Text report correction system and method fusing llm and ris processes

Abstract

The application discloses a text report error correction system and method fusing LLM and RIS processes, comprising a task triggering module, a rule engine module, a large language model scheduling module, a prompt word management module and a result processing and feedback module; the task triggering module automatically starts the error correction task at the key node of the report process, the rule engine module is responsible for rapid real-time rule checking, and the large language model scheduling module performs deep semantic analysis. The system dynamically selects a special prompt word template highly matched with the current scene through the prompt word management module, and significantly improves the accuracy of analysis. Finally, all results are integrated and interactively presented by the result processing and feedback module. The application organically combines real-time response and deep analysis, deeply integrates the existing workflow, reduces the interference on the doctor's work while ensuring medical safety, and improves the quality and efficiency of medical report generation.

Description

Technical Field

[0001] This invention relates to the field of medical information, specifically to a text report error correction system and method that integrates LLM and RIS processes. Background Technology

[0002] Currently, radiologists mostly use fixed text templates to copy / paste and then modify to complete diagnostic reports, which leads to numerous types of errors. These range from simple typos to errors in diagnostic logic, numbering in the dozens. In the past, these errors were typically detected using strict word matching or NLP (Neutral Language Processing) analysis with synonyms / near-synonyms. While this method is fast and can detect and alert on errors in real time, its drawback is that the richness and variety of report text make it impossible to match all possible errors, easily leading to omissions. Checking each error scenario individually would result in overly complex program logic, excessively high implementation costs, and impractical application. Even so, it cannot handle the ever-changing complexity of text, leading to a poor return on investment.

[0003] With an open-source LLM deployment, error correction in reports using prompt words can achieve broad matching of various synonyms / near-synonyms. One prompt word can manage many situations and can discover errors such as typos that are obvious to humans without relying on program rules. This greatly improves the performance-price ratio of text report error correction services.

[0004] However, LLM's response speed is relatively slow. Even when specifically deployed for a particular role, its feedback time typically takes 20 to 30 seconds; in scenarios with large-scale departmental operation, the feedback delay can easily exceed one minute. This significant delay makes it difficult to directly apply LLM's report correction function in real-time workflows. A common practice is for doctors to complete the report and close it, then review it later after the LLM analysis results are returned. However, this asynchronous processing method has significant drawbacks: when the LLM subsequently indicates a potential error, doctors face a dilemma of whether to revise the report and delay its release. Since LLM has a certain false alarm rate, over-reliance on its prompts will not only frequently interrupt doctors' workflows and waste their time, but will also directly affect the timeliness of report distribution to clinicians. Summary of the Invention

[0005] This invention provides a text report error correction method that integrates LLM and RIS workflows, solving the problem of how to efficiently integrate a large language model with slow response speed but powerful analysis capabilities into the workflow of a radiology information system with extremely high real-time requirements, and achieving multi-level, high-accuracy intelligent error correction of medical text reports without affecting the efficiency of report writing and review.

[0006] Firstly, this application provides a text report error correction system that integrates LLM and RIS processes, the system comprising:

[0007] The task triggering module is configured to respond to report status change events in the radiology information system and automatically generate error correction tasks at three preset process nodes: initial report submission, review report opening, and review report closing.

[0008] The rules engine module, connected to the task triggering module, is configured to call a pre-configured rule base to perform real-time analysis of the report text and generate a first correction result at the initial report submission node and the review report closing node.

[0009] The large language model scheduling module is connected to the task triggering module and is configured to manage the LLM analysis task queue and schedule task execution according to task priority. The tasks include in-depth analysis tasks triggered at the preliminary report submission node and timely analysis tasks triggered at the review report closing node.

[0010] The prompt word management module is connected to the large language model scheduling module and is configured to dynamically select matching prompt word templates from the prompt word template library based on the patient examination information associated with the report for use by the large language model scheduling module.

[0011] The result processing and feedback module is connected to the rule engine module and the large language model scheduling module, respectively. It is configured to receive the first error correction result and the second error correction result from the LLM, and control the user interface to perform corresponding interactive prompts and report status management based on the current process node.

[0012] A further optimization is that the task triggering module maintains a mapping relationship between report status identifiers and database stored procedures, and calls the corresponding stored procedure to create a task record when the status changes.

