AI-based medical image data labeling management method and system

Abstract

The application belongs to the technical field of medical informatics, and discloses an AI-based medical image data labeling management method and system; including: obtaining desensitization image data and desensitization report text, calling a pre-trained natural language processing model to extract medical entities in the desensitization report text and identify the logical relationship between the medical entities; associating and mapping the medical entities with the desensitization image data, and generating multi-dimensional classification labels for the desensitization image data; storing the desensitization image data and the desensitization report text in a structured database to establish an index, and based on the screening and combination of the multi-dimensional classification labels, performing screening and collection operations on the structured database to generate a structured medical data set of a specific topic; the application can quickly construct a high-quality structured medical data set of a specific topic, effectively solving the core bottleneck of low efficiency, inconsistent standards and inability to scale of the traditional manual labeling method.

Description

Technical Field

[0001] This invention relates to the field of medical informatics technology, and more specifically, to an AI-based method and system for the labeling and management of medical image data. Background Technology

[0002] With the rapid development of medical informatization, major hospitals and medical cloud platforms have accumulated massive amounts of medical imaging data and their corresponding diagnostic reports. However, most of this data, which is of great clinical and research value, exists in an unstructured form, like a huge "data black box," making it difficult to directly retrieve, statistically analyze, and process. Traditional data governance and the construction of disease-specific datasets heavily rely on radiologists to manually read, classify, and label images. This approach not only consumes extremely high human and time costs, exacerbating the strain on medical resources, but also, due to differences in the cognitive abilities and fatigue levels of different doctors, the results of manual annotation often suffer from problems such as non-standard formats, omissions of key information, and inconsistent terminology. As a result, the quality of the final datasets varies greatly, becoming a core bottleneck in releasing the value of medical data.

[0003] Furthermore, medical imaging data comprises two distinct modalities: visual images and text reports. Most existing medical data management systems are either purely Picture Archiving and Communication Systems (PACS) or general electronic medical record text retrieval tools, lacking the ability to deeply integrate and process the cross-modal characteristics of "image-text." Existing automated extraction tools often only perform simple keyword extraction from report text, failing to accurately extract complex medical entities and their logical relationships, let alone perform cross-modal semantic alignment between extracted text lesion information and specific sequence regions in the images. This results in unstructured data not being truly transformed into a structured knowledge base with rich dimensional labels, making it difficult to support subsequent advanced scientific disease screening and intelligent application development. Therefore, there is an urgent need for a medical imaging data labeling management system that can utilize artificial intelligence technology to extract key information from medical images and their text reports, perform cross-modal alignment, and automatically generate multi-dimensional labels.

[0004] In view of this, the present invention proposes an AI-based medical image data tagging management method and system to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of the existing technology and achieve the above objectives, the present invention provides the following technical solution: an AI-based method for labeling and managing medical image data, comprising: Obtain the hospital's original medical image data and corresponding unstructured text reports, perform de-identification processing on the original medical image data and unstructured text reports, generate de-identified image data and de-identified report text, and form a de-identified image report dataset; Based on the anonymized report text in the anonymized image report dataset, a pre-trained natural language processing model is invoked to extract medical entities and identify the logical relationships between each medical entity to form a structured semantic set. Based on the desensitized image data and structured semantic set in the desensitized image report dataset, medical entities are associated and mapped with the desensitized image data to form a cross-modal alignment result set; Based on the structured semantic set and the cross-modal alignment result set, multi-dimensional classification labels are automatically generated for desensitized image data, and the desensitized image report dataset is transformed into a labeled image dataset containing multi-dimensional classification labels. The labeled image dataset is stored in a pre-defined structured database and indexed. Based on the filtering and combination of multi-dimensional classification labels, the structured database is filtered and aggregated to generate a structured medical dataset for a specific topic.

[0006] Furthermore, methods for generating desensitized image report datasets include: The raw medical imaging data includes a unique examination code, a DICOM file header, and one or more image sequences. The DICOM file header contains patient information tags and technical parameter tags. The image sequence contains a sequence number, image modality type, examination site description, sequence description information, and multi-frame pixel data. The unstructured text report contains a unique examination code, examination description text, and diagnostic conclusion text. One set of medical imaging data and an unstructured text report corresponds to one imaging examination. A list of personal identification fields is pre-defined. For each set of original medical image data, all patient information tags in the corresponding DICOM file header are traversed. Patient information tags in the list of personal identification fields are replaced with predefined anonymous placeholders or deleted. The examination unique code and technical parameter tags are retained unchanged to obtain desensitized image data. A pre-defined set of text anonymization rules is used to scan the examination description text and diagnostic conclusion text for each group of unstructured text reports. The identified personal information fragments are replaced with general placeholders or deleted to obtain the anonymized report text. The unique examination code, desensitized image data, and desensitized report text corresponding to each image examination are integrated to form a desensitized image report dataset.

[0007] Furthermore, methods for extracting medical entities by calling pre-trained natural language processing models include: Multiple standard entity categories are preset, including disease name category, anatomical location category, imaging sign category, quantitative information category, and diagnostic conclusion category; for each de-identified report text, it is sequentially split into independent sentence units; a pre-built natural language processing model is constructed, and each sentence unit is sequentially input into the natural language processing model to output the corresponding candidate medical entity; each candidate medical entity includes the original text fragment, entity category, and initial confidence level; Negation range detection is performed on each candidate medical entity, and a corresponding negation label is set for each candidate medical entity based on the detection results; terminology standardization mapping is performed on each candidate medical entity to obtain the corresponding entity standard name and superordinate concept name; for candidate medical entities whose entity category is quantitative information category, quantitative attribute auxiliary information is extracted from the corresponding original text fragment; The entity standard name, entity category, original text fragment, negation mark, superordinate concept name and quantitative attribute information corresponding to the same candidate medical entity are integrated, and a unique entity identifier is assigned to each candidate medical entity to form a confirmed medical entity; all confirmed medical entities corresponding to the same de-identified report text are summarized to form an entity record list.

[0008] Furthermore, methods for forming structured semantic sets include: A pre-defined medical relationship type system is established, which includes multiple standard relationship types. The standard relationship types include location-lesion relationship, lesion-sign relationship, lesion-quantification relationship, sign-quantification relationship, and lesion-recommendation relationship. For the entity record list of the same anonymized report text, the relationship between any two medical entities is determined and the corresponding relationship record is generated. The relationship record includes a unique relation identifier, a head entity identifier, a tail entity identifier, a relation type, and a relation confidence level. All relation records corresponding to the same anonymized report text are summarized to form a relation record list. For each anonymized report text, perform semantic integrity checks on the entity record list to determine whether a semantic incompleteness flag is set. Integrate the unique check code, entity record list, relation record list, and semantic incompleteness flag corresponding to each anonymized report text to form a structured semantic set.

