Medical data intelligent management method and device based on medical image information large model

Abstract

The application discloses a medical data intelligent management method and device based on a medical image information large model. The method comprises the following steps: receiving medical image related data comprising medical image data, diagnosis and treatment information related to the medical image data, and report text; performing rule routing and management processing according to source information and data attributes; performing sequence-level splitting on the processed medical image data, obtaining a semantic classification result corresponding to a sequence-level data unit based on a medical image information large model; splitting the processed report text and extracting a semantic label corresponding to a sub-report; establishing an association relationship between the sequence-level data unit and the sub-report based on the semantic classification result and the semantic label, and storing the processed medical image related data, the semantic classification result and the association relationship. Thus, the standardized management of the medical image related data and the corresponding management of the image and the report are realized.

Description

Technical Field

[0001] This application relates to the field of medical information technology, and in particular to a method and device for intelligent governance of medical data based on a large model of medical image information. Background Technology

[0002] With the continuous advancement of medical informatization, hospitals generate a large amount of medical data in their daily clinical activities, especially medical imaging data from examinations such as CT, MRI, X-ray, and ultrasound, as well as related examination request information, patient basic information, and diagnostic report information. This data is typically stored in different operational systems such as PACS, HIS, and RIS, and forms a crucial foundation for clinical diagnosis, medical management, scientific research analysis, and the application of medical artificial intelligence. Therefore, how to effectively manage and utilize medical imaging-related data with high quality has become a widely concerned technical issue in the field of medical informatization.

[0003] In related technologies, medical institutions typically rely on existing information systems to collect, transmit, store, and query imaging data and related diagnostic and treatment information, supplemented by manual verification, rule validation, or routine data cleaning for data management. However, due to differences in the level of information system construction among different hospitals and variations in data formats, field definitions, coding rules, and storage methods among different vendors' equipment, issues such as missing data fields, non-standard filling, inconsistent terminology, and inconsistent coding systems easily arise in actual operations. Meanwhile, imaging data and corresponding diagnostic information often involve the flow and integration of multiple systems. In scenarios with large data volumes and complex business processes, problems such as inaccurate correspondences, insufficient consistency, or difficulties in subsequent utilization may also occur. Existing methods, when faced with massive, multi-source, and heterogeneous medical imaging-related data, typically suffer from limited automation, low processing efficiency, and insufficient ability to adapt to complex data situations, making it difficult to meet the actual needs of current medical data management and applications.

[0004] Therefore, in the governance of medical imaging data, the uneven quality of data, the difficulty in standardization, and the inability to guarantee the correspondence between images and diagnostic information have become urgent problems that need to be solved. Summary of the Invention

[0005] This application provides a medical data intelligent governance method and device based on a large model of medical image information, aiming to solve the problems of inconsistent data quality, difficulty in standardization processing, and difficulty in guaranteeing the correspondence between images and diagnostic information in the existing technology of medical image data governance.

[0006] Firstly, a method for intelligent governance of medical data based on a large model of medical image information, the method comprising: Receive medical imaging-related data, which includes medical imaging data, diagnostic and treatment information related to the medical imaging data, and report text; Based on the source information and data attributes of the medical image-related data, the medical image-related data is routed according to rules to obtain the medical image-related data after rule routing; The medical image-related data after the rule routing is processed to obtain processed medical image-related data; The medical image data in the processed medical image-related data is split into multiple sequence-level data units by sequence-level decomposition. Based on the large medical image information model, the sequence-level data units are classified to obtain the semantic classification results corresponding to the sequence-level data units. The report text in the processed medical image-related data is split to obtain at least one sub-report, and the semantic tags corresponding to the sub-report are extracted; Based on the semantic classification results of the sequence-level data units and the semantic tags corresponding to the sub-reports, the sequence-level data units and the sub-reports are matched to establish the association between the sequence-level data units and the sub-reports; The processed medical image-related data, the semantic classification results, and the association relationships are stored.

[0007] Optionally, in the above scheme, receiving medical image-related data includes: Receive DICOM medical image data from the PACS system; Receive patient information and treatment information from the HIS system; Receive inspection request information, report process information, and report text from the RIS system.

[0008] Optionally, in the above scheme, the step of performing rule-based routing on the medical image-related data according to the source information and data attributes of the medical image-related data to obtain the rule-based routed medical image-related data includes: Based on the hospital source, data type, and priority tags of the medical image-related data, a governance rule chain corresponding to the medical image-related data is determined; The medical image-related data is routed to the corresponding governance rule chain to obtain the rule-routed medical image-related data.

[0009] Optionally, in the above scheme, determining the governance rule chain corresponding to the medical image-related data based on the hospital source, data type, and priority tag of the medical image-related data includes: Obtain the hospital source, data type, and priority tag corresponding to the medical image-related data; Based on the hospital's source, data type, and priority label, the corresponding governance rule set is determined by matching it with the preset rule routing configuration. The governance rules in the governance rule set are compiled to obtain an executable rule component; The executable rule components are orchestrated using LiteFlow expressions to generate a governance rule chain corresponding to the medical image-related data.

[0010] Optionally, in the above scheme, routing the medical image-related data to the corresponding governance rule chain to obtain the rule-routed medical image-related data includes: Input the medical image-related data into the corresponding governance rule chain; The LiteFlow execution engine drives the execution of each executable rule component according to the orchestration logic of the governance rule chain; During the execution of each executable rule component, the medical image-related data and / or intermediate processing results are shared among the executable rule components through a context object to obtain the medical image-related data after rule routing.

