Medical literature vectorization retrieval device and method based on AI and RFID technologies

Abstract

The invention discloses a medical literature vectorization retrieval device and method based on AI and RFID technologies, and belongs to the field of medical literature retrieval, the medical literature vectorization retrieval device comprises a literature acquisition module used for acquiring multi-modal medical literature (including paper literature, electronic literature, medical images, operation videos, 3D animation generated by AI and foreign language data), performing physical marking on all data through an RFID tag, and synchronously inputting metadata; the 3D animation generation module is used for integrating an AI generation model and a medical knowledge base, supporting generation of a 3D animation of an operation process or a medical instrument according to a user demand, and outputting a standardized animation file containing an anatomical structure and an operation process; according to the method, a traditional passive query mode is broken through, customized 3D operation animation is actively generated through the AI generation module, dynamic demonstration content is covered, interactive 3D file output is supported, and literature content boundaries are expanded.

Description

Technical Field

[0001] This invention relates to the field of medical literature retrieval technology, and in particular to a vectorized retrieval device and method for medical literature based on AI and RFID technologies. Background Technology

[0002] With the surge in demand for dynamic and visualized data in medical education and clinical research, traditional literature retrieval faces new challenges:

[0003] Lack of dynamic data: The existing literature database is mainly composed of static text and images, lacking 3D animations that can intuitively show the surgical process, making it difficult for medical students to understand complex surgical procedures (such as robot-assisted radical prostatectomy).

[0004] AI-generated content has low credibility: Although AI can generate 3D surgical animations, due to biases in training data (such as ignoring anatomical variations) or errors in the process (such as reversing the order of steps), the generated content often does not match clinical reality and cannot be directly used for teaching.

[0005] Multimodal retrieval has limitations: traditional retrieval only supports "passive querying" of existing documents and cannot "actively generate" 3D animations that meet user needs, thus limiting the scalability of the document database.

[0006] Therefore, there is a need for a vectorized retrieval device and method for medical literature based on AI and RFID technologies to solve the problems that traditional retrieval methods cannot handle dynamic surgical demonstrations, the accuracy of AI-generated content is insufficient, and the retrieval scope is limited. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a vectorized retrieval device and method for medical literature based on AI (Artificial Intelligence) and RFID (Radio Frequency Identification) technologies. It ensures accuracy by generating 3D animations with AI and combining them with a medical knowledge verification mechanism. At the same time, it integrates RFID positioning and vectorized retrieval technologies to solve the problems of missing dynamic data, unreliable AI-generated content, and limited retrieval scope.

[0008] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a vectorized retrieval device for medical literature based on AI and RFID technologies, comprising:

[0009] Document Acquisition Module: Used to acquire multimodal medical literature (including paper documents, electronic documents, medical images, surgical videos, AI-generated 3D animations, and foreign language materials), physically marking all materials with RFID tags and synchronously entering metadata (including author, keywords, access level, document type, language type, and source of generation).

[0010] 3D Animation Generation Module: Integrates AI generation models (such as diffusion models and neural radiation field Nerf) with a medical knowledge base, supports the generation of 3D animations of surgical procedures or medical devices according to user needs, and outputs standardized animation files containing anatomical structures and operating procedures (this module is based on AI generation models and a medical knowledge base to realize the full automation of the "demand input - parameter extraction - dynamic generation" process of surgical / instrument 3D animation, and supports personalized animation customization).

[0011] The generation process of the 3D animation generation module includes:

[0012] It receives user input for the type of surgery (e.g., "laparoscopic cholecystectomy"), instrument model (e.g., "30° laparoscope"), and anatomical features (e.g., "gallbladder position variation").

[0013] Retrieve standard operating procedures (SOPs) and anatomical parameters (such as gallbladder size and blood vessel course) from the medical knowledge base;

[0014] Dynamic 3D animations containing surgical steps (exposure → separation → resection → hemostasis) are generated using AI generative models (such as 3D-GAN).

[0015] The output format is an interactive GLB / USDZ file, which supports rotation and scaling operations.

[0016] Medical knowledge verification module: Based on medical knowledge graph and expert annotation database, it performs structural verification (such as organ location and instrument size) and procedural verification (such as surgical step sequence) on AI-generated 3D animation, outputs verification report and drives animation optimization (through structural matching, procedural verification and expert collaboration mechanism, it ensures that AI-generated animation conforms to anatomical standards and clinical guidelines, and solves the pain point of traditional AI-generated content being "usable but unreliable").