[0009] Furthermore, methods for associating and mapping medical entities with desensitized image data include: From the entity record list, medical entities with the entity category of anatomical location are selected as anatomical location entities, forming a location entity set; from the entity record list, medical entities with the entity category of disease name are selected as lesion entities, forming a lesion entity set; from the relation record list, relation records with the relation type of location-lesion relationship are selected, forming a location association relation set. Based on the set of site associations, the corresponding anatomical site entity is determined for each lesion entity in the lesion entity set; based on the desensitized image data corresponding to each examination unique code, sequence-level alignment is performed to determine the associated sequence number and alignment status of each anatomical site entity; For anatomical entities that have been successfully aligned at the sequence level, perform layer-level localization to determine the corresponding associated layer range and alignment granularity; based on the attribution relationship between anatomical entities and lesion entities, inherit the associated sequence number, associated layer range, and alignment granularity of the lesion entity from the corresponding attribution entity; and set the alignment confidence level based on the alignment granularity and alignment status.

[0010] Furthermore, methods for forming cross-modal alignment result sets include: The associated sequence number, associated layer range, alignment granularity, and alignment confidence of each lesion entity are integrated to form the alignment result of the corresponding lesion entity; the associated sequence number, associated layer range, alignment granularity, and alignment confidence of each anatomical location entity are integrated to form the alignment result of the corresponding anatomical location entity; all alignment results under the same examination unique code are summarized to form the cross-modal alignment result; and all cross-modal alignment results are summarized to form the cross-modal alignment result set.

[0011] Furthermore, methods for automatically generating multi-dimensional classification labels for de-identified image data include: A multi-dimensional label classification system is pre-set, which includes disease classification dimension, anatomical location dimension, imaging feature dimension, severity dimension, and treatment recommendation dimension; Based on the entity record list, generate a set of disease classification labels for each imaging examination; based on the entity record list and the cross-modal alignment result set, generate a set of anatomical location labels for each imaging examination; based on the entity record list and the relation record list, generate a set of imaging sign labels for each imaging examination; based on the entity record list and the relation record list, infer the severity labels for each imaging examination; based on the entity record list, generate a set of treatment recommendation labels for each imaging examination. For all labels generated from each imaging examination, a label evidence chain is constructed; based on the initial confidence, semantic incompleteness markers, and alignment confidence of the medical entity on which each label is based, a comprehensive label quality score for each imaging examination is generated.

[0012] Furthermore, methods for converting desensitized image report datasets into labeled image datasets include: The unique examination code, disease classification label set, anatomical location label set, imaging sign label set, severity label, treatment suggestion label set, label evidence chain, and comprehensive label quality score corresponding to each imaging examination are integrated to form a multi-dimensional classification label set for the corresponding imaging examination. The multi-dimensional classification label set of each imaging examination is then linked and integrated with the corresponding desensitized image data and desensitized report text in the desensitized image report dataset to form a labeled image dataset.

[0013] Furthermore, methods for generating structured medical datasets on specific topics include: The system receives a dataset construction request, which includes a combination of filtering conditions and a dataset name. Based on the combination of filtering conditions, it performs a search query on the structured database to obtain a set of examination records that meet all filtering conditions. It then integrates the desensitized image data, desensitized report text, and multi-dimensional classification label set corresponding to each set of examination records in the examination record set to form a structured medical dataset labeled with the dataset name.

[0014] The AI-based medical image data labeling management system, implementing the aforementioned AI-based medical image data labeling management method, includes: The data anonymization module is used to acquire the hospital's original medical image data and corresponding unstructured text reports, perform de-identification processing on the original medical image data and unstructured text reports, generate anonymized image data and anonymized report text, and form an anonymized image report dataset. The semantic extraction module is used to extract medical entities from the desensitized report text in the desensitized image report dataset by calling a pre-trained natural language processing model, and to identify the logical relationships between the medical entities to form a structured semantic set. The image alignment module is used to associate and map medical entities with desensitized image data based on the desensitized image data and structured semantic set in the desensitized image report dataset, forming a cross-modal alignment result set; The label generation module is used to automatically generate multi-dimensional classification labels for desensitized image data based on the structured semantic set and the cross-modal alignment result set, and to transform the desensitized image report dataset into a labeled image dataset containing multi-dimensional classification labels. The data management module is used to store labeled image datasets into a preset structured database to create an index, and to perform filtering and aggregation operations on the structured database based on the filtering and combination of multi-dimensional classification labels to generate structured medical datasets on specific topics.

[0015] The technical effects and advantages of the AI-based medical image data tagging management method and system of this invention are as follows: By systematically removing patient information tags from the DICOM file header and personal information from the text report, patient privacy is protected at the source of data flow. A natural language processing model fine-tuned for medical text is invoked, combined with a negation range detection mechanism to accurately distinguish semantic differences between affirmative and exclusionary descriptions. Terminology standardization mapping unifies the differentiated terminology used by different doctors into standard medical terminology. Simultaneously, a cross-sentence relational reasoning mechanism captures the logical relationships between entities distributed across different sentences, overcoming the limitation of existing keyword extraction tools that can only perform simple single-sentence matching. Furthermore, by establishing a multi-granularity hierarchical alignment mechanism from the sequence level to the layer level, anatomical descriptions in the text are systematically mapped to the specific sequence and layer range of the DICOM image, achieving alignment between image and text. Cross-modal semantic linking; automatically generating structured labels covering five dimensions—disease classification, anatomical location, imaging features, severity, and treatment recommendations—based on a hierarchical multi-dimensional label classification system; introducing an exclusionary labeling mechanism to distinguish confirmed and excluded diseases; constructing a label evidence chain to achieve full-link traceability from each label to the original report text; and quantifying the label reliability of each set of examination data through a comprehensive label quality score; supporting flexible screening and combination based on multi-dimensional labels and label logic consistency verification, enabling researchers to quickly construct high-quality structured medical datasets on specific topics, fundamentally solving the core bottlenecks of traditional manual annotation methods—low efficiency, inconsistent standards, and inability to scale—and providing systematic technical support for the scientific research application and intelligent development of medical imaging big data. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the AI-based medical image data labeling management system according to Embodiment 1 of the present invention; Figure 2 This is a flowchart of the AI-based medical image data tagging management method according to Embodiment 2 of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1:

[0018] Please see Figure 1As shown in the figure, the AI-based medical image data labeling management system described in this embodiment includes a data desensitization module, a semantic extraction module, an image alignment module, a label generation module, and a data management module; each module is connected via wired and / or wireless means to realize data transmission between modules.

[0019] The data anonymization module is used to acquire the hospital's original medical image data and corresponding unstructured text reports, perform de-identification processing on the original medical image data and unstructured text reports, generate anonymized image data and anonymized report text, and form an anonymized image report dataset.

