A Cross-Modal Retrieval Method and System Based on a Large-Scale Pediatric Oncology Teaching Model

Abstract

This invention provides a cross-modal retrieval method and system based on a large-scale pediatric oncology teaching model, belonging to the field of large-scale model technology. First, it receives a user-initiated cross-modal retrieval request for pediatric oncology teaching, parses and obtains the retrieval trigger content and teaching stage information; it inputs the large-scale pediatric oncology teaching model to generate knowledge seed descriptions; it calls a multimodal resource library to extract initial resources, and generates a knowledge semantic evolution tree through large-scale model semantic evolution processing; it traverses the knowledge semantic evolution tree to construct cross-modal dynamic association chains; it filters cross-modal dynamic association chains based on teaching stage information, integrates resources to generate retrieval results and pushes them to the user, while simultaneously updating the knowledge association index of the multimodal resource library. This invention can accurately match the needs of different teaching stages, achieve effective integration and dynamic updating of cross-modal resources, and improve the efficiency and quality of pediatric oncology teaching.

Description

Technical Field

[0001] This invention relates to the field of large model technology, and more specifically, to a cross-modal retrieval method and system based on a large model for pediatric oncology teaching. Background Technology

[0002] In the field of pediatric oncology education, with the continuous accumulation of medical knowledge and the increasing diversification of teaching needs, traditional teaching methods and resource retrieval methods are facing many challenges.

[0003] On the one hand, the knowledge system of pediatric oncology is complex and highly specialized, covering multiple aspects such as basic pathology, diagnostic methods, treatment plans, and prognosis, with close connections and hierarchical relationships between different knowledge modules. However, existing teaching resource retrieval methods often rely solely on simple keyword matching, failing to deeply understand user search intent and accurately capture the specific needs of individual knowledge topics at different teaching stages within the field of pediatric oncology education. For example, when explaining early diagnostic methods for pediatric tumors, the required teaching resources should differ in depth and breadth for beginners and learners with some prior knowledge, but traditional retrieval methods cannot provide targeted resources based on the teaching stage information.

[0004] On the other hand, pediatric oncology teaching resources are multimodal, including textual materials, medical images, surgical videos, and audio explanations. Most existing retrieval systems can only search for resources of a single modality, failing to effectively link and integrate cross-modal resources. This forces teachers and learners to spend considerable time and effort switching and searching between different modalities, reducing teaching efficiency and learning outcomes. Furthermore, existing resource retrieval methods lack in-depth exploration of the semantic evolution of knowledge, failing to present the dynamic connections and developmental contexts between knowledge points, which is detrimental to learners building a complete knowledge system. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a cross-modal retrieval method based on a large-scale pediatric oncology teaching model, the method comprising:

[0006] The system receives a cross-modal retrieval request for pediatric oncology teaching initiated by a user, parses the request, and obtains the retrieval trigger content and teaching stage information. The retrieval trigger content is a single knowledge topic expression in the field of pediatric oncology teaching, and the teaching stage information includes the current knowledge level and teaching objective orientation of the teaching.

[0007] The retrieval trigger content and the teaching stage information are input into the pediatric oncology teaching model. The pediatric oncology teaching model generates a knowledge seed description corresponding to the retrieval trigger content. The knowledge seed description includes core knowledge modules and knowledge extension directions.

[0008] Based on the knowledge seed description, the pediatric oncology teaching multimodal resource library is invoked, and the initial multimodal resources associated with the core knowledge modules in the pediatric oncology teaching multimodal resource library are extracted. The semantic evolution processing of the initial multimodal resources is performed through the pediatric oncology teaching big model to generate a knowledge semantic evolution tree. The knowledge semantic evolution tree has the knowledge seed as the root node, the extended knowledge modules as child nodes, and semantic associations as branches.

[0009] Traverse the knowledge semantic evolution tree, extract the multimodal resources corresponding to each branch, construct a cross-modal dynamic association chain, the cross-modal dynamic association chain connects different modal resources in the order of knowledge evolution, and each resource node is labeled with the association logic with the knowledge seed;

[0010] Based on the teaching stage information, the association chain screening criteria are determined, cross-modal dynamic association chains that meet the association chain screening criteria are extracted, multimodal resources in the cross-modal dynamic association chains are integrated to generate cross-modal retrieval results, which are pushed to the user interaction terminal. At the same time, the knowledge semantic evolution tree generated this time and the cross-modal dynamic association chains are integrated into the knowledge association index of the pediatric oncology teaching multimodal resource library to update the basic knowledge evolution data for subsequent retrieval.

[0011] Furthermore, embodiments of the present invention also provide a cross-modal retrieval system based on a large-scale pediatric oncology teaching model, characterized in that it includes:

[0012] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned cross-modal retrieval method based on a large-scale pediatric oncology teaching model by executing the machine-executable instructions.

[0013] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, the processor of the cross-modal retrieval system based on the pediatric oncology teaching model reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the cross-modal retrieval system based on the pediatric oncology teaching model to execute the aforementioned cross-modal retrieval method based on the pediatric oncology teaching model.

[0014] Based on the above, by receiving and parsing user-initiated cross-modal retrieval requests for pediatric oncology teaching, the system obtains the retrieval trigger content and teaching stage information. This allows the system to understand the specific needs of users for a single knowledge topic at different teaching stages. The retrieval trigger content and teaching stage information are then input into a large-scale pediatric oncology teaching model to generate knowledge seed descriptions. These seed descriptions contain core knowledge modules and knowledge extension directions, making the retrieval process more targeted and systematic. Based on the knowledge seed descriptions, the system calls upon a multimodal resource library and extracts initial multimodal resources. Then, through semantic evolution processing performed by the large-scale model, a knowledge semantic evolution tree is generated. This tree-like structure presents the dynamic connections and development paths between knowledge points, helping learners build a complete knowledge system. Traversing the knowledge semantic evolution tree constructs cross-modal dynamic association chains, connecting different modal resources according to the knowledge evolution order and annotating the association logic. This achieves effective integration and association of cross-modal resources. Based on teaching stage information, association chain selection criteria are determined, and association chains that meet the criteria are extracted. Multimodal resources are integrated to generate retrieval results, which are then pushed to the user's interactive terminal. This meets the needs of different teaching stages, improving teaching efficiency and learning outcomes. Meanwhile, the generated knowledge semantic evolution tree and cross-modal dynamic association chain are integrated into the knowledge association index of the multimodal resource library to update the basic data of knowledge evolution for subsequent searches, thereby realizing dynamic updating and optimization of resources and improving the effectiveness of retrieval of pediatric oncology teaching resources. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the execution flow of the cross-modal retrieval method based on a large-scale pediatric tumor teaching model provided in an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of exemplary hardware and software components of a cross-modal retrieval system based on a large-scale pediatric oncology teaching model provided in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a cross-modal retrieval method based on a large-scale pediatric oncology teaching model provided in one embodiment of the present invention. The cross-modal retrieval method based on the large-scale pediatric oncology teaching model will be described in detail below.

