponding sentence sets. Each sentence set contains multiple clauses, and each clause constitutes an independent descriptive segment in the medical image report.

[0020] The rule engine's initial screening module communicates with the report segmentation module to receive the aforementioned set of sentence segments. The initial screening module uses its built-in rule engine to screen the sentence segment set, identifying clauses with logical conflicts, terminological errors, or missing descriptions, and outputs a set of candidate abnormal sentence segments. The rule engine includes organ-location correspondence rules, gender-organ logical conflict rules, description-conclusion consistency rules, and image feature integrity rules. The initial screening module includes an intelligent agent unit, which learns from historical data and updates the rule base within the rule engine.

[0021] The knowledge graph screening module communicates with the rule engine's initial screening module to obtain a set of candidate abnormal sentence segments. The knowledge graph screening module pre-stores a medical knowledge graph. It extracts entity data (including diseases, symptoms, organs, locations, and imaging signs) from clauses and maps this entity data to the medical knowledge graph, calculating the semantic confidence of each clause in the candidate abnormal sentence segment set. For clauses with a semantic confidence score lower than a preset first threshold (e.g., the preset first threshold is 0.6), the knowledge graph screening module filters and aggregates the corresponding clauses into a low-confidence sentence segment set.

[0022] The final review module communicates with the knowledge graph screening module to receive the aforementioned set of low-confidence sentence segments. The final review module integrates a large language model unit and a retrieval enhancement generation unit. The retrieval enhancement generation unit accesses an external medical knowledge base to extract similar reference report fragments. The large language model unit compares the clauses in the low-confidence sentence segment set with the reference report fragments, calculating the semantic reasonableness confidence rate of the corresponding clauses. When the semantic reasonableness confidence rate is lower than a preset second threshold (e.g., the preset second threshold is 0.8), the final review module determines the corresponding clause as erroneous content and outputs a quality control result including the error type, error location, and correction suggestions.

[0023] This invention provides an automated quality control method for medical image reports. The automated quality control method for medical image reports is executed based on the aforementioned automated quality control system for medical image reports, and includes: S1. The report segmentation module receives the medical image report text to be processed and performs segmentation operations on the text based on the punctuation marks within the text, generating a set of sentence segments. The punctuation marks used for segmentation include commas, pauses, and periods. The set of sentence segments contains multiple clauses, each clause constituting an independent descriptive fragment in the medical image report.

[0024] S2. The rule engine's initial screening module acquires a set of sentence segments and calls the rule engine to perform preliminary screening on each clause within the set. The rule engine is configured with rules for organ-location correspondence, logical conflict between gender and organ, consistency between description and conclusion, and integrity of image features. The rule engine's initial screening module identifies clauses with logical conflicts, terminological errors, or missing descriptions, outputting the identified clauses and forming a candidate abnormal sentence segment set. Simultaneously, the intelligent agent unit in the rule engine's initial screening module reads historical corpus data and updates the rule content within the rule engine based on this data.

[0025] S3. The knowledge graph screening module receives a set of candidate abnormal sentence segments and extracts entity data for each clause in the set. Entity data includes disease nodes, symptom nodes, organ nodes, location nodes, and imaging sign nodes. The knowledge graph screening module inputs the entity data into a medical knowledge graph for mapping and matching, calculating the semantic confidence of each clause with entities in the medical knowledge graph. The knowledge graph screening module filters clauses with semantic confidence less than a preset first threshold, combining the filtered clauses into a low-confidence sentence segment set. The preset first threshold is 0.6.

[0026] S4. The final review module receives a set of low-confidence sentence segments and controls its internal retrieval enhancement generation unit to access an external medical knowledge base to retrieve and extract reference report fragments related to clauses in the low-confidence sentence segment set. The final review module controls its internal, fine-tuned large language model unit to compare clauses with reference report fragments and calculate the semantic reasonableness confidence rate of each clause. The final review module determines whether the semantic reasonableness confidence rate is less than a preset second threshold. The preset second threshold is 0.8. When the semantic reasonableness confidence rate is less than the preset second threshold, the final review module determines the corresponding clause as an erroneous clause and outputs the corresponding quality control result. The quality control result includes the error type, error location, correction suggestions, and explanation of the cause.