[0020] Methods for obtaining raw medical imaging data and corresponding unstructured text reports from hospitals include: Raw medical image data is obtained from a Picture Archiving and Communication System (PACS, a medical information platform used to store, transmit, and manage medical image examination data). This raw medical image data is a collection of medical image files conforming to the DICOM standard format, with each set of raw medical image data corresponding to one imaging examination. Each set of raw medical image data contains a DICOM file header and one or more image sequences. The DICOM file header is a structured header area in the DICOM standard format used to store metadata information for the image files, including patient information tags and technical parameter tags. The patient information tags record information related to the patient's personal identity, such as the patient's name, ID number, date of birth, address, and contact number. The technical parameter category tags are used to record technical information related to image acquisition and content. Each set of raw medical image data has a unique examination code, which is a unique identification code for the image examination predefined in the technical parameter category tags. Each image sequence contains a sequence number, image modality type, examination site description, sequence description information, and multi-frame pixel data. The sequence number is the sequential identifier of the image sequence in the same examination. Image modality types include CT, MRI, X-ray, ultrasound, etc. The examination site description is the name of the human body part being examined recorded in the technical parameter category tags. The sequence description information is the scanning protocol name and parameter description recorded in the technical parameter category tags. The multi-frame pixel data is the set of image pixel values ​​for all scan layers in the corresponding image sequence. The unstructured text reports corresponding to each imaging examination are obtained from the Radiology Information System (RIS, an information management platform used to manage radiology examination processes and generate diagnostic reports). These unstructured text reports are natural language texts written by radiologists for each imaging examination, containing a unique examination code, an examination description, and a diagnostic conclusion. The examination description is a detailed account of the imaging findings by the radiologist; the diagnostic conclusion is the radiologist's diagnostic opinion and treatment recommendations based on the imaging findings. Each set of original medical imaging data is then matched with its corresponding unstructured text report based on the examination's unique code.

[0021] Methods for generating desensitized image report datasets include: A pre-defined list of personal identification fields is provided, which contains the names of all patient information tags in the DICOM file header that involve personally identifiable information. The list of personal identification fields is pre-set by a person skilled in the art based on the privacy tag definitions in data security regulations and the DICOM standard. For each set of original medical image data, all patient information tags in the corresponding DICOM file header are traversed, and the patient information tags existing in the list of personal identification fields are replaced with predefined anonymous placeholder values ​​(such as anonymous name, fake ID, null value, etc.) or deleted, while retaining the unique code and technical parameter tags unchanged, thus obtaining desensitized image data. A pre-defined text anonymization rule set is provided, which includes regular expressions and keyword patterns for identifying personal information fragments in unstructured text reports. The text anonymization rule set is pre-set by a person skilled in the art based on the writing conventions of Chinese medical reports. For each set of unstructured text reports, the text anonymization rule set is used to scan the examination description text and the diagnostic conclusion text, and the identified personal information fragments are replaced with general placeholders or deleted to obtain the anonymized report text. The unique examination code, desensitized image data, and desensitized report text corresponding to each image examination are integrated to form a desensitized image report dataset.

[0022] The semantic extraction module is used to extract medical entities from the desensitized report text in the desensitized image report dataset by calling a pre-trained natural language processing model, and to identify the logical relationships between the medical entities to form a structured semantic set.

[0023] Methods for extracting medical entities by calling pre-trained natural language processing models include: From the desensitized image report dataset, the desensitized report text corresponding to the unique code of each examination is obtained. For each desensitized report text, sentence segmentation is performed on the examination description text and the diagnostic conclusion text, splitting continuous text paragraphs into independent sentence units according to punctuation marks and semantic boundaries. Sentence segmentation is a conventional text preprocessing technique in this field and will not be elaborated on here. All the segmented sentence units are arranged according to their order of appearance in the corresponding examination description text or diagnostic conclusion text to form a sentence sequence, and the source paragraph type of each sentence unit is labeled. The source paragraph type is either examination description or diagnostic conclusion. A pre-built natural language processing (NLP) model is used. This NLP model is a medical text understanding model obtained by fine-tuning a pre-trained language model through domain adaptation on a large-scale medical text corpus. It has the ability to perform deep semantic understanding and structured information extraction of professional terms, anatomical structure descriptions, and clinical diagnostic expressions in medical diagnostic reports. Examples of pre-trained language models include BERT series models, RoBERTa series models, and ELECTRA series models. The fine-tuning training method for the NLP model is a conventional technique in this field and will not be elaborated on here. A pre-defined medical entity category system is established, which includes multiple standard entity categories. These standard entity categories include disease name categories, anatomical location categories, imaging feature categories, quantitative information categories, and diagnostic conclusion categories. Specifically, the disease name category identifies the name of a disease or lesion, such as pulmonary nodule, pneumonia, or liver cyst; the anatomical location category identifies the name of a human anatomical structure, such as the right upper lobe of the lung or the right lobe of the liver; the imaging feature category identifies imaging characteristics, such as ground-glass opacity, calcification, and enhancement; the quantitative information category identifies numerical descriptions related to the lesion, such as lesion size and density values; and the diagnostic conclusion category identifies diagnostic judgments and recommended treatment statements, such as recommendations for follow-up or consideration of malignancy. Each standard entity category is pre-set by a person skilled in the art based on the information organization characteristics of radiology reports. Each sentence unit is sequentially input into the natural language processing model. The natural language processing model performs sequence labeling on the text fragments in each sentence unit and outputs the candidate medical entities identified in each sentence unit. Each candidate medical entity includes the original text fragment, entity category, and initial confidence score. The original text fragment is the original text content of the candidate medical entity in the sentence unit. The initial confidence score is the degree of certainty that the natural language processing model is about the recognition result of the corresponding candidate entity.

[0024] Each candidate medical entity undergoes a negative range test, and a corresponding negative label is set for each candidate medical entity based on the test results. Specifically, in medical reports, doctors often use negative expressions to describe exclusionary observations. For example, in "no obvious abnormal enhancement seen," although "abnormal enhancement" is identified as a candidate medical entity, its actual semantic meaning is that the sign does not exist. A pre-defined set of negative trigger words is included, containing common negative expressions in medical reports, including but not limited to "not seen," "not reached," "none," "no obvious," "not accompanied," "not found," and "excluded." The set of negative trigger words is determined by a person skilled in the art based on Chinese radiology reports. Language habits are pre-set; for each candidate medical entity, it is checked whether any of the negation trigger words in the set of negation trigger words exist in the sentence unit where it is located; if a negation trigger word exists, the positional relationship between the original text fragment of the corresponding candidate medical entity and the negation trigger word in the corresponding sentence unit is further determined; if the original text fragment is located after the negation trigger word and there is no transition conjunction or new subject component between the two, the corresponding candidate medical entity is determined to be within the negation range, and the negation mark of the corresponding candidate medical entity is set to negated; if the above conditions are not met, the negation mark of the corresponding candidate medical entity is set to unnegated. It should be noted that the negation range detection mechanism, by combining the relative positional relationship between the negation trigger word and the candidate entity in the sentence, as well as the blocking effect of the adversative conjunction, can accurately distinguish between two completely different semantic situations: "no obvious space-occupying lesions were seen in the liver" and "space-occupying lesions were seen in the liver, but no obvious calcification was seen." This is different from the simple keyword matching method used in existing medical text processing tools.

