A method for building a medical image data set having a uniform structure

Abstract

The application provides a method for establishing a medical image data set with a unified structure, belonging to the technical field of medical image processing and analysis, comprising: based on unstructured image reports and corresponding DICOM format image data, unstructured image report texts, DICOM images and DICOM metadata are extracted; the unstructured image report texts are processed through a private AI model to obtain structured data, the structured data and the DICOM metadata are merged to obtain an SR file, the SR file comprises the DICOM metadata, medical feature codes and a unique identifier; based on the unique identifier, the SR file and the DICOM image are associated to form a structured data set; the medical feature codes in the SR file are associated with the unique identifier, and a user input is mapped to the medical feature codes through a BERT-NER model to query the structured data set.

Description

Technical Field

[0001] The present application belongs to the field of medical image processing and analysis technology, and specifically relates to a method for establishing a medical image dataset with a unified structure. Background Art

[0002] The standardized integration and efficient management of medical imaging data are the core challenges in achieving precision medicine and artificial intelligence-assisted diagnosis. The difficulty in data standardization lies in the arbitrariness of annotation and the lack of uniformity in data structure. In current clinical practice, imaging diagnostic reports are mostly stored independently in the RIS system in the form of unstructured free text, and there is a lack of an automated association mechanism between them and DICOM format image data. Traditional manual annotation methods have low processing efficiency, and the differences in understanding of text descriptions by different physicians lead to insufficient consistency in annotation results. In addition, existing image annotation work usually needs to be performed separately outside the diagnostic business process, further exacerbating the problems of low data annotation efficiency and uneven quality. These factors together lead to technical bottlenecks such as high construction costs, poor data consistency, and weak cross-modal analysis capabilities in medical imaging datasets, which seriously restrict the training effect and clinical application value of artificial intelligence models.

[0003] Existing technologies suffer from low efficiency in processing unstructured reports. Different doctors may use different terminology and descriptions when writing imaging reports, making it difficult to ensure data consistency and accuracy. There is a lack of effective automated linkage mechanisms between imaging data and corresponding diagnostic reports, leading to severe information silos. Furthermore, inconsistent terminology across different medical systems makes it difficult for traditional technologies to achieve automated linkage and integration of images and reports. Furthermore, while existing images and reports can help doctors learn about disease symptoms, existing technologies are unable to deliver images and reports of varying difficulty levels to doctors based on their proficiency, lacking personalized design for doctors' learning. Summary of the Invention

[0004] In order to solve at least one technical problem in the background technology, the present application provides a method for establishing a medical imaging dataset with a unified structure.

[0005] The technical solutions adopted in this application are:

[0006] A first embodiment of the present application provides a method for establishing a medical image dataset with a unified structure, comprising:

[0007] Based on the unstructured imaging report and the corresponding DICOM format image data, the unstructured imaging report text, DICOM image and DICOM metadata are extracted; the unstructured imaging report text is processed by a private AI model to obtain structured data, and the structured data is merged with the DICOM metadata to obtain an SR file, and the SR file includes the DICOM metadata, medical feature code and unique identifier; based on the unique identifier, the SR file is associated with the DICOM image to form a structured data set; the medical feature code in the SR file is associated with the unique identification code, and the user input is mapped to the medical feature code through the BERT-NER model to query the structured data set.

[0008] According to one embodiment of the present application, the unstructured imaging report text is processed by a private AI model to obtain structured data, and the structured data is merged with the DICOM metadata to obtain an SR file. The SR file includes the DICOM metadata, medical feature code and unique identifier, specifically:

[0009] Based on medical standard terminology, a structured template is defined; the private AI model is fine-tuned in a supervised manner using an manually annotated initial data set, and the output of the private AI model is iteratively optimized so that the output matches the medical standard terminology; the unstructured imaging report text is parsed into the structured data that conforms to the structured template by the private AI model; a data template is defined according to the DICOM SR standard, and it is determined that the structured data matches the unique identifier of the DICOM metadata; based on the DICOM SR standard, an SR file is initialized; basic information in the DICOM metadata is filled into the basic metadata field of the SR file, where the basic information includes at least one of patient information and examination parameters; the structured data is traversed to extract medical feature codes and related information; the medical feature codes are formatted according to the requirements of the DICOM SR standard; the formatted medical feature codes and related information are embedded into the content sequence of the SR file; in addition to the basic metadata fields, the DICOM metadata is also added to the relevant fields of the SR file; the fields are checked, and a DICOM validation tool or library is used to check whether the generated SR file conforms to the DICOM SR standard.

[0010] According to one embodiment of the present application, the SR file is associated with the DICOM image based on the unique identifier to form a structured data set, specifically:

[0011] Confirm that the unique identifier of the SR file is consistent with the unique identifier of the DICOM image; construct a structured report table to store relevant information of the SR file; construct a feature code table to store the medical feature code extracted from the SR file; upload the SR file to the PACS system, use the unique identifier to find and associate the corresponding DICOM image, or, use the unique identifier as a foreign key to establish an association between the structured report table and the DICOM image metadata; associate the DICOM image with the SR file through the unique identifier; associate the SR file with the medical feature code in the feature code table through the unique identifier; support cross-modal queries based on symptoms through the medical feature code, and establish a composite index on the feature code to accelerate queries based on the medical feature code.

[0012] According to one embodiment of the present application, the medical feature code in the SR file is associated with the unique identification code of the research instance, and the user input is mapped to the medical feature code through the BERT-NER model to query the structured dataset, specifically:

[0013] The BERT-NER model is used to parse the natural language description input by the user to identify key medical entities; the key medical entities are converted into corresponding medical feature codes; the medical feature codes are used to construct SQL query statements to retrieve the relevant unique identifiers; based on the unique identifiers, the corresponding SR files are searched in the structured report table; the same unique identifiers are used to search for the relevant DICOM images in the PACS system or DICOM image database; the SR files and the DICOM images are integrated to form a structured data set.

[0014] According to one embodiment of the present application, it further includes:

[0015] When obtaining the unstructured imaging report and the corresponding DICOM format image data, the doctor's learning identity is recorded; the doctor's basic ability is evaluated through standardized tests, and the test results are converted into proficiency scores through a machine learning model; the doctor's operational behavior data is recorded, and the proficiency score is updated in real time using an online learning algorithm; and structured templates and rules are preset for different proficiency scores.

[0016] According to one embodiment of the present application, the structured templates and rules are preset for different proficiency scores, specifically:

[0017] If the proficiency is the first level, the structured data is split into standardized fields and annotations are added; if the proficiency is the second level, the detail of the structured data is reduced, a summary structured report is provided, and key conclusions are retained.

[0018] According to one embodiment of the present application, the processing of the unstructured imaging report text by a private AI model to obtain structured data further includes:

[0019] Prompt words are dynamically generated according to the proficiency to control the level of detail of the private AI model output; if the proficiency is the first level, the private AI model is fine-tuned using preset granularity annotation data; if the proficiency is the second level, the private AI model is fine-tuned using summary annotation data.

[0020] According to one embodiment of the present application, merging the structured data with the DICOM metadata to obtain an SR file further includes:

[0021] When generating the SR file, a proficiency identification field is added to mark the structured degree of the SR file.

[0022] According to one embodiment of the present application, after associating the medical feature code in the SR file with the unique identification code and mapping the user input to the medical feature code through the BERT-NER model to query the structured dataset, the method further includes: recommending differentiated learning content based on the doctor's proficiency score, specifically:

[0023] If the proficiency level is the first level, structured reports and basic cases will be pushed first, combined with interactive learning tools; if the proficiency level is the second level, mixed content will be provided, and complex cases and diagnostic reasoning training will be introduced; cases of similar difficulty will be recommended based on the doctor's historical learning behavior.

[0024] According to one embodiment of the present application, after recommending differentiated learning content based on the doctor's proficiency score, the method further includes:

[0025] Behavioral indicators and ability indicators are used as quantitative indicators of learning effects, and the differentiated learning content is dynamically adjusted based on the quantitative indicators of learning effects; the search keyword weight is adjusted according to the doctor's proficiency, and the search result explanation is provided to doctors with second-level proficiency.