[0027] This invention provides an automated quality control method for medical image reports. The report segmentation step in the automated quality control method for medical image reports may include: The report segmentation module receives the raw medical image report transmitted from an external system and performs text preprocessing operations on the raw medical image report to remove invisible characters, thereby obtaining the medical image report text to be processed. The medical image report text contains a continuous string of medical descriptive text.

[0028] The report segmentation module iterates through the text and symbols in the medical image report, extracting commas, pauses, and periods as segmentation markers. Using these markers as sentence breakpoints, the module performs semantic segmentation on the medical image report, filtering out invalid empty characters generated during the segmentation process, and generating multiple separate clauses.

[0029] The report segmentation module records the original order of multiple clauses in the medical image report text, and combines the clauses according to the original order to construct a set of sentence segments. Variables in the formula This indicates the total number of clauses generated after the segmentation operation.

[0030] Sentence Collection Each clause in Each of these constitutes an independent descriptive segment in the original medical imaging report. An independent descriptive segment carries an anatomical description or pathological feature description of a single site. Variables in the formula It means greater than or equal to 1 and less than or equal to 1. A positive integer. The report segmentation module will construct the complete set of sentence segments. The data is transmitted to the rule engine's initial screening module for subsequent logical review operations.

[0031] The rule engine initial screening step in the automated quality control method for medical image reports may include: The rule engine's initial screening module receives the set of sentence segments transmitted by the report segmentation module. The system then activates its built-in intelligent agent unit. This unit pre-extracts historical medical image report corpora and manually modified records using natural language processing technology, mapping the medical entities and their relationships before and after modification. Based on these mappings, it generates logical decision templates containing conditional statements or regular expressions, and uses these templates to construct the initial rule engine.

[0032] In one alternative implementation, the intelligent agent unit can not only utilize manual modification of records, but also extract entities and relationships from the positive sample corpus of historical medical image reports using natural language processing technology combined with rule building tools, thereby supplementing and improving the logical judgment template containing conditional statements or regular expressions.

[0033] Intelligent agent unit extracts sentence segment set Each clause in and each clause The input rule engine performs feature matching. The rule engine is configured with rules for organ-location correspondence, rules for logical conflict between gender and organ, rules for consistency between description and conclusion, and rules for image feature integrity.

[0034] Based on the correspondence rules between organs and locations, the rule engine validates the clauses. The validity of the combination of anatomical site names and spatial location terms. Based on the gender and organ logical conflict rules, the rule engine extracts the patient gender identifier from the structured information fields of the medical image report and combines the patient gender identifier with clauses. The organ names in the system are mutually exclusive; when the gender identifier and the organ name have a logical conflict, the rule engine generates a conflict determination.

[0035] Based on the consistency rule between description and conclusion, the rule engine extracts the diagnostic conclusion content at the end of the medical imaging report text and compares the clauses. The consistency between the pathological description and the diagnostic conclusion in the image, including their positive / negative attributes and affirmative / negative judgments. Based on the image integrity rules, the rule engine detects clauses. Does the descriptive text for a specific lesion contain basic vocabulary indicating size, shape, margins, and density?

[0036] The rule engine's initial screening module gathers clauses that trigger any of the above rule decisions. The clause that will trigger the judgment. Defined as candidate anomalous clauses, and combined according to the order of the candidate anomalous clauses in the original text to generate a set of candidate anomalous sentence segments. Variables in the formula This indicates the number of candidate anomalous clauses that have potential logical conflicts, terminological errors, or missing descriptions. (Variable) For variables less than or equal to Integers.