[0018] Step S110: Receive a cross-modal retrieval request for pediatric oncology teaching initiated by a user, parse the cross-modal retrieval request for pediatric oncology teaching, and obtain the retrieval trigger content and teaching stage information. The retrieval trigger content is a single knowledge topic description in the field of pediatric oncology teaching, and the teaching stage information includes the knowledge level and teaching goal orientation of the current teaching.

[0019] In this embodiment, a user's search for knowledge related to "chemotherapy regimens for acute lymphoblastic leukemia in children" within a pediatric oncology teaching scenario is used as an example. The user submits a search request through a pre-defined interactive terminal interface. This search request includes search trigger content and teaching stage information. First, the received search request is parsed, and the search trigger content is extracted from the request data. This search trigger content is a single knowledge topic statement in the field of pediatric oncology teaching: "chemotherapy regimens for acute lymphoblastic leukemia in children." Simultaneously, teaching stage information is extracted from the search request. This teaching stage information includes the current knowledge level and teaching objective orientation, for example, the current knowledge level is "clinical application level," and the teaching objective orientation is "skill mastery orientation."

[0020] Step S120: Input the retrieval trigger content and the teaching stage information into the pediatric oncology teaching model, and generate a knowledge seed description corresponding to the retrieval trigger content through the pediatric oncology teaching model. The knowledge seed description includes core knowledge modules and knowledge extension directions.

[0021] After obtaining the search trigger content "chemotherapy regimen for childhood acute lymphoblastic leukemia" and the teaching stage information (knowledge level is clinical application level, teaching goal orientation is skill mastery orientation), the above two are input into the large-scale teaching model for pediatric oncology to generate corresponding knowledge seed descriptions.

[0022] Step S121: The large-scale pediatric oncology teaching model calls upon the pediatric oncology teaching knowledge system. The pediatric oncology teaching knowledge system is divided into knowledge levels, including a basic cognition layer, a principle analysis layer, a clinical application layer, and a comprehensive extension layer. Each knowledge level contains multiple knowledge modules, and each knowledge module contains preset core knowledge points and knowledge extension paths.

[0023] Upon receiving the retrieved content and teaching stage information, the comprehensive pediatric oncology teaching model first invokes its built-in pediatric oncology teaching knowledge system. This system is hierarchically structured, comprising a basic cognitive layer, a principle analysis layer, a clinical application layer, and a comprehensive extension layer. Each knowledge level contains multiple knowledge modules, each with pre-defined core knowledge points and extension pathways. For example, the clinical application layer includes modules such as "chemotherapy regimens for childhood acute lymphoblastic leukemia" and "surgical treatment of childhood neuroblastoma." Each module, such as the "chemotherapy regimens for childhood acute lymphoblastic leukemia" module, might have core knowledge points including types of chemotherapy drugs, routes of administration, and dosage adjustment principles, while extension pathways might involve managing adverse reactions to chemotherapy and evaluating its effectiveness.

[0024] Step S122: Analyze the teaching stage information, determine the knowledge level corresponding to the current teaching, and extract the knowledge depth requirements and knowledge extension scope under the knowledge level. The knowledge depth requirements include the level of detail of the knowledge points and the coverage of related knowledge. The knowledge extension scope includes vertical in-depth and horizontal association. Vertical in-depth refers to the in-depth expansion of the same module, and horizontal association refers to the related expansion of different modules.

[0025] Next, the input teaching stage information is analyzed. Based on the knowledge level identifiers contained in the teaching stage information, the current teaching knowledge level is determined to be the clinical application layer. Then, the knowledge depth requirements and knowledge extension scope corresponding to the clinical application layer are extracted from the pediatric oncology teaching knowledge system. Regarding knowledge depth requirements, for the topic of "chemotherapy regimens for childhood acute lymphoblastic leukemia," under the clinical application layer, the level of detail required for knowledge points should reach the implementation details of specific chemotherapy regimens, such as the combination of chemotherapy drugs corresponding to different risk groups, the order of administration, and the course of treatment; the coverage of related knowledge needs to include pre-chemotherapy assessment, monitoring indicators during chemotherapy, and follow-up plans after chemotherapy. Regarding the knowledge extension scope, vertical in-depth refers to in-depth expansion within the same module of "chemotherapy regimens for childhood acute lymphoblastic leukemia," such as delving into the mechanism of action of chemotherapy drugs and the mechanism of drug resistance; horizontal association refers to expansion through association with other related knowledge modules, such as association with knowledge modules such as "pathological diagnosis of childhood acute lymphoblastic leukemia" and "management of adverse reactions to chemotherapy in pediatric tumors."

[0026] Step S123: Divide the search trigger content into multiple search semantic units. Each search semantic unit is a text fragment that expresses a single knowledge point. Each search semantic unit contains professional terms and expression logic.

[0027] Subsequently, the search trigger content "chemotherapy regimen for childhood acute lymphoblastic leukemia" was broken down into multiple search semantic units. Each search semantic unit is a text fragment expressing a single knowledge point and contains professional terminology and expression logic. For example, the search trigger content can be split into two search semantic units: "childhood acute lymphoblastic leukemia" and "chemotherapy regimen." The search semantic unit "childhood acute lymphoblastic leukemia" contains the professional term "childhood acute lymphoblastic leukemia" and its expression logic points to a specific disease; the search semantic unit "chemotherapy regimen" contains the professional term "chemotherapy regimen" and its expression logic points to a specific regimen in the treatment of this disease.

