A multi-modal medical image knowledge graph construction system and method for an artificial intelligence undergraduate medical image AI practical training platform

Abstract

The application discloses a kind of multi-modal medical image knowledge graph construction system and method for AI practical teaching, belong to artificial intelligence and medical image teaching cross technical field, including: multi-modal medical image data acquisition and pretreatment module, for collecting multi-modal image data, diagnosis report, teaching case and student practical operation data;Medical entity recognition and relationship extraction module is used to identify medical entity and extract the relationship between entities;Multi-modal knowledge graph construction module is used to integrate entity, relationship and image features into structured knowledge graph;Knowledge graph query and reasoning module is used to provide knowledge services for automatic scoring, weak point positioning, intelligent question answering and practical training guidance, the technical problems that the present application solves are that the data source is single, not designed for practical training scene, disjointed with evaluation system and lack of dynamic updating mechanism, realize the deep integration of multi-modal medical image knowledge and practical teaching.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and medical imaging teaching, and in particular to a multimodal medical imaging knowledge graph construction system and method for an AI training platform for undergraduate medical imaging. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology in the field of medical imaging, medical imaging AI has become an important part of the curriculum for undergraduate AI majors. Traditional medical imaging teaching mainly relies on theoretical lectures and single-experiment teaching, leaving students with limited opportunities for systematic AI modeling and practical training with medical imaging data. In recent years, some universities have begun to build medical imaging AI training platforms, providing students with a complete operational environment encompassing image preprocessing, feature extraction, model building, training, and parameter tuning. During practical training, a pressing technical challenge is how to structure and correlate multimodal medical imaging data (CT, MRI, PET, etc.) and related diagnostic knowledge, imaging features, and anatomical information so that students can quickly retrieve, understand, and apply this knowledge during training, while also providing knowledge support for the platform's automatic scoring and weakness identification.

[0003] Knowledge graphs, as a knowledge representation method that organizes entities and their relationships using a graph structure, provide an effective way to solve the above problems. However, there is currently a lack of systematic research and mature technical solutions for constructing multimodal medical imaging knowledge graphs for undergraduate practical training scenarios.

[0004] Chinese invention patent CN120654797A discloses a teaching system for medical image reports, mainly including a structured processing module, a knowledge graph construction module, a knowledge base construction module, a knowledge graph association module, a user question retrieval module, and an answer content generation module. It describes the hierarchical features of medical image reports and associates feature entities, constructs a knowledge graph based on the relationships between feature entities, and then builds a disease database, an image feature database, and a report feature database. Finally, it uses a RAG model to implement a knowledge graph-based question-and-answer function. The existing technology has the following shortcomings: (1) The data source is singular, mainly relying on medical image reports as the source of knowledge, and does not make full use of multimodal medical image data (such as original images in DICOM format, image annotation data, etc.); (2) It is not designed for practical training scenarios, and is positioned as "medical image report teaching", which cannot support students' needs for knowledge query and correlation analysis of operations such as image preprocessing, feature extraction, and model training during practical training; (3) It lacks linkage with the practical training process and does not consider knowledge linkage and data interaction with other modules of the practical training platform; (4) The knowledge update mechanism is missing and no dynamic update mechanism for the knowledge graph is provided.

[0005] A Chinese invention patent, CN120706515A, discloses a method for constructing a medical multimodal knowledge graph. The method mainly includes: constructing a disease seed database based on the International Classification of Diseases (ICD) standard and selecting standardized disease terms; obtaining disease-related text data from a general knowledge platform using terms from the disease seed database as keywords; performing medical entity recognition on the text data using a medical named entity recognition tool; extracting medical knowledge triples based on a large language model-based instruction fine-tuning method; obtaining medical images using disease terms as keywords; and integrating and storing medical entities, triple relationships, and associated medical images into a graph database. Its core lies in fusing textual and image information from the disease seed database to enhance the multimodal expressive capabilities of the knowledge graph. The existing technology has the following shortcomings: (1) It is geared towards clinical auxiliary diagnosis rather than teaching, and does not take into account the special needs in practical training (such as high requirements for knowledge interpretability, and the need to include student operation data in the analysis); (2) The knowledge source depends on public data platforms, and in the practical training scenario, it is also necessary to integrate knowledge from specific sources such as teaching cases provided by teachers, historical student practical training data, and teaching outlines; (3) It lacks connection with the practical training evaluation system and does not consider how the knowledge graph can serve practical training scoring and weak point location; (4) The construction process is relatively heavy, relying on the fine-tuning of instructions from large language models, and has high computational resource requirements.