[0025] A terminology standardization mapping is performed on each candidate medical entity. Specifically, in actual clinical reports, different doctors may use different expressions for the same medical concept; for example, GGO, ground-glass opacity, and ground-glass opacity density all refer to the same imaging sign. A pre-defined medical terminology standardization dictionary is used, which contains multiple standard term entries. Each standard term entry includes a standard term name, entity category, a list of synonyms for the standard term, and a superordinate concept name. The synonym list includes all possible variations of the expression that may appear in the clinical report. The superordinate concept name is used to identify the higher-level classification concept to which the corresponding standard term belongs in the medical concept hierarchy. The medical terminology standardization dictionary is pre-set by those skilled in the art according to the International Classification of Diseases (ICD) coding standard and the medical imaging terminology specification. For each candidate medical entity, the corresponding original text fragment is matched with the standard term name and synonym list of each standard term entry in the medical terminology standardization dictionary. If the match is successful, the corresponding standard term name is used as the entity standard name of the corresponding candidate medical entity, and the corresponding superordinate concept name is obtained. If the match is unsuccessful, the corresponding original text fragment is retained as the entity standard name, and the superordinate concept name is set to null. For candidate medical entities whose entity category is quantitative information, quantitative attribute information is extracted from the corresponding original text fragments. This quantitative attribute information includes numerical values, units, and quantitative dimensions. Quantitative dimensions include size, density, and quantity. A pre-defined set of quantitative attribute extraction rules is used, pre-set by those skilled in the art based on the writing standards for quantitative descriptions in radiology reports. The entity standard name, entity category, original text fragment, negation marker, superordinate concept name, and quantitative attribute information corresponding to the same candidate medical entity are integrated, and a unique entity identifier is assigned to each candidate medical entity to form a confirmed medical entity. All confirmed medical entities corresponding to the same de-identified report text are summarized to form an entity record list.

[0026] Methods for identifying the logical relationships between medical entities include: A pre-defined medical relationship type system is established, which includes multiple standard relationship types. These standard relationship types include location-lesion relationship, lesion-sign relationship, lesion-quantification relationship, sign-quantification relationship, and lesion-recommendation relationship. Among them, the location-lesion relationship indicates that a lesion is located in a certain anatomical location; the lesion-sign relationship indicates that a lesion presents a certain imaging sign; the lesion-quantification relationship and the sign-quantification relationship respectively represent the measurement and description relationship between a lesion or sign and quantitative information; and the lesion-recommendation relationship represents the correspondence between a disease and a treatment recommendation. Each standard relationship type is pre-set by a person skilled in the art based on the information logical structure of a radiology report. For the entity record list of the same anonymized report text, the relationship between any two medical entities is determined. Specifically, for entity pairs appearing in the same sentence unit, the text fragment of the corresponding sentence unit is input into the relationship classification function of the natural language processing model. The natural language processing model outputs the relationship type and relationship confidence of the corresponding entity pair. If the relationship confidence is greater than a preset relationship confidence threshold, a relationship record is generated. The relationship confidence threshold is preset by those skilled in the art according to the relationship extraction accuracy requirements. Each relationship record contains a unique relationship identifier, a head entity identifier, a tail entity identifier, a relationship type, and a relationship confidence. For entity pairs distributed across different sentence units, cross-sentence relation reasoning is performed. Specifically, in medical reports, descriptions of the anatomical location and symptoms of the same lesion are often distributed across different sentences. For cross-sentence entity pairs (i.e., entity pairs distributed across different sentence units), it is determined whether the sentence units containing the corresponding two medical entities are adjacent in the sentence sequence. If adjacent, it is further determined whether the entity category combination of the two medical entities meets the head and tail entity category requirements of any standard relation type in the medical relation type system. If it meets the requirements, a relation record is generated, and the relation confidence is set to the preset default confidence of cross-sentence relations. The default confidence of cross-sentence relations is preset by a person skilled in the art based on the experience value of the accuracy of cross-sentence reasoning. If they are not adjacent or do not meet the requirements, it is determined that there is no logical relationship, i.e., no relation record is generated. The text corresponding to the same de-identified report is then... All relational records are summarized to form a relational record list. For example, the medical relational type system predefines a standard relational type for location-lesion relationships, which requires the head entity category to be the anatomical location category and the tail entity category to be the disease name category. Assuming that in a cross-sentence entity pair, the first medical entity has an anatomical location category (e.g., right upper lobe of the lung) and the second medical entity has a disease name category (e.g., ground-glass nodule), the corresponding entity category combination (anatomical location category → disease name category) meets the head-tail entity category requirements of the location-lesion relationship. Conversely, if the first medical entity is an imaging sign category (e.g., lobulation sign) and the second medical entity is an anatomical location category (e.g., right upper lobe of the lung), the corresponding entity category combination (imaging sign category → anatomical location category) does not meet the head-tail entity category requirements of any standard relational type in the medical relational type system.

[0027] Methods for forming structured semantic sets include: Semantic integrity verification is performed on the entity record list for each anonymized report text. Specifically, a set of essential entity categories is preset, which includes multiple essential entity categories and is pre-set by those skilled in the art according to the information integrity specifications of radiology reports. The entity record list is checked to see if it contains at least one medical entity corresponding to each essential entity category in the essential entity category set. If an essential entity category is missing, a semantic incompleteness mark is set for the corresponding entity record list. The unique code, entity record list, relation record list, and semantic incompleteness mark corresponding to each anonymized report text are integrated to form a structured semantic set.

[0028] The image alignment module is used to associate and map medical entities with desensitized image data based on the desensitized image data and structured semantic set in the desensitized image report dataset, forming a cross-modal alignment result set.