[0013] A further optimization is that the large language model scheduling module is configured as follows:

[0014] The task triggered by the initial report submission node is set as a deep analysis task and added to the regular task queue for asynchronous processing;

[0015] The task triggered by the closed audit report node is set as an immediate analysis task and is prioritized and inserted at the front of the task queue for immediate processing.

[0016] A further optimization involves the prompt word management module executing a structured query to match the device type field and / or diagnosis conclusion field in the patient information database with preset filtering conditions in the prompt word template library, in order to select an applicable prompt word template.

[0017] A further optimization is that, at the audit report closing node, the result processing and feedback module is configured as follows:

[0018] When a bug correction task is triggered, the report is placed in a buffered state;

[0019] The system receives the real-time analysis results from the rule engine module and the timely analysis results from the large language model scheduling module in parallel.

[0020] If neither the real-time analysis result nor the timely analysis result indicates an error, the report status will be automatically updated to the published status.

[0021] If either party's analysis results indicate an error, the report will remain in a buffer state and a warning will be displayed in the user interface. It can only be published after user confirmation.

[0022] A further optimization is that the rule base called by the rule engine module contains customizable terms and logical rules, which are used to perform at least one verification, including gender conflict, device type conflict, and critical value judgment.

[0023] Secondly, this application provides a text report error correction method that integrates LLM and RIS processes, applied to the text report error correction system that integrates LLM and RIS processes as described above. The method includes:

[0024] Monitor RIS report status changes and trigger error correction tasks at the nodes of initial report submission, review report opening, and review report closing.

[0025] Depending on the triggering node, the task is routed to the rule engine for real-time analysis and / or to the LLM task queue for asynchronous analysis;

[0026] For tasks routed to the LLM task queue, scheduling is performed based on their priority attributes, and prompt word templates are dynamically matched for the tasks to invoke the LLM service for analysis;

[0027] Integrate real-time analysis results from the rules engine and asynchronous analysis results from the LLM;

[0028] Based on the integrated results and the current node of the report, control the flow of the user interface and the report status.

[0029] A further optimization scheme is as follows: at the initial report submission node, the rule engine is triggered to perform real-time analysis and create an LLM in-depth analysis task; at the audit report opening node, the completed LLM analysis results are displayed; at the audit report closing node, the rule engine is triggered to perform real-time analysis and create an LLM timely analysis task, and the report release process is determined based on the analysis results of both.

[0030] A further optimization involves creating multiple LLM analysis tasks for the same report during the asynchronous analysis step, with each task using a different prompt word template to analyze different business dimensions.

[0031] Thirdly, this application provides a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above.

[0032] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0033] By deploying different types of error correction tasks on multiple business nodes, real-time rule verification is combined with asynchronous deep analysis of LLM, which not only ensures the timely discovery of critical errors, but also utilizes the time difference of the process to complete deep semantic analysis, thus avoiding the blocking of the process by LLM response delay.

[0034] By dividing LLM tasks into "timely tasks" and "deep analysis tasks" and setting different priorities, the system can ensure that the analysis results of key nodes are returned quickly, thereby optimizing the utilization efficiency of computing resources.

[0035] The prompt word management module dynamically selects dedicated prompt word templates based on information such as examination equipment, body parts, and diagnostic conclusions, which improves the accuracy and relevance of the large language model in specific medical scenarios and reduces the "hallucination" phenomenon.

[0036] The system's triggering and feedback mechanisms are deeply embedded in the RIS standard process, the error correction actions are naturally aligned with the doctor's operation nodes, and the prompt information is presented at the right time, greatly reducing interference with the doctor's work. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0038] Figure 1 A functional block diagram of a text report error correction system integrating LLM and RIS processes provided in an embodiment of this application;

[0039] Figure 2 Example diagram of the report quality control rule configuration interface provided in the embodiments of this application;

[0040] Figure 3Example diagram of the report page operation settings interface provided in the embodiments of this application;

[0041] Figure 4 Example diagrams illustrating the definition of prompt words provided in embodiments of this application;

[0042] Figure 5 Example diagram illustrating an LLM analysis task provided in this application embodiment;

[0043] Figure 6 This is an example diagram illustrating the RIS state triggering configuration provided in the embodiments of this application;

[0044] Figure 7 Example diagram of stored procedure logic provided in the embodiments of this application;

[0045] Figure 8 Example diagram of background service program configuration provided in the embodiments of this application;

[0046] Figure 9 Example diagram of the preliminary report submission node processing logic provided for embodiments of this application;

[0047] Figure 10 Example diagram of the audit report opening node interface provided in this application embodiment;

[0048] Figure 11 This is an example diagram of the audit report closing node report list provided in the embodiments of this application. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0050] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0051] RIS: Radiological Information System.