[0006] There is an existing Chinese invention patent CN121506405A, which discloses an AI-based intelligent teaching platform for medical imaging, mainly including an image acquisition unit, a question storage unit, a statistics unit, a comparison unit, and an analysis unit. Its core technical solution lies in determining the suitability of stored data for medical image subcategories based on contour difference characterization values, and achieving dynamic optimization management of medical image data in teaching resources by adjusting parameters such as acquisition cycle, storage capacity, and resolution. The existing technology has the following shortcomings: (1) It focuses on image resource management rather than the entire teaching process, mainly solving the problem of "efficiency of teaching resource management", and cannot cover the knowledge needs of students in the entire AI training process; (2) It does not involve the fusion processing of multimodal data, and the knowledge dimension is singular; (3) It does not include a knowledge graph construction module, and cannot provide structured knowledge organization and related retrieval services; (4) It is not linked with the training evaluation system; (5) It is not geared towards AI professional training scenarios.

[0007] There is an existing Chinese invention patent CN122117443A, which discloses a method for constructing a medical data analysis model, mainly including: constructing a medical knowledge graph based on medical data sources; constructing a multimodal data processing module to process and extract features from multimodal medical image data; and constructing a data analysis processing module to perform semantic association and constraint on medical image features based on the knowledge graph to obtain medical data analysis results. The existing technology has the following shortcomings: (1) it is geared towards clinical auxiliary diagnosis rather than teaching scenarios; (2) it lacks linkage with the practical training evaluation system; (3) the construction process is heavy and the deployment threshold is high, relying on complex technologies such as cross-modal feature alignment and fusion; (4) it does not involve the dynamic update mechanism of knowledge graph; (5) the knowledge source does not match the practical training scenario.

[0008] In summary, existing technologies all share the following common problems: First, the data sources are singular or mismatched with the training scenarios, failing to fully integrate multimodal medical imaging data with training data; second, none are designed for undergraduate AI training scenarios, failing to support the knowledge needs of students throughout the entire training process; third, they are disconnected from the training evaluation system, failing to provide knowledge support for automatic scoring and weak point identification; and fourth, the construction process is cumbersome, the deployment threshold is high, or they lack dynamic update capabilities, making it difficult to adapt to the resource constraints of the undergraduate teaching environment and the realistic needs of the continuous evolution of teaching content.

[0009] Therefore, there is an urgent need for a multimodal medical image knowledge graph construction system and method for an AI training platform for undergraduate medical imaging. This system would integrate image data from multiple modalities, such as CT, MRI, and PET, along with their text descriptions and DICOM metadata, into a knowledge graph through a multimodal medical image data acquisition and preprocessing module and an image feature embedding submodule. This would overcome the shortcomings of existing technologies that rely on a single data source and achieve true multimodal medical image knowledge fusion. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention provides a multimodal medical imaging knowledge graph construction system and method for an AI-based undergraduate medical imaging training platform. This system solves the technical problems of existing technologies, such as single data sources, lack of design for training scenarios, disconnect from evaluation systems, and lack of dynamic update mechanisms. It achieves deep integration of multimodal medical imaging knowledge with training and teaching.