[0026] Due to the adoption of the above technical solution, the beneficial effects achieved by this application are as follows:

[0027] This application ensures the extraction of all relevant information from the original medical imaging examination, including detailed imaging report text, the actual DICOM images, and DICOM metadata containing patient information and examination details. This provides a complete data foundation for subsequent data processing and analysis. By simultaneously extracting unstructured imaging report text, DICOM images, and metadata, data integrity and consistency are ensured, ensuring that each imaging report accurately matches its corresponding medical image. This application converts unstructured imaging reports into structured data, enabling the system to accurately identify and utilize key information (such as lesion location and mass size), significantly improving data operability and interpretability. By merging structured data with DICOM metadata to generate an SR file, the image data and report content are effectively integrated, facilitating unified management and retrieval. The use of medical feature encoding to standardize the representation of specific medical terminology not only improves interoperability between different systems but also provides strong support for subsequent data analysis and clinical decision support. This application utilizes unique identifiers (UIDs) to ensure that each SR file accurately matches its corresponding DICOM image, preventing data confusion or misassociation. Through the association relationship established by UID, users can quickly and easily query relevant SR files and their corresponding DICOM images as needed, which greatly facilitates the doctor's workflow. The structured data set formed not only contains detailed medical images, but also includes processed structured report content, further enhancing the value and usability of the data. This application uses the BERT-NER model, and users only need to enter natural language descriptions to quickly locate the required medical feature codes, and query the structured data set accordingly, which greatly simplifies the process of complex queries. The BERT-NER model can accurately identify and convert key medical entities in user input into corresponding medical feature codes, reducing possible errors in manual searches and improving the accuracy of query results. This query method based on natural language understanding and medical feature coding can provide customized search results based on the specific needs of different users, better meeting the actual work needs of clinicians. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0029] Figure 1 A flowchart of a method for establishing a medical imaging dataset with a unified structure provided in an embodiment of the present application;

[0030] Figure 2This is a schematic diagram of a method for extracting key information from unstructured reports based on a method for establishing a medical imaging dataset with a unified structure, provided in an embodiment of the present application. Figure 1 ;

[0031] Figure 3 This is a schematic diagram of a method for extracting key information from unstructured reports based on a method for establishing a medical imaging dataset with a unified structure, provided in an embodiment of the present application. Figure 2 . DETAILED DESCRIPTION

[0032] Example 1

[0033] like Figure 1 As shown, the first embodiment of the present application provides a method for establishing a medical image dataset with a unified structure, comprising:

[0034] S100 : extracting unstructured image report text, DICOM images, and DICOM metadata based on the unstructured image report and the corresponding DICOM format image data.

[0035] As mentioned above, to obtain unstructured imaging report text, relevant imaging reports are directly retrieved from the radiology information system (RIS). These reports typically exist in free-text format and contain information such as the physician's description of the imaging findings, analysis, and diagnosis. This unstructured content, meaning it lacks a fixed format or classification, requires subsequent processing to be effectively used for data analysis or machine learning model training. Regarding the acquisition of DICOM images, the image data corresponding to these reports is similarly obtained from the RIS system in the Digital Imaging and Communications in Medicine (DICOM) format. DICOM is a standard format used in the medical field for storing, exchanging, and transmitting medical images and related information. This process ensures the acquisition of actual medical image data, which is the foundation for medical diagnosis. This involves extracting DICOM metadata, i.e., extracting relevant metadata from the DICOM-formatted image data. This metadata includes, but is not limited to, patient personal information (such as name, age, and gender), examination parameters (such as scan type and scan time), and other information that facilitates the understanding and use of these image data. This metadata is crucial for determining the source of the image, understanding its context, and associating the image with a specific patient. By extracting this information, we can have a more comprehensive understanding of the context of each piece of image data, providing necessary support for subsequent data processing and analysis.

[0036] For example, the imaging report for a patient is retrieved from the radiology information system (RIS). This report is a physician's description of the mammographic findings, including information such as breast density, lesion location, mass shape, and margin status. These descriptions are unstructured, free text, such as "A round mass with well-defined margins was found in the right upper outer quadrant of the breast." The corresponding DICOM-formatted image data is retrieved from the same system. This means obtaining the actual mammographic image files, which were acquired under specific conditions, such as using a specific type of X-ray equipment and scanning according to set parameters. Metadata needs to be extracted from the DICOM image data. This metadata includes basic patient information such as name, age, gender, and medical record number; it also includes specific parameters regarding the examination, such as the scanning method used, scan time, and exposure conditions. For example, it can be determined that the patient is a 45-year-old woman, and the examination was performed on March 8, 2025, using Digital Breast Tomosynthesis (DBT). This process not only yields the detailed, unstructured imaging report text—the physician's written description of the mammography findings—but also the actual DICOM image itself, the mammography image file itself, along with its associated metadata. This metadata helps us understand the image context, such as the patient's personal information and the examination's technical parameters, providing a comprehensive foundation for subsequent data analysis, diagnostic support, or research. This effectively combines unstructured text descriptions with specific medical images and their associated detailed context, paving the way for further applications.

[0037] It should be noted that in specific implementation scenarios, advanced natural language processing technologies can be used, in addition to the above solution, to conduct in-depth analysis of unstructured imaging report text. For example, named entity recognition (NER) technology can be used to automatically identify and extract key medical terms and concepts, such as disease names, anatomical locations, and lesion characteristics, from these free texts. This helps convert previously unstructured information into structured data, facilitating subsequent data retrieval and analysis. Using standard terminology systems specific to the medical field (such as SNOMED CT and RadLex), the extracted key information is semantically annotated and converted into standard coding. This process not only improves the consistency and accuracy of information but also promotes interoperability between different systems, enabling easier sharing and integration of data from different medical institutions.

[0038] In specific implementation scenarios, building on the aforementioned approach, a knowledge graph for medical imaging can be constructed based on the extracted structured data and the encoded information. This knowledge graph can demonstrate the complex relationships between medical concepts, such as the relationship between diseases and symptoms and the comparison of treatment effectiveness, thus providing a strong knowledge foundation for clinical decision support.

[0039] S200. Process the unstructured imaging report text through a private AI model to obtain structured data, and merge the structured data with the DICOM metadata to obtain an SR file. The SR file includes the DICOM metadata, a medical feature code, and a unique identifier.

[0040] As described above, the process of processing unstructured imaging report text using a private AI model to generate structured data involves several key steps. First, the raw imaging reports obtained from the radiology information system (RIS) are in free text format, containing physicians' descriptions, analyses, and conclusions regarding the imaging findings. Because this information is unstructured, it is difficult to directly use it for data analysis or machine learning. This unstructured imaging report text is processed using a private AI model deployed on a local server. This private AI model is trained and fine-tuned using a large amount of annotated data. The specific process is as follows: Based on authoritative medical terminology systems (such as SNOMED CT and RadLex), and incorporating practical clinical diagnostic and treatment needs, a standardized extraction template is constructed. This template precisely defines the medical entities to be extracted from imaging reports and their logical relationships, forming a standardized framework for extracting key information from unstructured reports. This step ensures accurate identification and extraction of key information during subsequent processing. Based on this standardized template, prompt engineering is carried out to design specialized prompt words that are compatible with the large language model and clearly define the model's task instructions for processing unstructured reports. At the same time, some unstructured reports are manually annotated based on templates to generate an initial annotated dataset, providing a supervised learning foundation for model training. This step is crucial as it directly determines whether the model can correctly understand the task and efficiently execute it. Using the prompt words and the initial annotated dataset, the large language model undergoes the first round of SFT. This round of optimization enables the model to process a large number of unstructured medical imaging reports and generate large-scale structured reports, thereby forming a rich annotated dataset and further expanding the data scale for supervised learning. An iterative training strategy is employed, with multiple rounds of supervised fine-tuning repeated. Each iteration feeds the model with specialized prompt words designed in the previous round and the large-scale annotated data generated. By continuously calibrating the model output to conform to medical professional standards, model performance is gradually optimized, ultimately generating a target private AI model that meets the structured requirements of medical imaging reports. This approach helps continuously improve the model's accuracy and reliability on specific tasks. After multiple rounds of fine-tuning, the target model is deployed to a local server, establishing a local processing system. This system can structure acquired unstructured imaging reports and generate structured imaging reports that meet clinical standards, providing standardized data support for the construction of medical imaging datasets. The system is optimized specifically for structured processing of imaging reports. Specifically, it designs specialized prompts based on the characteristics of imaging reports to guide the model in understanding specific task requirements, such as identifying lesion locations and describing mass characteristics. During processing, medical terminology is converted into a unified coding format (such as using SNOMED CT or RadLex) to improve interoperability and data sharing efficiency across different medical systems.By continuously adjusting model parameters and reviewing results, a quality control mechanism is established to ensure the quality of the model output. Furthermore, the model is continuously improved based on feedback from real-world applications. The model understands medical terminology and can identify key information in imaging reports, such as lesion location, mass size, and morphology. The processing includes, but is not limited to, extracting specific medical entities and their logical relationships and converting unstructured text into a clearly structured data format. For example, information is organized according to predefined templates to ensure that each piece of data is clear and easy to parse. Once structured data is obtained, the next step is to merge it with metadata extracted from DICOM images. DICOM metadata includes basic patient information (such as name, age, and gender), examination parameters (such as scanning method and time), and other information that facilitates the understanding and use of image data. This allows the structured imaging report content to be linked to detailed patient information and the specific circumstances of the imaging acquisition, forming a more comprehensive dataset. An SR file is generated that conforms to the DICOM Structured Reporting (SR) standard. This SR file not only contains the aforementioned DICOM metadata but also incorporates medical feature codes and unique identifiers derived from the structured data. Among them, medical feature coding is the result of standardized representation of medical terms in reports using internationally or domestically recognized medical terminology standards (such as SNOMED CT, ICD, RadLex, etc.), which helps interoperability and data sharing between different medical systems.

[0041] Unique identifiers include:

[0042] The Patient ID (UID) uniquely identifies each patient. This ID links all patient activities and records within a healthcare facility (e.g., examinations, diagnoses, treatments, etc.). This ensures that all medical information related to a patient can be accurately archived and retrieved, avoiding confusion between records for different patients.