[0017] The verification logic of the medical knowledge verification module includes:

[0018] Structural verification: The organs / instruments in the generated animation are compared with the standard structures in medical atlases using a geometric matching algorithm (e.g., the long diameter of the gallbladder should be ≤8cm). Correction is triggered when the error exceeds 5%.

[0019] Process validation: Based on the time series analysis model, check whether the order of surgical steps in the animation conforms to clinical guidelines (e.g., "dissection of the Calot's triangle" should precede "transection of the cystic duct").

[0020] Expert Collaboration: Supports pushing unverified animations to expert terminals, receiving correction instructions from manual annotations (such as "adjust the angle of the electrocoagulation hook"), and driving iterative optimization of the AI ​​model.

[0021] Vectorization module: Integrates large deep learning models (such as DeepSeek) and computer vision models (such as CNN-LSTM, U-Net) to perform vectorization processing on text content, medical images, surgical videos, and AI-generated 3D animations. It supports automatic translation of multilingual documents and cross-language vectorization mapping (this module designs an "anatomy-operation" composite vectorization strategy for AI-generated 3D animations, extracting structural parameters "such as screw length" and motion trajectories "such as stapler firing angle" from the animation to generate high-dimensional features that can be matched with clinical case and surgical video vectors).

[0022] The retrieval module, based on vector matching algorithms (such as cosine similarity), receives multimodal input from users (text / image / video clips / animation requirements descriptions), calls the vector library generated by the vectorization processing module for matching, and outputs a set of candidate documents (including AI-generated 3D animations).

[0023] RFID positioning module: Communicates with the RFID tag bound to the document acquisition module to obtain the physical location information of the paper document (accuracy down to the centimeter level) and feeds the location data back to the retrieval module;

[0024] Encryption access control module: It uses asymmetric encryption (such as RSA) and national cryptographic algorithms (such as SM4) to encrypt sensitive data. It sets access permissions (public / internal / confidential) based on user roles (researchers / medical staff / medical students), and supports dynamic generation of decryption keys and matching of encryption state vectors.

[0025] According to another aspect of the present invention, more specifically, a vectorized retrieval method for medical literature based on AI and RFID technology, using the aforementioned vectorized retrieval device for medical literature based on AI and RFID technology, includes the following steps:

[0026] S1. Document Tagging and Acquisition (implemented through the document acquisition module)

[0027] Complete the following operations using the document acquisition module:

[0028] High-frequency RFID tags (operating frequency 13.56MHz) are affixed to paper documents. The tags store a unique document ID and basic metadata (including category, shelving time, and access level).

[0029] Automatically capture and extract metadata (including case number, examination type, and animation generation parameters) from electronic documents (including medical images from PACS system, live surgical videos, and AI-generated 3D animation files).

[0030] For foreign language materials, the OCR component built into the document acquisition module is used to recognize the text content and simultaneously input the language type metadata (such as English and Japanese).

[0031] The physical tags and electronic metadata of all materials are bound together in the document acquisition module and stored in the database.

[0032] S2. AI-generated 3D animation and knowledge verification (achieved through collaboration between the 3D animation generation module and the medical knowledge verification module).

[0033] When a user requests 3D animation generation:

[0034] User input: Users can input the type of surgery (e.g., "knee replacement surgery"), the model of the device (e.g., "posterior cruciate ligament-preserving prosthesis"), and special requirements (e.g., "show the bone cement filling process") through the search module.

[0035] Model generation: The 3D animation generation module calls the medical knowledge base to extract standard anatomical parameters (such as femoral condyle size) and operation procedures (such as "osteotomy → prosthesis installation → range of motion testing"), and outputs the initial 3D animation through AI-generated model;

[0036] Knowledge Verification: The medical knowledge verification module performs the following operations:

[0037] Structural verification: Compare the prosthesis dimensions in the animation with the patient's imaging measurements (error must be ≤2mm);

[0038] Process verification: Check whether "bone surface grinding" in the animation occurs before "prosthetic fitting";

[0039] If the verification passes, generate animation metadata (such as "Verification passed" and "Generation time"); if it fails, return correction suggestions (such as "Adjust the prosthesis rotation angle") and drive the model to be regenerated.