[0028] Step S124: Traverse the knowledge modules in the pediatric oncology teaching knowledge system, perform semantic association analysis between each retrieval semantic unit and the core knowledge points of the knowledge module, and determine the knowledge module to which each retrieval semantic unit belongs.

[0029] After breaking down the search trigger content into multiple search semantic units, it is necessary to perform affiliation analysis on each search semantic unit to determine its corresponding knowledge module. In this embodiment, the search trigger content "chemotherapy regimen for childhood acute lymphoblastic leukemia" is broken down into two search semantic units: "childhood acute lymphoblastic leukemia" and "chemotherapy regimen." The knowledge modules within the pediatric oncology teaching knowledge system are traversed, and semantic association analysis is performed on each search semantic unit with the core knowledge points of the knowledge module to determine the knowledge module to which each search semantic unit belongs.

[0030] Step S1241: Read the first knowledge level from the pediatric oncology teaching knowledge system, extract all knowledge modules under this knowledge level, and form a knowledge module list.

[0031] The first knowledge level, the basic cognitive level, is retrieved from the pediatric oncology teaching knowledge system. All knowledge modules under this level are extracted to form a knowledge module list. For example, the basic cognitive level includes modules such as "Basic Concepts of Pediatric Oncology," "Overview of Childhood Acute Lymphoblastic Leukemia," and "Classification of Pediatric Oncology." These modules are then arranged according to their storage order within the knowledge system to form the knowledge module list for the current knowledge level.

[0032] Step S1242: Take the first retrieval semantic unit, extract vocabulary from the retrieval semantic unit to obtain a retrieval vocabulary set. The retrieval vocabulary set includes professional terms, logical expressions, and knowledge attribute words in the field of pediatric oncology teaching. The knowledge attribute words include etiology, pathology, and treatment.

[0033] The first retrieval semantic unit, "acute lymphoblastic leukemia in children," is selected, and lexical extraction is performed on this unit. First, word segmentation is performed, resulting in the lexical sequence "children" and "acute lymphoblastic leukemia." From this sequence, specialized terms in the field of pediatric oncology are identified and extracted, with "acute lymphoblastic leukemia" being a specific term. Logical expression words are identified, but no obvious logical expression words are found in this retrieval semantic unit. Knowledge attribute words are identified; for example, "acute lymphoblastic leukemia" implicitly contains disease classification attributes, but these do not fully correspond to specific knowledge attribute words such as etiology, pathology, and treatment, so these specific knowledge attribute words are not extracted at this stage. The extracted specialized term "acute lymphoblastic leukemia" is then used as an element in the retrieval vocabulary set.

[0034] Step S1243: Read the first knowledge module in the knowledge module list, extract the core knowledge points of the knowledge module, and form a core knowledge point list. Each core knowledge point contains standard expression terms and knowledge attribute tags.

[0035] Read the first knowledge module in the basic cognitive layer knowledge module list, "Basic Concepts of Childhood Tumors." Extract its core knowledge points from this module to form a core knowledge point list. For example, core knowledge points include "Definition of Tumor," whose standard expression is "A tumor is a new growth formed by the abnormal proliferation of body cells," with the knowledge attribute tag "Concept"; "Classification of Tumor," whose standard expression is "Tumors are divided into benign tumors and malignant tumors," with the knowledge attribute tag "Classification"; and "Characteristics of Childhood Tumors," whose standard expression is "Childhood tumors mostly originate from embryonic remnants," with the knowledge attribute tag "Characteristics." List these core knowledge points, their standard expressions, and their knowledge attribute tags.

[0036] Step S1244: Perform an association analysis between the retrieval vocabulary set and the standard expression vocabulary of each core knowledge point, and record the number and type of overlapping words, including overlapping professional terms and overlapping knowledge attribute words.

[0037] Perform an association analysis between the technical term "acute lymphoblastic leukemia" in the search terminology set and the standard expression of the core knowledge point "definition of tumor" ("a tumor is a new growth formed by the abnormal proliferation of body cells"). Compare the two texts by word choice to find overlapping terms. One overlapping term, "tumor," is found. Analyzing the type of overlapping term, "tumor" is a technical term, so it is recorded as "technical term overlap." Further association analysis is performed on the core knowledge point "tumor classification" (standard expression "tumors are divided into benign and malignant tumors"). The overlapping term is still "tumor," with one instance, also classified as technical term overlap. Finally, association analysis is performed on the core knowledge point "characteristics of childhood tumors" (standard expression "childhood tumors mostly originate from embryonic remnants"). The overlapping terms are "child" and "tumor," with two instances. "Child" is a descriptive term, while "tumor" is a technical term. Therefore, the overlapping term types are recorded as "technical term overlap" and "other term overlap."

[0038] Step S1245: Analyze the fit between the expression logic of the retrieved semantic unit and the standard logic of the core knowledge point, extract the logical connectives between the two, including "because", "therefore", and "step", and record whether the semantic direction of the logical connectives is consistent.

[0039] Analyzing the semantic unit "childhood acute lymphoblastic leukemia" reveals that this semantic unit is a noun phrase without explicit logical connectors, directly pointing to the disease name. The core knowledge point "tumor definition" follows a standard definitional description without logical connectors; "tumor classification" follows a standard classification description without logical connectors; and "characteristics of childhood tumors" follows a standard characteristic description without logical connectors. Therefore, for the current semantic unit and the aforementioned core knowledge points, no logical connectors were extracted, and the semantic direction of these logical connectors is recorded as none.

[0040] Step S1246: Based on the number, type, and logical fit of overlapping words, determine the degree of association between the retrieval semantic unit and the current core knowledge point. If at least one core knowledge point has a degree of association with the retrieval semantic unit that is greater than the set degree, then the retrieval semantic unit is assigned to the current knowledge module. Otherwise, read the next knowledge module in the knowledge module list and repeat the above association analysis steps.