[0043] Each medical imaging examination has a unique StudyInstance UID, which is used throughout the entire examination process, including all imaging series and images. This ensures that all relevant data for each examination (including DICOM images and SR files) can be accurately linked together, supporting data consistency and integrity.

[0044] The Series Instance UID (SUID) identifies different image series within the same examination. For example, a chest CT scan may contain multiple series, each with its own SeriesInstance UID. This helps distinguish and manage images of different types or angles within the same examination, facilitating more precise data management and querying.

[0045] Each individual DICOM image has a unique SOPInstance UID, which identifies the specific image file. This ensures that each image file is unique within its series and examination, facilitating accurate location and management of individual image files. The SR Document Instance UID (SR Document Instance UID) assigns a unique identifier to each structured report (SR file). This allows each SR file to be uniquely identified in the system and linked to other associated DICOM images and other SR files.

[0046] For example, an unstructured imaging report for a patient is retrieved from the radiology information system (RIS). The report might read: "A well-defined, round mass, approximately 1.2 x 0.6 cm, suspected to be benign calcification, was detected in the right upper outer quadrant of the breast." This unstructured imaging report is processed using a private AI model deployed on a local server. This model, trained and fine-tuned using a large amount of labeled data, is specifically designed to identify and extract key information from medical imaging reports. In this example, the model identifies the following: Lesion Location: Right Upper Outer Quadrant, Mass Shape: Round, Margin Status: Clear, Size: 1.2 x 0.6 cm, Possible Characteristics: Benign Calcification. This information is converted into a structured format, with fields such as "Lesion Location" and "Mass Shape" as separate data fields, and each feature is represented using standard medical terminology codes (such as RadLex). Simultaneously, metadata is extracted from the DICOM image files for the same patient. This metadata includes, but is not limited to: Patient Name: Zhang San, Patient ID: 000123456, Gender: Female, Age: 45, Examination Date: March 8, 2025, Scanning Method: Digital Breast Tomosynthesis (DBT). Next, the structured data obtained through the proprietary AI model is combined with the DICOM metadata to generate a Structured Report (SR) file that complies with the DICOM SR standard. This SR file not only contains the aforementioned DICOM metadata but also includes medical feature codes and unique identifiers. For example, for the key information mentioned above, "lesion location" is encoded with a specific RadLex code, and "lump shape" is also encoded accordingly. Furthermore, the SR file includes an SR file instance UID, ensuring that each SR file has a unique identity within the entire dataset, facilitating subsequent data management and retrieval.

[0047] It should be noted that, in specific implementation scenarios, the above solution can also be used to develop a rule-based automated review system. This system can check whether the generated structured data meets predefined standards (such as consistency in medical terminology and accuracy of logical relationships). If any anomalies or non-compliance are found, the system will flag them for manual review. A feedback loop can be used to continuously improve the model. Whenever errors or adjustments are made, this information can be fed back into the training dataset, allowing the private AI model to be fine-tuned and improve its accuracy and robustness.

[0048] S300: Based on the unique identifier, associate the SR file with the DICOM image to form a structured data set.

[0049] As mentioned above, the SR file instance UID is a key identifier that ensures each SR file has a unique identity within the entire dataset. The UID follows specific encoding rules, enabling accurate location and differentiation of each report during subsequent data management and retrieval. DICOM images also have unique UIDs, which contain key data points such as patient information and the exam instance. The exam instance UID allows precise matching of one or more SR files with the corresponding DICOM images. When an SR file is generated, each file is assigned a unique UID. This UID, the SR file instance UID, identifies the SR file itself and contains information about the associated DICOM image. Similarly, DICOM images also have a built-in exam instance UID, which is automatically generated when the image is created and persists throughout the medical workflow. The exam instance UID is used to establish data associations in PACS (Picture Archiving and Communication Systems). Specifically, the exam instance UID is used to link SR files to the corresponding DICOM images. This association ensures that whenever a patient's imaging data is accessed, a detailed structured report for that image is also available. Once the SR file is successfully associated with the DICOM image, a rich, structured dataset is formed. This dataset includes not only the original medical image but also the processed, structured report, which contains detailed medical findings, measurements, and other important information. Furthermore, because all data is standardizedly identified and linked using UIDs, data consistency and traceability are greatly improved.

[0050] For example, after a mammography examination, the system generates image data in DICOM format. This image data contains detailed medical images and metadata, such as the patient's name (Li Hua), gender (female), age (48 years old), medical record number (123456789), and examination date (April 20, 2025). Each DICOM image has a unique identifier (examination instance UID). For example, the primary UID for this examination might be "1.3.6.1.4.1.9590.100.1.2.1234567890." The doctor writes an unstructured imaging report based on the examination results, describing key findings, such as "A well-defined, round mass, approximately 1.2 x 0.6 cm, was found in the upper outer quadrant of the right breast, suspected to be benign calcification." This report is then converted into a structured format using a proprietary AI model, generating an SR file that complies with the DICOM SR standard. This SR file also has a unique identifier (exam instance UID), such as "2.16.840.1.113669.632.20.123456789012345678901." In the PACS system, this unique identification information is used to associate the generated SR file with the corresponding DICOM image. Specifically, because both contain the same exam instance UID, they can be easily identified and matched. In this example, the exam instance UID "1.3.6.1.4.1.9590.100.1.2.1234567890" exists in both the DICOM image and the SR file, ensuring accurate correspondence. Once this association is established, a structured dataset containing rich information is formed. This dataset includes not only the original mammographic images but also detailed structured report content, such as detailed descriptions of lesion location, mass size, margin status, and possible features. All of this information is centrally managed and stored, facilitating subsequent query and analysis.

[0051] It should be noted that in specific implementation scenarios, based on the above solution, automated scripts or tools can be developed to automatically check whether the examination instance UID of the SR file and the DICOM image match each time a link is established, and to confirm that all necessary metadata fields are present and correct. A detailed log is created for each data linking operation, including information such as the operation time, performer, and linking results. This helps track potential issues and ensure data integrity and traceability.

[0052] S400 , associating the medical feature code in the SR file with the unique identification code, and mapping the user input to the medical feature code through a BERT-NER model to query the structured dataset.

[0053] As mentioned above, when generating an SR file that complies with the DICOM SR standard, each medical feature (such as lesion location and mass size) is assigned a specific medical feature code. These codes are typically based on internationally or domestically recognized standard terminology systems (such as SNOMED CT and RadLex). Furthermore, both the SR file itself and its associated DICOM image have a unique identifier (examination instance UID) that uniquely identifies them within the entire dataset. Within the SR file, the medical feature code not only records the specific medical information but also links it to the examination instance information through the examination instance UID. For example, an SR file might contain a description of "a round mass with clear margins was found in the upper outer quadrant of the right breast." Key features such as "right breast," "upper outer quadrant," "clear margins," and "round mass" are converted into corresponding medical feature codes, which are then closely linked to the examination instance UID of the SR file. When users want to query a specific medical image or report, they may enter natural language keywords, such as "right breast mass." The system then uses the BERT-NER model to process the user input. The BERT-NER model is an advanced natural language processing technology that identifies and extracts named entities from text and converts them into a predefined encoding format. Specifically, the BERT-NER model analyzes natural language keywords entered by the user, identifies key medical terms, and maps these terms to corresponding medical feature codes based on the trained model parameters. For example, "right breast lump" can be parsed into multiple medical features, including "right breast" and "lump," and corresponding codes are found, such as "RID29896" for "right breast" and "RID39055" for "lump." Once the medical feature codes corresponding to the user input are obtained, the system searches the structured database based on these codes. The system first searches the feature table (sr_features) for records containing all the target codes. These features are then aggregated using the lesion table (sr_lesions) to ensure that all features belong to the same lesion (lesion_id). The system then links the structured report identifier (sr_id) of the lesion to the report table (sr_reports) to locate the specific SR file and its corresponding DICOM image (via the study_instance_uid field). For example, if the user enters a query for "right breast lump," the BERT-NER model will process the resulting codes (such as "RID29896" and "RID39055"). The system then searches the sr_features table for records containing this set of codes and, using the lesion_id to which the features belong, selects lesions that share these features within the same lesion structure.Next, the sr_id in the sr_lesions table is linked to the sr_reports table to further locate the study_instance_uid corresponding to the report, obtain the complete structured report content and its corresponding DICOM image, and finally return the required information to the user.

