A Medical Image Report Quality Control System Based on a Local Large Language Model
By using a medical image report quality control system based on a local large language model, automated, traceable, and accurate summarization and semantic quality control have been achieved. This has solved the problems of cognitive load and semantic errors in the medical image report generation process, and improved the efficiency and accuracy of quality control.
Patent Information
- Application Number
- CN202611353778.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-02
- Publication Date
- 2026-11-17
AI Technical Summary
The medical image report generation process suffers from problems such as excessive cognitive load, difficulty in capturing semantic errors, and a large workload for manual comparison that depends on clinical expertise. Existing report quality control systems cannot effectively address these issues.
Design a medical image report quality control system based on a local large language model, including a global interaction platform, a data source layer, a core control and large language model processing layer, and a data persistence layer. It realizes automated, traceable, and accurate summarization, semantic quality control, and long-term conclusion verification, and has data security, full-process automation, and high configurability.
It alleviates the problems of existing technologies, provides a fully automated quality control solution, improves the efficiency and accuracy of report quality control, reduces reliance on manual labor, and has data security and low-cost deployment capabilities.
Smart Images

Figure CN123120179A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medical image report quality control, specifically to a medical image report quality control system based on a local large language model. Background Technology
[0002] The process of generating a medical imaging report typically includes: the radiologist reading the patient's historical medical records in the Hospital Information System (HIS) to obtain clinical context, reviewing the images and writing the report's findings and conclusions, submitting it for review by a senior physician, and finally issuing it. Within this process, there are three long-standing problems that cannot be systematically resolved manually: First, the cognitive load is too high. Hundreds of thousands of words of unstructured text cannot be read completely in a limited time, and key information that has been out of the public for a long time will inevitably be missed. Second, existing report quality control systems can only perform mechanical comparisons at the string level. However, the real risks in medical imaging reports come from semantic errors, which are difficult to capture by traditional string matching systems. Third, the traditional method relies on manual comparison one by one, which requires reviewing report snapshots and subsequent medical records at the same time, resulting in a huge workload (the retrospective analysis of a report often takes 20-30 minutes). Routine retrospective analysis is equivalent to reviewing each historical report again, and it requires a very high level of clinical skills and workload from the executor. Summary of the Invention
[0003] This application provides a medical image report quality control system based on a local large language model. Based on a locally deployed lightweight large language model (LLM), a medical image report quality control system architecture is designed. This system encompasses automated, traceable, and accurate summarization of clinical background and pre-screening of patient consultations, automated semantic quality control of report documents, and automated verification of the accuracy of long-term conclusion predictions. Furthermore, the system allows for data and algorithm reuse across all levels, alleviating the problems associated with existing technologies. It also possesses numerous advantages, including data security, end-to-end automation, traceability of content generated by the large language model, high configurability, cross-modal transferability, low deployment cost, and model replaceability, providing excellent tool support for medical image reporting.
[0004] This application provides a medical image report quality control system based on a local large language model. The system includes: A global interactive platform enables human-computer interaction in the doctor's report workbench; The data source layer connects to the hospital information system to obtain the patient's full historical medical record data and connects to the imaging report system to obtain medical imaging report data as data input; The core control and large language model processing layer includes a task scheduling engine, a large language model processing pipeline, basic shared components, and a large language model service. The task scheduling engine performs nine-level priority task queue management, version-aware dependency chain management, and day / night execution window management. The large language model processing pipeline includes, in sequence, a summary module for processing medical history summaries based on data input, a quality control module for implementing multi-dimensional semantic report quality control, and an evaluation module for implementing retrospective assessment. The basic shared components include a fingerprint engine for processing data content fingerprints, an incremental medical history extraction pipeline for extracting incremental medical history, a department memory bank for maintaining department-level experience rules, and a privacy desensitization module for desensitizing privacy data. The large language model service provides the summary module, quality control module, and evaluation module with a locally deployed large language model inference interface. The data persistence layer stores core data in a local system database for reading and writing by the large language model processing pipeline and for front-end querying and display.
[0005] In one exemplary embodiment, the human-computer interaction functions include displaying analysis results, monitoring task queues, operating the quality control management panel, and providing feedback and corrections from doctors. The patient's complete historical medical record data includes outpatient documents, inpatient documents, laboratory reports, and examination reports; The core data includes structured traceable medical history, structured traceable incremental medical history, snapshots of quality control results from previous tests, retrospective analysis results, and fingerprint tables recorded in the overall hospital information system.
[0006] In yet another exemplary embodiment, for the summary module, the simplified path strategy corresponding to a total word count less than a threshold includes: The patient's complete historical medical record data, including outpatient documents, inpatient documents, laboratory reports and examination reports, the traceability ID of the patient's complete historical medical record data, and the clinical history in medical imaging report data, are structured and merged into a single input. A single input is submitted to the model for processing to generate a structured medical history summary and pre-screening judgment, which are output as the medical history summary. At the same time, the fact fragment source ID is output. Then, the fact fragments corresponding to the fact fragment source ID and the medical history summary are submitted to the model again for fact fragment matching. The pre-screening judgment is used to pre-screen the sufficiency of the medical history in two cases: sufficient and necessary to take a medical history.