[0040] Storage and Archiving: The document acquisition module assigns RFID tags (for physical archiving) or electronic metadata (for digital storage) to the verified 3D animations and stores them in the database. (This step adds a "generation-verification" step before document acquisition to ensure that the 3D animations stored in the database are both innovative and accurate.)

[0041] S3. Multimodal vectorization processing (implemented through the vectorization processing module)

[0042] The vectorization processing module performs the following sub-steps:

[0043] Text vectorization: Call a large deep learning model (such as DeepSeek) to perform word segmentation and semantic encoding on Chinese text or translated foreign text, and generate high-dimensional semantic vectors (dimension ≥ 768) containing contextual relationships.

[0044] Non-text vectorization:

[0045] Medical imaging (CT / MRI): The lesion area is segmented using a U-Net network, and anatomical feature vectors (such as lesion size and location) are extracted using a CNN.

[0046] Surgical video: Keyframe features (such as surgical instruments and tissue morphology) are extracted using CNN, and temporal operation features (such as suturing sequence) are captured by LSTM model to generate a temporal vector containing surgical steps;

[0047] AI generates 3D animations: extracting features such as motion trajectories (e.g., prosthesis installation angle) and structural parameters (e.g., screw length) from the animations, mapping them to standard operating procedure vectors in the medical knowledge base, and generating composite vectors containing both "anatomy-operation" semantics.

[0048] Cross-language vectorization: Foreign language materials are translated into Chinese using the built-in medical NMT translation model (such as MedicalmBERT) in the vectorization processing module, and then semantic vectors in the same vector space as Chinese documents are generated.

[0049] S4. Vector retrieval and positioning (achieved through collaboration between the retrieval module and the RFID positioning module)

[0050] The retrieval module receives multimodal queries (text / image / video clip / animation requirement description) from the user and performs the following operations:

[0051] Input the query content into the vectorization processing module to generate a query vector;

[0052] Use vector matching algorithms (such as cosine similarity) to calculate the similarity between the query vector and the document vector in the vector library, and filter candidate documents (including AI-generated 3D animations) with a threshold (such as 0.85) or higher.

[0053] For paper documents among the candidate documents, the physical location information (such as "5th compartment on the 2nd floor of shelf B on the 3rd floor") is obtained by calling the RFID positioning module through the retrieval module.

[0054] For electronic documents (including AI-generated 3D animations), the storage path is directly linked to generate preliminary search results containing links to electronic documents, interactive addresses for 3D animations, and locations of printed documents.

[0055] S5. Access Control and Result Output (implemented through the encrypted access control module)

[0056] The encryption access control module processes the initial search results as follows:

[0057] Verify user roles (such as resident physician / chief physician / medical student) and permission levels (public / internal / confidential).

[0058] Literature on permission matching:

[0059] Publicly available materials (such as published surgical videos): Decrypted via AES-128, returning the full text of the electronic document, image preview, or 3D animation playback link (supports web-based interaction).

[0060] Internal-level data (such as AI-generated instructional animations): Decrypted using the SM4 algorithm combined with the user's dynamic key, returning the animation interaction address (supporting annotation of key steps);

[0061] Confidential information (such as unpublished innovative technique animations): Homomorphic encryption technology is used to complete vector matching in the encrypted state, only returning a "related literature exists" prompt, without revealing the specific content;

[0062] For documents with mismatched permissions, return an "Access denied" message and log the access information.

[0063] Finally, the system outputs the user with filtered search results (including links to electronic documents, directions to print documents, 3D animation interactive addresses, and content previews).

[0064] The beneficial effects of the medical literature vectorization retrieval device and method based on AI and RFID technology of this invention are as follows:

[0065] (1) This invention breaks through the traditional “passive query” mode and actively generates customized 3D surgical animations through the AI ​​generation module, covering dynamic demonstration content, supporting interactive 3D file output, and realizing the expansion of literature content boundaries.

[0066] (2) This invention uses a three-level verification mechanism of structural verification, process verification and expert collaboration to ensure that the AI-generated animation conforms to anatomical standards and clinical guidelines, and to ensure the accuracy of the generated content.

[0067] Multimodal vectorization processing enables cross-modal semantic alignment, and combined with RFID positioning technology, it significantly improves retrieval efficiency and the speed of locating paper documents.