[0054] For example, after a mammography examination, the doctor's written imaging report is converted into a structured format using a proprietary AI model, generating an SR file that complies with the DICOM SR standard. This SR file details the finding of a "circular mass with clear margins, approximately 1.2 x 0.6 cm, located in the upper outer quadrant of the right breast." During this process, key medical features (such as "right breast," "upper outer quadrant," "clear margins," and "circular mass") are converted into specific medical feature codes (e.g., "RID29896" for "right breast," "RID5707" for "clear and smooth," and "RID34240" for "high density"). The SR file also has a unique identifier (UID), such as "2.16.840.1.113669.632.20.123456789012345678901." In the PACS system, the generated SR file is associated with the corresponding DICOM image based on the examination instance's UID. This allows access to both the raw mammographic images and the detailed structured report content through the same exam instance UID. A doctor wants to find case data related to a "right breast lump." They enter the natural language keyword "right breast lump." Upon receiving this query, the system uses a BERT-NER model to analyze the input natural language keywords. The BERT-NER model identifies the key medical terms "right breast" and "lump" and maps them to corresponding medical feature codes. For example, "right breast" is mapped to "RID29896," while "lump" may correspond to multiple codes, such as "RID39055." Querying the structured dataset: With these medical feature codes, the system searches the sr_features table for records matching "RID29896" and "RID39055." These features are then aggregated by lesion_id to further filter out lesion records that contain all target codes within the same lesion structure. The system then links to the sr_reports table through the sr_id field in the sr_lesions table to locate the corresponding report content and its associated DICOM image data. In this example, the system will return all relevant information about Ms. Wang's mammography, including specific image files and detailed structured reports. Ultimately, the doctor can see Ms. Wang's mammography images and also review the detailed structured report to understand detailed information such as the specific location of the lump (right upper outer quadrant of the breast), morphology (round), edge status (clear), and size (1.2×0.6 cm). This allows doctors to quickly and accurately obtain the information they need, supporting them in making more accurate diagnostic or treatment decisions.

[0055] It should be noted that in specific implementation scenarios, the above solutions can also be combined with image content analysis (such as analyzing DICOM images through deep learning models) and text descriptions (i.e., reports in SR files) to provide a more comprehensive contextual understanding. For example, when processing the query "right breast lump," in addition to identifying keywords in the text description, features in the actual image can also be analyzed to ensure more accurate results. Leveraging knowledge graphs or semantic networks to enrich the understanding of medical terminology allows the system to not only identify specific entities but also understand the relationships between them. For example, the relationship between "right breast" and "benign calcification" can help the system better understand and respond to complex query requests.

[0056] In some embodiments of the present application, the unstructured imaging report text is processed by a private AI model to obtain structured data, and the structured data is merged with the DICOM metadata to obtain an SR file. The SR file includes the DICOM metadata, the medical feature code, and the unique identifier, specifically:

[0057] Based on medical standard terminology, a structured template is defined; supervised fine-tuning of the private AI model is performed using a manually annotated initial data set, and the output of the private AI model is iteratively optimized so that the output matches the medical standard terminology; the unstructured imaging report text is parsed into the structured data that conforms to the structured template by the private AI model; a data template is defined according to the DICOM SR standard, and it is determined that the structured data matches the unique identifier of the DICOM metadata; based on the DICOM SR standard, an SR file is initialized; basic information in the DICOM metadata is filled into the basic metadata field of the SR file; the structured data is traversed to extract medical feature codes and related information; the medical feature codes are formatted according to the requirements of the DICOM SR standard; the formatted medical feature codes and related information are embedded into the content sequence of the SR file;

[0058] In addition to the basic metadata fields, the DICOM metadata is also added to the relevant fields of the SR file; the fields are checked and a DICOM validation tool or library is used to check whether the generated SR file complies with the DICOM SR standard.

[0059] As described above, a structured template is defined based on authoritative medical standard terminology systems (such as SNOMED CT and RadLex). This template details the key information to be extracted from unstructured imaging reports and its format, such as lesion location, mass size, and morphology. This step ensures that subsequent processing accurately identifies and extracts key information and converts it into a clearly structured data format. A private AI model is supervised fine-tuned using a manually annotated initial dataset. This process involves manually annotating a subset of unstructured imaging reports to generate an initial annotated dataset. This annotated data is then used to perform a first round of fine-tuning (Supervised Fine-Tuning, SFT) on the large language model. The optimized model is then used to process a large number of unstructured medical imaging reports and generate large-scale structured reports. An iterative training strategy is employed, with multiple rounds of supervised fine-tuning repeated to continuously calibrate the model output to medical professional standards and gradually optimize model performance, ultimately generating a target private AI model that meets the structured requirements of medical imaging reports. The fine-tuned private AI model is then used to parse the unstructured imaging report text and convert it into structured data that conforms to the structured template defined above. For example, a description such as "A round mass with clear margins was found in the right upper outer quadrant of the breast" is converted into structured data containing specific fields such as "Lesion location: right upper outer quadrant of the breast," "Mass shape: round," and "Margin status: clear." The DICOM SR standard defines a data template to ensure that the structured data matches the unique identifiers in the DICOM metadata. This means that when generating an SR file, all relevant information (examination instance UID) must have consistent and unique identifiers to facilitate subsequent data management and retrieval. A new SR file is initialized based on the DICOM SR standard. This file adheres to the DICOM standard data storage format, ensuring that every byte and field conforms to the standard's definitions and specifications. Basic information from the DICOM metadata is populated into the basic metadata fields of the SR file. This information includes, but is not limited to, the patient's name, age, gender, medical record number, examination date, and scanning method. This metadata is crucial for understanding the image's source, context, and relevance to the patient. The structured data is traversed to extract medical feature codes and related information. For example, descriptions such as "right breast" and "round mass" are converted into specific medical feature codes (e.g., "RID29897" for "right breast" and "RID39055" for "mass"). The extracted medical feature codes are formatted according to the DICOM SR standard. This step ensures that the codes are correctly embedded in the content sequence of the SR file and enables interoperability between different systems. The formatted medical feature codes and related information are embedded in the content sequence of the SR file.In this way, the SR file not only contains basic metadata, but also detailed medical findings, measurement data and other important information. In addition to the basic metadata fields, other relevant DICOM metadata is added to the relevant fields of the SR file. This further enriches the content of the SR file, making it a comprehensive document reflecting the patient's examination situation. Finally, the fields are checked using DICOM validation tools or libraries to check whether the generated SR file meets the DICOM SR standard. This step ensures the quality and consistency of the SR file, enabling it to be seamlessly shared and used in different medical information systems.

[0060] In some embodiments of the present application, based on the unique identifier, the SR file is associated with the DICOM image to form a structured data set, specifically:

[0061] Confirm that the unique identifier (study instance UID) of the SR file is consistent with the unique identifier (study instance UID) of the DICOM image; build a structured report table to store relevant information of the SR file; build a feature code table to store the medical feature codes extracted from the SR file; upload the SR file to the PACS system, use the unique identifier (study instance UID) to find and associate the corresponding DICOM image, or use the unique identifier (study instance UID) as a foreign key to establish an association between the structured report table and the DICOM image metadata; associate the DICOM image with the SR file through the unique identifier (study instance UID); associate the SR file with the medical feature codes in the feature code table through the unique identifier (study instance UID); support disease-based cross-modality queries through the medical feature codes, and establish a composite index on the feature codes to accelerate queries based on the medical feature codes.

[0062] As mentioned above, it is necessary to verify that the unique identifier (exam instance UID) in the SR file matches the unique identifier (exam instance UID) in the DICOM image. These UIDs are assigned to each individual exam instance upon generation and remain unchanged throughout the medical process. By comparing the exam instance UID, it is ensured that the SR file and the corresponding DICOM image truly belong to the same exam. Since an SR report may include multiple lesions, the system must construct a lesion table (sr_lesions) to manage structured information at the lesion level. This table records the structured location and affiliation of each lesion in the report. Its key fields include: lesion unique identifier (lesion_id), structured report ID (sr_id), lesion order number (lesion_order), and lesion text description (description, optional). This table establishes a foreign key relationship with the structured report table via sr_id, ensuring that each lesion can be accurately traced back to its source report. The system also requires a feature code table (sr_features) to store all medical feature codes extracted from the SR file. For example, descriptions such as "right breast" and "round lump" will be converted into standardized medical feature codes (such as "RID29896" for "right breast" and "RID39055" for "lump"). The sr_features table is linked to the lesion table (sr_lesions) through a foreign key to ensure that each feature code can be traced back to its corresponding lesion structure (through lesion_id) and that each feature code can be traced back to the SR file from which it originated. Next, the SR file is uploaded to the PACS system. During the upload process, a unique identifier (examination instance UID) is used to find and associate the corresponding DICOM image. This means that when the SR file is uploaded, the system will automatically find the matching DICOM image based on the examination instance UID and establish an association between the two. The purpose of this is to ensure that every time a patient's imaging data is accessed, a detailed structured report of the image can be obtained synchronously. By checking the instance UID, not only can a direct association be established between the DICOM image and the SR file, but an indirect association can also be established between the SR file and the medical feature code in the feature coding table. Specifically:

[0063] DICOM images and SR files association: By checking the instance UID, it can be determined which DICOM images are related to a specific SR file. SR file and feature code association: By checking the instance UID as a foreign key, a link can be established between the structured report table and the feature code table, so that each medical feature code contained in the SR file can be tracked and queried. In order to support disease-based cross-modality queries, it is very important to establish a composite index on the feature code. The composite index usually includes multiple fields, such as medical feature code (radlex_code) and examination instance UID (study_instance_uid). This index structure greatly improves the speed and efficiency of queries based on specific medical features, allowing users to quickly locate all relevant records containing specific medical features, whether they are images or structured reports. Through the above steps, especially by establishing a composite index in the feature code table, queries based on medical feature codes can be significantly accelerated. For example, if a doctor wants to find all cases with "benign calcification", the system can quickly locate all records containing the corresponding feature code through the index, and return the relevant DICOM images and SR files, thereby providing strong support for clinical decision-making.