[0052] LLM: Large Language Model;

[0053] NLP: Natural Language Processing;

[0054] JSON: JavaScript Object Notation, a lightweight data interchange format;

[0055] SQL: Structured Query Language;

[0056] ID: Identification;

[0057] CT: Computed Tomography;

[0058] MR: Magnetic Resonance;

[0059] QC: Quality Control.

[0060] AI: Artificial Intelligence.

[0061] Firstly, such as Figure 1 As shown, this application provides a text report error correction system that integrates LLM and RIS processes, including a task triggering module 10, a rule engine module 20, a large language model scheduling module 30, a prompt word management module 40, and a result processing and feedback module 50.

[0062] The task triggering module 10 is used to listen for and obtain report status change events, and automatically trigger error correction tasks at preset initial report submission nodes, audit report opening nodes, and audit report closing nodes;

[0063] The rule engine module 20 is communicatively connected to the task triggering module 10. After receiving tasks from the preliminary report submission node and the review report closing node, it calls a preset configurable term library and logical rules to perform real-time matching analysis on the report text and obtain the matching results.

[0064] The large language model scheduling module 30 is communicatively connected to the task triggering module 10, and is used to receive task messages, manage the large language model analysis task queue, schedule tasks according to task priority, and generate analysis results.

[0065] The prompt word management module 40 is communicatively connected to the large language model scheduling module 30. It is used to dynamically query and select matching prompt word templates based on the scanning site and equipment type in the patient's examination information, and provide the selected templates to the large language model scheduling module.

[0066] The result processing and feedback module 50 is communicatively connected to the rule engine module 20 and the large language model scheduling module 30, respectively. It is used to receive the matching results sent by the rule engine module and the analysis results output by the large language model scheduling module, and to control the user interface to present error correction information to the user in different interactive ways according to different trigger nodes.

[0067] The text report error correction system provided in this application, which integrates LLM and RIS workflows, utilizes the natural time interval between the initial report and the review report (typically tens of minutes to several hours), allowing ample time to initiate multiple different LLM analysis tasks on the same report. This design allows complex error correction requirements to be broken down into multiple focused cue word tasks that can be executed sequentially or in parallel, avoiding model comprehension bias or "illusion" phenomena that may result from using a single complex cue word, thereby improving accuracy and reliability while ensuring analytical depth.

[0068] In one embodiment, the task triggering module 10 is configured to automatically trigger error correction tasks at three preset key business nodes—the preliminary report submission node, the review report opening node, and the review report closing node—by monitoring changes in the report status of the radiology information system, thereby achieving seamless integration with the RIS process.

[0069] This task triggering module configures a mapping between status identifiers and stored procedures. When a status change occurs, it calls the corresponding stored procedure to write a task record to the task management database. Specifically, it can configure an InterRis trigger table (e.g., T_I_STATUS_SQL) to bind specific report statuses (e.g., report creation F_STATUS_ID:90005, report review F_STATUS_ID:90008) to the stored procedure P_ADD_AI_TASK. When a status change event occurs, the system automatically calls this stored procedure to add a task, where parameter 1 is the patient examination ID and parameter 2 is the trigger status ID.

[0070] In one embodiment, upon receiving tasks from the preliminary report submission node and the review report closing node, the rule engine module 20 calls a preset configurable term library and logical rules to perform real-time matching analysis on the report text and obtain the matching results;

[0071] More specifically, the configurable term library content upon which the rule engine module 20 depends can be customized and maintained by the user, including but not limited to the following types of rules:

[0072] Critical value reporting: life-threatening conditions such as acute pulmonary embolism, acute aortic dissection, DeBakey type I / II dissection, pericardial tamponade, massive pleural effusion / gas / blood, tracheal foreign body, acute cerebral infarction / hemorrhage, brain hernia, gastrointestinal perforation, perforation of hollow viscera, strangulating intestinal obstruction, etc.;