[0068] (3) The present invention sets three levels of permissions based on user roles, and adopts national cryptographic algorithms and homomorphic encryption technology to ensure the security of sensitive data, meet the compliance requirements of medical data, and strengthen data security compliance. Attached Figure Description

[0069] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0070] Figure 1 This is a schematic diagram of the structure of the vectorized retrieval device for medical literature based on AI and RFID technology of the present invention;

[0071] Figure 2This is a schematic diagram of the structure of the vectorized retrieval method for medical literature based on AI and RFID technology of the present invention. Detailed Implementation

[0072] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0073] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0074] Reference Figures 1-2 A device and method for vectorized retrieval of medical literature based on AI and RFID technologies, including a literature acquisition module, a 3D animation generation module, a medical knowledge verification module, a vectorization processing module, a retrieval module, an RFID positioning module, and an encrypted access control module.

[0075] Example 1

[0076] Taking the scenario of "medical students learning laparoscopic cholecystectomy" as an example, the implementation process of a vectorized retrieval device and method for medical literature based on AI and RFID technology is as follows:

[0077] S1. Document Tagging and Acquisition (Document Acquisition Module)

[0078] The literature acquisition module is used to acquire multimodal medical literature; it is also used to physically mark data using RFID tags; and it is also used to synchronously input metadata.

[0079] Paper documents: "Laparoscopic Cholecystectomy" is affixed with an RFID tag (ID: BOOK-001) and the document ID and metadata are stored. The metadata is entered as "Surgical Guideline", "General Surgery", and "Public Level".

[0080] Electronic document: Real surgical video (name: LC_202405) was automatically captured and stored, and metadata was extracted. The metadata was entered as "Laparoscopic cholecystectomy", "General Surgery", and "Internal level".

[0081] Foreign language materials: OCR recognizes text and inputs language type metadata;

[0082] Data binding: Physical tags are bound to electronic metadata and entered into the database.

[0083] S2, AI-generated 3D animation and knowledge verification (3D animation generation module + medical knowledge verification module)

[0084] The 3D animation generation module integrates AI generation models and medical knowledge bases, generates 3D animations according to user needs, and outputs standardized animation files.

[0085] AI-generated animation: Based on the medical student's requirement to "show a 3D animation of the gallbladder triangle separation steps", the generation process is as follows: input requirements → extract parameters → generate animation → output file.

[0086] Input Requirements: Enter the type of surgery, instrument model, and special requirements. For example, a medical student can enter "laparoscopic cholecystectomy" and "highlight gallbladder triangle dissection" into the search module.

[0087] Model generation: The 3D animation generation module calls the medical knowledge base to extract the anatomical parameters of the Calot's triangle (the length of the gallbladder duct is 2-4cm and the diameter of the common bile duct is 6-8mm) and the operation process ("electric hook separation of the serosa → exposure of Calot's triangle → confirmation of the relationship between the three ducts"), and generates the initial animation through the AI ​​model 3D-GAN.

[0088] The medical knowledge verification module performs structural and process verification based on knowledge graphs and standard databases; it also supports expert collaborative correction. The verification logic is: structural comparison → process check → expert correction. After successful verification, data is stored and archived: tags / metadata are assigned and stored in the database.

[0089] Knowledge verification:

[0090] Structural verification: The length of the cystic duct in the animation (3cm) was checked to match the standard range (2-4cm), with an error of 0mm;

[0091] Procedure validation: It was confirmed that "separating the serous membrane" in the animation precedes "exposing Calot's triangle," which is in line with clinical guidelines;

[0092] Once verification is successful, metadata ("Verification Successful", "Teaching Level") will be generated.

[0093] Storage and archiving: Animation files (name: LC_Calot_3D) are stored in the database through the literature acquisition module and associated with metadata ("3D animation", "general surgery", "teaching grade").

[0094] S3. Multimodal vectorization processing (vectorization processing module)

[0095] The vectorization module integrates deep learning and computer vision models to process multimodal data vectorization, supporting multilingual translation and vectorization mapping. Processed content includes, but is not limited to: text, images, videos, animations, and foreign languages.

[0096] Medical image vectorization: U-Net segmentation + CNN feature vector extraction;

[0097] 3D animation vectorization: Extract the motion trajectory of gallbladder triangle separation (electric hook angle 45°-60°), structural parameters (gallbladder duct diameter 3mm) and other features to generate a composite vector V_3D containing "anatomical structure-operation action";

[0098] Cross-language vectorization: Generating semantic vectors after medical NMT translation;

[0099] Surgical video vectorization: CNN extracts the exposure degree of Calot triangle (90%) and instrument operation angle (50°) in keyframes, and then uses LSTM to capture temporal features to generate temporal vector V_video;

[0100] Text vectorization: Paper guide vectorization: The text "When separating the gallbladder triangle, care should be taken to protect the common bile duct" is semantically encoded using a deep learning large model to generate the vector V_text.