[0064] In some embodiments of the present application, the association of the medical feature code in the SR file with the study instance unique identifier is achieved by mapping user input to the medical feature code using a BERT-NER model to query the structured data set, specifically:

[0065] The BERT-NER model is used to analyze the natural language description input by the user and identify key medical entities; the key medical entities are converted into corresponding medical feature codes; the medical feature codes are used to construct SQL query statements to retrieve relevant unique identifiers; based on the unique identifiers, the corresponding SR files are searched in the structured report table; using the same unique identifiers, the relevant DICOM images are searched in the PACS system or DICOM image database; the SR files and DICOM images are integrated to form a structured data set.

[0066] As mentioned above, when a user enters a natural language description (for example, "right breast lump"), the system uses the BERT-NER (Bidirectional Encoder Representations from Transformers - Named Entity Recognition) model to parse this description. The BERT-NER model identifies and extracts key medical entities from the text, such as "right breast" and "lump." These key medical entities are the core content of the user's query and represent the specific medical features they are looking for. The system then converts the identified key medical entities into corresponding medical feature codes. For example, "right breast" might be converted to "RID29896," while "lump" might be converted to "RID39055." These codes are typically based on internationally or domestically recognized standard terminology schemes (such as SNOMED CT and RadLex), ensuring interoperability and consistency across different systems. Using the converted medical feature codes, the system constructs a structured SQL query to retrieve lesion records containing all target codes and ultimately returns the corresponding study instance UID (study_instance_uid). Specifically, the system searches the feature code table (sr_features) for records containing the specified RadLex code. Using the lesion_id aggregation, it determines whether any lesion has all of the specified codes. If the conditions are met, the system further retrieves the corresponding sr_id from the lesion table (sr_lesions) and links it to the structured report table (sr_reports) to extract its study_instance_uid. For example, if the query conditions are "RID29896" (right breast) and "RID39055" (lump), the system retrieves all records containing both codes within the same lesion structure and returns their corresponding study instance UIDs. Based on the study instance UID obtained from the previous step, the system searches the structured report table (sr_reports) for the corresponding SR file. Because each SR file has a unique study instance UID, this step accurately locates all SR files relevant to the user's query. These SR files contain detailed imaging report content, such as lesion location, mass size, and margin status. At the same time, using the same exam instance UID, the system searches for related DICOM images in a PACS system or DICOM image database. Because the exam instance UID remains consistent throughout the medical process, it ensures that a DICOM image that exactly matches a specific SR file is found. This step allows physicians to view not only a detailed structured report but also the actual medical images, providing more comprehensive information to support clinical decision-making.Finally, the system integrates the found SR files and DICOM images to form a complete structured dataset. This dataset includes not only the original medical images but also detailed structured reports and all medical feature codes extracted from them. This integration significantly improves data usability and accessibility, allowing doctors to quickly and accurately obtain the information they need, supporting more efficient diagnosis and treatment.

[0067] In some embodiments of the present application, it also includes: when obtaining the unstructured imaging report and the corresponding DICOM format image data, recording the doctor's learning identity; evaluating the doctor's basic ability through standardized tests, and converting the test results into a proficiency score through a machine learning model; recording the doctor's operational behavior data, and using an online learning algorithm to update the proficiency score in real time; and presetting structured templates and rules for different proficiency scores.

[0068] As mentioned above, when obtaining unstructured imaging reports and corresponding DICOM-formatted image data, the physician's identity information (such as ID or username) is recorded. This step ensures that all subsequent operations can be traced back to the specific physician, providing a basis for personalized evaluation and feedback. A user authentication module is integrated into the system to automatically record the performer of each operation. This information can be stored in log files or databases for subsequent analysis. Standardized tests are used to assess physicians' basic competencies, such as imaging interpretation skills and diagnostic accuracy. This helps understand physicians' professional proficiency and provides a basis for personalized training and development. Regular online tests or simulated case analyses are conducted, covering multiple areas such as medical knowledge and imaging interpretation skills. Test results are analyzed using machine learning models to convert quantitative scores into more intuitive proficiency scores. For example, a model trained on historical data can predict a physician's actual performance based on their performance in the field. Physicians' actual operational data (such as report writing quality, diagnostic speed, and patient feedback) is recorded and their proficiency scores are updated in real time using online learning algorithms. This approach reflects physicians' true capabilities and progress, allowing for timely identification of areas for improvement. The system continuously collects physician operational data, including but not limited to report generation time, terminology accuracy, and patient treatment outcomes. This data is fed into a pre-trained online learning model, which continuously adjusts the doctor's proficiency score based on new data. For example, if a doctor's diagnostic accuracy has significantly improved in recent times, their score will increase accordingly.

[0069] Based on the proficiency scores of different doctors, the system presets specific structured templates and rules for them. For doctors with less experience, more detailed and guiding templates can be provided; for those with more experience, more concise and efficient templates can be used. Develop a multi-level structured template system, with each level corresponding to a different proficiency score range. For example, junior doctors may use templates with more prompts and instructions, while senior doctors can use concise templates. In addition to templates, different sets of rules can also be set for doctors with different proficiency levels. For example, for novice doctors, the system may require them to confirm the report twice before submitting it; for senior doctors, they are allowed to submit it directly.

[0070] In some embodiments of the present application, structured templates and rules are preset for different proficiency scores, specifically:

[0071] If the proficiency is the first level, the structured data is split into standardized fields and annotations are added; if the proficiency is the second level, the detail of the structured data is reduced, a summary structured report is provided, and key conclusions are retained.

[0072] As mentioned above, for physicians at the first level of proficiency (typically new or less experienced physicians), the system provides a highly detailed, structured template with detailed annotations next to each field. This helps them better understand the significance of each entry and ensures the accuracy and completeness of their reports.

[0073] Split structured data into multiple standardized fields. For example, information such as "lesion location", "mass size", and "margin status" in the imaging report are listed separately instead of being described as a whole. Add detailed annotations to each field to help doctors understand the specific meaning of each field and its filling requirements. For example, add a note next to the "lesion location" field: "Please select the correct site based on standard anatomical terms, such as the right upper outer quadrant of the breast." Provide specific examples or guidance statements to help doctors fill in each field correctly. For example, "If a lump is found, please measure its maximum diameter and record it here."

[0074] Assuming Dr. Li is a new junior doctor, he uses a template designed for the first level of proficiency when writing a mammography image report. This template explicitly lists all the fields that need to be filled in, with detailed annotations and examples provided next to each field to help Dr. Li complete the report accurately. For doctors at the second level of proficiency, who are typically experienced and can quickly and accurately complete tasks, the system provides a simplified structured template that highlights key conclusions. This template reduces unnecessary details, allowing doctors to complete reports more efficiently while ensuring that key information is not missed.

[0075] Compared to the first level template, the second level template reduces the level of detail in structured data. For example, instead of filling in detailed anatomical locations and specific dimensions one by one, a comprehensive description is given directly.

[0076] While reducing the level of detail, the most important conclusion part is still retained. For example, only "There is a round mass in the outer upper quadrant of the right breast with clear edges, suspected benign calcification" needs to be pointed out in the report, without the need for detailed description of each measurement value.

[0077] The system can automatically fill in some commonly used information based on historical data and common patterns, further reducing the workload of doctors. For example, the system can automatically generate some common diagnostic suggestions based on previous cases.

[0078] Assuming Dr. Zhang is an experienced doctor, he uses a template designed for the second level of proficiency when writing a mammography image report. This template only requires him to fill in a few key fields such as lesion location, mass shape, and preliminary diagnosis conclusion, while other details are automatically processed or simplified by the system. This allows Dr. Zhang to complete the report more quickly and focus on complex case analysis.

[0079] Through this hierarchical structured template design, the system can provide personalized support according to the proficiency of different doctors:

[0080] For new doctors, detailed field splitting and annotation instructions help them understand and learn, ensuring the quality and accuracy of the report.

[0081] For experienced doctors, simplified summary reports improve work efficiency, allowing them to complete routine tasks more quickly and focus on complex problems.

[0082] This approach not only enhances the physician experience but also optimizes the efficiency and quality of the entire healthcare process. It embodies the concept of personalized education and support, leveraging technology to help physicians maximize support and growth opportunities at each stage of their career development. Furthermore, this design helps medical institutions better manage and cultivate talent, ultimately improving overall service levels.

[0083] In some embodiments of the present application, the unstructured imaging report text is processed by a private AI model to obtain structured data, and further includes: dynamically generating prompt words according to proficiency to control the level of detail output by the private AI model; if the proficiency is the first level proficiency, fine-tuning the private AI model using preset granularity annotation data; if the proficiency is the second level proficiency, fine-tuning the private AI model using summary annotation data.

[0084] As mentioned above, by dynamically generating different prompt words based on the doctor's proficiency, it is ensured that the content output by the private AI model can adapt to the needs of doctors at different levels. Prompt words can guide the model to process unstructured imaging report text more accurately and generate appropriate structured data. For novice or less experienced doctors (first level of proficiency), prompt words will be more detailed and specific, helping the model to generate structured data with more details. For example, prompt words may require the model to provide a specific anatomical description when identifying the location of the lesion and annotate the specific size of each feature. Example prompt words: "Please describe the location, size, morphology and marginal status of the lesion in detail. For example, a round mass with a diameter of 1.2 cm and clear margins was found in the upper outer quadrant of the right breast."