[0037] The rule engine's initial screening module sets up a collection of candidate abnormal sentence segments. The data is sent to the knowledge graph screening module. Simultaneously, the intelligent agent unit receives external quality control feedback signals, analyzes new error characteristics based on these signals, and adds new rules to the rule engine based on these error characteristics, thus completing the adaptive update of the rule base within the rule engine.

[0038] The knowledge graph screening step in automated quality control methods for medical imaging reports may include: The knowledge graph screening module receives a set of candidate abnormal sentence segments sent by the rule engine's initial screening module. The knowledge graph screening module pre-builds and stores a medical knowledge graph (KG). The medical knowledge graph KG contains multiple medical entity nodes and semantic association edges connecting these nodes. Specifically, the medical entity nodes include disease nodes, symptom nodes, organ nodes, location nodes, and imaging sign nodes. The semantic association edges specifically indicate the inclusion relationships, concurrency relationships, and anatomical location relationships between the medical entity nodes.

[0039] The knowledge graph screening module traverses the set of candidate abnormal sentence segments. Each candidate exception clause Candidate anomaly clauses are extracted using named entity recognition algorithms. The text entity data is mapped to corresponding medical entity nodes in the medical knowledge graph (KG), and a text embedding algorithm is used to convert the text entity data and the corresponding medical entity nodes into text entity vectors and medical entity node vectors, respectively. Specifically, the text embedding algorithm employs a language model pre-trained on medical domain corpora, such as ClinicalBERT or BioBERT. The variables in the formula... It means greater than or equal to 1 and less than or equal to 1. Positive integers.

[0040] The knowledge graph screening module calculates candidate anomaly clauses based on the mapping results and vector transformation results. semantic confidence Call the confidence calculation formula Perform the calculation. In the confidence level calculation formula, Indicates candidate exception clauses The extracted text entity vectors, This represents the vector of medical entity nodes matched in the medical knowledge graph (KG). This represents the vector similarity calculation function, which specifically includes the cosine similarity function. Semantic confidence. Used to characterize candidate exception clauses The degree of matching between the medical descriptions in the database and the internal logic of the medical knowledge graph (KG).

[0041] The knowledge graph screening module will calculate the semantic confidence. The value is compared with a preset first threshold; the preset first threshold is set to 0.6. When a candidate exception clause... semantic confidence When the value is less than 0.6, the knowledge graph screening module identifies candidate abnormal clauses. The medical logic check failed, and candidate abnormal clauses that failed the medical logic check were extracted. .

[0042] The knowledge graph screening module aggregates all semantic confidence scores. Candidate outlier clauses with a confidence level less than 0.6 are used to construct a set of low-confidence sentence segments. And the completed set of low-confidence sentence segments The data is sent to the final review module for subsequent semantic reasonableness evaluation. Variables in the formula. Indicates the total number of candidate abnormal clauses that failed the medical logic review; variable Less than or equal to Positive integers.

[0043] The final review step in the automated quality control method for medical image reports may include: The final review module receives a set of low-confidence sentence segments transmitted by the knowledge graph screening module. It then activates the internally configured retrieval enhancement generation unit and the large language model unit. The large language model unit is pre-processed using medical image annotation corpus through instruction fine-tuning and stores instruction-following weight parameters in the medical field.

[0044] The final review module extracts a set of low-confidence sentence segments. Low confidence clauses in ; variables in the formula It means greater than or equal to 1 and less than or equal to 1. Positive integers. The retrieval enhancement generation unit obtains low-confidence clauses. The text features are analyzed and transformed into retrieval vectors, which are then input into a pre-built external medical knowledge base for retrieval matching. The external medical knowledge base is a vector database storing a large number of standard medical knowledge fragments. The retrieval enhancement generation unit uses a vector similarity algorithm to recall clauses with low confidence from the external medical knowledge base. Reference report fragments with text similarity are used as contextual data for medical reasoning.