[0007] In yet another exemplary embodiment, for the summary module, the batching path strategy corresponding to a total word count not less than a threshold includes: During the initial draft generation stage, outpatient and inpatient documents are combined to obtain the initial draft of the medical history. During the supplementary extraction phase, under the batch parallel processing mechanism, outpatient documents, inpatient documents, and draft medical histories are submitted to the model for processing in batches to extract relevant clinical findings that were missed in the draft medical histories, forming fact fragments with traceability IDs, and obtaining supplementary medical history findings. In the final draft stage of medical history, the initial draft of medical history and the supplementary findings of medical history are merged to obtain a complete final draft of medical history, while the factual fragments with traceability IDs that are deemed valuable are selected. During the anomaly screening phase, under the batch parallel processing mechanism, with the final draft of the complete medical history as the background, the test reports and examination reports are submitted to the model for processing in batches to screen positive and abnormal results related to the current diagnosis. The output format is fact fragments with traceability ID. In the final summary stage, the complete medical history, positive results, and abnormal results are submitted to the model for further refinement, and corresponding structured medical history summaries and pre-screening judgments are generated. In the source ID restoration stage, the factual fragments with source IDs output from the structured medical history summary stage, the final draft of the medical history stage, and the factual fragments with source IDs output from the anomaly screening stage are submitted to the model to generate source fragments for locating the structured medical history summary. The clinical history in the medical imaging report data is included in separate chapters in each of the six stages.
[0008] In yet another exemplary embodiment, the amount of input during the draft generation stage is controlled by a first preset word limit, and if it exceeds the limit, it is truncated in reverse order of the medical record date. In the concurrent batch control strategy of the batch parallel processing mechanism, the capacity of each batch of text is controlled by the upper limit of the second preset number of characters, and there is no shared state between batches. After the supplementary extraction and anomaly screening phases are completed, if the total output exceeds the preset word limit, a parallel compression call will be initiated. Each stage of the model call requires output in structured JSON format and declaration of required fields. If parsing fails, an automatic repair tool is invoked to handle common formatting errors.
[0009] In yet another exemplary embodiment, the quality control module includes: During the submission review and quality control stage, the initial quality control review is triggered after the report status changes to the submission review status and the cooling-off period is met. The corresponding historical version quality control records are entered, the evaluation conclusion is output, and then the re-review mechanism is initiated. In the audited quality control stage, a second quality control audit is triggered after the report status changes to audited and the record submitted for audited quality control reaches the final state. The corresponding historical version of the quality control record is input, the evaluation conclusion is output, and then the re-review mechanism is initiated. The processing of the submitted quality control stage and the audited quality control stage are independent of each other.
[0010] In yet another exemplary embodiment, the submission review quality control stage and the reviewed quality control stage employ an incremental medical history extraction mechanism based on content fingerprints to obtain medical record data after the inspection date, including: During the data retrieval phase, the start and end points of the time window are determined. The start point of the time window extends backward from the inspection date, and the end point of the time window corresponds to the submission time for the review and quality control phase, specifically the submission time for review, and to the review and quality control phase, specifically the time when the review is passed. During the fingerprint difference filtering stage, a fingerprint table of the hospital information system is maintained. The records retrieved this time are compared with the fingerprint table of the hospital information system, and only records that do not exist in the fingerprint table of the hospital information system are returned. In the semantic extraction stage, the existing medical history summary is used as the baseline, and the original hospital information system text is added as the incremental input. The model is called to extract incremental clinical information not covered in the existing summary. During the unified append storage phase, successfully added incremental content is extracted, appended to the currently checked incremental medical history storage field with the date as the title; During the fingerprint persistence stage, after successful extraction, the newly recorded fingerprint content is written into the fingerprint record table of the hospital information system. The submission review quality control stage, the reviewed quality control stage, and the evaluation module all use the same three shared extraction functions, but with different time window parameters.
[0011] In yet another exemplary embodiment, quality control of medical image report data includes: In the classification stage, the image layout type is identified and the lesion area is located; During the cropping stage, the lesion area is cropped into independent sub-images based on the layout type classification results from the classification stage. In the single-graph evaluation phase, each independent subgraph is evaluated independently. In the summary phase, the evaluation results of all independent subgraphs are combined to generate the final conclusion, which is then merged with the text quality control conclusion, depending on whether merging is configured.
[0012] In yet another exemplary embodiment, the evaluation module includes: Once the quality control report is completed, a retrospective analysis is triggered to generate retrospective records; Retrospective analysis includes the following processes: In the baseline medical history construction phase, the medical history summary output by the summary module, all incremental medical histories extracted by the quality control module, and the clinical medical history in the medical imaging report data are merged to obtain the baseline medical history. In the subsequent stage of acquiring new medical history, the shared extraction function of the quality control module is reused to extract the incremental medical history. In the model comparison and judgment phase, the baseline medical history known at the time and the subsequent incremental medical history that appeared afterward are input into the model to judge whether the report conclusion at that time was accurate in predicting the nature of the main lesions, and to obtain the judgment result.
[0013] From the above, it can be concluded that this application has the following beneficial effects: Under the goal of quality control of medical imaging reports, this application designs a medical imaging report quality control system architecture based on a locally deployed lightweight large language model. It covers automated, traceable, and accurate summarization of clinical background and pre-screening of consultations, automated semantic quality control of report documents, and automated verification of the accuracy of long-term conclusion predictions. The system also features data and algorithm reuse between each level, which alleviates the problems associated with existing technologies. It has many advantages such as data security, full-process automation, traceability of content generated by the large language model, high configurability, cross-modal transferability, low deployment cost, and model replaceability, providing good tool support for medical imaging reporting. Attached Figure Description
[0014] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0015] Figure 1 This is a schematic diagram of a system architecture for a medical image report quality control system based on a local large language model, as described in this application. Detailed Implementation
[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0017] Under the goal of medical image report quality control, this application provides a medical image report quality control system based on a local large language model, referencing... Figure 1 The diagram shown represents a system architecture of the medical image report quality control system based on a local large language model, which may specifically include a global interaction platform, a data source layer, a core control and large language model processing layer, and a data persistence layer.
[0018] Next, we will introduce the components of these major systems in detail: 1) Global Interaction Platform A global interactive platform enables human-computer interaction in the doctor's report workbench.
[0019] Understandably, in the context of medical image report quality control, the global interactive platform serves as a doctor's report workbench for doctors, where all outputs from large language models ultimately converge for review and decision-making by doctors.