[0041] Based on the quantity, type, and logical fit of overlapping words, the degree of relevance between the retrieved semantic unit and the current core knowledge point is calculated. The formula for calculating the degree of relevance is: the degree of relevance equals the number of overlapping professional terms multiplied by the weight coefficient of the professional terms, plus the number of overlapping other words multiplied by the weight coefficient of other words, plus the logical connective word matching score. In this embodiment, the weight coefficient of professional terms is set to 0.6, the weight coefficient of other words is set to 0.2, and the logical connective word matching score is increased by 0.4 if there is a logical connective word and the semantic direction is consistent, otherwise it is increased by 0. For the retrieved semantic unit "acute lymphoblastic leukemia in children" and the core knowledge point "characteristics of childhood tumors," the number of overlapping professional terms is 1 ("tumor"), and the professional term overlap score equals 1 multiplied by 0.6 equals 0.6; the number of overlapping other words is 1 ("children"), and the other word overlap score equals 1 multiplied by 0.2 equals 0.2; the logical connective word matching score is 0; the degree of relevance equals 0.6 plus 0.2 plus 0 equals 0.8. For the core knowledge points "definition of tumor" and "classification of tumor," the number of overlapping professional terms is 1, and the professional term overlap score equals 1 multiplied by 0.6, which equals 0.6; the number of overlapping other words is 0, and the other word overlap score equals 0; the logical conjunction matching score is 0; the relevance value equals 0.6 plus 0 plus 0, which equals 0.6. The relevance threshold is set at 0.7. The relevance value of 0.8 with the core knowledge point "characteristics of childhood tumors" is greater than the threshold of 0.7, while the relevance value of 0.6 with "definition of tumor" and "classification of tumors" is less than the threshold of 0.7. Since at least one core knowledge point ("characteristics of childhood tumors") has a relevance value greater than the set threshold, the retrieval semantic unit should ideally be classified into the current knowledge module "basic concepts of childhood tumors." However, further analysis reveals that although the relevance value reaches the threshold, "children" and "tumor" in the core knowledge point "characteristics of childhood tumors" are broad expressions and do not precisely match the core disease type "acute lymphoblastic leukemia" of the retrieval semantic unit. Therefore, a stricter judgment mechanism needs to be introduced, namely, a requirement for complete matching of core professional terms.

[0042] Step S1247: Based on the correlation value being greater than the set threshold, further check whether there is a complete match of core professional terms. If there is a complete match of core professional terms, confirm the attribution; otherwise, continue traversing.

[0043] The search query for the semantic unit "acute lymphoblastic leukemia in children" checks whether the core term "acute lymphoblastic leukemia" completely matches any part of the standard terminology for the core knowledge point "characteristics of childhood tumors." The standard terminology "childhood tumors mostly originate from embryonic remnants" does not contain the complete term "acute lymphoblastic leukemia," but only the broader concept of "tumor," thus failing to form a complete match for the core term. Therefore, even though the association score of 0.8 is greater than the threshold of 0.7, this semantic unit is not classified into the current knowledge module "basic concepts of childhood tumors." The search continues to the next knowledge module in the knowledge module list, repeating the above association analysis steps.

[0044] Step S1248: Read the next knowledge module in the knowledge module list and continue to perform the above association analysis steps.

[0045] The next knowledge module in the basic cognitive layer knowledge module list, "Overview of Acute Lymphoblastic Leukemia in Children," is retrieved. The core knowledge points of this module are extracted to form a core knowledge point list. These core knowledge points include: "Definition of Acute Lymphoblastic Leukemia," whose standard expression is "Acute lymphoblastic leukemia is the most common hematologic malignancy in children," tagged with "definition"; "Epidemiological Characteristics," whose standard expression is "The incidence of acute lymphoblastic leukemia in children accounts for approximately 75% of all childhood leukemia cases," tagged with "epidemiology"; and "Clinical Manifestations," whose standard expression is "Fever, anemia, bleeding, hepatosplenomegaly, and lymphadenopathy," tagged with "clinical manifestations." A correlation analysis is then performed between the search term set "acute lymphoblastic leukemia" and the standard expressions of the aforementioned core knowledge points.

[0046] For the core knowledge point "definition of acute lymphoblastic leukemia," the overlapping term is "acute lymphoblastic leukemia," with a term overlap count of 1 and a term overlap score of 1 multiplied by 0.6, which equals 0.6. Other term overlap counts are 0, and other term overlap scores are 0. The logical conjunction matching score is 0, and the correlation degree value is 0.6. This correlation degree value of 0.6 is less than the threshold of 0.7. However, checking the complete match of the core term reveals that the complete term "acute lymphoblastic leukemia" perfectly matches the standard expression "acute lymphoblastic leukemia." Therefore, based on the complete match rule for core terminology, the attribution is directly determined without relying on the correlation degree value threshold.

[0047] For the core knowledge point "epidemiological characteristics," the overlapping term is "acute lymphoblastic leukemia," with a term overlap of 1, and a term overlap score of 1 multiplied by 0.6 equals 0.6; other terms have 0 overlaps; the correlation score is 0.6, indicating a complete match of core terms, and the attribution is determined. For the core knowledge point "clinical manifestations," the overlapping term is "acute lymphoblastic leukemia," with a term overlap of 1, and a correlation score of 0.6, indicating a complete match of core terms, and the attribution is determined.

[0048] Based on the above analysis, it was determined that the retrieval semantic unit "Childhood Acute Lymphoblastic Leukemia" has a complete match with several core professional terms in the knowledge module "Overview of Childhood Acute Lymphoblastic Leukemia". Therefore, this retrieval semantic unit is classified into the current knowledge module "Overview of Childhood Acute Lymphoblastic Leukemia".

[0049] Step S1249: If none of the knowledge modules under the current knowledge level can form a belonging association with the search semantic unit, then read the next knowledge level and repeat the above knowledge module traversal and association analysis steps until a corresponding knowledge module is found for each search semantic unit. If there is no matching knowledge module in all knowledge levels, then mark the search semantic unit as a knowledge unit to be supplemented, and subsequently indicate the type of knowledge module to be expanded in the knowledge seed description.