[0073] Device type contradiction: for example, CT reports usually should not describe MR-related terms such as "diffusion restriction", "signal", "long T / short T", etc.; MR reports usually should not describe CT-related terms such as "density", "CT value", "HU", etc.;

[0074] Gender contradiction: for example, descriptions of female-specific organs such as "uterus", "ovary", etc. should not appear in reports of male patients; descriptions of male-specific organs such as "prostate", "seminal vesicle", etc. should not appear in reports of female patients;

[0075] Age contradiction: for example, rationality verification of the same patient age for common diseases in infants (neuroblastoma, retinoblastoma, rhabdomyosarcoma) and common diseases in adolescents (Ewing's sarcoma, osteosarcoma);

[0076] Measurement value contradiction: for example, unit consistency check (CM and MM cannot be used simultaneously), rationality judgment of measurement value range, identification of unreasonable measurement units (such as M);

[0077] Typo detection: for example, common typos such as "cone", "diaphragm", "mediastinum", "course", "meridian", etc.;

[0078] And duplicate character detection.

[0079] These rules can be added, deleted, or modified through the configuration interface to achieve flexible customized quality control. This type of analysis is mainly implemented based on conventional NLP techniques such as strict word matching or synonym / near-synonym search, with a fast response speed, enabling sub-second feedback, and being able to prompt the reporting doctor to immediately correct the detected simple errors, usually arranged to run synchronously at each report operation node.

[0080] In one embodiment, the large language model scheduling module 30 is used to receive task messages and manage the large language model analysis task queue, classifying tasks into two priorities: immediate tasks and in-depth analysis tasks;

[0081] Immediate tasks are given the highest priority and are configured to be preferentially inserted to the front of the task queue for priority execution. Such tasks are usually large language model prompt analysis tasks that need to be processed with the highest priority to ensure that the analysis results of key nodes can be quickly returned;

[0082] Deep analysis tasks have a relatively low priority and are mostly large language model cue word analysis tasks, but they adopt a first-in-first-out scheduling strategy, waiting in a regular queue for the large language model hardware resources to process them.

[0083] In one embodiment, whether it's the analysis by the rule engine module 20 based on "program logic" or the analysis results from the large language model scheduling module 30 using LLM prompt words, the final output is standardized into a unified structured data format (such as JSON). Based on the different timeliness requirements of the tasks, the system divides the analysis tasks into the following three categories:

[0084] Real-time tasks: These are typically executed by the rule engine module 20. They perform matching analysis based on a preset configurable term library and logical rules, resulting in high efficiency and sub-second feedback. These tasks are usually scheduled to run synchronously at key nodes in each report operation (such as initial report submission and review report closure) to enable the immediate detection and prevention of critical errors.

[0085] Timely tasks: These are typically LLM prompt word analysis tasks and are given the highest priority. At critical junctures such as when audit reports are closed, these tasks are inserted at the head of the message queue to ensure they are consumed and executed first. The goal is to return analysis results within 1-2 minutes to support rapid report release decisions.

[0086] Deep analysis tasks: These are most likely LLM cue word analysis tasks. These tasks are first-in, first-out, and queued up for analysis by LLM hardware.

[0087] In one embodiment, the prompt word management module 40 is used to dynamically query and select the most matching prompt word template based on patient examination information (such as scan site and device type), and provide the selected template to the large language model scheduling module 30. This module matches fields in the patient information database (such as device type and diagnostic conclusion) with preset filtering conditions in the prompt word template library by executing predefined structured query language statements; specifically, it involves creating a prompt word definition table (e.g., T_LLM_PROMPT_LIST), which contains key fields: prompt word primary key ID (F_PROMPT_ID), prompt word name (F_PROMPT_NAME), prompt word content (F_PROMPT_TEXT), task type (F_TYPE_ID, such as 1 for immediate tasks and 2 for in-depth tasks), and custom SQL conditions (F_PROMPT_FITER). The F_PROMPT_FITER field is used to configure the filtering logic, thereby achieving dynamic and accurate matching of prompt word templates.