[0045] The final review module will use low-confidence clauses The input sequence is concatenated with the reference context data to form a composite input sequence containing prompt words; the task instructions specify the output format as a scoring field containing values ​​between 0 and 1. The large language model unit receives the composite input sequence and compares it with low-confidence clauses. The semantic consistency status between the response text and the reference context data is determined, and the prompt words and semantic consistency status are combined to generate a response text containing text evaluation content and a numerical score field. The final review module parses the numerical score field in the response text using regular expression matching rules and extracts this numerical score field as the semantic reasonableness confidence rate for low-confidence clauses. .

[0046] The final review module will calculate the semantic reasonableness confidence rate. The value is compared with a preset second threshold; the preset second threshold is set to 0.8. When the semantic reasonableness confidence rate... When the confidence level is less than 0.8, the final review module determines the clause to be of low confidence. It identifies erroneous clauses and controls the large language model unit to generate structured quality control results for these erroneous clauses.

[0047] The structured quality control results include error type, error location, correction suggestions, and explanation of the cause. Error types specifically cover logical errors, location conflicts, and descriptive inconsistencies. The final review module aggregates the structured quality control results corresponding to all error clauses and performs a centralized output operation, completing the final review quality control process for medical image report texts.

[0048] The automated quality control system for medical image reports operates within an electronic device, which may include a processor, memory, and a communication interface. The processor, memory, and communication interface are electrically connected and have data communication links with each other via a system bus.

[0049] The memory is configured to store the computer program and the data required for system operation. The computer program consists of a series of computer-executable software codes. Logically, the software code is divided into the aforementioned report segmentation module, rule engine initial screening module, knowledge graph screening module, and final review module.

[0050] The processor accesses memory, reads and executes the computer program, thereby completing each processing step in the automated quality control method for medical image reports described in the foregoing embodiments. The processor controls the communication interface to receive the raw medical image report sent by the external medical information system, and controls the communication interface to output the generated structured quality control results to the peripheral terminal.

[0051] The present invention also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program is read and executed by a processor in an electronic device, it drives the electronic device to perform the aforementioned automated quality control method for medical image reports, which includes report segmentation, rule engine initial screening, knowledge graph screening, and final review.

[0052] Computer-readable storage media are non-transient machine-readable media. Machine-readable media specifically include read-only memory, random access memory, magnetic hard disks, solid-state drives, or optical disks.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for medical image reporting automatic quality control based on multi-agent, characterized in that, Includes the following steps: Receive the medical image report text to be processed, and perform a segmentation operation based on the punctuation marks in the medical image report text to generate a sentence set containing multiple clauses; The built-in intelligent agent unit calls the rule engine to perform preliminary screening of each clause in the sentence segment set, identify clauses with logical conflicts, terminological errors or missing descriptions, and output the identified clauses as candidate abnormal clauses to form a candidate abnormal sentence segment set. Extract entity data of each candidate abnormal clause from the candidate abnormal sentence segment set, input the entity data into a medical knowledge graph for mapping and matching, calculate the semantic confidence of each candidate abnormal clause with the entity of the medical knowledge graph, and filter and combine candidate abnormal clauses with semantic confidence less than a preset first threshold into a low confidence sentence segment set, the low confidence sentence segment set containing multiple low confidence clauses; By accessing an external medical knowledge base through a retrieval enhancement generation unit, reference report fragments related to the low-confidence clauses in the set of low-confidence sentence segments are retrieved and extracted; By using a large language model unit that has been fine-tuned by instructions, the low-confidence clauses are compared with the reference report fragments to calculate the semantic reasonableness confidence rate of each low-confidence clause. When the semantic reasonableness confidence rate is less than a preset second threshold, the corresponding low-confidence clause is determined to be an erroneous clause, and the corresponding structured quality control result is output.