[0020] Furthermore, the human-computer interaction functions involved may specifically include functions or specific human-computer interaction operations such as displaying analysis results, monitoring task queues, operating the quality control management panel, and providing feedback and corrections from doctors.
[0021] 2) Data Source Layer The data source layer connects to the hospital information system (HIS) to obtain the patient's full historical medical record data and connects to the imaging report system to obtain medical imaging report data as data input.
[0022] Understandably, the data input for this application includes not only medical image report data, but also the patient's complete historical medical record data. High-quality medical image quality control processing is carried out through these two categories of source data.
[0023] In practice, all patient historical medical records and medical imaging report data can be encapsulated using the adapter pattern, without relying on the implementation details of any specific vendor.
[0024] Furthermore, the patient's complete historical medical record data may include outpatient documents, inpatient documents, laboratory reports and examination reports, etc. Medical imaging report data specifically refers to PET / CT image report data, CT image report data, MRI image report data, ultrasound image report data or DR image report data, which are used as quality control targets.
[0025] Among them, CT, PET / CT, MRI and DR are common or mature imaging types in clinical practice. Their specific full names in Chinese and English are Computed Tomography (CT), Positron Emission Tomography / Computed Tomography (PET / CT), Magnetic Resonance Imaging (MRI) and Digital Radiography (DR), respectively.
[0026] In current processing, medical image report data typically only involves one of the above-mentioned types of images.
[0027] At the same time, it should be noted that medical imaging report data can include not only basic report images / pictures, but also relevant examination information and report text (such as subsequent clinical history) and other related data.
[0028] In addition, it is understandable that the patient's full historical medical records and medical imaging reports contain different types of data at the detail level, which usually also include basic information about the patient such as gender.
[0029] 3) Core Control and Large Language Model Processing Layer The core control and large language model processing layer includes a task scheduling engine, a large language model processing pipeline, basic shared components, and a large language model service. The task scheduling engine performs nine-level priority task queue management, version-aware dependency chain management, and day / night execution window management. The large language model processing pipeline includes, in sequence, a summary module for processing medical history summaries based on data input, a quality control module for implementing multi-dimensional semantic report quality control, and an evaluation module for implementing retrospective assessment. The basic shared components include a fingerprint engine for processing data content fingerprints, an incremental medical history extraction pipeline for extracting incremental medical history, a department memory bank for maintaining department-level experience rules, and a privacy desensitization module for desensitizing privacy data. The large language model service provides the summary module, quality control module, and evaluation module with a locally deployed large language model inference interface.
[0030] It is readily apparent that the core control and large language model processing layer constitute the core of this application's system, specifically comprising four sub-layers: a task scheduling engine, a large language model processing pipeline, basic shared components, and a large language model service. These sub-layers correspond to: The task scheduling engine manages task queues with nine priority levels (P0-P8), corresponding to the core control and large language model processing layers, which are processed as different tasks. The priority of each task can be freely adjusted through the administrator interface. For example, within the subsequent quality control module's submission review quality control stage and the reviewed quality control stage, the priority is divided into recent and historical sub-priorities based on whether the inspection date is more than 3 days ago, to ensure that the current report is processed first.
[0031] In addition, the task scheduling engine is also used to manage version-aware dependency chains, dispatching tasks in sequence according to the progressive dependency relationship of disease history summary → submission for review and quality control → reviewed quality control → retrospective verification (corresponding to the specific workflow of the large language model processing pipeline).
[0032] In addition, the task scheduling engine supports a configurable nighttime execution window, which restricts low-priority, high-volume tasks (such as retrospective verification) to nighttime execution to avoid daytime resource contention.
[0033] The large language model processing pipeline, according to the workflow or the aforementioned progressive dependency relationship, can sequentially include a summary module for processing medical history summaries based on data input, a quality control module for implementing multidimensional semantic report quality control, and an evaluation module for implementing retrospective assessment. These are the three core modules involved in the system of this application, and their specific work content will be described in detail later.
[0034] The basic shared components serve to provide components for invocation, offering common capabilities to upper-layer modules. Specifically, they can provide components such as a fingerprint engine for processing data content fingerprints (which may involve SHA-256 content hashing and difference algorithms), an incremental medical history extraction pipeline for extracting incremental medical history (which may involve three-layer reuse), a department memory bank for maintaining department-level experience rules (which may involve large language model-driven automatic selection and injection), and a privacy desensitization module for desensitizing privacy data (optional automatic desensitization), etc. The large language model service can be understood using a locally deployed open-source model inference platform, such as the Qwen3.5 27B model (the specific model can be adjusted), thus providing the necessary large language model inference capabilities for all large language model tasks.
[0035] 4) Data persistence layer The data persistence layer stores core data in a local system database for reading and writing by the large language model processing pipeline and for front-end querying and display.
[0036] The data persistence layer (Sink) is easy to see. Based on the system's local database (such as SQLite database), it implements the system's data persistence function to meet the data usage needs of functions such as reading and writing in the large language model processing pipeline (corresponding to the core control and large language model processing layer) and front-end query and display (corresponding to the global interactive platform).
[0037] Furthermore, the core data involved in the processing of dynamics / results may include structured traceable medical history, structured traceable incremental medical history, snapshots of previous quality control results, retrospective analysis results, and fingerprint tables recorded in the overall hospital information system.
[0038] Next, we will elaborate on the core control and large language model processing layer, which are the core of this application system, in order to better understand the high-quality control performance of this application system.