[0029] Methods for associating and mapping medical entities with desensitized image data include: Obtain the entity record list and relation record list corresponding to the unique code of each examination from the structured semantic set; filter out medical entities of the anatomical location category from the entity record list and mark them as anatomical location entities to form a location entity set; filter out medical entities of the disease name category from the entity record list and mark them as lesion entities to form a lesion entity set; filter out relation records of the location-lesion relationship type from the relation record list to form a location association relationship set. Based on the set of location associations, the corresponding anatomical location entity is determined for each lesion entity in the lesion entity set. Specifically, for each lesion entity, a relationship record is searched in the set of location associations, using the unique entity identifier of the corresponding lesion entity as the tail entity identifier. If a relationship record is found, the head entity identifier in the corresponding relationship record is obtained, and the corresponding anatomical location entity is obtained from the set of location entities as the belonging location entity of the corresponding lesion entity. If no relationship record is found, the belonging location entity is marked as undefined.

[0030] Based on the desensitized image data corresponding to the unique codes of each examination in the desensitized image report dataset, sequence-level alignment is performed. Specifically, the purpose of sequence-level alignment is to match and associate each anatomical entity in the entity set with the corresponding image sequence in the desensitized image data. A pre-set anatomical location-sequence description mapping table is used, which contains multiple mapping rules. These rules are pre-set by those skilled in the art based on the scanning protocol naming conventions for each image modality and knowledge of human anatomy. Each mapping rule contains a set of standard anatomical location keywords and sequence matching keywords. For each anatomical location entity, obtain the corresponding standard entity name and match it with the standard anatomical location keywords of each mapping rule in the anatomical location-sequence description mapping table. If a match is successful, obtain the corresponding sequence matching keyword set. Traverse all image sequences in the corresponding desensitized image data, and perform inclusion matching between the examination location description and sequence description information of each image sequence and the sequence matching keyword set. If the examination location description or sequence description information of the image sequence contains any keyword in the sequence matching keyword set, it is determined that there is a sequence-level alignment relationship between the corresponding image sequence and the corresponding anatomical location entity, and the corresponding sequence number is recorded as the associated sequence number. If no image sequence is matched, the alignment status of the corresponding anatomical location entity is marked as sequence alignment failure.

[0031] For anatomical entities that have been successfully aligned at the sequence level, layer-level localization is performed. Specifically, the purpose of layer-level localization is to further determine the scanning layer range corresponding to the anatomical entity in the matched image sequence. A pre-defined anatomical entity-layer range mapping table is used, which contains multiple layer-level localization rules. These rules are pre-set by those skilled in the art based on knowledge of human sectional anatomy and the standard coverage range of each modal scanning protocol. Each layer-level localization rule includes standard anatomical entity keywords, image modality type, and standardized layer position ratio range. The standardized layer position ratio range starts from the first frame and ends at the last frame in the image sequence, using the start ratio and end ratio to represent the layer range typically covered by the corresponding anatomical entity. For each anatomical entity that has completed sequence-level alignment, obtain the corresponding entity standard name and image modality type, and search for the corresponding layer positioning rule in the anatomical entity-layer range mapping table. If a layer positioning rule is found, obtain the corresponding start ratio and end ratio, and obtain the total number of frames from the multi-frame pixel data of the corresponding image sequence. Calculate the product of the start ratio and the total number of frames to obtain the start frame number; calculate the product of the end ratio and the total number of frames to obtain the end frame number; use the range from the start frame number to the end frame number as the associated layer range of the corresponding anatomical entity, and mark the alignment granularity as fine-grained; if no layer positioning rule is found, use the entire frame range of the image sequence as the associated layer range of the corresponding anatomical entity, and mark the alignment granularity as coarse-grained.

[0032] Based on the attribution relationship between anatomical site entities and lesion entities, the associated sequence number, associated layer range, and alignment granularity of the lesion entity are inherited from the corresponding attribution site entity. Specifically, for each lesion entity, if the corresponding attribution site entity has completed sequence-level alignment and layer-level positioning, the associated sequence number, associated layer range, and alignment granularity of the corresponding attribution site entity are directly assigned to the corresponding lesion entity; if the alignment status of the attribution site entity is sequence alignment failure or the attribution site entity is marked as unclear, the alignment status of the corresponding lesion entity is marked as alignment unreachable. The alignment confidence level is set according to the alignment granularity and alignment status. Specifically, if the alignment granularity is fine-grained, the alignment confidence level is set to a preset high confidence value; if the alignment granularity is coarse-grained, the alignment confidence level is set to a preset medium confidence value; if the alignment status is unreachable, the alignment confidence level is set to a preset low confidence value. The high confidence value, medium confidence value, and low confidence value are all preset by those skilled in the art based on their experience with the reliability of the alignment results.

[0033] Methods for generating cross-modal alignment result sets include: The associated sequence number, associated layer range, alignment granularity, and alignment confidence of each lesion entity are integrated to form the alignment result of the corresponding lesion entity; the associated sequence number, associated layer range, alignment granularity, and alignment confidence of each anatomical location entity are integrated to form the alignment result of the corresponding anatomical location entity; all alignment results under the same examination unique code are summarized to form the cross-modal alignment result; and all cross-modal alignment results are summarized to form the cross-modal alignment result set.

[0034] The label generation module is used to automatically generate multi-dimensional classification labels for desensitized image data based on the structured semantic set and the cross-modal alignment result set, and to transform the desensitized image report dataset into a labeled image dataset containing multi-dimensional classification labels.