[0011] The above-mentioned technical objective of the present invention is achieved through the following technical solution: A multimodal medical image knowledge graph construction system for AI practical training includes a multimodal medical image data acquisition and preprocessing module, which is used to collect multimodal medical image data and its metadata, diagnostic report text, teaching case annotation data and students' historical practical training operation data from the data source of the training platform, and perform standardized preprocessing. The medical entity recognition and relation extraction module is used to identify medical entities from preprocessed text data, map entities to a standard medical terminology system, extract relations between entities, and establish multimodal associations between entities and image data. The multimodal knowledge graph construction module is used to organize entities and relations into knowledge triples according to predefined graph patterns, extract image feature vectors using deep learning models as entity attributes and embed them into the graph, and fuse and deduplicate knowledge extracted from different data sources and store it in the graph database. The knowledge graph query and reasoning module is used to provide structured query, semantic retrieval, knowledge reasoning and path discovery services for the upper-level applications of the training platform; The knowledge graph dynamic update module is used to monitor data changes on the training platform and trigger incremental updates when new data is added, provide a manual review interface, and record version logs.

[0012] Furthermore, the multimodal medical image data acquisition and preprocessing module includes: The image data acquisition submodule is used to acquire medical image data and its DICOM metadata of various modalities such as CT, MRI, and PET via the DICOM protocol or file import method. The image preprocessing submodule is used to perform format unification, desensitization, grayscale normalization, and resampling preprocessing operations on the acquired medical images. The text data acquisition submodule is used to collect text data related to images, including diagnostic reports, image findings, anatomical structure annotations, and teaching case descriptions. The training data acquisition submodule is used to collect image preprocessing parameters, feature extraction results, model training logs, and training result data of students in the training platform.

[0013] Furthermore, the medical entity recognition and relation extraction module includes: The medical named entity recognition submodule is used to identify medical entities such as disease names, anatomical structures, imaging features, examination methods, and equipment types from text data using a deep learning-based Chinese medical named entity recognition model. The entity standardization submodule is used to map the identified entities to the UMLS, SNOMED CT, or ICD standard medical terminology system; The relation extraction submodule is used to extract the relationships between entity pairs from text data. The relationships include disease-imaging sign relationships, anatomical structure-imaging manifestation relationships, disease-examination method relationships, sign-severity relationships, and pre- and post-relationships between training operation steps. The multimodal association submodule is used to establish the association between medical entities identified in the text and their corresponding image regions or features.

[0014] Furthermore, the multimodal knowledge graph construction module includes: The graph pattern definition submodule is used to define the entity types, relation types, and attribute types of the knowledge graph. The entity types include diseases, anatomical structures, imaging signs, examination modalities, equipment parameters, training operation steps, and evaluation indicators. The relation types include representation, location, use, pre-position, post-position, association, and dependency. The knowledge triple generation submodule is used to convert the results of entity recognition and relation extraction into triples of <entity, relation, entity>. The image feature embedding submodule is used to extract features from medical images using a deep learning model and store the generated feature vectors as attributes of the corresponding entities in the knowledge graph. The graph fusion submodule is used to fuse and deduplicat knowledge triples extracted from different data sources such as image reports, teaching cases and training logs. The graph storage submodule is used to store the constructed knowledge graph in a graph database.

[0015] Furthermore, the knowledge graph query and reasoning module includes: The structured query submodule provides an API interface based on a graph query language, enabling upper-layer applications to perform precise queries on the knowledge graph. The semantic retrieval submodule is used to support natural language-based semantic retrieval, transforming user input into knowledge graph query statements; The knowledge reasoning submodule is used to perform reasoning based on entities and relationships in the knowledge graph to discover implicit knowledge connections; The path discovery submodule is used to discover the associated paths between two entities.

[0016] Furthermore, the knowledge graph dynamic update module includes: The incremental update submodule is used to automatically trigger incremental updates of the knowledge graph when data changes, updating only the affected parts; The manual review submodule provides an interface for teachers to manually review and confirm automatically updated content. The version management submodule is used to record each update of the knowledge graph, and supports backtracking and historical comparison.