[0050] In this embodiment, the first retrieval semantic unit "Childhood Acute Lymphoblastic Leukemia" has already been found to belong to the knowledge module "Overview of Childhood Acute Lymphoblastic Leukemia" at the basic cognitive layer. Next, the second retrieval semantic unit "Chemotherapy Regimen" is processed. The knowledge module list is traversed starting from the basic cognitive layer, including modules such as "Basic Concepts of Childhood Tumors," "Overview of Childhood Acute Lymphoblastic Leukemia," and "Classification of Childhood Tumors." For the "Basic Concepts of Childhood Tumors" knowledge module, its core knowledge points are extracted, and the degree of association with the retrieval semantic unit "Chemotherapy Regimen" is calculated. The retrieval term set includes the professional term "Chemotherapy Regimen." Association analysis with the core knowledge point "Tumor Definition" shows no overlapping words, 0 overlapping professional terms, 0 overlapping other words, and a correlation value of 0. Association analysis with "Tumor Classification" shows no overlapping words and a correlation value of 0. Association analysis with "Characteristics of Childhood Tumors" shows no overlapping words and a correlation value of 0. No core professional term is completely matched, and the correlation values ​​are all less than the threshold of 0.7, therefore, it cannot be assigned.

[0051] For the knowledge module "Overview of Acute Lymphoblastic Leukemia in Children," the standard vocabulary for its core knowledge points "Definition of Acute Lymphoblastic Leukemia," "Epidemiological Characteristics," and "Clinical Manifestations" does not include any content related to "chemotherapy regimens." The number of overlapping terms is 0, the correlation value is 0, and it cannot be assigned. For the knowledge module "Classification of Childhood Tumors," its core knowledge points, such as the standard vocabulary for "Classification of Leukemia," may involve "acute lymphoblastic leukemia," but still do not include "chemotherapy regimens." The correlation value is 0, and it cannot be assigned. After traversing all knowledge modules at the basic cognitive level, no attribution association can be formed.

[0052] The next knowledge level, the principle analysis layer, is then read. Under this layer, knowledge modules include "Principles of Chemotherapy for Pediatric Tumors," "Chemotherapy Drugs for Pediatric Tumors," and "Design of Treatment Plans for Pediatric Tumors." These modules are traversed, and the semantic unit "chemotherapy regimen" is analyzed for its affiliation. For the "Principles of Chemotherapy for Pediatric Tumors" module, its core knowledge point, such as the standard expression for "mechanism of action of chemotherapy," is "chemotherapy drugs kill tumor cells by interfering with the cell cycle and inhibiting DNA synthesis." This is analyzed for its association with the semantic unit "chemotherapy regimen." The overlapping term is "chemotherapy," with a term overlap count of 1, resulting in a term overlap score of 1 multiplied by 0.6, which equals 0.6. Other terms have an overlap count of 0, and the association score is 0.6. A complete match of core terms is checked. "Chemotherapy regimen" does not completely match "chemotherapy" in "mechanism of action of chemotherapy," and there is no corresponding term "regimen." Therefore, the core terms are not completely matched, and the association score of 0.6 is less than the threshold of 0.7, making a proper affiliation impossible.

[0053] For the knowledge module on "chemotherapy drugs for pediatric tumors", the core knowledge points, such as the standard expression of "commonly used chemotherapy drugs", are "commonly used chemotherapy drugs for pediatric tumors include vincristine, cyclophosphamide, doxorubicin, etc." In the correlation analysis with "chemotherapy regimen", the overlapping word is "chemotherapy". The professional terminology overlap score is 0.6 and the correlation degree value is 0.6. The core professional terms are not completely matched and cannot be assigned.

[0054] For the knowledge module "Design of Pediatric Cancer Treatment Plans," its core knowledge points include "Principles of Chemotherapy Regimen Composition," with the standard terminology being "Chemotherapy regimens typically consist of stages such as induction remission, consolidation and intensification, and maintenance therapy"; "Introduction to Standard Chemotherapy Regimens," with the standard terminology being "Commonly used regimens for childhood acute lymphoblastic leukemia include VDLP and COG"; and "Basis for Adjusting Chemotherapy Regimens," with the standard terminology being "Adjusting chemotherapy regimens based on the child's risk grouping and treatment response." A correlation analysis with the core knowledge point "Principles of Chemotherapy Regimen Composition" revealed that the overlapping term was "chemotherapy regimen," with a terminology overlap count of 1 and a terminology overlap score of 0.6; other terms had 0 overlaps; and the correlation strength value was 0.6. A check for complete matching of core professional terms showed that "chemotherapy regimen" completely matched the standard terminology "chemotherapy regimen," therefore, according to the rule for complete matching of core professional terms, the attribution was determined. Association analysis with the core knowledge point "Introduction to Standard Chemotherapy Regimens" revealed the overlapping term "regimen," with a professional terminology overlap score of 0.6. However, "regimen" and "chemotherapy regimen" did not perfectly match. Furthermore, the standard terminology included specific regimen names, such as the VDLP regimen, where "regimen" had a semantic connection with the retrieval semantic unit "chemotherapy regimen." However, a complete match of the core professional terminology requires a complete match of "chemotherapy regimen," which was not achieved here. The association score remained at 0.6, less than the threshold of 0.7. However, since other core knowledge points had already achieved a complete match, this did not affect the attribution determination. Association analysis with the core knowledge point "Basis for Adjusting Chemotherapy Regimens" also revealed the overlapping term "chemotherapy regimen," forming a complete match of the core professional terminology, and the attribution was similarly determined. Therefore, the retrieval semantic unit "chemotherapy regimen" was assigned to the "Pediatric Oncology Treatment Plan Design" knowledge module in the principle analysis layer.

[0055] At this point, both retrieval semantic units have found their corresponding knowledge modules, and there is no need to mark them as knowledge units to be supplemented. If, after traversing all knowledge levels, there are still retrieval semantic units that cannot find their belonging knowledge module, then they are marked as knowledge units to be supplemented, and the type of knowledge module to be expanded is noted in the subsequent knowledge seed description.

[0056] Step S125: Statistically analyze the knowledge modules to which all retrieval semantic units belong, merge the statistical results of the belonging to the same knowledge module, and determine the core knowledge module corresponding to the retrieval trigger content. The core knowledge module is the knowledge module to which the most retrieval semantic units belong.

[0057] The knowledge modules to which all search semantic units belong are statistically analyzed. For example, the search semantic unit "childhood acute lymphoblastic leukemia" belongs to the knowledge module "clinical diagnosis and treatment of childhood acute lymphoblastic leukemia," and the search semantic unit "chemotherapy regimen" belongs to the knowledge module "chemotherapy regimen for childhood acute lymphoblastic leukemia." The statistical results of belonging to the same knowledge module are merged. It is found that the knowledge module "chemotherapy regimen for childhood acute lymphoblastic leukemia" is belonged to the same search semantic unit, and the knowledge module "clinical diagnosis and treatment of childhood acute lymphoblastic leukemia" is also belonged to the same search semantic unit. Since the number of belongings for both is the same, and considering that the knowledge level in the teaching stage information is the clinical application level, the knowledge module "chemotherapy regimen for childhood acute lymphoblastic leukemia" better meets the knowledge needs of the clinical application level. Therefore, the core knowledge module corresponding to the search trigger content is determined to be "chemotherapy regimen for childhood acute lymphoblastic leukemia."