[0088] Furthermore, when the timely task is triggered at the closing node of the audit report, the prompt word management module or related logic can also dynamically select a more targeted prompt word template based on the existing historical LLM analysis results of the report, thereby realizing the iteration and deepening of the analysis process;

[0089] Different prompt word templates encapsulate different analysis rules and focuses. Through the above mechanism, the system can use prompt words to report and correct errors, broadly covering various synonyms / near-synonyms, and can discover errors such as typos that are obvious to humans but difficult to detect through fixed rules, thereby significantly improving the coverage and intelligence level of error correction.

[0090] In one embodiment, the result processing and feedback module 50 is used to receive real-time matching results from the rule engine and structured analysis results from the large language model, and control the user interface to present error correction information to the user in a corresponding interactive manner according to different trigger nodes. Specifically, this is implemented as follows:

[0091] At the initial report submission node, when the report is saved, this module will control the interface to pop up a prompt box based on the real-time analysis results of the rules engine, requiring the reporting doctor to immediately address any simple errors found.

[0092] At the point where the review report is opened, the results processing and feedback module will automatically check if there is a Large Language Model (LLM) deep analysis task associated with and completed for that report. If such a result exists, the module will proactively prompt the reviewer through the user interface. These analysis results originate from the deep analysis performed in the background by LLM through an asynchronous task queue after the initial report submission, aiming to reveal potential deep-seated errors or inconsistencies;

[0093] At the report closing node, when the report is saved after review, the system first calls the rules engine for real-time verification and displays a pop-up prompt. The report is then moved to the background buffer and not published. The system executes two analyses in parallel: a rapid review by the rules engine, and a high-priority LLM timely task initiated based on the historical Large Language Model (LLM) analysis results, dynamically selecting a dedicated prompt word library (requiring feedback within 1-2 minutes). If all analyses are normal, the report is automatically published; if any analysis finds a potential error, the system will keep the report in a "buffered" state and highlight it in the report list (e.g., in red), pending manual confirmation and publication by the reviewing physician after handling the current patient.

[0094] This process, through the collaboration of a rules engine and an LLM, balances real-time performance with in-depth analysis capabilities, and utilizes state control and interface prompts to ensure that doctors have the final say on abnormal results.

[0095] Secondly, this application provides a text report error correction method that integrates LLM and RIS processes, applied to the text report error correction system that integrates LLM and RIS processes as described above; the method includes the following steps:

[0096] Step S1: By monitoring the report status changes in the radiology information system, generate corresponding error correction task events at key nodes such as preliminary report submission, review report opening, and review report closing;

[0097] Step S2: Based on the type of the error correction task event, route it to the real-time rule processing flow or the large language model asynchronous analysis flow;

[0098] Step S3: For task events routed to the real-time rule processing flow, trigger the rule matching operation on the report text and generate real-time matching results;

[0099] Step S4: For task events routed to the asynchronous analysis process of the large language model, place them into the task queue for scheduling according to their priority attributes;

[0100] Step S5: For the task events to be executed in the task queue, dynamically select an appropriate prompt word template, and combine the report content with the prompt word template to form a large language model analysis instruction;

[0101] Step S6: Submit the large language model analysis command to the large language model service and obtain the returned structured analysis results; typically, the structured analysis results are saved to the relevant fields in the task table;

[0102] Step S7: Combine the real-time matching results and structured analysis results with the current process node of the report to determine the final information prompt method and control the flow of the report status.

[0103] This application, through tight integration with the RIS system and the deployment of different types of error analysis tasks at three key business nodes, ensures that the application of large language models no longer impacts the efficiency of business processes. This approach allows the system to continuously improve report quality without interfering with doctors' existing work patterns.

[0104] By organically combining various error correction functional modules with the RIS process and decomposing the report error correction task into multiple nodes, with each node focusing on different types of error correction, the deep analysis capabilities of the large language model are fully utilized without affecting business efficiency.

[0105] Since there is usually a gap of tens of minutes or even hours between the preliminary report and the review report, the backend has ample time to conduct in-depth analysis of the report text through concurrent large language model analysis tasks, making it possible to use multiple prompt words to correct errors in the same report.

[0106] By fully utilizing information such as scan site, equipment type, and diagnosis type to dynamically select different prompt word libraries, the accuracy of the analysis is improved. For example, the complexity varies greatly for different diseases in different sites. This solution dynamically selects prompt word templates based on provided information such as scan site, equipment type, and diagnostic conclusion. Based on the aforementioned multi-task analysis framework, it can select personalized prompt word combinations for analysis based on multiple diagnostic conclusions. This simplifies the difficulty of prompt word design and improves the accuracy of the analysis.