[0035] Methods for automatically generating multi-dimensional classification labels for desensitized image data include: A multi-dimensional tag classification system is pre-defined, containing multiple tag dimensions. These dimensions include disease classification, anatomical location, imaging findings, severity, and treatment recommendations. The disease classification dimension employs a hierarchical structure with three levels: the first level classifies human body systems, including but not limited to the respiratory, digestive, and nervous systems; the second level lists the major disease categories within each system; and the third level lists the specific disease names. Similarly, the anatomical location dimension also uses a hierarchical structure: the first level classifies large areas of the human body, the second level classifies specific organs, and the third level classifies... Subregions within organs; each dimension—imaging features, severity, and treatment recommendations—has a corresponding set of label values. The label value set for the imaging features dimension is pre-set by those skilled in the art according to the radiological imaging feature classification standards. The label value set for the severity dimension includes normal, mildly abnormal, moderately abnormal, severely abnormal, and critical. The label value set for the treatment recommendations dimension includes no treatment required, regular follow-up, further examination, treatment recommendation, and emergency treatment. The multi-dimensional label classification system is pre-set by those skilled in the art according to the International Classification of Diseases (ICD) coding standards and radiological reporting guidelines. Based on the entity record list in the structured semantic set, disease classification labels for each imaging examination are generated. Specifically, all medical entities whose entity category is the disease name category and whose negation mark is not negated are obtained from the entity record list. For each obtained medical entity, the corresponding entity standard name and superordinate concept name are obtained, and hierarchical matching is performed in the disease classification dimension of the multi-dimensional label classification system. If the entity standard name matches the third level successfully, the successfully matched third-level label, the corresponding second-level label, and the first-level label are used as a set of disease classification labels. If the entity standard name does not match the third level but the superordinate concept name matches the second level successfully, the second-level label and the first-level label are used as a set of disease classification labels. The corresponding first-level labels are used as a set of disease classification labels, and the hierarchical precision is marked as coarse-grained. If no match is found, the entity standard name is used as a custom disease label, and the hierarchical precision is marked as unclassified. For medical entities corresponding to disease name categories that are negatively marked as negative in the entity record list, the corresponding entity standard name is written as an exclusion label into the disease classification label set and marked as an exclusion label type. The exclusion label is used to clearly identify diseases that have been reported and excluded in the corresponding imaging examination, so as to distinguish them from unmentioned diseases and confirmed diseases. All disease classification labels corresponding to the same imaging examination are summarized to form a disease classification label set. Based on the entity record list in the structured semantic set and the cross-modal alignment result set, anatomical site labels for each imaging examination are generated. Specifically, all anatomical site entities are obtained from the entity record list. For each anatomical site entity, its corresponding standard name is obtained, and hierarchical matching is performed in the anatomical site dimension of the multi-dimensional label classification system to generate anatomical site labels. The alignment results of the corresponding anatomical site entities are obtained from the cross-modal alignment result set. If the alignment confidence in the alignment results is greater than or equal to a preset effective alignment threshold, image localization information is added to the anatomical site label. The image localization information includes the associated sequence number and the associated layer range. The effective alignment threshold is preset by those skilled in the art based on the reliability requirements of the alignment results. If the alignment confidence is less than the effective alignment threshold, no image localization information is added. All anatomical site labels corresponding to the same imaging examination are summarized to form an anatomical site label set. Based on the entity record list and relation record list in the structured semantic set, image sign labels for each imaging examination are generated. All medical entities whose entity category is image sign and whose negation flag is not negated are retrieved from the entity record list and marked as image sign entities. For each image sign entity, the corresponding entity standard name is obtained and matched with the label value set of the image sign dimension to generate an image sign label. Simultaneously, relation records with a relation type of sign-quantification relationship and associated with the corresponding image sign entity are searched from the relation record list. If a relation record is found, the quantification attribute information of the entity at the other end of the relation record is retrieved and appended to the corresponding image sign label. All image sign labels corresponding to the same imaging examination are summarized to form an image sign label set. Based on the entity record list and relation record list in the structured semantic set, the severity label of each imaging examination is inferred. Specifically, a set of severity inference rules is preset, which contains multiple inference rules, pre-set by those skilled in the art according to clinical diagnostic grading standards. Each inference rule includes a trigger condition and a severity label value (i.e., the label value in the set of label values ​​corresponding to the severity dimension). The trigger condition is defined based on a combination of entity category, entity standard name, and quantitative attribute information. For each imaging examination, the inference rules are matched sequentially according to their order in the severity inference rule set, and the severity label value corresponding to the first matched inference rule is taken as the severity label of the corresponding imaging examination. If all inference rules cannot be matched, the severity label is set to mild abnormality. Based on the entity record list in the structured semantic set, treatment suggestion tags for each imaging examination are generated. Specifically, all medical entities whose entity category is the diagnostic conclusion category and whose negation mark is not negated are obtained from the entity record list and marked as diagnostic conclusion entities. For each diagnostic conclusion entity, the corresponding entity standard name is obtained and matched with the tag value set of the treatment suggestion dimension to generate treatment suggestion tags. All treatment suggestion tags corresponding to the same imaging examination are summarized to form a treatment suggestion tag set. For all tags generated from each image examination, a tag evidence chain is constructed. This chain records the basis for each tag's generation, enabling end-to-end traceability from the tag to the original report text. Specifically, for each generated tag, a corresponding entity identifier list and relationship identifier list are recorded and labeled as the source entity identifier list and source relationship identifier list, respectively. The tag is then associated with its corresponding source entity identifier list and source relationship identifier list to form the corresponding tag evidence chain record. Finally, the evidence chain records for all tags corresponding to the same image examination are summarized to form the tag evidence chain. Based on the initial confidence, semantically incomplete markers, and alignment confidence of the medical entities upon which each label is based, a comprehensive label quality score is generated for each imaging examination. Specifically, the average entity confidence is obtained by averaging the initial confidence of the medical entities upon which all labels for the corresponding imaging examinations are based; the average alignment confidence is obtained by averaging the alignment confidence of all alignment results for the corresponding imaging examination from the cross-modal alignment result set; and corresponding quality weights are assigned to the average entity confidence and the average alignment confidence, respectively. These quality weights are determined by those skilled in the art based on the contribution ratio of text extraction and image alignment to label quality. The process involves: pre-setting parameters; calculating a weighted sum of the average entity confidence and the average alignment confidence based on quality weights to obtain a basic quality score; obtaining semantic incompleteness markers for the corresponding image examination from the structured semantic set; if semantic incompleteness markers exist, calculating the product of the basic quality score and a preset incompleteness penalty factor to obtain the comprehensive label quality score for the corresponding image examination; wherein, the incompleteness penalty factor is preset by those skilled in the art based on the degree of influence of information completeness on label accuracy; if no semantic incompleteness markers exist, the basic quality score is directly used as the comprehensive label quality score for the corresponding image examination.

[0036] Methods for converting de-identified image report datasets into labeled image datasets include: The unique examination code, disease classification label set, anatomical location label set, imaging sign label set, severity label, treatment suggestion label set, label evidence chain, and comprehensive label quality score corresponding to each imaging examination are integrated to form a multi-dimensional classification label set for the corresponding imaging examination. The multi-dimensional classification label set of each imaging examination is then linked and integrated with the corresponding desensitized image data and desensitized report text in the desensitized image report dataset to form a labeled image dataset.

[0037] The data management module is used to store labeled image datasets into a preset structured database to create an index, and to perform filtering and aggregation operations on the structured database based on the filtering and combination of multi-dimensional classification labels to generate structured medical datasets on specific topics.

[0038] Methods for storing tagged image datasets into a pre-defined structured database and creating an index include: A pre-defined structured database is used as a data management platform to store labeled image data and support multi-dimensional retrieval. The labeled image dataset is centralized, and the data corresponding to each image examination is written into the structured database. Specifically, the unique examination code of each image examination is used as the primary key; the desensitized image data is stored as a file reference with the storage path recorded; the desensitized report text is stored as a text field; and the data corresponding to each label dimension in the multi-dimensional classification label set is stored in the corresponding structured fields. Database indexes are established for disease classification, anatomical location, imaging features, severity, and treatment recommendations. A numerical index is also established for the comprehensive label quality score. The establishment of database indexes and numerical indexes are standard database management techniques in this field and will not be elaborated upon further. Methods for generating structured medical datasets on specific topics include: The system receives a dataset construction request, which includes a combination of filtering criteria and a dataset name. Each filtering criterion is composed of one or more criteria combined using logical operators. Each criterion specifies a label dimension and a corresponding matching value or range. Logical operators include AND, OR, and NAND. Based on the filtering criterion combination, the system performs a search query on the structured database to obtain a set of examination records that satisfy all filtering criteria. The system then integrates the anonymized image data, anonymized report text, and multi-dimensional classification label set corresponding to each set of examination records in the examination record set to form a structured medical dataset labeled with the dataset name.