[0101] S4. Vector Retrieval and Positioning (Retrieval Module + RFID Positioning Module)

[0102] The retrieval module is based on a vector matching algorithm. It receives multimodal input, calls the vector library for matching, and outputs a set of candidate documents.

[0103] The RFID positioning module communicates with the RFID tag to obtain the physical location of the paper document and feeds back the location data to the retrieval module.

[0104] Multimodal query: Input text, images, video clips, or animation descriptions. For example, a medical student might input: "3D animation of gallbladder triangle separation during laparoscopic cholecystectomy".

[0105] Generate query vector: The vectorization processing module generates a query vector V_query (containing the keywords "gallbladder triangle", "separation", and "3D animation").

[0106] Vector matching: Calculate the cosine similarity between V_query and V_3D (0.95 > 0.85 threshold), filter candidate documents, and match AI-generated 3D animations;

[0107] Location processing: RFID locates paper documents and associates them with the storage path of electronic documents. For example, it associates the storage path of animations (server address: / 3D_animations / LC_Calot_3D.glb); if a paper guide with the same name (ID: BOOK-001) exists, the RFID positioning module returns its location ("2nd floor, A section, shelf 3, 2nd compartment").

[0108] S5. Access Control and Result Output (Encrypted Access Management Module)

[0109] The encryption access control module employs asymmetric encryption and Chinese national cryptographic algorithms, sets permissions based on user roles, and supports dynamic key and encryption state matching. Access levels include: Public (AES-128), Internal (SM4 dynamic key), and Confidential (homomorphic encryption).

[0110] The medical student's role is "Clinical Intern" (permission level ≥ "Teaching Level");

[0111] The encryption access management module decrypts 3D animation files using AES-128 and returns the web-based interactive address (supports rotating to view the gallbladder triangle structure).

[0112] The output includes: links to electronic documents, locations of printed documents, and addresses for 3D animation interactions. If permissions do not match, a "No access permission" message will be returned and logged.

[0113] 3D animation interactive links (key steps such as "separation of the film layer" can be annotated);

[0114] Paper-based location guides (e.g., "2nd floor, A section, shelf 3, 2nd shelf");

[0115] Related surgical video clips (showing the Calot triangle separation process in a real surgery).

[0116] This embodiment expands the content boundaries of medical literature by generating 3D animations using AI, ensures the accuracy of the generated content by combining it with a medical knowledge verification mechanism, and achieves full-scenario coverage of "passive query + active generation" through composite vectorized retrieval, providing a more efficient and reliable literature retrieval solution for medical education and clinical research.

[0117] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A vectorized retrieval device for medical literature based on AI and RFID technologies, characterized in that, include: Document Acquisition Module: Used to acquire multimodal medical literature, physically tag all documents with RFID tags, and synchronously input metadata; 3D animation generation module: Integrates AI generation model and medical knowledge base, supports the generation of 3D animations of surgical procedures or medical devices according to user needs, and outputs standardized animation files containing anatomical structures and operation procedures; Medical knowledge verification module: Based on medical knowledge graph and expert annotation database, it performs structural and process verification on AI-generated 3D animation, outputs verification report and drives animation optimization; Vectorization processing module: Integrates deep learning large models and computer vision models to perform vectorization processing on text content, medical images, surgical videos and AI-generated 3D animations, and supports automatic translation of multilingual documents and cross-language vectorization mapping; The retrieval module, based on a vector matching algorithm, receives multimodal input from users, calls the vector library generated by the vectorization processing module for matching, and outputs a set of candidate documents. RFID positioning module: Communicates with the RFID tag bound to the document acquisition module to obtain the physical location information of the paper document and feeds the location data back to the retrieval module; Encryption access control module: It uses asymmetric encryption and national cryptographic algorithms to encrypt sensitive data, sets access permissions based on user roles, and supports dynamic generation of decryption keys and matching of encryption state vectors.