[0011] Optionally, in the above scheme, the step of processing the medical image-related data after the rule routing to obtain processed medical image-related data includes: The quality of the medical image-related data after the rule routing is evaluated to obtain the quality evaluation results. Based on the quality assessment results, data cleaning and standardization transformation are performed on the medical image-related data after the rule routing to obtain the processed medical image-related data.

[0012] Optionally, in the above scheme, the medical image data in the processed medical image-related data is split into multiple sequence-level data units, including: The DICOM examination data in the processed medical image-related data are split at the sequence level based on the unique identifier of the sequence instance to obtain multiple sequences; The multiple sequences are grouped by combining imaging modality, examination site, and sequence description to obtain multiple sequence-level data units; The classification of the sequence-level data units based on the large medical image information model to obtain the semantic classification results corresponding to the sequence-level data units includes: Select representative slices from the sequence-level data units; Based on the large medical image information model, reasoning is performed on the representative slices to obtain the anatomical location, image modality confirmation result and image quality level corresponding to the sequence-level data unit; The anatomical location, the image modality confirmation result, and the image quality level are used as the semantic classification result.

[0013] Optionally, in the above scheme, the step of splitting the report text in the processed medical image-related data to obtain at least one sub-report, and extracting the semantic tags corresponding to the sub-report, includes: The report text is subjected to part identification and paragraph segmentation to obtain at least one sub-report; Extract the semantic anatomical site labels corresponding to each of the sub-reports as semantic tags for the sub-reports; The process of matching the sequence-level data units with the sub-reports based on the semantic classification results of the sequence-level data units and the semantic tags corresponding to the sub-reports, and establishing the association between the sequence-level data units and the sub-reports, includes: Based on the semantic classification results of the sequence-level data units and the semantic anatomical site labels corresponding to each sub-report, the sequence-level data units and the sub-reports are matched to obtain matching results; When the confidence level of the matching result is higher than or equal to a preset threshold, an association relationship between the sequence-level data unit and the sub-report is established based on the matching result; When the confidence level of the matching result is lower than a preset threshold, the matching result is pushed to manual confirmation; The association between the sequence-level data unit and the sub-report is established based on the results of manual confirmation.

[0014] Optionally, in the above scheme, storing the processed medical image-related data, the semantic classification results, and the association relationships includes: The structured data in the processed medical image-related data is written into a relational database; The medical image data in the processed medical image-related data is written into distributed storage; The semantic classification results and the associations are cached and / or indexed.

[0015] Secondly, a medical data intelligent governance device based on a large model of medical image information, the device comprising: The receiving module is used to receive medical image-related data, which includes medical image data, diagnosis and treatment information related to the medical image data, and report text. The rule routing module is used to perform rule routing on the medical image-related data based on the source information and data attributes of the medical image-related data, so as to obtain the rule-routed medical image-related data. The governance and processing module is used to process the medical image-related data after the rule routing to obtain the processed medical image-related data. The sequence-level splitting module is used to perform sequence-level splitting on the medical image data in the processed medical image-related data to obtain multiple sequence-level data units; The classification module is used to classify the sequence-level data units based on the large medical image information model to obtain the semantic classification results corresponding to the sequence-level data units; The report processing module is used to split the report text in the processed medical image-related data to obtain at least one sub-report, and extract the semantic tags corresponding to the sub-report; The association establishment module is used to match the sequence-level data unit with the sub-report based on the semantic classification result of the sequence-level data unit and the semantic tag corresponding to the sub-report, and establish the association relationship between the sequence-level data unit and the sub-report; The storage module is used to store the processed medical image-related data, the semantic classification results, and the association relationships.

[0016] Compared with the prior art, this application has at least the following beneficial effects: This application, based on further analysis and research of existing technical problems, recognizes that existing technologies in medical image data governance suffer from inconsistent data quality, difficulties in standardization, and challenges in ensuring the correspondence between images and diagnostic information. By receiving medical image-related data and performing rule-based routing according to its source information and data attributes, medical image-related data from different sources and with different attributes can enter appropriate processing paths. Based on this, the rule-based routing data undergoes governance processing. First, the medical image-related data itself is standardized and regulated. Then, the medical image data within the processed data is split into sequences, and semantic classification results corresponding to the sequence-level data units are obtained based on a large-scale medical image information model. Simultaneously, the report text within the processed medical image-related data is split, and semantic tags corresponding to sub-reports are extracted. Furthermore, this process utilizes… The semantic classification results of sequence-level data units are matched with the semantic tags corresponding to the sub-reports to establish the association between sequence-level data units and sub-reports. Finally, the processed medical image-related data, semantic classification results, and association relationships are stored. Since this application does not simply save or roughly associate the original mixed data, but first processes, then splits, then classifies, and then matches, it can, on the one hand, process medical image-related data in an orderly manner and improve data consistency and manageability. On the other hand, it can refine medical image data from coarse-grained to sequence-level data units and then semantically correspond it with the report text, thereby making the processing results of medical image-related data more standardized and the correspondence between report text and medical image data clearer. This can solve the problems of inconsistent data quality, difficulty in standardization processing, and difficulty in guaranteeing the correspondence between images and diagnostic information in the background technology. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a medical data intelligent governance method based on a large medical image information model, provided as an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] In one embodiment, such as Figure 1 As shown, a method for intelligent governance of medical data based on a large model of medical image information is provided, including the following steps: Receive medical imaging-related data, which includes medical imaging data, diagnostic and treatment information related to the medical imaging data, and report text; Based on the source information and data attributes of the medical image-related data, the medical image-related data is routed according to rules to obtain the medical image-related data after rule routing; The medical image-related data after the rule routing is processed to obtain processed medical image-related data; The medical image data in the processed medical image-related data is split into multiple sequence-level data units by sequence-level decomposition. Based on the large medical image information model, the sequence-level data units are classified to obtain the semantic classification results corresponding to the sequence-level data units. The report text in the processed medical image-related data is split to obtain at least one sub-report, and the semantic tags corresponding to the sub-report are extracted; Based on the semantic classification results of the sequence-level data units and the semantic tags corresponding to the sub-reports, the sequence-level data units and the sub-reports are matched to establish the association between the sequence-level data units and the sub-reports; The processed medical image-related data, the semantic classification results, and the association relationships are stored.