[0039] This embodiment systematically removes patient information tags from the DICOM file header and personal information from the text report, ensuring patient privacy at the data flow source. By invoking a medical text-tuned natural language processing model and combining it with a negation range detection mechanism, it accurately distinguishes the semantic differences between affirmative and exclusionary descriptions. Furthermore, through terminology standardization mapping, it unifies the differentiated terminology used by different doctors into standard medical terminology. Simultaneously, it captures the logical relationships between entities distributed across different sentences through a cross-sentence relational reasoning mechanism, overcoming the limitation of existing keyword extraction tools that can only perform simple single-sentence matching. Finally, by establishing a multi-granularity hierarchical alignment mechanism from the sequence level to the layer level, it systematically maps anatomical descriptions in the text to the specific sequence and layer range of the DICOM image, achieving image-text alignment. It enables cross-modal semantic links; automatically generates structured labels covering five dimensions—disease classification, anatomical location, imaging features, severity, and treatment recommendations—based on a hierarchical multi-dimensional label classification system; introduces an exclusionary labeling mechanism to distinguish between confirmed and excluded diseases; constructs a label evidence chain to achieve full-link traceability from each label to the original report text; and quantifies the label reliability of each set of examination data through a comprehensive label quality score; by supporting flexible screening and combination of multi-dimensional labels and label logic consistency verification, it enables researchers to quickly construct high-quality structured medical datasets on specific topics, fundamentally solving the core bottlenecks of traditional manual annotation methods, such as low efficiency, inconsistent standards, and inability to scale, and providing systematic technical support for the scientific research application and intelligent development of medical imaging big data. Example 2:

[0040] Please see Figure 2 As shown, the parts not described in detail in this embodiment are described in Embodiment 1. An AI-based method for labeling and managing medical image data is provided, the method including: Obtain the hospital's original medical image data and corresponding unstructured text reports, perform de-identification processing on the original medical image data and unstructured text reports, generate de-identified image data and de-identified report text, and form a de-identified image report dataset; Based on the anonymized report text in the anonymized image report dataset, a pre-trained natural language processing model is invoked to extract medical entities and identify the logical relationships between each medical entity to form a structured semantic set. Based on the desensitized image data and structured semantic set in the desensitized image report dataset, medical entities are associated and mapped with the desensitized image data to form a cross-modal alignment result set; Based on the structured semantic set and the cross-modal alignment result set, multi-dimensional classification labels are automatically generated for desensitized image data, and the desensitized image report dataset is transformed into a labeled image dataset containing multi-dimensional classification labels. The labeled image dataset is stored in a pre-defined structured database and indexed. Based on the filtering and combination of multi-dimensional classification labels, the structured database is filtered and aggregated to generate a structured medical dataset for a specific topic. Example 3:

[0041] This application also provides an electronic device. The electronic device may include one or more processors and one or more memories. The memories store computer-readable code, which, when executed by the one or more processors, can perform the AI-based medical image data tagging management method described above.

[0042] The methods or systems according to embodiments of this application can also be implemented using the architecture of the electronic device shown in this application. The electronic device may include a bus, one or more CPUs, ROM, RAM, a communication port connected to a network, input / output, a hard disk, etc. Storage devices in the electronic device, such as ROM or hard disk, may store the AI-based medical image data tagging management method provided in this application. Furthermore, the electronic device may also include a user interface. Of course, the architecture shown in this application is merely exemplary; when implementing different devices, one or more components in the electronic device shown in this application may be omitted according to actual needs. Example 4:

[0043] One embodiment of this application discloses a computer-readable storage medium. The computer-readable storage medium stores computer-readable instructions. When the computer-readable instructions are executed by a processor, the AI-based medical image data tagging management method according to the embodiment of this application, described with reference to the above figures, can be performed. The storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0044] Furthermore, according to embodiments of this application, the processes described in the above-referenced flowcharts can be implemented as computer software programs. For example, this application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be executed by a processor to perform instructions corresponding to the method steps provided in this application, such as an AI-based medical image data tagging management method. When this computer program is executed by a central processing unit (CPU), it performs the functions defined in the method of this application.

[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0046] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0047] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An AI-based method for labeling and managing medical image data, characterized in that, include: Obtain the hospital's original medical image data and corresponding unstructured text reports, perform de-identification processing on the original medical image data and unstructured text reports, generate de-identified image data and de-identified report text, and form a de-identified image report dataset; Based on the anonymized report text in the anonymized image report dataset, a pre-trained natural language processing model is invoked to extract medical entities and identify the logical relationships between each medical entity to form a structured semantic set. Based on the desensitized image data and structured semantic set in the desensitized image report dataset, medical entities are associated and mapped with the desensitized image data to form a cross-modal alignment result set; Based on the structured semantic set and the cross-modal alignment result set, multi-dimensional classification labels are automatically generated for desensitized image data, and the desensitized image report dataset is transformed into a labeled image dataset containing multi-dimensional classification labels. The labeled image dataset is stored in a pre-defined structured database and indexed. Based on the filtering and combination of multi-dimensional classification labels, the structured database is filtered and aggregated to generate a structured medical dataset for a specific topic.

2. The AI-based medical image data tagging management method according to claim 1, characterized in that, Methods for generating desensitized image report datasets include: The raw medical imaging data includes a unique examination code, a DICOM file header, and one or more image sequences. The DICOM file header contains patient information tags and technical parameter tags. The image sequence contains a sequence number, image modality type, examination site description, sequence description information, and multi-frame pixel data. The unstructured text report contains a unique examination code, examination description text, and diagnostic conclusion text. One set of medical imaging data and an unstructured text report corresponds to one imaging examination. A list of personal identification fields is pre-defined. For each set of original medical image data, all patient information tags in the corresponding DICOM file header are traversed. Patient information tags in the list of personal identification fields are replaced with predefined anonymous placeholders or deleted. The examination unique code and technical parameter tags are retained unchanged to obtain desensitized image data. A pre-defined set of text anonymization rules is used to scan the examination description text and diagnostic conclusion text for each group of unstructured text reports. The identified personal information fragments are replaced with general placeholders or deleted to obtain the anonymized report text. The unique examination code, desensitized image data, and desensitized report text corresponding to each image examination are integrated to form a desensitized image report dataset.