[0107] In one embodiment, step S1: By monitoring changes in the report status of the Radiology Information System (RIS), a corresponding error correction task event is generated at a preset key node, specifically including the following steps:

[0108] Step S11: At the initial report submission node, upon clicking the report submission button, the background uses "program logic" to immediately perform error correction analysis within seconds. This type of analysis is usually based on conventional NLP techniques such as strict word matching or synonym / near-synonym search, which can require the doctor making the initial report to immediately correct simple errors;

[0109] Step S12: At the "Open Review Report" node, when the review report is opened, a notification will appear indicating deep-seated errors discovered by the LLM (Low-Level Management). These errors are analyzed and processed by the LLM in the background after the initial report submission. This function provides important reference information for the reviewing physicians.

[0110] Step S13: At the closed node of the audit report, the rule engine is first triggered to perform real-time analysis and display the results; then, the report is placed in the background buffer and not published temporarily, while the rule engine reviews and LLM timely task analysis are performed in parallel. If the analysis is satisfactory, it is submitted directly; if there are problems, the audit doctor is reminded to confirm.

[0111] When a task submitted for review is placed in the buffer, the following two types of error correction analysis are performed:

[0112] First, review the report again using "program logic";

[0113] Secondly, based on the results of the previous LLM analysis in the background, a prompt word library specifically for review reports is dynamically selected, and this task is made the highest priority and immediately put into the LLM for judgment, ensuring that feedback is received within 1-2 minutes.

[0114] In one embodiment, step S2: based on the type of the error correction task event, route it to different processing flows.

[0115] Step S21: Route some tasks that require real-time response, such as the preliminary report submission and the review report closing node, to the real-time rule processing flow.

[0116] Step S22: Allow task events for in-depth analysis, such as preliminary report submission and review report opening nodes, to be routed to the asynchronous analysis process of the large language model.

[0117] In one embodiment, step S3: For the task event routed to the real-time rule processing flow, a fast rule matching operation is triggered on the report text, and a real-time matching result is generated.

[0118] In one embodiment, step S4: For the task events routed to the asynchronous analysis process of the large language model, schedule them according to their attributes.

[0119] Step S41: Determine task priority. Mark tasks requiring rapid feedback, such as the audit report closing node, as immediate tasks; mark tasks that can be delayed and are triggered after the initial report is submitted as in-depth analysis tasks.

[0120] Step S42: Insert timely tasks into the head of the LLM task queue for priority scheduling; place deep analysis tasks into the tail of the queue for sequential scheduling.

[0121] Optionally, in asynchronous analysis task scheduling, it is supported to initiate multiple prompt word template call requests for different analysis dimensions (such as typos, logical contradictions, and diagnostic consistency) sequentially or in parallel for the same report, so as to realize multi-angle in-depth analysis.

[0122] In one embodiment, step S5: dynamically prepare analysis instructions for the LLM task event to be executed.

[0123] Step S51: Dynamically select an appropriate prompt word template based on the patient's examination information (such as the scanning site, equipment type, and diagnosis conclusion).

[0124] Step S52: Combine the report text content with the selected prompt word template to form the final large language model analysis instruction.

[0125] In one embodiment, step S6: Submit the large language model analysis command to the large language model service and obtain the returned structured analysis results (usually in JSON format). The analysis results are saved to the corresponding field of the task management table for subsequent retrieval and display.

[0126] In one embodiment, step S7 is specifically implemented as follows:

[0127] By combining the real-time matching results and the structured analysis results, and considering the specific process node of the current report, the final information prompt method is determined, and the final flow of the report status is controlled.

[0128] This application achieves the following significant advantages and positive effects through the above system and method:

[0129] By deploying different types (real-time, timely, and in-depth) error correction tasks on multiple business nodes, the time difference between business processes is cleverly utilized, so that the in-depth analysis of LLM no longer blocks the real-time business process.

[0130] By prioritizing tasks and managing queues, the use of limited LLM computing resources was optimized.

[0131] By dynamically selecting dedicated prompt word templates based on patient and device information, the accuracy and relevance of LLM analysis in specific medical scenarios are significantly improved;

[0132] The system's triggering and feedback mechanisms are deeply embedded in the RIS standard process, minimizing interference with doctors' work.