[0020] In some implementations, the intelligent governance method for medical data based on a large-scale medical imaging information model first receives medical imaging-related data. This data includes medical imaging data, related diagnostic and treatment information, and report texts. Medical imaging data can be DICOM format data generated from examinations such as CT, MRI, X-ray, and ultrasound; diagnostic and treatment information can include basic patient information, outpatient information, inpatient information, medical orders, examination request information, report status information, and reviewing physician information; and report texts can be free-text image reports corresponding to the medical imaging data. By uniformly incorporating these different types of data into the medical imaging-related data, subsequent processing steps can revolve around the same set of governance targets.

[0021] In some implementations, medical image-related data is routed according to its source information and data attributes, resulting in rule-routed medical image-related data. Source information may include hospital source, system source, and access source; data attributes may include data type, business category, and priority tag. Through rule-route, medical image-related data from different hospitals, with different data types, and different priorities can be sent to different governance rule chains for execution. The rule-routed medical image-related data then enters the governance processing stage, and the output of the governance processing is the processed medical image-related data.

[0022] In some implementations, the medical image-related data after rule-based routing undergoes governance processing, which may include quality assessment, data cleaning, and standardization transformation. Quality assessment can be used to determine the completeness, standardization, and timeliness of the rule-based routing data; data cleaning can be used to correct formatting errors, filter abnormal characters, and standardize field formats; standardization transformation can be used to convert localized field values, localized terminology, and mixed encodings into a unified data representation. After governance processing, processed medical image-related data suitable for further analysis is obtained.

[0023] In some implementations, medical image data within the processed medical image-related data is decomposed at the sequence level to obtain multiple sequence-level data units. The object of sequence-level decomposition is the processed DICOM examination data, and each sequence-level data unit corresponds to a relatively independent image acquisition sequence. Subsequently, the sequence-level data units are classified based on a large-scale medical image information model to obtain semantic classification results corresponding to the sequence-level data units. The semantic classification results may include anatomical location, image modality confirmation results, and image quality level. Because the classification object has been refined from the examination level to the sequence-level data unit, the obtained semantic classification results can more accurately correspond to the actual image sequence content.

[0024] In some implementations, the report text in the processed medical image-related data is split into at least one sub-report, and the semantic tags corresponding to the sub-reports are extracted. For cases where multiple body parts or examination contents are described simultaneously in the same report text, multiple sub-reports can be formed through body part identification and paragraph segmentation, with each sub-report corresponding to a semantic tag. Subsequently, based on the semantic classification results of sequence-level data units and the semantic tags corresponding to the sub-reports, the sequence-level data units and sub-reports are matched to establish the association between them. After the association is established, the processed medical image-related data, semantic classification results, and associations are stored for subsequent querying, review, statistical analysis, and training data construction.

[0025] This embodiment can incorporate received medical image data, diagnosis and treatment information, and report text into a unified governance process. Through continuous processing of rule routing, governance processing, sequence-level splitting, semantic classification, report splitting, and association establishment, a fine-grained correspondence is formed between medical image data and report text. The processed medical image-related data, semantic classification results, and association relationships are uniformly stored, thus providing a complete foundation for subsequent data management and data utilization.

[0026] In this embodiment, receiving medical image-related data includes: Receive DICOM medical image data from the PACS system; Receive patient information and treatment information from the HIS system; Receive inspection request information, report process information, and report text from the RIS system.

[0027] In some implementations, when receiving medical image-related data, DICOM medical image data from a PACS system is received. For a PACS system, DICOM medical image data can be received via DICOM communication, and the received content may include image pixel data and metadata fields from the DICOM file. The received DICOM medical image data can be either real-time pushed data or batch imported data.

[0028] In some implementations, patient information and treatment information are received from the HIS system. The information output by the HIS system can be used to characterize the patient's identity and treatment context. Patient information may include the patient's name, visit number, hospital number, etc., while treatment information may include outpatient records, inpatient records, and medical orders. Examination request information, report flow information, and report text are received from the RIS system. Examination request information can be used to identify specific imaging examination requests, report flow information can be used to characterize the report generation, review, and issuance status, and report text can be directly used as input for subsequent report splitting. The data from these different systems can be uniformly packaged after receipt to form medical image-related data used for subsequent rule routing and governance processing.