[0017] A method for constructing a multimodal medical image knowledge graph for AI practical training includes the following steps: Step S1: Collect multimodal medical image data and its DICOM metadata, diagnostic report text, teaching case annotation data and students' historical training operation data from the data source of the training platform. Perform format unification, desensitization and normalization preprocessing on the collected image data, and perform word segmentation and noise reduction preprocessing on the text data. Step S2: Use a medical named entity recognition model to identify medical entities from the preprocessed text data, map the entities to the standard medical terminology system, and use a combination of rule-based and deep learning methods to extract the relationships between entities and establish multimodal associations between entities and corresponding image regions or features. Step S3: Organize entities and relations into knowledge triples according to the predefined graph pattern, use a deep learning model to extract image feature vectors as entity attributes and embed them into the graph, fuse and deduplicate knowledge extracted from different data sources, and store the final knowledge graph in the graph database. Step S4: Encapsulate the constructed knowledge graph into a RESTful API service for use by the training platform's automatic scoring module, weak point location module, intelligent question answering module, and training guidance module; Step S5: Monitor data changes on the training platform, automatically trigger incremental updates when data changes, confirm the content of automatic updates through the teacher review interface, and record the version log of all update operations; Furthermore, the student historical training operation data mentioned in step S1 includes image preprocessing parameters, feature extraction results, model training logs, and training result data.

[0018] Furthermore, the entity types defined in the graph pattern in step S3 include diseases, anatomical structures, imaging signs, examination modalities, equipment parameters, training operation steps, and evaluation indicators, and the relationship types include manifestation, location, use, preceding, following, association, and dependency.

[0019] Furthermore, in step S4, the automatic scoring module automatically scores the students' training results based on the standard answers and knowledge associations in the knowledge graph; the weakness point identification module identifies the students' weak knowledge points based on their performance in each stage of the training and the knowledge dependencies in the knowledge graph; the intelligent question-answering module answers the questions raised by the students during the training based on the knowledge graph; and the training guidance module provides students with real-time training guidance based on the operational process knowledge in the knowledge graph.

[0020] In summary, the beneficial technical effects of the present invention are as follows: 1. Enhance the multimodal fusion capability of knowledge organization. Through the multimodal medical image data acquisition and preprocessing module and the image feature embedding submodule, image data from multiple modalities such as CT, MRI, and PET, along with their text descriptions and DICOM metadata, are uniformly integrated into the knowledge graph. This overcomes the shortcomings of existing technologies that rely on a single data source and achieves true multimodal medical image knowledge fusion.

[0021] 2. Achieve deep integration of knowledge graph and practical training process. By incorporating students' practical training operation data (preprocessing parameters, feature extraction results, model training logs, etc.) into the construction and reasoning process of knowledge graph through the practical training data collection submodule, the knowledge graph can directly serve practical training scoring and weak point identification, solving the problem of the disconnect between existing technology and practical training evaluation system.

[0022] 3. Lowering the deployment threshold for knowledge graph construction. Through modular design and lightweight technical implementation path (using lightweight model architectures such as BERT-BiLSTM-CRF for NER, without necessarily relying on large language model fine-tuning), the system of this invention can be deployed and run at low cost in undergraduate teaching environments, overcoming the shortcomings of existing technologies such as heavy construction process and high computing power requirements, and has good scalability.

[0023] 4. Ensure the continuous effectiveness of the knowledge graph. The dynamic update module enables incremental updates and version management of the knowledge graph, allowing it to continuously evolve as teaching content is updated, maintaining the timeliness of knowledge and addressing the deficiency of existing technologies that lack a dynamic update mechanism.