[0058] Step S126: Based on the knowledge depth requirements, extract the key knowledge points under the core knowledge module. The key knowledge points are the content that the core knowledge module needs to cover in the current knowledge level.

[0059] Based on the knowledge depth requirements extracted in step S122, key knowledge points are extracted under the core knowledge module "Chemotherapy Regimen for Acute Lymphoblastic Leukemia in Children". These key knowledge points are the contents that this core knowledge module needs to cover in the clinical application layer. For example, they include the combination of chemotherapy drugs for different risk groups (low risk, intermediate risk, high risk), the method of calculating the dosage, the choice of the route of administration, the division of the treatment course (induction remission period, consolidation and strengthening period, maintenance treatment period), and the basis for adjusting chemotherapy drugs at each stage.

[0060] Step S127: Based on the knowledge extension scope, determine the extension direction of the core knowledge module. The vertical in-depth direction corresponds to the subdivided knowledge points of the core knowledge module, and the horizontal association direction corresponds to other knowledge modules that have a logical connection with the core knowledge module.

[0061] Based on the scope of knowledge extension, the extension directions of the core knowledge module "Chemotherapy Regimens for Acute Lymphoblastic Leukemia in Children" were determined. Vertically, the direction of in-depth study corresponds to the subdivided knowledge points of the core knowledge module, such as the pharmacokinetic characteristics of chemotherapy drugs, and the detection and management strategies for chemotherapy drug resistance. Horizontally, the direction of association corresponds to other knowledge modules that are logically related to the core knowledge module, such as the knowledge modules "Management of Adverse Reactions to Chemotherapy for Acute Lymphoblastic Leukemia in Children" (the association logic is causal, chemotherapy may cause adverse reactions) and "Efficacy Evaluation of Chemotherapy for Acute Lymphoblastic Leukemia in Children" (the association logic is application, efficacy evaluation is required after chemotherapy).

[0062] Step S128: Integrate core knowledge modules, key knowledge points, and extension directions to form a knowledge seed description. The knowledge seed description also includes the modal resource requirements corresponding to each extension direction. The modal resource requirements are determined based on the knowledge presentation format, with conceptual knowledge corresponding to text resources and operational knowledge corresponding to video resources.

[0063] The core knowledge module "Chemotherapy Regimens for Childhood Acute Lymphoblastic Leukemia," extracted key knowledge points (drug combinations in chemotherapy regimens for different risk groups, etc.), and identified extension directions (vertical in-depth study of the pharmacokinetic characteristics of chemotherapy drugs, etc., and horizontal association of the management of adverse reactions to chemotherapy, etc.) are integrated to form a knowledge seed description. Simultaneously, the knowledge seed description includes the modal resource requirements corresponding to each extension direction, determined based on the knowledge presentation format. For example, for conceptual knowledge such as the pharmacokinetic characteristics of chemotherapy drugs in the vertical in-depth direction, the corresponding modal resource requirement is text resources; for operational knowledge such as the procedures for managing adverse reactions to chemotherapy in the horizontal association direction, the corresponding modal resource requirement is video resources.

[0064] Step S130: Based on the knowledge seed description, call the pediatric oncology teaching multimodal resource library, extract the initial multimodal resources associated with the core knowledge module in the pediatric oncology teaching multimodal resource library, perform semantic evolution processing of the initial multimodal resources through the pediatric oncology teaching big model, and generate a knowledge semantic evolution tree. The knowledge semantic evolution tree has the knowledge seed as the root node, the extended knowledge module as the child node, and the semantic association as the branch.

[0065] After generating the knowledge seed description, the pediatric oncology teaching multimodal resource library is called based on the knowledge seed description. The initial multimodal resources associated with the core knowledge module "chemotherapy regimen for childhood acute lymphoblastic leukemia" are extracted from the resource library. Then, the initial multimodal resources are subjected to semantic evolution processing through the pediatric oncology teaching big model to generate a knowledge semantic evolution tree.

[0066] Step S131: Call the resource retrieval interface of the Pediatric Oncology Teaching Multimodal Resource Library, input the core knowledge module name and key knowledge points in the knowledge seed description, and execute the resource retrieval.

[0067] The resource retrieval interface of the multimodal resource library for pediatric oncology teaching is called. The core knowledge module name "chemotherapy regimen for acute lymphoblastic leukemia in children" in the knowledge seed description and the extracted key knowledge points (drug combinations of chemotherapy regimens for different risk groups, etc.) are input as retrieval parameters into the resource retrieval interface to perform the resource retrieval operation.

[0068] Step S132: Traverse all multimodal resources in the resource library through the resource retrieval interface, and extract the knowledge module tags and knowledge point tags for each resource. The knowledge module tags and knowledge point tags are attribute information labeled based on the pediatric oncology teaching knowledge system.

[0069] After receiving the search parameters, the resource retrieval interface iterates through all multimodal resources in the pediatric oncology teaching multimodal resource library. For each multimodal resource, it extracts its pre-labeled knowledge module tags and knowledge point tags. These tags are attribute information labeled based on the pediatric oncology teaching knowledge system. For example, a text resource might have the knowledge module tag "chemotherapy regimen for childhood acute lymphoblastic leukemia" and knowledge point tags such as "low-risk group chemotherapy drug combination" and "dosage during induction remission"; a video resource might have the knowledge module tag "chemotherapy regimen for childhood acute lymphoblastic leukemia" and knowledge point tags such as "chemotherapy drug administration route and operation".

[0070] Step S133: Select multimodal resources whose knowledge module tags are consistent with the core knowledge module name and whose knowledge point tags contain key knowledge points, as initial multimodal resources. The initial multimodal resources include text resources, image resources, and video resources.