[0085] For experienced doctors who can complete the task quickly and accurately (Level 2 proficiency), prompts are more concise, emphasizing key conclusions. This helps the model generate concise, summary-style structured data. An example prompt is: "Please summarize the key findings and provide a preliminary diagnosis. For example, there is a round mass suspected to be benign calcification in the right breast." The level of detail in the private AI model's output is adjusted based on the doctor's proficiency, ensuring that the generated structured data meets the doctor's actual needs while improving work efficiency.

[0086] For doctors at the first level of proficiency, the model output should be as detailed as possible. For example, when processing a mammography report, the model not only needs to identify the "right breast lump", but also needs to provide detailed information such as the specific size, shape, and edge status. Output example: "A round lump with a diameter of 1.2 cm was found in the upper outer quadrant of the right breast with clear edges and suspected benign calcification." For doctors at the second level of proficiency, the model output should be more concise and focus on key conclusions. For example, only the main findings and preliminary diagnostic conclusions need to be indicated. Output example: "There is a round lump in the right breast that is suspected to be benign calcification." Depending on the proficiency of different doctors, the private AI model is fine-tuned using labeled data of different granularities to ensure that the model can better adapt to the needs of specific users.

[0087] For novice doctors, the model is fine-tuned using detailed annotated data. These annotated data usually include very detailed fields such as lesion location, mass size, margin status, etc. This fine-tuning method ensures that the model can generate detailed structured data to help novice doctors understand and learn. Example annotated data: "Legion location: right upper outer quadrant of the breast; mass size: 1.2 cm in diameter; morphology: round; margin status: clear; preliminary diagnosis: suspected benign calcification." For senior doctors, the model is fine-tuned using more concise summary annotated data. These annotated data only contain key conclusions and reduce unnecessary details. This fine-tuning method improves the model's ability to generate summary structured data, enabling senior doctors to complete reports faster. Example annotated data: "Main findings: There is a round mass suspected of benign calcification in the right breast." Suppose Dr. Li is a novice doctor who has just joined the job, and Dr. Zhang is an experienced senior doctor. When they write mammography imaging reports, the system will automatically adjust the behavior of the private AI model according to their proficiency:

[0088] Dr. Li (Level 1 Proficiency): The system generates detailed prompts to guide the model in generating structured data containing detailed information. The model is fine-tuned using annotated data of a preset granularity to ensure that the output is comprehensive and easy to understand. Output example: "A round mass with a diameter of 1.2 cm was found in the upper outer quadrant of the right breast with clear margins, suspected to be benign calcification." Dr. Zhang (Level 2 Proficiency): The system generates concise prompts to emphasize key conclusions. The model is fine-tuned using summary annotated data to ensure that the output is concise and clear. Output example: "A round mass suspected to be benign calcification was found in the right breast." By dynamically adjusting prompts and fine-tuning strategies based on the doctor's proficiency, the system can better support the work needs of doctors at different levels. For novice doctors, detailed prompts and detailed fine-tuning of annotations help them understand and learn, ensuring the quality and accuracy of their reports. For experienced doctors, concise prompts and fine-tuning of summary annotations improve work efficiency, allowing them to complete routine tasks more quickly and focus on complex issues.

[0089] In some embodiments of the present application, the merging of the structured data with the DICOM metadata to obtain the SR file further includes: when generating the SR file, adding a proficiency identification field to mark the degree of structuring of the SR file.

[0090] As mentioned above, by adding a dedicated field to the SR file to mark its structured level, it is possible to clearly indicate the level of physician proficiency based on which the file was generated. This not only helps to understand the quality and level of detail of the report, but also supports personalized data management and application.

[0091] When generating an SR file, the system automatically adds a "Proficiency Level" field based on the proficiency score of the physician who processed the imaging report. This field can be a simple numeric value (such as 1 for first-level proficiency, 2 for second-level proficiency) or a descriptive label (such as "beginner" or "senior"). Example field name: "proficiency_level" Example values: First-level proficiency: "1" or "beginner", Second-level proficiency: "2" or "senior".

[0092] By marking the level of structure of SR files, medical institutions can help manage and utilize these files more effectively. For example, when conducting data analysis, clinical decision support or research, appropriate files can be selected according to different levels of structure. If the SR file is generated based on the first level of proficiency, then its level of structure will be very high, containing detailed information such as anatomical location, specific size, morphological characteristics, etc. This type of file is suitable for situations where in-depth understanding of case details is required. Example content: "Lesion location: right upper outer quadrant of breast; mass size: 1.2 cm in diameter; morphology: round; edge status: clear; preliminary diagnosis: suspected benign calcification." If the SR file is generated based on the second level of proficiency, then its level of structure is relatively low, retaining only key conclusions and main findings. This type of file is suitable for quick browsing and preliminary judgment. Example content: "Main finding: There is a round mass suspected of benign calcification in the right breast."

[0093] Suppose a hospital's PACS system stores a large number of SR files generated by physicians of varying proficiency levels. The hospital now wants to categorize and further analyze these files. The hospital can use the "Proficiency Level Identifier" field in the SR files to sort all files into two categories: detailed reports generated by less experienced physicians (Level 1 Proficiency) and concise reports generated by more experienced physicians (Level 2 Proficiency). This categorization helps quickly find the information needed during daily work. For example, during routine exams, concise reports can be prioritized for efficiency, while detailed reports can be accessed in depth for complex case discussions. For research or quality control projects, researchers can filter specific SR file types based on the "Proficiency Level Identifier" field for analysis. For example, they can compare the accuracy, completeness, and consistency of reports generated by different proficiency levels. This analysis helps identify potential problem areas and provide targeted training and support to physicians.

[0094] The system can provide personalized feedback and educational suggestions to each doctor based on the "Proficiency Indicator" field in the SR file. For example, for novice doctors, annotations and guidance in the detailed report can help them improve their skills; for experienced doctors, they can be encouraged to simplify their reports while maintaining high quality standards.

[0095] By adding a proficiency indicator field and marking the degree of structure when generating SR files, the system can better adapt to the needs of different doctors and provide greater flexibility and accuracy for subsequent data management and application. This approach not only improves the doctor's work experience but also optimizes the efficiency and quality of the entire medical process. It embodies the concept of personalized education and support, using technology to help doctors obtain maximum support and growth opportunities at each stage of their career development. Furthermore, this design helps medical institutions better manage and cultivate talent, improving overall service levels.

[0096] In some embodiments of the present application, after associating the medical feature code in the SR file with the unique identification code and mapping the user input to the medical feature code through the BERT-NER model to query the structured data set, it also includes: recommending differentiated learning content based on the doctor's proficiency score, specifically: if the proficiency is the first level proficiency, then structured reports and basic cases are pushed preferentially, combined with interactive learning tools; if the proficiency is the second level proficiency, then mixed content is provided, and complex cases and diagnostic reasoning training are introduced; and cases of similar difficulty are recommended based on the doctor's historical learning behavior.

[0097] As mentioned above, by analyzing the doctors' proficiency scores, the system is able to recommend learning content that best suits their current level for each doctor, thereby promoting their professional growth and development. For doctors at the first level of proficiency, the system will prioritize pushing structured reports and basic cases containing detailed information. These reports usually include detailed anatomical location, specific dimensions, morphological characteristics and other information to help doctors understand the meaning of each content and ensure the accuracy and completeness of the report. Example learning content: Structured report: "Language location: right upper outer quadrant of the breast; Lump size: 1.2 cm in diameter; Morphology: round; Margin status: clear; Preliminary diagnosis: suspected benign calcification." Basic case: "A 45-year-old female patient was found to have a round mass with a diameter of approximately 1.2 cm and clear margins in the right breast."

[0098] To help novice doctors better understand and master the knowledge, the system also provides a series of interactive learning tools. These tools include simulation exercises, real-time feedback mechanisms, and a case discussion platform. For example, doctors can practice image interpretation in a virtual environment, and the system will provide instant feedback and suggestions. When writing reports, the system can provide instant prompts and corrections based on the doctor's input, helping them improve accuracy.

[0099] For doctors at the second level of proficiency, the system provides more diverse learning content, including both basic and complex cases. This mixed content is designed to maintain doctors' mastery of foundational knowledge while challenging them to apply their skills in more complex scenarios. Example learning content: Complex case: "A 60-year-old male patient with a long history of smoking presented with a chest CT scan showing multiple pulmonary nodules, some with spiculated margins, suspected of being malignant." This mixed content example combines basic and complex cases to help doctors transition between different levels of difficulty. In addition to providing complex cases, the system also introduces diagnostic reasoning training modules. These training modules are designed to enhance doctors' clinical thinking and problem-solving skills, enabling them to make more accurate judgments when faced with complex conditions. Example training: Doctors are required to derive possible diagnoses based on provided medical records and imaging results, and explain their rationale. The system provides feedback and guidance based on the doctors' responses. By analyzing doctors' historical learning behavior, the system can more accurately recommend learning cases that match their current level. This not only improves learning efficiency but also enhances doctors' interest and motivation. The system records each doctor's learning activity, including the types of cases they review, the number and quality of exercises they complete, and the discussion topics they participate in. Based on this data, the system can analyze doctors' learning preferences and areas of expertise. Example data point: Doctor A frequently reviews breast cancer cases and performs well in related exercises. Doctor B focuses more on cardiovascular disease diagnosis and excels in diagnostic reasoning training in this area.