[0024] 5. Enhance the intelligence level of practical training. By providing knowledge support for functions such as automatic scoring, weak point identification, intelligent question answering, and practical training guidance, comprehensively enhance the intelligence level of medical imaging AI practical training and support the reform of undergraduate process-based teaching assessment. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention; Figure 2 This is a schematic diagram of the module structure of the system of the present invention; Figure 3 This is a schematic diagram of the data flow of the system of the present invention; Figure 4 This is a schematic diagram of the workflow of the system of the present invention. Detailed Implementation

[0026] The present invention will now be clearly and completely described in conjunction with embodiments.

[0027] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0028] In the description of the embodiments, unless otherwise explicitly specified and limited, the terms "set," "connect," etc., should be interpreted broadly. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or a connection through an intermediate medium, or it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0029] See appendix Figure 1-4 The diagram illustrates a multimodal medical image knowledge graph construction system and method for AI practical training provided by the present invention, specifically including the following implementation methods: Implementation Method 1: System Deployment and Initialization In this embodiment, the system of the present invention is deployed in the medical imaging AI training platform of an undergraduate major in artificial intelligence at a university. The system server is configured with: 16 CPU cores, 64GB of memory, one NVIDIA A10 GPU, and 4TB of storage space. The graph database used is Neo4j version 5.x. The following operations are performed during the initialization phase: First, the system administrator imports the initial teaching dataset through the multimodal medical image data acquisition and preprocessing module, including: (1) 100 CT images and corresponding DICOM metadata; (2) 100 MRI images and corresponding DICOM metadata; (3) 50 PET images and corresponding DICOM metadata; (4) the diagnostic report text and image description text corresponding to the above images; (5) the anatomical structure annotations and knowledge point descriptions of 50 teaching cases; Secondly, the medical entity recognition and relation extraction module processes the imported text data. In this implementation, the NER model uses a BERT-BiLSTM-CRF-based Chinese medical named entity recognition model, trained on pre-annotated medical corpus. The model identifies entities such as disease names, anatomical structures, imaging features, and examination methods from the text, and maps these entities to the ICD-11 and SNOMED CT standard terminology systems through the entity normalization submodule. Relation extraction employs a method combining rule templates and remote supervision to extract relationships between entity pairs. Then, the multimodal knowledge graph construction module generates knowledge triples based on predefined graph patterns. In this embodiment, the entity types defined by the graph patterns include: Disease, Anatomy, Sign, Modality, Step, and Metric; the relation types include: manifest_as, located_at, use, precede, follow, relate_to, and depend_on. The image feature embedding submodule uses the ResNet-50 model to extract feature vectors (2048 dimensions) from CT and MRI images, which are stored in the knowledge graph as attributes of the corresponding image entities. The graph fusion submodule performs deduplication and fusion on triples extracted from different data sources based on entity alignment and relation conflict resolution. The final knowledge graph is stored in the Neo4j graph database.

[0030] Implementation Method Two: Knowledge Graph Service-Oriented Application and Practical Training In this implementation, the knowledge graph query and reasoning module encapsulates the constructed knowledge graph into a RESTful API service, which can be called by various functional modules of the training platform. Automatic scoring scenario: After a student completes a classification model training exercise for a lung nodule CT image, the automatic scoring module calls the API through the structured query submodule to retrieve standard imaging features related to "lung nodules" (such as "lobulation sign", "spiculation sign", "calcification", etc.) and the corresponding standard diagnostic procedures from the knowledge graph. The module then compares the student's model output with the standard knowledge in the knowledge graph to achieve automatic scoring. Weakness identification scenario: After students' operational data in each stage of practical training (such as incorrect preprocessing parameter settings, incomplete feature extraction, etc.) are entered into the system through the practical training data acquisition submodule, the weakness identification module combines the knowledge dependency relationships in the knowledge graph (such as "feature extraction" depends on "image preprocessing", "model training" depends on "feature extraction", etc.) to locate the students' weak knowledge points and their positions in the knowledge graph, and generate a personalized weakness diagnosis report. Intelligent question answering scenario: Students input "What are the CT imaging features of lung nodules?" through the question answering interface of the training platform. The semantic retrieval submodule transforms the natural language question into a knowledge graph query statement, retrieves relevant entities and relationships from the knowledge graph, and returns a structured answer. Practical training guidance scenario: When students are performing image preprocessing operations, the practical training guidance module provides real-time step guidance and prompts based on the operation process knowledge in the knowledge graph (the post-step of "image preprocessing" is "feature extraction", and the pre-step of "feature extraction" is "image preprocessing", etc.).