[0029] This embodiment can obtain medical imaging data, patient information and diagnosis and treatment information, examination application information, reporting process information and report text from PACS system, HIS system and RIS system respectively, so that the subsequent treatment process has a clear and complete input source, thereby ensuring the consistency and traceability of the objects processed in each subsequent step.

[0030] In this embodiment, the step of performing rule-based routing on the medical image-related data according to the source information and data attributes of the medical image-related data to obtain rule-based routed medical image-related data includes: Based on the hospital source, data type, and priority tags of the medical image-related data, a governance rule chain corresponding to the medical image-related data is determined; The medical image-related data is routed to the corresponding governance rule chain to obtain the rule-routed medical image-related data.

[0031] In some implementations, when routing medical imaging-related data according to its source information and data attributes, the first step is to determine the governance rule chain corresponding to the medical imaging-related data based on its hospital source, data type, and priority label. The hospital source can distinguish data from different hospitals or different campuses; the data type can distinguish between medical imaging data, diagnostic information, and report texts; and the priority label can distinguish between high-priority governance tasks and routine governance tasks.

[0032] In some implementations, after determining the governance rule chain, medical image-related data is routed to the corresponding governance rule chain to obtain rule-routed medical image-related data. The routing action can be understood as: sending medical image-related data into the determined governance rule chain for execution. For different hospital sources, different data types, and different priority tags, the system can use different rule chains to apply different governance paths to different medical image-related data.

[0033] This embodiment enables medical image-related data to be differentiated based on source, type, and priority before entering the processing stage, thereby making subsequent processing more targeted and distinguishing different processing paths from each other.

[0034] In this embodiment, determining the governance rule chain corresponding to the medical image-related data based on the hospital source, data type, and priority tag of the medical image-related data includes: Obtain the hospital source, data type, and priority tag corresponding to the medical image-related data; Based on the hospital's source, data type, and priority label, the corresponding governance rule set is determined by matching it with the preset rule routing configuration. The governance rules in the governance rule set are compiled to obtain an executable rule component; The executable rule components are orchestrated using LiteFlow expressions to generate a governance rule chain corresponding to the medical image-related data.

[0035] In some implementations, when determining the governance rule chain corresponding to medical image-related data based on the hospital source, data type, and priority label of the medical image-related data, the hospital source, data type, and priority label corresponding to the medical image-related data can be obtained first. The acquisition method may include reading information such as hospital code, service category identifier, and priority field from the access data packet, or it may be determined based on the access port, source system identifier, or service configuration.

[0036] In some implementations, the corresponding governance rule set is determined by matching the hospital source, data type, and priority label with a preset rule routing configuration. The preset rule routing configuration can pre-establish a mapping relationship between the hospital source, data type, priority label, and the governance rule set. After determining the governance rule set, the governance rules within the set are compiled to obtain executable rule components. Governance rules can be defined using Groovy scripts or QLExpress expressions, and after compilation, they form executable rule components that can be directly called by the execution engine. Subsequently, LiteFlow expressions are used to orchestrate the executable rule components, generating a governance rule chain corresponding to the medical image-related data. LiteFlow expressions can represent serial, parallel, conditional branch, and loop relationships between executable rule components.

[0037] This embodiment can establish a configurable mapping relationship between hospital source, data type and priority label and governance rule chain, and generate directly executable governance rule chain by compiling and orchestrating governance rule set, thereby providing a clear and adjustable rule execution structure for subsequent routing execution.

[0038] In this embodiment, routing the medical image-related data to the corresponding governance rule chain to obtain the rule-routed medical image-related data includes: Input the medical image-related data into the corresponding governance rule chain; The LiteFlow execution engine drives the execution of each executable rule component according to the orchestration logic of the governance rule chain; During the execution of each executable rule component, the medical image-related data and / or intermediate processing results are shared among the executable rule components through a context object to obtain the medical image-related data after rule routing.

[0039] In some implementations, when routing medical image-related data to the corresponding governance rule chain, the medical image-related data can be first input into the corresponding governance rule chain. The input governance rule chain is then executed by the LiteFlow execution engine according to the orchestration logic of the governance rule chain, driving each executable rule component to run. Each executable rule component can run in a serial, parallel, conditional branch, or loop manner as predefined by LiteFlow expressions.

[0040] In some implementations, during the execution of each executable rule component, medical image-related data and / or intermediate processing results are shared among the executable rule components through a context object. That is, the processing result obtained by the previous executable rule component can be directly passed to the next executable rule component through the context object, serving as input or auxiliary judgment criteria for the next executable rule component. After the governance rule chain is executed, the medical image-related data after the governance rule chain execution is output, resulting in the rule-routed medical image-related data.

[0041] This embodiment enables multiple executable rule components in the governance rule chain to work collaboratively according to a unified orchestration logic, and to share medical image-related data and intermediate processing results during execution, thereby giving the medical image-related data after rule routing a unified processing context and a clear basis for the next step input.

[0042] In this embodiment, the step of processing the medical image-related data after the rule routing to obtain processed medical image-related data includes: The quality of the medical image-related data after the rule routing is evaluated to obtain the quality evaluation results. Based on the quality assessment results, data cleaning and standardization transformation are performed on the medical image-related data after the rule routing to obtain the processed medical image-related data.

[0043] In some implementations, when processing rule-routed medical image data, a quality assessment is first performed on the data to obtain a quality assessment result. This result can be a breakdown of individual assessments or a comprehensive assessment. Following the quality assessment, data cleaning and standardization transformation are then performed on the rule-routed medical image data to obtain processed medical image data.