2. The automatic quality control method for medical image reports based on multi-agent technology according to claim 1, characterized in that, The process involves receiving the medical image report text to be processed and performing segmentation based on the punctuation marks within the text to generate a set of sentence segments containing multiple clauses, including: Perform text preprocessing on the original medical image report to remove invisible characters and obtain the medical image report text to be processed; Traverse the text and symbols in the medical image report text, and extract the commas, pauses and periods in the medical image report text as segmentation markers; Using the location of the segmentation marker as the sentence break node, a semantic segmentation operation is performed on the medical image report text, while filtering out invalid empty characters generated by the segmentation operation to generate multiple separate clauses; Record the original arrangement order of multiple clauses in the medical image report text, and combine multiple clauses according to the original arrangement order to construct the sentence set. 3.The multi-agent based medical image reporting automatic quality control method of claim 1, wherein, The rule engine is configured with organ-location correspondence rules, gender-organ logical conflict rules, description-conclusion consistency rules, and image sign integrity rules. The invocation rule engine performs a preliminary screening of each clause within the sentence segment set, including: The built-in intelligent agent unit is activated to extract each clause from the sentence segment set, and each clause is input into the rule engine for feature matching.

4. The automatic quality control method for medical image reports based on multi-agent technology according to claim 3, characterized in that, The step of inputting each clause into the rule engine for feature matching includes: Based on the organ-location correspondence rule, verify the effectiveness of the combination of anatomical part names and spatial location terms in the clause; Based on the aforementioned gender and organ logical conflict rules, the patient gender identifier is extracted from the structured information field of the medical image report, and the patient gender identifier is mutually exclusive with the organ name in the clause. When the patient gender identifier and the organ name have a logical conflict, a conflict determination is generated. Based on the consistency rule between description and conclusion, the diagnostic conclusion content at the end of the medical imaging report text is extracted, and the consistency between the pathological description content in the clause and the positive / negative attribute and the affirmative / negative judgment of the diagnostic conclusion content is compared. Based on the image feature integrity rules, it is detected whether the descriptive text of the clause for a specific lesion contains basic element words representing size, shape, edge and density. 5.The multi-agent based medical image reporting automatic quality control method of claim 1, wherein, Before the rule engine performs preliminary screening on each clause in the sentence segment set, the method further includes the step of constructing a rule engine using the intelligent agent unit: The intelligent agent unit extracts the mapping relationship between historical medical image report corpus and medical entities in manually modified records and before and after modification through natural language processing technology. Based on the mapping relationship, it generates a logical judgment template containing conditional judgment statements or regular expressions, and constructs the initial rule engine through the logical judgment template. During the screening process, the intelligent agent unit receives externally input quality control feedback signals, analyzes new error characteristics based on the quality control feedback signals, and adds new rules to the rule engine based on the error characteristics, thereby completing the adaptive update of the rule base within the rule engine. The medical knowledge graph contains multiple medical entity nodes and semantic association edges connecting the medical entity nodes; the medical entity nodes specifically include disease nodes, symptom nodes, organ nodes, location nodes and imaging sign nodes, and the semantic association edges specifically indicate the inclusion relationship, concurrency relationship and anatomical location relationship between the medical entity nodes. The step of extracting entity data for each candidate abnormal clause from the candidate abnormal sentence segment set and inputting the entity data into a medical knowledge graph for mapping and matching includes: The text entity data in the candidate abnormal clauses is extracted using a named entity recognition algorithm, and the text entity data is mapped to the corresponding medical entity nodes in the medical knowledge graph. The text embedding algorithm is invoked to convert the text entity data and the corresponding medical entity nodes into text entity vectors and medical entity node vectors, respectively.

6. The automatic quality control method for medical image reports based on multi-agent technology according to claim 5, characterized in that, The step of constructing a rule engine using the intelligent agent unit further includes: The intelligent agent unit uses natural language processing technology and rule building tools to extract entities and relationships from positive sample corpora of historical medical image reports, and generates logical judgment templates for regular expressions as auxiliary benchmarks for building or supplementing the initial rule engine.