[0039] In the core control and large language model processing layer, the summary module in the large language model processing pipeline can realize the traceable and accurate summary and pre-screening of ultra-long medical histories specifically designed in this application. Specifically, this can be achieved through the following two specific settings: 1.1) For the summary module in the large language model processing pipeline, based on the local large language model and under the adaptive splitting mechanism, the simplified path strategy for total character counts less than a threshold (e.g., 6000 characters) may include the following specific processing content: The patient's complete historical medical record data, including outpatient documents, inpatient documents, laboratory reports and examination reports, the traceability ID of the patient's complete historical medical record data, and the clinical history in medical imaging report data, are structured and merged into a single input. The single input is structured and split before being submitted to the model for processing to generate a structured medical history summary and pre-screening judgment as the output of the medical history summary. At the same time, the fact fragment source ID is also output. Then, the fact fragments corresponding to the fact fragment source ID and the medical history summary are submitted to the model again for fact fragment matching. The pre-screening judgment is used to pre-screen the sufficiency of the medical history based on two situations: sufficient and necessary to take a medical history.
[0040] Here, ID stands for Identification, or identifier.
[0041] It should be noted that, regarding the local large language model mentioned here, in practice, different models calling the service may use the same model or different models, which can be adjusted according to the actual situation.
[0042] 1.2) For the summary module in the large language model processing pipeline, based on the local large language model and under the adaptive splitting mechanism, the corresponding batching path strategy with a total word count not less than a threshold, the corresponding processing content is a five-step pipeline processing, which may include: In the initial draft generation stage, outpatient and inpatient documents are combined to obtain an initial draft of medical history that provides a basic framework for subsequent model processing; During the supplementary extraction phase, under the batch parallel processing mechanism, outpatient documents, inpatient documents, and draft medical histories are submitted to the model for processing in batches to extract relevant clinical findings that were missed in the draft medical histories, forming fact fragments with traceability IDs, and obtaining supplementary medical history findings. In the final draft stage of medical history, the initial draft of medical history and the supplementary findings of medical history are merged to obtain a complete final draft of medical history, while the factual fragments with traceability IDs that are deemed valuable are selected. During the anomaly screening phase, under the batch parallel processing mechanism, with the final draft of the complete medical history as the background, the test reports and examination reports are submitted to the model for processing in batches to screen positive and abnormal results related to the current diagnosis. The output format is fact fragments with traceability ID. In the final summary stage, the complete medical history, positive results, and abnormal results are submitted to the local large language model for further refinement, and corresponding structured medical history summaries and pre-screening judgments are generated.
[0043] In the source ID restoration stage, the factual fragments with source IDs output from the structured medical history summary, the final draft of the medical history, and the anomaly screening stage are submitted to the model to generate source fragments for locating the structured medical history summary. When the front-end browser clicks on these fragments, it will jump to the original information in the medical record system (such as the original report). Among them, the supplementary extraction stage and the anomaly screening stage, which involve a batch parallel processing mechanism, are executed concurrently internally, and the number of concurrent processes is configurable.
[0044] Understandably, the key advantage of the aforementioned adaptive triage mechanism is that simple cases (small data volume) directly return results with zero additional overhead, while complex cases (large data volume) automatically enter batch deep processing to extract medical history information more fully. This achieves a concise and accurate extraction effect of medical history information, especially for the accurate summarization of extremely long medical histories.
[0045] As for the pre-screening judgment, the pre-screening of medical history is to be achieved. In layman's terms, no matter which path is taken, the model will output a judgment that a medical history is sufficient or necessary. When the output is that a medical history is necessary, the front-end list page can mark the corresponding patient with a yellow badge. In this way, the on-duty physician can pay priority attention to the patient and complete the supplementary medical history before the patient leaves the department.
[0046] The automatic pre-screening function for the adequacy of medical history is something that traditional manual processes cannot achieve, because previously it could only rely on the doctor's ad-hoc judgment when writing the report.
[0047] Therefore, as can be seen from this setup, in addition to providing background for report writing, this application also solves another previously impossible need: identifying patients with incomplete medical history in advance and reminding on-duty physicians to conduct on-site consultations promptly. In the traditional process, because on-duty physicians cannot read the complete medical history of hundreds of patients in a single day, the adequacy of patient medical history can only be judged by the physician on the spot when writing the report. By this time, the patient has usually left the examination room. However, this application's system automatically completes a full medical history analysis within minutes of the patient's information being entered into the system. When it is determined that the medical history information is insufficient to support the diagnostic conclusion (e.g., missing key laboratory tests, or unavailable imaging reports from other hospitals), it immediately marks the situation as requiring consultation. This effectively motivates on-duty physicians to conduct supplementary inquiries in a timely manner before the patient leaves the department, avoiding subsequent report delays or diagnostic errors due to incomplete medical history.
[0048] In addition, the clinical history in the medical imaging report data is injected as an independent chapter in each of the six stages, which ensures that the clinical information recorded by the ordering physician and the attending physician is not overwhelmed by the massive amount of data in the hospital information system.
[0049] Furthermore, for the batch path strategy under the above adaptive traffic splitting mechanism, this application can also configure the following key detail optimization settings: Dynamic text stage: The amount of input during the initial draft generation stage is controlled by the upper limit of the first preset word count (e.g., 10,000 words). If it exceeds the limit, it will be truncated in reverse order of the medical record date to ensure that the latest information is included in the initial medical record draft first. Concurrent batch control: In the concurrent batch control strategy of the batch parallel processing mechanism, the capacity of each batch of text is controlled by the upper limit of the second preset number of characters (e.g., 3000 characters). Smaller context is more conducive to the model discovering low-frequency key medical history. There is no shared state between batches, they are completely independent and naturally suitable for parallel processing. Intermediate result compression: After the supplementary extraction stage and the anomaly screening stage are completed, if the total output exceeds the preset word limit, parallel compression is initiated. Specifically, the model is required to further refine the effective information based on the initial draft of the medical history, so as to avoid subsequent processing from exceeding the local large language model context window. Output structure constraints and repair: Each stage of model call requires output in structured JSON format and declaration of required fields. If parsing fails, an automatic repair tool is called to handle common format errors. Each stage can be retried a maximum of 3 times.