[0044] In some implementations, data cleaning may include field mapping, abnormal character filtering, date format standardization, data validation, and format repair; standardization conversion processing may include standardization mapping of local terms, semantic normalization of non-standard terms and abbreviations, conversion of diagnostic codes from different versions, and unified mapping of image terms. For problems identified during quality assessment, targeted repair or conversion actions can be performed to ensure that the data output from the governance process remains highly usable in subsequent steps.

[0045] This embodiment can first identify problems in the medical image-related data after rule routing through quality assessment, and then perform data cleaning and standardization transformation based on the quality assessment results, so that the processed medical image-related data is more suitable for subsequent sequence-level splitting, large-scale medical image information classification, and report text splitting and matching.

[0046] In this embodiment, the medical image data in the processed medical image-related data is split into multiple sequence-level data units, including: The DICOM examination data in the processed medical image-related data are split at the sequence level based on the unique identifier of the sequence instance to obtain multiple sequences; The multiple sequences are grouped by combining imaging modality, examination site, and sequence description to obtain multiple sequence-level data units; The classification of the sequence-level data units based on the large medical image information model to obtain the semantic classification results corresponding to the sequence-level data units includes: Select representative slices from the sequence-level data units; Based on the large medical image information model, reasoning is performed on the representative slices to obtain the anatomical location, image modality confirmation result and image quality level corresponding to the sequence-level data unit; The anatomical location, the image modality confirmation result, and the image quality level are used as the semantic classification result.

[0047] In some implementations, when performing sequence-level splitting of medical imaging data within processed medical imaging-related data, DICOM examination data can be split into multiple sequences based on unique sequence instance identifiers. The unique sequence instance identifier distinguishes different acquisition sequences within the same examination. After splitting the DICOM examination data, the multiple sequences are grouped by imaging modality, examination site, and sequence description, resulting in multiple sequence-level data units. Imaging modality can be used to distinguish between different imaging methods such as CT, MRI, X-ray, and ultrasound; examination site can be used to distinguish between different anatomical sites such as the chest, abdomen, and skull; and sequence description can reflect the specific acquisition protocol and scanning sequence characteristics. These grouped sequence-level data units can then serve as input for subsequent classification processing.

[0048] In some implementations, when classifying sequence-level data units based on a large-scale medical image information model, representative slices can be selected for each sequence-level data unit first. Representative slices can be extracted from the sequence-level data units according to a preset sampling interval, or slices that reflect the main anatomical structures and imaging features can be selected based on the image content. Subsequently, the large-scale medical image information model is used to infer the representative slices to obtain the anatomical location, image modality confirmation result, and image quality level corresponding to the sequence-level data unit; and the anatomical location, image modality confirmation result, and image quality level are used as semantic classification results. The semantic classification results can be directly used for subsequent sub-report matching, or as a basis for image quality screening and result verification.

[0049] This embodiment first refines medical image data from examination-level granularity to sequence-level data units, and then obtains corresponding semantic classification results through a large medical image information model. This refines the objects and matching criteria for subsequent sub-reports to the sequence-level granularity, thereby establishing a clearer semantic basis for the subsequent association between sequence-level data units and report text.

[0050] In this embodiment, the step of splitting the report text in the processed medical image-related data to obtain at least one sub-report, and extracting the semantic tags corresponding to the sub-report, includes: The report text is subjected to part identification and paragraph segmentation to obtain at least one sub-report; Extract the semantic anatomical site labels corresponding to each of the sub-reports as semantic tags for the sub-reports; The process of matching the sequence-level data units with the sub-reports based on the semantic classification results of the sequence-level data units and the semantic tags corresponding to the sub-reports, and establishing the association between the sequence-level data units and the sub-reports, includes: Based on the semantic classification results of the sequence-level data units and the semantic anatomical site labels corresponding to each sub-report, the sequence-level data units and the sub-reports are matched to obtain matching results; When the confidence level of the matching result is higher than or equal to a preset threshold, an association relationship between the sequence-level data unit and the sub-report is established based on the matching result; When the confidence level of the matching result is lower than a preset threshold, the matching result is pushed to manual confirmation; The association between the sequence-level data unit and the sub-report is established based on the results of manual confirmation.

[0051] In some implementations, when splitting the report text in the processed medical image-related data, site identification and paragraph segmentation can be performed on the report text to obtain at least one sub-report. For single-site report text, site identification and paragraph segmentation can result in only one sub-report; for multi-site mixed report text, site identification and paragraph segmentation can result in multiple sub-reports. Subsequently, the semantic anatomical site labels corresponding to each sub-report are extracted as semantic tags for the sub-report. Semantic anatomical site labels can be used to characterize the anatomical object mainly described in the sub-report.

[0052] In some implementations, when matching sequence-level data units with sub-reports based on the semantic classification results of the sequence-level data units and the semantic tags corresponding to each sub-report, the matching can be performed according to the semantic classification results of the sequence-level data units and the semantic anatomical site tags corresponding to each sub-report, resulting in a matching result. The matching result may include the matching object and its corresponding confidence level. When the confidence level of the matching result is higher than or equal to a preset threshold, the association between the sequence-level data unit and the sub-report is directly established based on the matching result; when the confidence level of the matching result is lower than the preset threshold, the matching result is pushed to manual confirmation, and then the association between the sequence-level data unit and the sub-report is established based on the manual confirmation result.