[0100] Based on the above analysis results, the system recommends cases of similar difficulty and type to each doctor's historical learning behavior. For example, if a doctor has mainly handled basic breast cancer cases in the past few months and performed well, the system may begin to recommend some slightly more complex breast cancer cases to gradually improve their skills. Example recommendation: For Doctor A, the system may recommend some complex cases involving multiple breast nodules or other complications. For Doctor B, the system may recommend some cardiovascular disease cases involving acute myocardial infarction or other emergency situations.

[0101] By recommending differentiated learning content based on physician proficiency scores, the system can better support physicians' professional growth and development. For novice physicians, detailed structured reports and basic case studies, combined with interactive learning tools, help them lay a solid foundation and gradually improve their skills. For experienced physicians, mixed content and diagnostic reasoning training are provided to both consolidate fundamental knowledge and challenge their application in complex scenarios. By recommending cases of similar difficulty based on historical learning behavior, learning content is more tailored to physicians' actual needs, improving learning efficiency and interest. This approach not only enhances the physician's work experience but also optimizes the efficiency and quality of the entire medical process. It embodies the concept of personalized education and support, using technology to help physicians obtain maximum support and growth opportunities at each stage of their career development. Furthermore, this design helps medical institutions better manage and cultivate talent, thereby improving overall service levels.

[0102] In some embodiments of the present application, after recommending differentiated learning content based on the doctor's proficiency score, the method further includes:

[0103] Taking behavioral indicators and ability indicators as quantitative indicators of learning effects, and dynamically adjusting the differentiated learning content based on the quantitative indicators of learning effects;

[0104] Adjust the search keyword weight according to the doctor's proficiency, and provide search result explanations for doctors with second-level proficiency.

[0105] As mentioned above, by analyzing physicians' behavioral indicators (such as the number of exercises completed and the frequency of discussions) and competency indicators (such as diagnostic accuracy and report quality), the system can quantify physicians' learning outcomes and dynamically adjust recommended learning content accordingly. The system records each physician's specific learning activities, including the number of cases reviewed, the number of exercises completed, and the online discussions they participated in. Example behavioral indicators: Physician A completed 10 basic case exercises and participated in 3 online discussions in the past week. Physician B completed 5 complex case exercises and submitted 2 detailed case analysis reports in the past month. The system also assesses physicians' practical skills, such as diagnostic accuracy and report writing quality. Example competency indicators: Physician A's diagnostic accuracy in the basic case exercises was 90%. Physician B's report quality score in the complex case exercises was 85 out of 100.

[0106] Based on the aforementioned behavioral and competency indicators, the system dynamically adjusts recommended learning content. If a physician excels in a particular area, the system may reduce the amount of basic content recommended in that area and increase more challenging cases. Conversely, if a physician performs poorly in a particular area, the system may increase basic exercises and guidance in that area. Example adjustments: If physician A performs exceptionally well on basic breast cancer cases, the system may begin recommending slightly more complex breast cancer cases or introduce diagnostic reasoning training. If physician B performs averagely on complex cardiovascular disease cases, the system may increase basic knowledge review and basic exercises in that area. Search keyword weighting is adjusted based on the physician's proficiency level, ensuring that search results are more tailored to their specific needs and expertise. This helps improve the efficiency and accuracy of information acquisition. For physicians at Level 1 proficiency, the system prioritizes materials and cases related to basic content during the search process. For example, when a physician enters "breast lump," the system prioritizes basic case reports and structured templates that provide detailed information such as anatomical location, specific dimensions, and morphological features. Example search result: "A 45-year-old female patient was found to have a round mass with a diameter of approximately 1.2 cm and clear margins in the right breast." Structured template: "Lesion location: right upper outer quadrant of breast; Mass size: 1.2 cm in diameter; Shape: round; Margin status: clear."

[0107] For doctors with second-level proficiency: the system will prioritize displaying materials related to complex cases and advanced diagnoses during the search process. In addition, the system will provide explanations of the search results to help doctors better understand and apply this information. Example search results and explanations: Search results: "A 60-year-old male patient with a long history of smoking, chest CT showed multiple lung nodules, some with burr-like margins, suspected of being malignant tumors." Explanation: "This case involves multiple lung nodules with burr-like margins, which usually indicate a high possibility of malignancy. It is recommended to make a comprehensive judgment based on the patient's medical history and other imaging examination results."

[0108] For doctors with the second level of proficiency, the system not only provides search results, but also provides concise explanations to help doctors quickly understand and apply this information. The system can automatically generate short explanatory texts based on the search results, highlighting key conclusions and precautions. This explanation can be based on a predefined knowledge base or automatically generated through natural language processing technology. Example explanation: When a doctor searches for "lung nodules", in addition to returning relevant cases, the system will also provide an explanation: "In this case, the patient has multiple lung nodules, some with burr-like margins, suggesting a high possibility of malignancy. Further PET-CT scans or tissue biopsies are recommended to confirm the diagnosis."

[0109] By using behavioral and competency indicators as quantitative indicators of learning outcomes and dynamically adjusting differentiated learning content based on these indicators, the system can more accurately meet doctors' learning needs and promote their professional growth and development. At the same time, the system adjusts search keyword weights based on the doctor's proficiency and provides search result explanations for doctors at the second level of proficiency, making information acquisition more efficient and targeted.

[0110] Example 2

[0111] The technical details of the method and device are now described in detail in conjunction with a specific embodiment - the construction of a mammographic image dataset, so as to fully demonstrate its innovative value and practical feasibility.

[0112] Step 1: Image and report acquisition. Obtain the mammography image report and corresponding DICOM-formatted image data from the RIS system. The image report is in unstructured, free-text format. Mammography image data in DICOM format from the RIS system. Unstructured mammography image report from the RIS system.

[0113] Step 2: Report Structuring. Structuring the imaging report is performed based on a private AI model deployed on a local server. The local private AI model is obtained through supervised fine-tuning of labeled data, and the private AI model is set with specific task instructions for structuring imaging reports. Step 2 is divided into the following steps:

[0114] S1: Based on the authoritative standard terminology system in the medical field (consensus on mammography examination and diagnosis), combined with the actual needs of clinical diagnosis and treatment, a template for extracting key information from mammography reports is constructed. This template accurately defines the medical entities to be extracted and their logical relationships, forming a standardized framework for extracting key information from unstructured reports ( Figure 2 、 Figure 3 );

[0115] S2: Based on the standardized templates in S1, we conducted a prompting project, designing specialized prompts that are compatible with the large language model and clarifying the model's task instructions for processing unstructured reports. Simultaneously, we manually annotated some unstructured reports based on the templates to generate an initial annotated dataset, providing a supervised learning foundation for model training. Because different doctors use different descriptive conventions and terminology when writing imaging reports, we created a comparison table of mammography report templates based on our hospital's situation and used this comparison table to standardize the prompts for our private AI model.

[0116] S3: Use the prompt words and initial annotation dataset of S2 to perform the first round of SFT on the large language model. The optimized model is used to process a large number of unstructured medical image reports to generate a large-scale structured report and form a rich annotation dataset, expanding the scale of supervised learning data;

[0117] S4: Adopt an iterative training strategy, adapt the specialized prompt words of S2 to the large language model, input the task instructions for the model to process unstructured reports and the large-scale annotation data generated by S3 into the model, and repeat multiple rounds of supervised fine-tuning. By continuously calibrating the consistency of the model output with medical professional standards, the model performance is gradually optimized, and finally the target private AI model that meets the structured requirements of medical image reports is generated;

[0118] S5: Deploy the target model after multiple rounds of fine-tuning to a local server to build a localized processing system. Use the final model deployed on the local server to structure the breast X-ray photography unstructured image reports obtained in step one, and obtain breast X-ray photography structured image reports.