[0071] The extracted multimodal resources with knowledge module tags and knowledge point tags were screened. The screening criteria were that the knowledge module tag was consistent with the core knowledge module name "Chemotherapy Regimen for Acute Lymphoblastic Leukemia in Children," and the knowledge point tag contained key knowledge points (such as drug combinations for chemotherapy regimens in different risk groups, dosage calculation methods, etc.). The multimodal resources obtained after screening were used as initial multimodal resources. These initial multimodal resources included text resources (such as chemotherapy regimen guidelines and drug instruction manuals), image resources (such as schematic diagrams of the mechanism of action of chemotherapy drugs and bone marrow aspiration images), and video resources (such as videos of chemotherapy drug preparation and administration routes).

[0072] Step S134: Classify the initial multimodal resources according to modal type to obtain the initial text resource group, the initial image resource group, and the initial video resource group.

[0073] The initial multimodal resources obtained from the screening were classified according to modality type, resulting in initial text resource groups, initial image resource groups, and initial video resource groups. The initial text resource group contains initial multimodal resources of all text types, such as the text document "Clinical Guidelines for Chemotherapy Regimens in Children with Acute Lymphoblastic Leukemia V2.0" and the text file "Manual for Calculating Dosage of Commonly Used Chemotherapy Drugs." The initial image resource group contains initial multimodal resources of all image types, such as "Acute Lymphoblastic Leukemia Cell Morphology Image Collection" and "Schematic Diagram of Chemotherapy Drug Targets." The initial video resource group contains initial multimodal resources of all video types, such as "Video of Intravenous Infusion of Chemotherapy Drugs for Children with Acute Lymphoblastic Leukemia" and "Demonstration Video of Emergency Management of Adverse Reactions to Chemotherapy."

[0074] Step S135: Input the initial text resource group into the text semantic evolution unit of the pediatric oncology teaching model, perform semantic decomposition of the text resources, extract the core semantic components of each text resource, extend and expand the core semantic components based on the extension direction described by the knowledge seed, and generate text semantic extension branches. Each text semantic extension branch corresponds to an extended knowledge point. The core semantic components include concepts, principles, and cases.

[0075] The initial set of text resources is input into the text semantic evolution unit of the large-scale pediatric oncology teaching model. The text semantic evolution unit first performs semantic decomposition on each text resource. For example, the text document "Clinical Guidelines for Chemotherapy Regimens for Acute Lymphoblastic Leukemia in Children V2.0" is semantically divided by chapters and paragraphs, extracting its core semantic components. These core semantic components include concepts (such as the concept of "induction remission therapy"), principles (such as the principle of synergistic effects of chemotherapy drugs), and cases (such as "a case study of the application of a chemotherapy regimen in a child with intermediate-risk disease"). Then, based on the extension direction described by the knowledge seeds, the core semantic components are extended and expanded. For example, for the core semantic component of "induction remission therapy," combined with the vertical extension direction, the extended knowledge point "drug metabolism pathways during induction remission therapy" is derived; for the core semantic component "the principle of synergistic effects of chemotherapy drugs," combined with the vertical extension direction, the extended knowledge point "molecular mechanisms of synergistic effects of different chemotherapy drugs" is derived. Each extended knowledge point corresponds to a text semantic extension branch.

[0076] Step S1351: Take the first text resource in the initial text resource group, divide it into multiple text paragraph units, and each text paragraph unit is a continuous semantically complete text.

[0077] Take the first text resource in the initial text resource group, such as the text document "Clinical Guidelines for Chemotherapy Regimens for Acute Lymphoblastic Leukemia in Children V2.0", and divide the text resource into multiple text paragraph units. Each text paragraph unit is a continuous semantically complete text. For example, the first paragraph of the "1.1 Induction Remission Treatment" section of the document, "Induction remission treatment aims to rapidly reduce the leukemia cell burden. Commonly used drugs include vincristine, prednisone, daunorubicin, etc.", is considered as a text paragraph unit.

[0078] Step S1352: Perform semantic annotation on each text paragraph unit. The annotation content is the semantic type corresponding to the paragraph unit. The semantic type includes concept type, principle type, and case type.

[0079] Each text segment unit is semantically labeled, and its corresponding semantic type is marked. Semantic types include concept type, principle type, and case type. For example, the text segment unit "Induction remission treatment aims to rapidly reduce the leukemia cell burden, and commonly used drugs include vincristine, prednisone, daunorubicin, etc." is labeled as a concept type; the segment "Vincristine exerts its anti-tumor effect by inhibiting microtubule polymerization and preventing cell mitosis" is labeled as a principle type; and the segment "A 5-year-old boy was diagnosed with acute lymphoblastic leukemia (intermediate-risk group) and underwent induction remission treatment using the VDLP regimen. On the 28th day of treatment, the bone marrow morphology showed complete remission" is labeled as a case type.

[0080] Step S1353: Extract the definition, attribute description, and classification information from the concept-type paragraph units to form the concept semantic components; extract the mechanism explanation, causal relationship description, and process description from the principle-type paragraph units to form the principle semantic components; extract the case background, treatment process, and result description from the case-type paragraph units to form the case semantic components.

[0081] For paragraphs labeled as conceptual, extract the following: definition (e.g., "Induction remission treatment refers to the treatment process in the initial stage of chemotherapy aimed at rapidly reducing the number of leukemia cells"), attribute description (e.g., "Induction remission treatment is characterized by a short treatment course and high drug dosage"), and classification information (e.g., "According to different risk groups of children, induction remission treatment regimens are divided into standard regimens and intensive regimens"). Combine these extracted contents to form conceptual semantic components. For paragraphs labeled as principle-based, extract the following: mechanism explanation (e.g., "Daunorubicin inhibits tumor cell proliferation by interfering with DNA synthesis and transcription through embedding in the DNA double strand"), causal relationship description (e.g., "Because prednisone can inhibit lymphocyte proliferation, its application in the early stages of induction remission can rapidly control the growth of leukemia cells"), and process description (e.g., "The induction remission treatment process is: vincristine is administered on day 1, prednisone is administered from days 1 to 28, and aunorubicin is administered on days 8, 15, and 22"). Combine these extracts to form principle semantic components. For case-based paragraph units, extract the case background (e.g., "Basic information of the child: age 6 years, weight 20kg, diagnosed as low-risk acute lymphoblastic leukemia"), treatment process (e.g., "VDLP chemotherapy regimen was administered: vincristine 1.5mg / m², once a week for 4 weeks; daunorubicin 30mg / m², once a week for 3 weeks; asparaginase 6000U / m², every other day for 8 weeks; prednisone 60mg / m², orally, days 1-28"), and result description (e.g., "Bone marrow examination after 4 weeks of treatment showed leukemia cell percentage <5%, achieving the complete remission standard") to form the semantic components of the case.