[0053] This embodiment can first split the report text into sub-reports that correspond to different anatomical sites or different examination contents, and then match the semantic classification results of the sequence-level data units with the semantic tags corresponding to the sub-reports. Based on the combination of automatically establishing associations with high confidence and manually confirming associations with low confidence, the final association relationship between the sequence-level data units and the sub-reports is formed.

[0054] In this embodiment, storing the processed medical image-related data, the semantic classification results, and the association relationships includes: The structured data in the processed medical image-related data is written into a relational database; The medical image data in the processed medical image-related data is written into distributed storage; The semantic classification results and the associations are cached and / or indexed.

[0055] In some implementations, when storing processed medical image-related data, semantic classification results, and relationships, the structured data in the processed medical image-related data can first be written into a relational database. Structured data may include patient information, diagnosis and treatment information, examination request information, report process information, quality assessment results, field mapping results, terminology standardization results, and encoding conversion results. The relational database can be a single database or a sharded database.

[0056] In some implementations, the medical image data from the processed medical image-related data is written to distributed storage. Distributed storage can be used to store DICOM examination data and its related files. Simultaneously, semantic classification results and associations are cached and / or indexed. Cached storage supports high-frequency access and fast retrieval, while indexed storage supports retrieval, aggregation, and statistical analysis based on criteria such as patient, examination type, site label, treatment status, and time range.

[0057] This embodiment can adopt different storage strategies for different types of data objects, so that structured data, medical image data, semantic classification results and relationships can be stored in suitable storage media, thereby providing a data foundation for subsequent querying, review, statistics, tracing and reuse.

[0058] In some implementations, the entire intelligent medical data governance system adopts a layered architecture, including at least a data access layer, a rule engine layer, a data processing layer, a distributed task processing layer, and a storage layer. The data access layer is responsible for interfacing with PACS, HIS, RIS systems, and other external data sources; the rule engine layer is responsible for loading, compiling, orchestrating, and executing governance rules; the data processing layer is responsible for implementing quality assessment, data cleaning, standardization transformation, sequence-level splitting, large-scale medical image information model classification, report text splitting, and association establishment; the distributed task processing layer is responsible for distributing, scheduling, concurrently executing, and retrying governance tasks; and the storage layer is responsible for classifying and storing different types of data objects. Each layer can be deployed on different service nodes or implemented in a centralized manner.

[0059] In some implementations, once medical image-related data enters the governance process, a globally unique governance serial number can be assigned to each piece of data. This governance serial number can be used throughout all steps of rule routing, governance processing, sequence-level splitting, classification, report processing, association establishment, and storage. Information such as input parameters, output results, execution time, processing status, and operating account generated during each step can be linked to the governance serial number and recorded. Thus, when auditing, problem localization, process backtracking, or statistical analysis is required, the processing trajectory of a piece of medical image-related data throughout the entire governance process can be reconstructed based on the governance serial number.

[0060] In some implementations, quality assessment may include at least completeness assessment, standardization assessment, and timeliness assessment. Completeness assessment can be achieved by detecting whether pre-defined mandatory tag fields are missing. These pre-defined mandatory tag fields may include key fields such as patient identifier, examination date, examination modality, and examination request number. Standardization assessment can be achieved by checking the validity of the tag field's value range, enumerated value validity, and format validity. Timeliness assessment can be achieved by comparing whether the time difference between the image acquisition time and the report issuance time falls within a pre-defined time window. Furthermore, completeness scores, standardization scores, and timeliness scores can be obtained separately, and a comprehensive score can be calculated according to pre-defined weights. Based on the comprehensive score, the medical image-related data can then be divided into different quality levels. Data of different quality levels can then proceed to different subsequent processing paths.

[0061] In some implementations, data cleaning may include field mapping, data validation, format conversion, special character filtering, and error value correction. Field mapping can map local fields from different hospitals or systems to unified fields; data validation can detect value ranges, enumerated values, and regular expression matching relationships; format conversion can unify different date formats, character encoding formats, or text representations. Standardization conversion processes may include medical terminology standardization and encoding unification conversion. Medical terminology standardization can use medical natural language processing models to normalize non-standard terms, synonyms, and abbreviations, for example, mapping localized expressions to unified terminology; encoding unification conversion may include conversion between different versions of diagnostic codes and mapping between imaging terminology systems.

[0062] In some implementations, the report text may be a single-site report or a multi-site mixed report. For multi-site mixed reports, the different sites involved can be identified first, and then paragraph segmentation can be performed based on paragraph boundaries, semantic boundaries, or rule boundaries to form multiple sub-reports. Each sub-report can have its corresponding semantic anatomical site label extracted. During matching, the anatomical sites of sequence-level data units can be matched with the corresponding semantic anatomical site labels of the sub-reports, or a combination of image modality confirmation results and text content can be considered for matching. The confidence level obtained from the matching can be used to distinguish between automatically established associations and manually confirmed associations.

[0063] In some implementations, quality assessment tasks, large-scale model inference tasks, and repair tasks generated during the intelligent governance of medical data can be written into a task queue, which is then uniformly distributed and executed by the task scheduling module. The task queue can be divided into different channels based on hospital origin, data type, and priority. Tasks requiring large-scale medical image information models can be distributed to inference nodes with graphics processing capabilities. During task execution, concurrent consumption can be achieved through a thread pool, and thread pool parameters can be adjusted according to the computational load. For tasks that fail, retrying can be performed using an exponential backoff mechanism; tasks that still fail after retrying can be transferred to a dead-letter queue and trigger manual processing or an alarm.