3. The AI-based medical image data tagging management method according to claim 2, characterized in that, Methods for extracting medical entities by calling pre-trained natural language processing models include: Multiple standard entity categories are preset, including disease name category, anatomical location category, imaging sign category, quantitative information category, and diagnostic conclusion category; for each de-identified report text, it is sequentially split into independent sentence units; a pre-built natural language processing model is constructed, and each sentence unit is sequentially input into the natural language processing model to output the corresponding candidate medical entity; each candidate medical entity includes the original text fragment, entity category, and initial confidence level; Negation range detection is performed on each candidate medical entity, and a corresponding negation label is set for each candidate medical entity based on the detection results; terminology standardization mapping is performed on each candidate medical entity to obtain the corresponding entity standard name and superordinate concept name; for candidate medical entities whose entity category is quantitative information category, quantitative attribute auxiliary information is extracted from the corresponding original text fragment; The entity standard name, entity category, original text fragment, negation mark, superordinate concept name and quantitative attribute information corresponding to the same candidate medical entity are integrated, and a unique entity identifier is assigned to each candidate medical entity to form a confirmed medical entity; all confirmed medical entities corresponding to the same de-identified report text are summarized to form an entity record list.

4. The AI-based medical image data tagging management method according to claim 3, characterized in that, Methods for forming structured semantic sets include: A pre-defined medical relationship type system is established, which includes multiple standard relationship types. The standard relationship types include location-lesion relationship, lesion-sign relationship, lesion-quantification relationship, sign-quantification relationship, and lesion-recommendation relationship. For the entity record list of the same anonymized report text, the relationship between any two medical entities is determined and the corresponding relationship record is generated. The relationship record includes a unique relation identifier, a head entity identifier, a tail entity identifier, a relation type, and a relation confidence level. All relation records corresponding to the same anonymized report text are summarized to form a relation record list. For each anonymized report text, perform semantic integrity checks on the entity record list to determine whether a semantic incompleteness flag is set. Integrate the unique check code, entity record list, relation record list, and semantic incompleteness flag corresponding to each anonymized report text to form a structured semantic set.

5. The AI-based medical image data tagging management method according to claim 4, characterized in that, Methods for associating and mapping medical entities with desensitized image data include: From the entity record list, medical entities with the entity category of anatomical location are selected as anatomical location entities, forming a location entity set; from the entity record list, medical entities with the entity category of disease name are selected as lesion entities, forming a lesion entity set; from the relation record list, relation records with the relation type of location-lesion relationship are selected, forming a location association relation set. Based on the set of site associations, the corresponding anatomical site entity is determined for each lesion entity in the lesion entity set; based on the desensitized image data corresponding to each examination unique code, sequence-level alignment is performed to determine the associated sequence number and alignment status of each anatomical site entity; For anatomical entities that have been successfully aligned at the sequence level, perform layer-level localization to determine the corresponding associated layer range and alignment granularity; based on the attribution relationship between anatomical entities and lesion entities, inherit the associated sequence number, associated layer range, and alignment granularity of the lesion entity from the corresponding attribution entity; and set the alignment confidence level based on the alignment granularity and alignment status.

6. The AI-based medical image data tagging management method according to claim 5, characterized in that, Methods for generating cross-modal alignment result sets include: The associated sequence number, associated layer range, alignment granularity, and alignment confidence of each lesion entity are integrated to form the alignment result of the corresponding lesion entity; the associated sequence number, associated layer range, alignment granularity, and alignment confidence of each anatomical location entity are integrated to form the alignment result of the corresponding anatomical location entity; all alignment results under the same examination unique code are summarized to form the cross-modal alignment result; and all cross-modal alignment results are summarized to form the cross-modal alignment result set.

7. The AI-based medical image data tagging management method according to claim 6, characterized in that, Methods for automatically generating multi-dimensional classification labels for desensitized image data include: A multi-dimensional label classification system is pre-set, which includes disease classification dimension, anatomical location dimension, imaging feature dimension, severity dimension, and treatment recommendation dimension; Based on the entity record list, generate a set of disease classification labels for each imaging examination; based on the entity record list and the cross-modal alignment result set, generate a set of anatomical location labels for each imaging examination; based on the entity record list and the relation record list, generate a set of imaging sign labels for each imaging examination; based on the entity record list and the relation record list, infer the severity labels for each imaging examination; based on the entity record list, generate a set of treatment recommendation labels for each imaging examination. For all labels generated from each imaging examination, a label evidence chain is constructed; based on the initial confidence, semantic incompleteness markers, and alignment confidence of the medical entity on which each label is based, a comprehensive label quality score for each imaging examination is generated.

8. The AI-based medical image data tagging management method according to claim 7, characterized in that, Methods for converting desensitized image report datasets into labeled image datasets include: The unique examination code, disease classification label set, anatomical location label set, imaging sign label set, severity label, treatment suggestion label set, label evidence chain, and comprehensive label quality score corresponding to each imaging examination are integrated to form a multi-dimensional classification label set for the corresponding imaging examination. The multi-dimensional classification label set of each imaging examination is then linked and integrated with the corresponding desensitized image data and desensitized report text in the desensitized image report dataset to form a labeled image dataset.

9. The AI-based medical image data tagging management method according to claim 8, characterized in that, Methods for generating structured medical datasets on specific topics include: The system receives a dataset construction request, which includes a combination of filtering conditions and a dataset name. Based on the combination of filtering conditions, it performs a search query on the structured database to obtain a set of examination records that meet all filtering conditions. It then integrates the desensitized image data, desensitized report text, and multi-dimensional classification label set corresponding to each set of examination records in the examination record set to form a structured medical dataset labeled with the dataset name.

10. An AI-based medical image data labeling management system, implementing the AI-based medical image data labeling management method according to any one of claims 1-9, characterized in that, include: The data anonymization module is used to acquire the hospital's original medical image data and corresponding unstructured text reports, perform de-identification processing on the original medical image data and unstructured text reports, generate anonymized image data and anonymized report text, and form an anonymized image report dataset. The semantic extraction module is used to extract medical entities from the desensitized report text in the desensitized image report dataset by calling a pre-trained natural language processing model, and to identify the logical relationships between the medical entities to form a structured semantic set. The image alignment module is used to associate and map medical entities with desensitized image data based on the desensitized image data and structured semantic set in the desensitized image report dataset, forming a cross-modal alignment result set; The label generation module is used to automatically generate multi-dimensional classification labels for desensitized image data based on the structured semantic set and the cross-modal alignment result set, and to transform the desensitized image report dataset into a labeled image dataset containing multi-dimensional classification labels. The data management module is used to store labeled image datasets into a preset structured database to create an index, and to perform filtering and aggregation operations on the structured database based on the filtering and combination of multi-dimensional classification labels to generate structured medical datasets on specific topics.

Images