[0050] JSON, or JavaScript Object Notation, is an open standard file and data exchange format.
[0051] Finally, with the assistance of factual fragments with traceable IDs, the generated structured medical history is annotated, enabling the traceability of medical history details (click to jump to the corresponding content in the hospital information system). The batch-extraction-merging method described above in this system effectively solves the problem of information dilution in long contexts by focusing layer by layer and filtering step by step.
[0052] At the same time, the system automatically pre-screens out patients with insufficient medical history, enabling on-duty physicians to supplement the consultation in time before the patient leaves. Previously, this judgment could only be made remedially by telephone consultation when the doctor was writing the report.
[0053] For the quality control objectives, this application designs a three-tier cascaded quality control architecture, which is implemented using the summary module and quality control module in the large language model processing pipeline.
[0054] The output of the previous summary module, namely the first level of processing medical history summaries, provides a structured clinical background for the second level of quality control submitted for review and the third level of quality control that has been reviewed, while also completing the pre-screening of consultation needs.
[0055] For the quality control module in the large language model processing pipeline, the specific processing content may include: 2.1) During the submission review and quality control stage, after the report status changes to the submission review status and the cooling-off period is met (wait for a period of time after the report is modified to avoid evaluating text that is frequently edited), the initial quality control review is triggered. The input includes the medical history summary, incremental medical history (first complete the period up to the report submission time), the current report and the historical version quality control record of the submission review and quality control stage, and outputs the evaluation conclusion of passing, risky or incorrect (which may also include detailed modification suggestions). Then, the re-review mechanism is initiated to judge whether there is excessive quality control in the evaluation conclusion, thereby reducing the false alarm rate.
[0056] In the submission review and quality control stage, historical version quality control records are attached, which are the findings, conclusions and quality control opinions of the reports generated in each submission review and quality control stage assessment. This allows the model to perceive the version evolution of the report and detect whether the new version has introduced errors that have been corrected before, or whether the problems marked in the old version have not been properly handled.
[0057] 2.2) In the audited quality control stage, after the report status changes to the audited status and the record submitted for the audited quality control stage reaches the final state, a second quality control audit is triggered. The input includes the medical history summary, incremental medical history (first complete the period up to the report submission time), the current report and the historical version of the audited quality control record, and the output evaluation conclusion (which may also include detailed modification suggestions). Then, the re-review mechanism is initiated to judge whether there is over-quality control in the evaluation conclusion, thereby reducing the false alarm rate. The processing of the submitted quality control stage and the audited quality control stage are independent of each other.
[0058] Similar to 2.1), the audited quality control stage will also include all historical quality control records as context.
[0059] It is important to note that the audited quality control stage is not simply a reuse of the results submitted for audit quality control. Instead, it involves independently pulling data, conducting independent evaluations, and generating independent evaluation conclusions.
[0060] Furthermore, both the submission for review and the reviewed quality control stage require obtaining new medical record information that appears after the inspection date before quality control. However, directly pulling this information would result in a large amount of already processed duplicate data. Therefore, in this application, which focuses on optimizing the design of incremental medical history, the submission for review and the reviewed quality control stage can also adopt an incremental medical history extraction mechanism based on content fingerprints to obtain medical record data after the inspection date. This includes the following processing: 3.1) During the data retrieval phase, determine the start and end points of the time window. The start point of the time window extends backward from the inspection date, and the end point of the time window is the report version time. Specifically, the submission time for review and quality control is the submission time for review, and the time of approval is the time of approval for the already reviewed quality control phase. Setting the start point of the time window helps ensure that boundary data caused by discrepancies in the medical record date records in the system is not missed. Setting the end point of the time window precisely pins it to the report version time, which helps to restore the doctor's perspective on the medical history. Moreover, it does not depend on the current system time, which helps to ensure that the same retrieval range can be obtained by clearing and rerunning at any time.
[0061] 3.2) In the fingerprint difference filtering stage, maintain a fingerprint table of hospital information system records. For each hospital information system record, store the corresponding source inspection identifier, original record identifier, record type and content hash value. Compare the record retrieved this time with the fingerprint table of hospital information system records, and only return records that do not exist in the fingerprint table of hospital information system records. The content fingerprint here obviously corresponds to the application of a fingerprint engine that processes data content fingerprints. Specifically, it can be equipped with specific hash algorithms such as SHA-256 to calculate the hash value of the corresponding data content as the content fingerprint.
[0062] The records returned to the hospital information system that do not exist in the fingerprint table are the records that were actually added or whose content has been changed.
[0063] 3.3) In the semantic extraction stage, the existing medical history summary is used as the baseline, and the original hospital information system text is added as the incremental input. The local large language model is called to extract incremental clinical information not covered in the existing summary. It is important to note that this stage / step is not a simple text concatenation process. The model needs to understand the semantics of the two texts and extract only the truly new clinical facts, rather than repeating existing content or merely supplementing the details and precision of existing facts.
[0064] 3.4) During the unified append storage phase, successfully added incremental content is extracted and appended to the currently checked incremental medical history storage field with the date as the title. The incremental medical history storage field is a cumulative text that continues to grow over time and with version iterations, forming a complete incremental medical history chain. The date is preferably accurate to the minute, that is, at the minute level / granularity.
[0065] 3.5) During the fingerprint persistence stage, after successful extraction, the newly recorded fingerprint content is written into the fingerprint record table of the hospital information system.
[0066] The fingerprint persistence operation performed immediately after successful extraction is understandable, as it helps prevent repeated processing in subsequent steps.