[0133] Both the rule base and the prompt word templates are configurable, allowing the system to be continuously optimized as medical knowledge is updated and error correction experience is accumulated.

[0134] To illustrate the implementation of the present invention in more detail, the following describes the practical application of the above system architecture and method in conjunction with a typical embodiment.

[0135] A text report error correction method integrating LLM and RIS processes is implemented as follows:

[0136] Step 1: Implement program logic by configuring customizable rules for maintaining term content, specifically including critical value reporting, equipment type conflicts, gender conflicts, age conflicts, measurement value conflicts, typo detection, and duplicate word detection; the report quality control rule configuration interface is as follows: Figure 2 As shown;

[0137] Step 2: Report page operation settings include prompt method, prompt content, judgment result display, continue operation, detection order, description, trigger time, conditions, and execution time, etc.; the report page operation settings interface is as follows: Figure 3 As shown;

[0138] Step 3: Create a prompt word definition table T_LLM_PROMPT_LIST, containing fields such as prompt word primary key ID, name, content, task type, and custom SQL conditions; See the image for an example of the prompt word definition table. Figure 4 As shown;

[0139] Step 4: Create the task table T_R_QC_REPORT_AIINFO, containing fields such as task primary key, prompt word ID, user ID, patient ID, task status, returned image manifestations, and diagnostic results; LLM analysis task table diagram, for example. Figure 5 As shown;

[0140] Step 5: Configure the InterRis trigger table T_I_STATUS_SQL to add tasks by executing stored procedures for specific report states; RIS state trigger configuration table diagram is shown below. Figure 6 As shown;

[0141] Step 6: Add a task using the stored procedure P_ADD_AI_TASK. Determine the task type based on the trigger status ID and retrieve the matching prompt word template based on the patient information. An example diagram of the stored procedure logic is shown below. Figure 7 As shown;

[0142] Step 7: Configure the interface address for the backend service program, prioritize the execution of timely tasks, replace variables, submit the LLM, and save the returned results; See the example diagram for the backend service program configuration. Figure 8 As shown;

[0143] Step 8: At the initial report submission node, a pop-up window will display the error correction results in real time, and an LLM task will be added simultaneously; the processing logic for the initial report submission node is as follows: Figure 9 As shown;

[0144] Step 9: In the "Open Report" section, display the completed LLM analysis results for the doctor to process; an example of the "Open Report" interface is shown below. Figure 10 As shown;

[0145] Step 10: At the review report closing node, a prompt is first issued based on the real-time analysis results of the rules engine; then the report enters a buffer state, triggering parallel rule engine review and high-priority LLM timely task analysis. Finally, based on all analysis results, a decision is made on whether to automatically publish or mark as abnormal and await doctor confirmation; an example image of the report list at the review report closing node is shown below. Figure 11 As shown.

[0146] In summary, this solution successfully transforms the powerful analytical capabilities of large language models into a stable, efficient, and non-disruptive productivity tool for medical report quality control practices.

[0147] The implementation of each step in the above method corresponds to each module in the above system embodiment, and the implementation process will not be described in detail here.

[0148] Thirdly, embodiments of this application provide a text report error correction device that integrates LLM and RIS processes. The text report error correction device that integrates LLM and RIS processes can be a personal computer (PC), laptop computer, server, or other device with data processing capabilities.

[0149] In this embodiment of the application, the text report error correction device that integrates LLM and RIS processes may include a processor, a memory, a communication interface, and a communication bus.

[0150] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0151] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces for interconnecting devices within the text report correction device that integrates LLM and RIS processes, as well as interfaces for interconnecting the text report correction device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0152] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0153] The processor can be a general-purpose processor, which can call a text report error correction program that integrates LLM and RIS processes stored in memory and execute the text report error correction method for integrating LLM and RIS processes provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the text report error correction program for integrating LLM and RIS processes is called can be referred to the various embodiments of the text report error correction method for integrating LLM and RIS processes in this application, and will not be repeated here.

[0154] Fourthly, embodiments of this application also provide a readable storage medium.

[0155] The present application stores a text report error correction program that integrates LLM and RIS processes on a readable storage medium, wherein when the text report error correction program that integrates LLM and RIS processes is executed by a processor, it implements the steps of the text report error correction method that integrates LLM and RIS processes as described above.