[0064] In some implementations, semantic classification results, associations, quality assessment results, and governance logs can be indexed for retrieval by patient, examination type, site label, governance status, and time range. Based on the indexed results, aggregated statistics can also be performed to generate data quality statistics across time periods, hospital sources, or examination types. Combined with the end-to-end information from the aforementioned governance serial number records, the governance process can be tracked and audited to pinpoint problems and trace responsibility in the event of mismatches, omissions, or abnormal governance results.

[0065] In some implementations, in addition to PACS, HIS, and RIS systems, historical or incremental data can be received via batch file import, SFTP transfer, or HTTP interface. This allows the system to support both real-time access processing and offline historical data governance. For different modalities of medical imaging data, such as X-ray, CT, MRI, ultrasound, or other medical image types, it is only necessary to adapt the corresponding data format at the receiving layer, extend the corresponding rules at the governance processing layer, and adapt the corresponding model in the classification module to incorporate them into the same governance process. This enables the entire intelligent medical data governance method and device to adapt to application scenarios with different hospital sources, different image modalities, and different governance scales.

[0066] In one embodiment, a medical data intelligent governance device based on a large model of medical image information is provided, comprising: The receiving module is used to receive medical image-related data, which includes medical image data, diagnosis and treatment information related to the medical image data, and report text. The rule routing module is used to perform rule routing on the medical image-related data based on the source information and data attributes of the medical image-related data, so as to obtain the rule-routed medical image-related data. The governance and processing module is used to process the medical image-related data after the rule routing to obtain the processed medical image-related data. The sequence-level splitting module is used to perform sequence-level splitting on the medical image data in the processed medical image-related data to obtain multiple sequence-level data units; The classification module is used to classify the sequence-level data units based on the large medical image information model to obtain the semantic classification results corresponding to the sequence-level data units; The report processing module is used to split the report text in the processed medical image-related data to obtain at least one sub-report, and extract the semantic tags corresponding to the sub-report; The association establishment module is used to match the sequence-level data unit with the sub-report based on the semantic classification result of the sequence-level data unit and the semantic tag corresponding to the sub-report, and establish the association relationship between the sequence-level data unit and the sub-report; The storage module is used to store the processed medical image-related data, the semantic classification results, and the association relationships.

[0067] The specific implementation details of each module can be found in the above description of the limitations of the intelligent governance method for medical data based on a large model of medical image information, and will not be repeated here.

[0068] In some implementations, the intelligent medical data governance device based on a large-scale medical image information model includes a receiving module, a rule routing module, a governance processing module, a sequence-level splitting module, a classification module, a report processing module, an association establishment module, and a storage module. The receiving module receives medical image-related data; the rule routing module performs rule routing on the medical image-related data according to its source information and data attributes to obtain rule-routed medical image-related data; the governance processing module processes the rule-routed medical image-related data to obtain processed medical image-related data.

[0069] In some implementations, a sequence-level splitting module is used to split the medical image data in the processed medical image-related data into multiple sequence-level data units; a classification module is used to classify the sequence-level data units based on a large medical image information model to obtain the semantic classification results corresponding to the sequence-level data units; a report processing module is used to split the report text in the processed medical image-related data to obtain at least one sub-report and extract the semantic tags corresponding to the sub-report; an association establishment module is used to match the sequence-level data units with the sub-reports based on the semantic classification results of the sequence-level data units and the semantic tags corresponding to the sub-reports to establish the association relationship between the sequence-level data units and the sub-reports; and a storage module is used to store the processed medical image-related data, the semantic classification results, and the association relationship.

[0070] In some implementations, each module can be implemented by a processor executing program instructions from memory, by a dedicated hardware unit, or by a combination of software and hardware. Modules can be deployed on the same server node or distributed across multiple nodes. Specifically, the classification module can be deployed on an inference node with graphics processing capabilities, the rule routing module and governance processing module can be deployed on application service nodes, and the storage module can connect to relational databases, distributed storage, caching systems, and indexing systems.

[0071] This embodiment implements the functional steps of receiving, rule routing, governance processing, sequence-level splitting, classification, report processing, association establishment, and storage in a modular form, thereby providing a corresponding device foundation for the automated governance of medical image-related data.

[0072] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for intelligent governance of medical data based on a large model of medical image information, characterized in that, The method includes: Receive medical imaging-related data, which includes medical imaging data, diagnostic and treatment information related to the medical imaging data, and report text; Based on the source information and data attributes of the medical image-related data, the medical image-related data is routed according to rules to obtain the medical image-related data after rule routing; The medical image-related data after the rule routing is processed to obtain processed medical image-related data; The medical image data in the processed medical image-related data is split into multiple sequence-level data units by sequence-level decomposition. Based on the large medical image information model, the sequence-level data units are classified to obtain the semantic classification results corresponding to the sequence-level data units. The report text in the processed medical image-related data is split to obtain at least one sub-report, and the semantic tags corresponding to the sub-report are extracted; Based on the semantic classification results of the sequence-level data units and the semantic tags corresponding to the sub-reports, the sequence-level data units and the sub-reports are matched to establish the association between the sequence-level data units and the sub-reports; The processed medical image-related data, the semantic classification results, and the association relationships are stored.