[0067] Furthermore, under the three-layer reuse mechanism of the incremental medical history extraction pipeline, the submission review quality control stage, the reviewed quality control stage, and the evaluation module only pass different time window parameters when using the same three shared extraction functions. Thus, when the reviewed quality control stage is called, since the submission review quality control stage has already written the processed records into the fingerprint table, the fingerprint difference set will naturally filter out these records and only return the hospital information system data that appears after the submission review quality control stage is executed. The incremental content of the subsequent retrospective analysis is not added to the cumulative field, but is stored independently because its time window belongs to the category of post-event review and should not be mixed with the incremental medical history of the quality control period.
[0068] Furthermore, under the equivalence design of the time window that ensures consistent results even after clearing and rerunning at any time, in addition to the previously mentioned loose starting point and precise ending point settings (i.e., the starting point is extended forward (relying on the fingerprint difference set for natural deduplication), and the ending point is fixed at a fixed time node (version time or review day arrival time), not dependent on the current system time), there can also be a feature to simultaneously clear fingerprint settings during reruns, namely: When a rerun is triggered (a complete reconstruction triggered by manual reassessment, or a doctor's feedback correction that requires reconstruction in subsequent versions), the accumulated content of incremental medical history and fingerprint records are cleared, and the incremental medical history is reconstructed from zero. The window remains unchanged, and the fingerprints are reconstructed from empty. The result is consistent with the first execution.
[0069] As for the department memory bank in the basic shared components, it can be manually maintained by the preset senior doctors, that is, experienced senior doctors, to store the experience rules accumulated in image report writing and quality control.
[0070] Furthermore, in the system design of this application, the quality control of the image portion of medical image reports can be carried out using a four-step process: classification, cropping, single-image evaluation, and summarization. In detail, this can include the following processing steps: 4.1) In the classification stage, identify the image layout type and locate the lesion area; The image layout types can specifically include four-view diagrams, nine-grid diagrams, and fused diagrams.
[0071] 4.2) In the cropping stage, based on the layout type classification results of the classification stage in 4.1), the lesion area is cropped into independent sub-images; 4.3) In the single-graph evaluation phase, the independent subgraphs in 4.2) are evaluated independently; In the independent evaluation process for each individual sub-image, specific checks can be made on image quality, clarity of lesion display, artifacts, etc.
[0072] 4.4) In the summary stage, the evaluation results of all independent subgraphs are combined to generate the final conclusion, which is then merged with the text quality control conclusion, depending on whether merging is configured.
[0073] If it is necessary to merge with the text quality control conclusions, for quality control conclusions with the same content, the one with the highest severity can be taken as the final merged output.
[0074] Next, we will continue to focus on the evaluation module. For the evaluation module in the large language model processing pipeline, the specific processing content may include: Once the quality control report is completed, a retrospective analysis is triggered, causing the scheduler to automatically generate retrospective records for each completed quality control report at three time points: 30 days, 60 days, and 90 days. These records are queued and executed sequentially upon their respective trigger times. Each retrospective record is uniquely constrained by the inspection identifier and the number of retrospective days.
[0075] The setting that each retrospective record is uniquely constrained by an inspection identifier and the number of days of review helps prevent duplicate creation in practical applications.
[0076] The retrospective analysis triggered in terms of details may specifically include the following processing: 5.1) In the baseline medical history construction stage, the medical history summary output by the summary module, all incremental medical histories extracted by the quality control module, and the clinical medical history in the medical imaging report data are merged to obtain the baseline medical history; The combined result of these three factors, namely the baseline medical history, constitutes all the clinical information known to the physician when the report is written and reviewed.
[0077] 5.2) In the subsequent incremental medical history acquisition stage, the shared extraction function of the quality control module is reused to extract the subsequent incremental medical history; The window extends forward by 3 days to cover any new medical records that may be added during the report writing and review process. The window ends at the time point when the review period is reached. Corresponding to the previous settings, the time window can be 3 days before the report is completed to 30 days, 60 days, or 90 days after the report is completed.
[0078] 5.3) In the model comparison and judgment stage, the corresponding known baseline medical history and the corresponding subsequent incremental medical history are input into the local big language model to judge whether the report conclusion at that time accurately predicted the nature of the main lesions, and obtain the judgment result of accurate prediction, incorrect prediction or inability to judge (which may also include detailed analysis of the reasons).
[0079] In addition, the following can be included in more detailed aspects of the above system design.
[0080] Based on the fingerprint engine, the fingerprint difference filtering stage, which involves the submission review and quality control stages and the already reviewed quality control stages, can have the following core logic in its fingerprint difference function: 6.1) Load the set of identifiers for all known records currently being examined from the fingerprint table in the hospital information system; This includes the original record identifier and the record type; 62) Iterate through the newly retrieved record list and check each record one by one. If the identifier of a record is not in the identifier set, mark it as a new record and add it to the output list. 6.3) Content hash values (such as SHA-256 hash values) are used to record content integrity verification, but the difference determination uses the record identifier as the primary key. Records in the same hospital information system will maintain the same original identifier when the content is updated. At this time, due to the different content hashes, both the old and new versions will be recognized. The fingerprint saving function uses batch writing and automatically skips existing records based on the database's unique constraints, thus preventing duplicates.
[0081] Furthermore, under the incremental medical history appending format setting, taking the incremental medical history accumulation field from May 11, 2026 as an example, the corresponding storage format can be: ## Incremental Medical History (2026-05-11 09:19) - Added content item 1 - Added content item 2 ## Incremental Medical History (2026-05-15 14:30) - Additional content from another quality control execution The above format allows the front end to automatically split by date title, and each increment is displayed independently with a date label (no longer distinguishing between the quality control stage from the submission review stage and the quality control stage after review), so that users see a unified timeline.