2. The medical literature data vectorization retrieval device based on AI and RFID technology according to claim 1, characterized in that: The generation process of the 3D animation generation module includes: Receives user input regarding the type of surgery, instrument model, and anatomical features; Retrieve standard operating procedures and anatomical parameters from the medical knowledge base; Dynamic 3D animations containing surgical steps are generated using AI-generated models; The output format is an interactive GLB / USDZ file, which supports rotation and scaling operations.

3. The medical literature data vectorization retrieval device based on AI and RFID technology according to claim 2, characterized in that: The verification logic of the medical knowledge verification module includes: Structural verification: The generated organs / instruments in the animation are compared with the standard structures in medical atlases using a geometric matching algorithm. Correction is triggered when the error exceeds 5%. Process validation: Based on a time series analysis model, check whether the sequence of surgical steps in the animation conforms to clinical guidelines; Expert Collaboration: Supports pushing unverified animations to expert terminals, receiving correction instructions from manual annotations, and driving iterative optimization of AI models.

4. A method for vectorized retrieval of medical literature based on AI and RFID technology, using the vectorized retrieval device for medical literature based on AI and RFID technology as described in claim 3, includes the following steps: S1. Document Marking and Collection: Complete the following operations using the document acquisition module: High-frequency RFID tags are affixed to paper documents, and the tags store a unique document ID and basic metadata. Automatically capture and extract metadata from electronic documents; For foreign language materials, the OCR component built into the document acquisition module is used to recognize the text content and simultaneously input the language type metadata. The physical tags and electronic metadata of all materials are bound together in the document acquisition module and stored in the database; S2, AI-generated 3D animation and knowledge verification: When a user requests 3D animation generation: User input: Users input the type of surgery, instrument model, and special requirements through the search module; Model generation: The 3D animation generation module calls the medical knowledge base to extract standard anatomical parameters and operating procedures, and generates an initial 3D animation through AI to generate a model; Knowledge Verification: The medical knowledge verification module performs the following operations: Structural verification: Compare the prosthesis dimensions in the animation with the patient's imaging measurements; Process verification: Check whether "bone surface grinding" in the animation occurs before "prosthetic fitting"; If the verification passes, animation metadata is generated; if it fails, correction suggestions are returned and the model is regenerated. Storage and Archiving: The document acquisition module assigns RFID tags or electronic metadata to verified 3D animations and stores them in the database; S3. Multimodal Vectorization Processing: The vectorization processing module performs the following sub-steps: Text vectorization: A large deep learning model is used to segment and semantically encode Chinese text or translated foreign text, generating high-dimensional semantic vectors that contain contextual relationships; Non-text vectorization: Medical imaging: The lesion region is segmented using a U-Net network, and anatomical feature vectors are extracted using a CNN. Surgical video: Keyframe features are extracted using CNN, and temporal operation features are captured by LSTM model to generate a temporal vector containing surgical steps; AI-generated 3D animation: Extract motion trajectories and structural parameter features from the animation, establish a mapping with standard operating procedure vectors in the medical knowledge base, and generate a composite vector containing the dual semantics of "anatomy-operation"; Cross-language vectorization: Foreign language materials are translated into Chinese using the medical NMT translation model built into the vectorization processing module, and then semantic vectors in the same vector space as Chinese documents are generated; S4. Vector retrieval and positioning: The retrieval module receives multimodal queries input by the user and performs the following operations: Input the query content into the vectorization processing module to generate a query vector; The vector matching algorithm is invoked to calculate the similarity between the query vector and the document vector in the vector database, and candidate documents above the threshold are filtered out. For paper documents among the candidate documents, the physical location information is obtained by calling the RFID positioning module through the retrieval module; For electronic documents, the storage path is directly linked to generate preliminary search results that include links to electronic documents, 3D animation interactive addresses, and the location of printed documents. S5. Access Control and Result Output: The encryption access control module processes the initial search results as follows: Verify user roles and permission levels; Literature on permission matching: Publicly available materials: Decrypted via AES-128, returning the full text of the electronic document, image preview, or 3D animation playback link; Internal-level data: Decrypted using the SM4 algorithm combined with the user's dynamic key, returning the animation interaction address; Confidential documents: Homomorphic encryption is used to complete vector matching in the encrypted state, and only a "Related documents exist" prompt is returned without revealing the specific content; For documents with mismatched permissions, return an "Access denied" message and log the access information. Finally, the system outputs the filtered search results to the user.

Images