2. The method according to claim 1, characterized in that, The received medical image-related data includes: Receive DICOM medical image data from the PACS system; Receive patient information and treatment information from the HIS system; Receive inspection request information, report process information, and report text from the RIS system.

3. The method according to claim 1, characterized in that, The step of performing rule-based routing on the medical image-related data based on the source information and data attributes of the medical image-related data to obtain rule-based routed medical image-related data includes: Based on the hospital source, data type, and priority tags of the medical image-related data, a governance rule chain corresponding to the medical image-related data is determined; The medical image-related data is routed to the corresponding governance rule chain to obtain the rule-routed medical image-related data.

4. The method according to claim 3, characterized in that, The step of determining the governance rule chain corresponding to the medical image-related data based on the hospital source, data type, and priority tag of the medical image-related data includes: Obtain the hospital source, data type, and priority tag corresponding to the medical image-related data; Based on the hospital's source, data type, and priority label, the corresponding governance rule set is determined by matching it with the preset rule routing configuration. The governance rules in the governance rule set are compiled to obtain an executable rule component; The executable rule components are orchestrated using LiteFlow expressions to generate a governance rule chain corresponding to the medical image-related data.

5. The method according to claim 4, characterized in that, The step of routing the medical image-related data to the corresponding governance rule chain to obtain the rule-routed medical image-related data includes: Input the medical image-related data into the corresponding governance rule chain; The LiteFlow execution engine drives the execution of each executable rule component according to the orchestration logic of the governance rule chain; During the execution of each executable rule component, the medical image-related data and / or intermediate processing results are shared among the executable rule components through a context object to obtain the medical image-related data after rule routing.

6. The method according to claim 1, characterized in that, The process of processing the medical image-related data after the rule routing to obtain processed medical image-related data includes: The quality of the medical image-related data after the rule routing is evaluated to obtain the quality evaluation results. Based on the quality assessment results, data cleaning and standardization transformation are performed on the medical image-related data after the rule routing to obtain the processed medical image-related data.

7. The method according to claim 1, characterized in that, The medical image data in the processed medical image-related data is split into multiple sequence-level data units, including: The DICOM examination data in the processed medical image-related data are split at the sequence level based on the unique identifier of the sequence instance to obtain multiple sequences; The multiple sequences are grouped by combining imaging modality, examination site, and sequence description to obtain multiple sequence-level data units; The classification of the sequence-level data units based on the large medical image information model to obtain the semantic classification results corresponding to the sequence-level data units includes: Select representative slices from the sequence-level data units; Based on the large medical image information model, reasoning is performed on the representative slices to obtain the anatomical location, image modality confirmation result and image quality level corresponding to the sequence-level data unit; The anatomical location, the image modality confirmation result, and the image quality level are used as the semantic classification result.

8. The method according to claim 1, characterized in that, The process involves splitting the report text in the processed medical image-related data to obtain at least one sub-report, and extracting the semantic tags corresponding to the sub-report, including: The report text is subjected to part identification and paragraph segmentation to obtain at least one sub-report; Extract the semantic anatomical site labels corresponding to each of the sub-reports as semantic tags for the sub-reports; The process of matching the sequence-level data units with the sub-reports based on the semantic classification results of the sequence-level data units and the semantic tags corresponding to the sub-reports, and establishing the association between the sequence-level data units and the sub-reports, includes: Based on the semantic classification results of the sequence-level data units and the semantic anatomical site labels corresponding to each sub-report, the sequence-level data units and the sub-reports are matched to obtain matching results; When the confidence level of the matching result is higher than or equal to a preset threshold, an association relationship between the sequence-level data unit and the sub-report is established based on the matching result; When the confidence level of the matching result is lower than a preset threshold, the matching result is pushed to manual confirmation; The association between the sequence-level data unit and the sub-report is established based on the results of manual confirmation.

9. The method according to claim 1, characterized in that, The storage of the processed medical image-related data, the semantic classification results, and the association relationships includes: The structured data in the processed medical image-related data is written into a relational database; The medical image data in the processed medical image-related data is written into distributed storage; The semantic classification results and the associations are cached and / or indexed and stored.

10. A medical data intelligent management device based on a large model of medical image information, characterized in that, The device includes: The receiving module is used to receive medical image-related data, which includes medical image data, diagnosis and treatment information related to the medical image data, and report text. The rule routing module is used to perform rule routing on the medical image-related data based on the source information and data attributes of the medical image-related data, so as to obtain the rule-routed medical image-related data. The governance and processing module is used to process the medical image-related data after the rule routing to obtain the processed medical image-related data. The sequence-level splitting module is used to perform sequence-level splitting on the medical image data in the processed medical image-related data to obtain multiple sequence-level data units; The classification module is used to classify the sequence-level data units based on the large medical image information model to obtain the semantic classification results corresponding to the sequence-level data units; The report processing module is used to split the report text in the processed medical image-related data to obtain at least one sub-report, and extract the semantic tags corresponding to the sub-report; The association establishment module is used to match the sequence-level data unit with the sub-report based on the semantic classification result of the sequence-level data unit and the semantic tag corresponding to the sub-report, and establish the association relationship between the sequence-level data unit and the sub-report; The storage module is used to store the processed medical image-related data, the semantic classification results, and the association relationships.

Images