[0082] Furthermore, regarding the task scheduling process, the scheduler can operate continuously at a high frequency (e.g., every 20 seconds per cycle), and the corresponding processing steps can include: 7.1) Synchronous Image Information System: Retrieve the report list for the most recent time period. For newly created reports, and for existing reports whose status has changed, create an independent quality control record for the corresponding new version, while the historical records of previous versions are not affected. 7.2) Task of collecting medical history summaries: Records that have not yet been completed or failed in the last time should be sorted by examination date and then added to the team; 7.3) Collection, submission, review, and quality control phase tasks: Records that have been submitted for review, have completed medical history summaries, and have met the cooling-off period are categorized into two priority groups, recent and historical, based on whether the inspection date is more than 3 days ago. 7.4) Collecting tasks that have been reviewed and approved: Records that have been reviewed and approved, medical history summaries that have been completed, and all submitted review and approval quality control stage records that have been completed. Similar to collecting submitted review and approval quality control stage tasks, they are divided into two priority categories: recent and historical. 7.5) Collect retrospective tasks: Automatically generate three types of retrospective records for 30 days, 60 days and 90 days for records that meet the conditions, and enqueue them when the trigger time arrives; 7.6) Collect and process feedback tasks: Records of doctor feedback pending processing are queued with the highest priority; 7.7) Dispatch and execute: Retrieve tasks from the queue according to priority, control the number of concurrent tasks through semaphores, and execute each task in an independent thread.
[0083] In addition, regarding resets and rebuilds, when a quality control rerun is required (manually triggered re-evaluation, or subsequent version rebuilds caused by version changes), the following procedures should be performed: 8.1) Clear the status and evaluation results of the target quality control records, but retain report snapshots and version information; 8.2) Clear the accumulated information of newly added medical history; 8.3) Delete the corresponding fingerprint record.
[0084] The above three steps are committed in the same transaction. In the next scheduling cycle, the quality control records with an empty status will be automatically collected and queued again and re-executed. Since the fingerprint has been cleared, the traceable structured medical history and incremental medical history will be reconstructed from zero.
[0085] Meanwhile, regarding the key database table structure, it may involve the hospital information system record fingerprint table and the retrospective analysis result table. The hospital information system record fingerprint table is the core data structure used to drive incremental deduplication. Each row records the fingerprint information of one hospital information system record, including: examination identifier, original record identifier, record type (outpatient / inpatient / laboratory / examination), and content hash value, with the examination identifier + original record identifier + record type as the unique constraint. Each row of the retrospective analysis result table records the result of one retrospective analysis, including: examination identifier, number of days of review (30 / 60 / 90), final report version time, trigger time, execution status, snapshot of report findings / conclusions, full text of baseline medical history, newly added medical history content (structured incremental data extracted by the model), judgment result (accurate prediction / incorrect prediction / unable to judge), and detailed analysis text, with the examination identifier + number of days of review as the unique constraint.
[0086] Furthermore, regarding the technical details of model invocation, the system can also have the following configurations: 9.1) Output structure constraints: All model calls must output in structured JSON format and declare the required output fields; 9.2) Structure Validation and Repair: When parsing the model output, verify the existence of required fields. If parsing fails, call the automatic repair tool to handle common JSON format errors. 9.3) Retry mechanism: Each call can be retried a maximum of 3 times; 9.4) Inference Mode Control: Some open-source models enable the inference chain mode by default, which will generate a lengthy internal inference process before the final output. For quality control assessment tasks that require complex semantic analysis, the system retains the inference chain to ensure assessment quality. For simple parallel subtasks (such as the supplementary screening step in batch medical history summarization), the system disables the inference chain to reduce latency and token consumption. This control strategy can be configured independently at the prompt word level. 9.5) Privacy Desensitization: The system has a built-in optional privacy desensitization mechanism that automatically replaces sensitive information such as patient name, ID, and hospital name with desensitized placeholders.
[0087] In conclusion, regarding the above solutions, under the goal of medical image report quality control, this application designs a medical image report quality control system architecture based on a locally deployed lightweight large language model. Thanks to this architecture, from automated, traceable, and accurate summarization of clinical background and pre-screening of consultations, to automated semantic quality control of report documents, and automated verification of the accuracy of long-term conclusion predictions, and the reuse of data and algorithms between each level, it alleviates the problems associated with existing technologies while possessing many advantages such as data security, full-process automation, traceability of content generated by the large language model, high configurability, cross-modal transferability, low deployment cost, and model replaceability, providing excellent tool support for medical image reporting work.
[0088] The above provides a detailed description of the medical image report quality control system based on a local large language model provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of this application; furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A medical image report quality control system based on a local large language model, characterized in that, The system includes: A global interactive platform enables human-computer interaction in the doctor's report workbench; The data source layer connects to the hospital information system to obtain the patient's full historical medical record data and connects to the imaging report system to obtain medical imaging report data as data input; The core control and large language model processing layer includes a task scheduling engine, a large language model processing pipeline, basic shared components, and a large language model service. The task scheduling engine performs nine-level priority task queue management, version-aware dependency chain management, and day / night execution window management. The large language model processing pipeline sequentially includes a summary module for processing medical history summaries based on the data input, a quality control module for implementing multi-dimensional semantic report quality control, and an evaluation module for implementing retrospective assessment. The basic shared components include a fingerprint engine for processing data content fingerprints, an incremental medical history extraction pipeline for extracting incremental medical history, a department memory bank for maintaining department-level experience rules, and a privacy desensitization module for desensitizing privacy data. The large language model service provides the summary module, the quality control module, and the evaluation module with a locally deployed large language model inference interface. The data persistence layer stores core data uniformly in the system's local database, which is used for reading and writing by the large language model processing pipeline and for front-end querying and display.
2. The system according to claim 1, characterized in that, The human-computer interaction functions include display of analysis results, monitoring of task queues, operation of the quality control management panel, and doctor feedback and correction. The patient's complete historical medical record data includes outpatient documents, inpatient documents, laboratory reports, and examination reports; The core data includes structured traceable medical history, structured traceable incremental medical history, snapshots of quality control results from previous tests, retrospective analysis results, and fingerprint tables recorded in the overall hospital information system.