[0031] Implementation Method 3: Dynamic Update of Knowledge Graph In this implementation, when a teacher imports a new teaching case (including new CT images, corresponding diagnostic reports, and explanations of teaching knowledge points) into the training platform, the knowledge graph dynamic update module automatically triggers the incremental update process. The incremental update submodule first identifies entities and relationships in the newly added data, compares them with the existing knowledge graph, and only writes the newly added or changed parts to the graph database, rather than rebuilding the entire graph. New content officially takes effect after being confirmed by the teacher through the manual review submodule. The version management submodule records information such as the time, content, and operator of this update, forming a version log that supports subsequent retrospective analysis and historical comparison. Through the above implementation methods, the present invention system realizes the construction, service-oriented and dynamic updating of a multimodal medical image knowledge graph for AI training and teaching, effectively supporting the intelligent operation of the entire process of medical image AI training and teaching.

[0032] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A multimodal medical image knowledge graph construction system for AI practical training, characterized in that: It includes a multimodal medical image data acquisition and preprocessing module, which is used to collect multimodal medical image data and its metadata, diagnostic report text, teaching case annotation data and students' historical training operation data from the data source of the training platform, and perform standardized preprocessing. The medical entity recognition and relation extraction module is used to identify medical entities from preprocessed text data, map entities to a standard medical terminology system, extract relations between entities, and establish multimodal associations between entities and image data. The multimodal knowledge graph construction module is used to organize entities and relations into knowledge triples according to predefined graph patterns, extract image feature vectors using deep learning models as entity attributes and embed them into the graph, and fuse and deduplicate knowledge extracted from different data sources and store it in the graph database. The knowledge graph query and reasoning module is used to provide structured query, semantic retrieval, knowledge reasoning and path discovery services for the upper-level applications of the training platform; The knowledge graph dynamic update module is used to monitor data changes on the training platform and trigger incremental updates when new data is added, provide a manual review interface, and record version logs.

2. The multimodal medical image knowledge graph construction system for AI practical training as described in claim 1, characterized in that: The multimodal medical image data acquisition and preprocessing module includes: The image data acquisition submodule is used to acquire medical image data and its DICOM metadata of various modalities such as CT, MRI, and PET via the DICOM protocol or file import method. The image preprocessing submodule is used to perform format unification, desensitization, grayscale normalization, and resampling preprocessing operations on the acquired medical images. The text data acquisition submodule is used to collect text data related to images, including diagnostic reports, image findings, anatomical structure annotations, and teaching case descriptions. The training data acquisition submodule is used to collect image preprocessing parameters, feature extraction results, model training logs, and training result data of students in the training platform.

3. The multimodal medical image knowledge graph construction system for AI practical training as described in claim 1, characterized in that: The medical entity recognition and relation extraction module includes: The medical named entity recognition submodule is used to identify medical entities such as disease names, anatomical structures, imaging features, examination methods, and equipment types from text data using a deep learning-based Chinese medical named entity recognition model. The entity standardization submodule is used to map the identified entities to the UMLS, SNOMED CT, or ICD standard medical terminology system; The relation extraction submodule is used to extract the relationships between entity pairs from text data. The relationships include disease-imaging sign relationships, anatomical structure-imaging manifestation relationships, disease-examination method relationships, sign-severity relationships, and pre- and post-relationships between training operation steps. The multimodal association submodule is used to establish the association between medical entities identified in the text and their corresponding image regions or features.