7. The automatic quality control method for medical image reports based on multi-agent technology according to claim 6, characterized in that, The calculation of the semantic confidence of each of the candidate abnormal clauses with the entities in the medical knowledge graph includes: The text entity vector and the medical entity node vector are input into the vector similarity calculation function to perform calculation operations, and a semantic confidence value is obtained to characterize the degree of internal logical matching between the medical description content in the candidate abnormal clause and the medical knowledge graph; wherein, the vector similarity calculation function specifically includes a cosine similarity function.

8. The automatic quality control method for medical image reports based on multi-agent technology according to claim 1, characterized in that, The control retrieval enhancement generation unit accesses an external medical knowledge base to retrieve and extract reference report fragments related to the low-confidence clauses in the set of low-confidence sentence segments, including: The text features of the low-confidence clause are obtained and converted into retrieval vectors. The retrieval vectors are then input into a pre-built external medical knowledge base to perform retrieval matching. The external medical knowledge base is a vector database that stores a large number of standard medical knowledge fragments. Using a vector similarity algorithm, reference report fragments with textual similarity to the low-confidence clause are retrieved from the external medical knowledge base, and these reference report fragments are used as reference context data for medical reasoning.

9. The automatic quality control method for medical image reports based on multi-agent technology according to claim 8, characterized in that, The large language model unit, fine-tuned by control instructions, compares the low-confidence clauses with the reference report fragments, calculates the semantic reasonableness confidence rate of each low-confidence clause, and outputs the corresponding structured quality control results, including: The low-confidence clause is concatenated with the reference context data to form a composite input sequence containing prompt words; the prompt words contain task instructions that constrain the large language model unit to perform medical logical reasoning and structured output, and the task instructions limit the output format to a scoring field containing numerical values; The large language model unit receives the composite input sequence, compares the semantic consistency status between the low-confidence clause and the reference context data, and integrates the prompt words with the semantic consistency status to generate a response text containing text evaluation content and numerical score fields. The numerical rating field in the response text is parsed using regular expression matching rules, and the numerical rating field is extracted as the semantic reasonableness confidence rate for the low confidence clause. When the low-confidence clause is determined to be an erroneous clause, the large language model unit is controlled to generate the structured quality control result for the erroneous clause. The structured quality control result includes the error type, error location, correction suggestions, and reason explanation.

10. A multi-agent-based automatic quality control system for medical image reports, based on the multi-agent-based automatic quality control method for medical image reports as described in any one of claims 1-9, characterized in that, include: The report segmentation module is used to receive the medical image report text to be processed, and perform segmentation operation according to the punctuation marks in the medical image report text to generate a set of sentence segments containing multiple clauses; The rule engine initial screening module is communicatively connected to the report segmentation module. It is used to call the rule engine through the built-in intelligent agent unit to perform initial screening on each clause in the sentence segment set, identify clauses with logical conflicts, terminological errors or missing descriptions, and output the identified clauses as candidate abnormal clauses to form a candidate abnormal sentence segment set. The knowledge graph screening module is communicatively connected to the rule engine initial screening module. It is used to extract entity data of each candidate abnormal clause in the candidate abnormal sentence segment set, input the entity data into the medical knowledge graph for mapping and matching, calculate the semantic confidence of each candidate abnormal clause with the entity of the medical knowledge graph, and filter and combine candidate abnormal clauses with semantic confidence less than a preset first threshold into a low confidence sentence segment set. The low confidence sentence segment set contains multiple low confidence clauses. The final review module, which communicates with the knowledge graph screening module, accesses an external medical knowledge base through an internally configured retrieval enhancement generation unit. It retrieves and extracts reference report fragments related to the low-confidence clauses in the low-confidence sentence set. Through an internally configured, finely tuned large language model unit, it compares the low-confidence clauses with the reference report fragments, calculates the semantic reasonableness confidence rate of each low-confidence clause, and determines that the corresponding low-confidence clause is an erroneous clause when the semantic reasonableness confidence rate is less than a preset second threshold, and outputs the corresponding structured quality control results.

Images