3. The system according to claim 1, characterized in that, For the summary module, the simplified path strategy for total word count less than a threshold includes: The patient's complete historical medical record data, including outpatient documents, inpatient documents, laboratory reports, and examination reports, the source ID of the patient's complete historical medical record data, and the clinical history in the medical imaging report data are structured and merged into a single input. The single input is submitted to the model for processing to generate a structured medical history summary and a pre-screening judgment, which are output as the medical history summary. At the same time, the fact fragment source ID is output. Then, the fact fragment corresponding to the fact fragment source ID and the medical history summary are submitted to the model again for fact fragment matching. The pre-screening judgment is used to pre-screen the sufficiency of the medical history in two cases: sufficient and necessary to take a medical history.
4. The system according to claim 3, characterized in that, For the summary module, the batching path strategy corresponding to a total word count not less than the threshold includes: During the initial draft generation stage, the outpatient documents and the inpatient documents are combined to obtain the initial draft of the medical history. In the supplementary extraction stage, under the batch parallel processing mechanism, the outpatient documents, inpatient documents and the draft medical history are submitted to the model for processing in batches to extract relevant clinical findings that were missed in the draft medical history, form the fact fragment with the traceability ID, and obtain supplementary medical history findings. In the final draft stage of medical history, the initial draft of medical history and the supplementary findings of medical history are merged to obtain a complete final draft of medical history, and at the same time, the fact fragments with the source ID that are deemed valuable are selected. During the anomaly screening phase, under the batch parallel processing mechanism, with the final draft of the complete medical history as the background, the test reports and examination reports are submitted to the model for processing in batches to screen positive and abnormal results related to the current diagnosis. The output format is the fact fragment with the traceability ID. In the final summary stage, the complete medical history, the positive results, and the abnormal results are submitted to the model for further refinement, and the corresponding structured medical history summary and pre-screening judgment are generated. In the source ID restoration stage, the structured medical history summary, the factual fragment with the source ID output from the final draft stage of the medical history, and the factual fragment with the source ID output from the anomaly screening stage are respectively submitted to the model to generate a source fragment for locating the structured medical history summary. The clinical history in the medical imaging report data is injected as a separate chapter in each of the six stages.
5. The system according to claim 4, characterized in that, The input amount during the initial draft generation stage is controlled by the upper limit of the first preset word count; if it exceeds the limit, it will be truncated in reverse order of the medical record date. In the concurrent batch control strategy of the batch parallel processing mechanism, the capacity of each batch of text is controlled by the upper limit of the second preset number of characters, and there is no shared state between batches. After the supplementary extraction stage and the anomaly screening stage are completed, if the total output exceeds the preset word limit, a parallel compression call is initiated. Each stage of the model call requires output in structured JSON format and declaration of required fields. If parsing fails, an automatic repair tool is invoked to handle common formatting errors.
6. The system according to claim 1, characterized in that, The quality control module includes: During the submission review and quality control stage, the initial quality control review is triggered after the report status changes to the submission review status and the cooling-off period is met. The corresponding historical version quality control records are entered, the evaluation conclusion is output, and then the re-review mechanism is initiated. In the audited quality control stage, after the report status changes to the audited status and the record in the audited quality control stage reaches the final state, a second quality control audit is triggered. The corresponding historical version quality control record is input, the evaluation conclusion is output, and then the re-review mechanism is started. The processing of the audited quality control stage and the audited quality control stage are independent of each other.
7. The system according to claim 6, characterized in that, The submission review and quality control stage and the reviewed quality control stage employ an incremental medical history extraction mechanism based on content fingerprinting to obtain medical record data after the examination date, including: During the data retrieval phase, the start and end points of the time window are determined. The start point of the time window extends backward from the inspection date, and the end point of the time window corresponds to the submission review and quality control phase, specifically the submission review time, and corresponds to the review and quality control phase, specifically the review approval time. During the fingerprint difference filtering stage, a fingerprint table of the hospital information system is maintained. The records retrieved this time are compared with the fingerprint table of the hospital information system, and only records that do not exist in the fingerprint table of the hospital information system are returned. In the semantic extraction stage, the existing medical history summary is used as the baseline, and the original hospital information system text is added as the incremental input. The model is called to extract incremental clinical information not covered in the existing summary. During the unified append storage phase, successfully added incremental content is extracted, appended to the currently checked incremental medical history storage field with the date as the title; During the fingerprint persistence stage, after successful extraction, the newly recorded fingerprint content is written into the fingerprint record table of the hospital information system; The submission review quality control stage, the reviewed quality control stage, and the evaluation module all use the same three shared extraction functions, but with different time window parameters.
8. The system according to claim 1, characterized in that, Quality control of the aforementioned medical imaging report data includes: In the classification stage, the image layout type is identified and the lesion area is located; During the cropping stage, the lesion region is cropped into independent sub-images based on the layout type classification results from the classification stage. In the single-graph evaluation phase, each independent subgraph is evaluated independently. In the summary phase, the evaluation results of all the independent subgraphs are combined to generate the final conclusion, which is then merged with the text quality control conclusion, depending on whether merging is configured.
9. The system according to claim 1, characterized in that, The evaluation module includes: Once the quality control report is completed, a retrospective analysis is triggered to generate retrospective records; The retrospective analysis included: In the baseline medical history construction stage, the medical history summary output by the summary module, all the incremental medical histories extracted by the quality control module, and the clinical medical history in the medical imaging report data are merged to obtain the baseline medical history. In the subsequent stage of acquiring new medical history, the shared extraction function of the quality control module is reused to extract the incremental medical history. In the model comparison and judgment phase, the baseline medical history known at the time and the subsequent incremental medical history that appeared afterward are respectively input into the model to judge whether the report conclusion at that time was accurate in predicting the nature of the main lesions, and to obtain the judgment result.