4. The multimodal medical image knowledge graph construction system for AI practical training as described in claim 1, characterized in that: The multimodal knowledge graph construction module includes: The graph pattern definition submodule is used to define the entity types, relation types, and attribute types of the knowledge graph. The entity types include diseases, anatomical structures, imaging signs, examination modalities, equipment parameters, training operation steps, and evaluation indicators. The relation types include representation, location, use, pre-position, post-position, association, and dependency. The knowledge triple generation submodule is used to convert the results of entity recognition and relation extraction into triples of <entity, relation, entity>. The image feature embedding submodule is used to extract features from medical images using a deep learning model and store the generated feature vectors as attributes of the corresponding entities in the knowledge graph. The graph fusion submodule is used to fuse and deduplicat knowledge triples extracted from different data sources such as image reports, teaching cases and training logs. The graph storage submodule is used to store the constructed knowledge graph in a graph database.

5. The multimodal medical image knowledge graph construction system for AI practical training as described in claim 1, characterized in that: The knowledge graph query and reasoning module includes: The structured query submodule provides an API interface based on a graph query language, enabling upper-layer applications to perform precise queries on the knowledge graph. The semantic retrieval submodule is used to support natural language-based semantic retrieval, transforming user input into knowledge graph query statements; The knowledge reasoning submodule is used to perform reasoning based on entities and relationships in the knowledge graph to discover implicit knowledge connections; The path discovery submodule is used to discover the associated paths between two entities.

6. The multimodal medical image knowledge graph construction system for AI practical training as described in claim 1, characterized in that: The knowledge graph dynamic update module includes: The incremental update submodule is used to automatically trigger incremental updates of the knowledge graph when data changes, updating only the affected parts; The manual review submodule provides an interface for teachers to manually review and confirm automatically updated content. The version management submodule is used to record each update of the knowledge graph, and supports backtracking and historical comparison.

7. A method for constructing a multimodal medical image knowledge graph for AI practical training, characterized in that, Includes the following steps: Step S1: Collect multimodal medical image data and its DICOM metadata, diagnostic report text, teaching case annotation data and students' historical training operation data from the data source of the training platform. Perform format unification, desensitization and normalization preprocessing on the collected image data, and perform word segmentation and noise reduction preprocessing on the text data. Step S2: Use a medical named entity recognition model to identify medical entities from the preprocessed text data, map the entities to the standard medical terminology system, and use a combination of rule-based and deep learning methods to extract the relationships between entities and establish multimodal associations between entities and corresponding image regions or features. Step S3: Organize entities and relations into knowledge triples according to the predefined graph pattern, use a deep learning model to extract image feature vectors as entity attributes and embed them into the graph, fuse and deduplicate knowledge extracted from different data sources, and store the final knowledge graph in the graph database. Step S4: Encapsulate the constructed knowledge graph into a RESTful API service for use by the training platform's automatic scoring module, weak point location module, intelligent question answering module, and training guidance module; Step S5: Monitor data changes on the training platform, automatically trigger incremental updates when data changes, confirm the content of the automatic updates through the teacher review interface, and record the version log of all update operations.

8. The method according to claim 7, characterized in that, The student historical training operation data mentioned in step S1 includes image preprocessing parameters, feature extraction results, model training logs, and training result data.

9. The method according to claim 7, characterized in that, The entity types defined in the graph pattern in step S3 include diseases, anatomical structures, imaging features, examination modalities, equipment parameters, training operation steps, and evaluation indicators. The relationship types include manifestation, location, use, preceding, following, association, and dependency.

10. The method according to claim 7, characterized in that, In step S4, the automatic scoring module automatically scores the students' training results based on the standard answers and knowledge associations in the knowledge graph; the weakness point identification module identifies the students' weak knowledge points based on their performance in each stage of the training and the knowledge dependencies in the knowledge graph; the intelligent question-answering module answers the questions raised by the students during the training based on the knowledge graph; and the training guidance module provides students with real-time training guidance based on the operational process knowledge in the knowledge graph.

Images