[0156] The method implemented when the text report error correction program integrating LLM and RIS processes is executed can be referred to in the various embodiments of the text report error correction method integrating LLM and RIS processes in this application, and will not be repeated here.

[0157] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A text report error correction system integrating LLM and RIS processes, characterized in that, The system includes: The task triggering module is configured to respond to report status change events in the radiology information system and automatically generate error correction tasks at three preset process nodes: initial report submission, review report opening, and review report closing. The rules engine module, connected to the task triggering module, is configured to call a pre-configured rule base to perform real-time analysis of the report text and generate a first correction result at the initial report submission node and the review report closing node. The large language model scheduling module is connected to the task triggering module and is configured to manage the LLM analysis task queue and schedule task execution according to task priority. The tasks include in-depth analysis tasks triggered at the preliminary report submission node and timely analysis tasks triggered at the review report closing node. The prompt word management module is connected to the large language model scheduling module and is configured to dynamically select matching prompt word templates from the prompt word template library based on the patient examination information associated with the report for use by the large language model scheduling module. The result processing and feedback module is connected to the rule engine module and the large language model scheduling module, respectively. It is configured to receive the first error correction result and the second error correction result from the LLM, and control the user interface to perform corresponding interactive prompts and report status management based on the current process node.

2. The text report error correction system integrating LLM and RIS processes according to claim 1, characterized in that, The task triggering module maintains a mapping relationship between report status identifiers and database stored procedures, and calls the corresponding stored procedure to create a task record when the status changes.

3. The text report error correction system integrating LLM and RIS processes according to claim 1, characterized in that, The large language model scheduling module is configured as follows: The task triggered by the initial report submission node is set as a deep analysis task and added to the regular task queue for asynchronous processing; The task triggered by the closed audit report node is set as an immediate analysis task and is prioritized and inserted at the front of the task queue for immediate processing.

4. The text report error correction system integrating LLM and RIS processes according to claim 1, characterized in that, The prompt word management module executes structured query statements to match the device type field and / or diagnosis conclusion field in the patient information database with preset filtering conditions in the prompt word template library in order to select an applicable prompt word template.

5. The text report error correction system integrating LLM and RIS processes according to claim 1, characterized in that, At the audit report closing node, the result processing and feedback module is configured as follows: When a bug correction task is triggered, the report is placed in a buffered state; The system receives the real-time analysis results from the rule engine module and the timely analysis results from the large language model scheduling module in parallel. If neither the real-time analysis result nor the timely analysis result indicates an error, the report status will be automatically updated to the published status. If either party's analysis results indicate an error, the report will remain in a buffer state and a warning will be displayed in the user interface. It can only be published after user confirmation.

6. The text report error correction system integrating LLM and RIS processes according to claim 1, characterized in that, The rule engine module calls a rule base containing customizable terms and logical rules, which are used to perform at least one verification, including gender conflict, device type conflict, and critical value judgment.

7. A text report error correction method integrating LLM and RIS processes, characterized in that, The method, applied in a text report error correction system integrating LLM and RIS processes as described in any one of claims 1-6, comprises: Monitor RIS report status changes and trigger error correction tasks at the nodes of initial report submission, review report opening, and review report closing. Depending on the triggering node, the task is routed to the rule engine for real-time analysis and / or to the LLM task queue for asynchronous analysis; For tasks routed to the LLM task queue, scheduling is performed based on their priority attributes, and prompt word templates are dynamically matched for the tasks to invoke the LLM service for analysis; Integrate real-time analysis results from the rules engine and asynchronous analysis results from the LLM; Based on the integrated results and the current node of the report, control the flow of the user interface and the report status.

8. The text report error correction method integrating LLM and RIS processes according to claim 7, characterized in that, At the initial report submission node, the rule engine is triggered to perform real-time analysis and create an LLM deep analysis task; The completed LLM analysis results are displayed in the audit report open node; At the closed node of the audit report, the rule engine is triggered to perform real-time analysis and create an LLM timely analysis task, and the report release process is determined based on the analysis results of both.

9. The text report error correction method integrating LLM and RIS processes according to claim 7, characterized in that, In the asynchronous analysis step, multiple LLM analysis tasks are created for the same report, each using a different prompt word template to analyze different business dimensions.

10. A non-volatile computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 7 to 9.

Images