[0119] Step three: Data association and SR generation: Use specific data integration scripts to associate and integrate the breast X-ray photography structured medical image reports output by the private AI model with the metadata in the DICOM format breast X-ray photography images (such as patient information, examination parameters, etc. Data can be used as association basis) to generate structured report files that comply with DICOM SR standard specifications. The files use DICOM standard data storage format (i.e. file extension.dcm). Step three is divided into the following small steps:

[0120] S6: Generate a series of key basic metadata for the SR file. Among them, the unique identifier is a key identification that ensures each SR file has a unique identity in the entire dataset. It follows specific coding rules to accurately locate and distinguish each report in subsequent data management and retrieval;

[0121] S7: Carefully read the breast X-ray photography structured image report content and its nested structure generated in S5 step. This requires in-depth analysis of every detail in the report, including various medical findings, related measurement data, and their logical relationships, etc. Then, these contents are accurately added to the corresponding positions in the SR file to ensure the integrity and accuracy of the information. In this process, in order to realize the standardization and interoperability of medical information, each medical term in the report content is uniformly represented by RadLex term coding. By using these standard term codes, different medical systems and software can more accurately understand and exchange medical information, avoiding misunderstandings and errors caused by inconsistent terminology;

[0122] S8: Extract key metadata information from the DICOM format mammographic image data obtained in step 1. These metadata include but are not limited to patient information (such as name, age, gender, medical record number, etc.), examination parameters (such as scanning method, scanning time, scanning parameter settings, etc.) and other data that can be used as a basis for association. These metadata are crucial for understanding the source, background and relationship of the image with the patient. Then, these extracted metadata information are accurately referenced to the SR file. By establishing this data association, the SR file not only contains detailed structured report content, but can also be closely linked to the corresponding DICOM image data to form a complete medical information unit. This association relationship provides a more comprehensive and accurate data foundation for subsequent medical image analysis, diagnostic decision-making and data mining;

[0123] S9: Finally, based on the work completed in the previous steps, a mammography structured report file is generated that fully complies with the DICOM SR standard. This process strictly adheres to the data storage format requirements for SR files in the DICOM standard, ensuring that every byte and every field of the file conforms to the standard's definitions and specifications. The final file extension is .dcm, which is the file format specified by the DICOM standard and can be recognized and processed by a wide range of medical devices, image processing software, and information systems. By generating such a standard-compliant structured report file, standardized storage and exchange of medical imaging data and related reports is achieved, providing solid data support for information sharing, quality control, and clinical research in the medical field.

[0124] Step 4: Dataset Construction and Association. The mammography structured report file, which complies with the DICOM SR standard and is stored in the .dcm format, is transferred back to the PACS system. By establishing data associations in the PACS system (based on unique identifiers such as patient ID and examination time), a medical imaging dataset is formed that links the medical images (in DICOM format) and the medical imaging structured report (which complies with the DICOM SR standard and is stored in the .dcm format). Step 4 is divided into the following sub-steps:

[0125] S10: Association database table structure design. In the original medical image system, a structured report table sr_reports, a lesion table sr_lesions, and a feature coding table sr_features are added. The three are associated through the primary foreign key to build a hierarchical association of "report -> lesion -> feature". The sr_reports table stores the SR file metadata generated in step three (such as SOPInstance UID, Study Instance UID), complete text content, and creation timestamp, ensuring the traceability of the report. Each report is uniquely identified by the primary key sr_id. The sr_lesions table is used to record the multiple lesion structure information extracted from the report. Each record is associated with the sr_reports table through the foreign key sr_id, realizing a one-to-many mapping relationship between the report and the lesion. The table can contain lesion number, lesion location, automatic abstract, etc. The sr_features table is used to store the standardized medical feature information extracted under each lesion. Each record in the table is associated with the sr_lesions table through the foreign key lesion_id, so that all feature information has a clear structural attribution relationship and can be traced back to the original SR file.

[0126] S11: Composite index creation. Create a composite index idx_sr_features in the sr_features table, using a B + tree index structure to speed up feature retrieval. The index fields are radlex_code and lesion_id.

[0127] S12: Build a three-level association system of "image -> report -> feature" to support cross-modal joint query. Associate the image data through the study_instance_uid of the sr_reports table, and associate the feature dimension through the radlex_code of the sr_features table.

[0128] Step five: Image retrieval based on disease features. Use the NER model based on the BERT architecture to realize the accurate conversion of natural language keywords to feature codes, and use the converted codes to perform association query to obtain the target image. Step five is divided into the following small steps:

[0129] S13: "Natural language - coding" mapping engine: The BERT-based named entity recognition model realizes the accurate mapping of natural language keywords to RadLex codes through three stages: first, use large-scale text such as medical literature and clinical reports to pre-train BERT model parameters to learn general medical semantic representation; then fine-tune the model on the annotated data set, optimize the parameters by minimizing the cross-entropy loss function to make the conditional probability distribution Approximate the true encoding distribution; finally, the input keyword sequence , the model is calculated by Output the RadLex code with the highest probability, achieving end-to-end conversion from unstructured text to standardized feature coding, providing structured input for image retrieval;

[0130] S14: The sr_features table records in detail the RadLex encoding information for the disease features corresponding to each examination instance. After natural language is converted into RadLex encoding using the BERT-based named entity recognition model, these encodings can be used to search the sr_features table to accurately locate the corresponding examination instance. The images table records the identifier of the examination instance to which each image file belongs. After the examination instance is identified in the first association step, the identification of the examination instance is used to search the images table to find the corresponding image file. This two-level association allows for efficient and accurate identification of corresponding medical image files based on disease features described in natural language, providing strong data support for clinical diagnosis and research.

Claims

1. A method for establishing a medical image dataset with a unified structure, characterized in that: include: Based on the unstructured imaging report and the corresponding DICOM format image data, the unstructured imaging report text, DICOM images and DICOM metadata are extracted; The unstructured image report text is processed through a private AI model to obtain structured data. The structured data is merged with the DICOM metadata to obtain an SR file. When generating the SR file, a proficiency identification field is added to mark the structured degree of the SR file. The SR file includes the DICOM metadata, medical feature code and unique identifier, specifically: Define structured templates based on standard medical terminology; Performing supervised fine-tuning on the private AI model using a manually annotated initial dataset, and iteratively optimizing the output of the private AI model so that the output matches the medical standard terminology; Parsing the unstructured imaging report text into the structured data conforming to the structured template through the private AI model; Determining, according to the DICOM SR standard, a data template, that the structured data matches the unique identifier of the DICOM metadata; Initialize the SR file based on the DICOM SR standard; Filling basic information in the DICOM metadata into the basic metadata field of the SR file, wherein the basic information includes at least one of patient information and examination parameters; Traversing the structured data to extract medical feature codes and related information; Formatting the medical feature codes according to the requirements of the DICOM SR standard; Embedding the formatted medical feature code and related information into the content sequence of the SR file; In addition to the basic metadata fields, the DICOM metadata is also added to the relevant fields of the SR file; Check the fields and use DICOM validation tools or libraries to check whether the generated SR file complies with the DICOM SR standard; Based on the unique identifier, the SR file is associated with the DICOM image to form a structured data set, specifically; confirming that the unique identifier of the SR file is consistent with the unique identifier of the DICOM image; Constructing a structured report table to store relevant information of the SR file; Constructing a feature coding table to store the medical feature codes extracted from the SR file; Uploading the SR file to a PACS system, using the unique identifier to find and associate the corresponding DICOM image, or establishing an association between the structured report form and DICOM image metadata using the unique identifier as a foreign key; Associating the DICOM image with the SR file through the unique identifier; Associating the SR file with the medical feature code in the feature code table through the unique identifier; Supporting cross-modal queries based on disease symptoms through the medical feature codes, and establishing a composite index on the feature codes to accelerate queries based on the medical feature codes; Associating the medical feature code in the SR file with the unique identifier, and mapping the user input to the medical feature code through a BERT-NER model to query the structured dataset; When obtaining the unstructured imaging report and the corresponding DICOM format image data, recording the learning identity of the doctor; Assess doctors' basic competencies through standardized tests, and convert test results into proficiency scores through machine learning models; Recording the doctor's operating behavior data and using an online learning algorithm to update the proficiency score in real time; Preset structured templates and rules for different proficiency scores; It also includes: dynamically generating prompt words based on proficiency to control the level of detail output by the private AI model; If the proficiency is the first level, fine-tuning the private AI model using the annotated data of a preset granularity; If the proficiency is the second level of proficiency, the private AI model is fine-tuned using summary-type annotated data.

2. The method according to claim 1, characterized in that The step of associating the medical feature code in the SR file with the unique identifier and mapping the user input to the medical feature code through the BERT-NER model to query the structured dataset is as follows: Use the BERT-NER model to parse the natural language description entered by the user and identify key medical entities; Converting the key medical entity into the corresponding medical feature code; Using the medical feature code, construct an SQL query statement to retrieve the related unique identifier; Based on the unique identifier, searching the structured report table for the corresponding SR file; Using the same unique identifier to search for the related DICOM image in a PACS system or a DICOM image database; The SR file and the DICOM image are integrated to form a structured data set.

3. The method according to claim 1, characterized in that The structured templates and rules are preset for different proficiency scores, specifically: If the proficiency is the first level proficiency, split the structured data into standardized fields and add annotations; If the proficiency level is the second level, the detail level of the structured data is reduced, a summary structured report is provided, and key conclusions are retained.

4. The method according to claim 1, wherein After associating the medical feature code in the SR file with the unique identifier and mapping the user input to the medical feature code through the BERT-NER model to query the structured dataset, the method further includes: recommending differentiated learning content based on the doctor's proficiency score, specifically: If the proficiency level is the first level, structured reports and basic cases will be pushed first, combined with interactive learning tools; If the stated proficiency level is Level 2, provide mixed content and introduce complex cases and diagnostic reasoning training; Recommend cases of similar difficulty based on the doctor's historical learning behavior.

5. The method according to claim 4, characterized in that After recommending differentiated learning content based on the doctor's proficiency score, it also includes: Taking behavioral indicators and ability indicators as quantitative indicators of learning effects, and dynamically adjusting the differentiated learning content based on the quantitative indicators of learning effects; Adjust the search keyword weight according to the doctor's proficiency, and provide search result explanations for doctors with second-level proficiency.

Images