[0082] Step S1354: Collect conceptual semantic components, principle semantic components, and case semantic components to form a set of core semantic components of the current text resource.

[0083] Collect conceptual semantic components, principle semantic components, and case semantic components extracted from the current text resource, and integrate these semantic components to form the core semantic component set of the current text resource. For example, the core semantic component set of the text document "Clinical Guidelines for Chemotherapy Regimens for Childhood Acute Lymphoblastic Leukemia V2.0" includes the above-extracted conceptual semantic components about induction remission therapy, principle semantic components about the mechanism of action of chemotherapy drugs, and case semantic components about chemotherapy cases.

[0084] Step S1355: Read the extension direction described by the knowledge seed. If the extension direction is vertical in-depth, then for each component in the core semantic component set, extract the subdivided content of that component. The subdivided content includes the subdivided types of concepts and the subdivided mechanisms of principles, and generate vertically extended knowledge points.

[0085] Read the extension direction determined in the knowledge seed description. If the extension direction is vertical in-depth, such as the semantic component of the concept of "induction of remission therapy" in the core semantic component set, extract the sub-content of this component, such as the sub-type of the concept (e.g., "standard induction of remission therapy" and "enhanced induction of remission therapy"). For the semantic component of the principle of "synergistic effect of chemotherapy drugs", extract its sub-mechanism (e.g., "synergistic mechanism of drug A and drug B at different phases of the cell cycle" and "synergistic mechanism of drug C enhancing the cell uptake of drug D"). Based on these sub-contents, generate vertically extended knowledge points, such as "applicable population of standard induction of remission therapy regimen" and "molecular mechanism of synergistic effect of drug A and drug B".

[0086] Step S1356: If the extension direction is horizontal association, then based on the core semantic component set, find the knowledge points of other knowledge modules in the pediatric oncology teaching knowledge system that are logically related to the component, and generate horizontally extended knowledge points.

[0087] If the extension direction is horizontal association, based on the components in the core semantic component set, find knowledge points in other knowledge modules within the pediatric oncology teaching knowledge system that are logically related to that component. For example, the semantic component of "chemotherapy drug dosage adjustment" in the core semantic component set is logically related to the knowledge module "calculation of body surface area of ​​pediatric cancer patients" in the pediatric oncology teaching knowledge system (dosage adjustment needs to be based on body surface area calculation). From this knowledge module, extract knowledge points such as "body surface area calculation method (Mosteller formula)" and "body surface area correction method for children with special body types" to generate the horizontally extended knowledge point "application of body surface area calculation in chemotherapy drug dosage adjustment".

[0088] Step S1357: Arrange the vertically extended knowledge points or horizontally extended knowledge points in a semantically related order. Each knowledge point corresponds to a text resource fragment. The text resource fragment comes from the current text resource or the text resource of the related knowledge module, forming a text semantic extension branch.

[0089] The generated vertically or horizontally extended knowledge points are arranged according to semantic association. For example, the concept of "induction remission therapy" can be extended to "standard induction remission therapy" and "intensified induction remission therapy," and further extended to their respective applicable populations, drug combinations, etc., forming a hierarchical semantic association order. Each knowledge point corresponds to a text resource fragment. These text resource fragments can come from the current text resource (such as the descriptive paragraph about this knowledge point in "Clinical Guidelines for Chemotherapy Regimens for Acute Lymphoblastic Leukemia in Children V2.0") or the text resources of related knowledge modules (such as the descriptive paragraph about body surface area calculation methods in "Handbook for Calculating Body Surface Area in Pediatric Cancer Patients"). The arranged knowledge points and corresponding text resource fragments are combined to form semantic extension branches.

[0090] Step S136: Input the initial image resource group into the image semantic evolution unit of the pediatric oncology teaching model, perform visual semantic transformation of the image resources, convert the visual features in the images into textual visual semantic descriptions, combine the extension direction of the knowledge seed description, deduce the extended knowledge points based on the visual semantic descriptions, and generate image semantic extension branches. The visual features include tumor morphology and cell structure.

[0091] The initial image resource set is input into the image semantic evolution unit of the pediatric oncology teaching model. The image semantic evolution unit first performs visual semantic transformation on the image resources. For example, for bone marrow smear images from the "Acute Lymphoblastic Leukemia Cell Morphology Image Set," it identifies visual features such as tumor cell morphology (size, shape, nucleocytoplasmic ratio), cell structure (nuclear morphology, chromatin distribution, number and location of nucleoli), etc., and converts these visual features into textual visual semantic descriptions, such as "A large number of primitive and immature lymphocytes are visible in the field of view. The cells are small in size, have a high nucleocytoplasmic ratio, round or oval nuclei, fine chromatin, and 1-2 nucleoli are visible." Then, combining the extension direction of the knowledge seed description, extended knowledge points are derived based on the visual semantic description. For example, combining the lateral correlation extension direction, the extended knowledge point "correlation between leukemia cell morphology characteristics and chemotherapy efficacy" is derived from the above visual semantic description, and this knowledge point corresponds to the generation of an image semantic extension branch.

[0092] Step S1361: Take the first image resource in the initial image resource group, convert it into a digital image format, perform image enhancement processing, and use a region segmentation algorithm to divide the enhanced image into regions, identify regions containing visual information related to the core knowledge module, including tumor lesion regions and cell structure regions, as key visual regions.

[0093] The first image resource in the initial image resource group, such as "acute lymphoblastic leukemia bone marrow smear image," is taken and converted into a preset digital image format (such as DICOM format). Image enhancement processing is performed on the converted digital image, such as adjusting contrast, brightness, and denoising, to enhance the clarity of key visual information in the image. Next, a region segmentation algorithm (such as threshold-based segmentation algorithm or edge detection-based segmentation algorithm) is used to divide the enhanced image into regions, identifying regions containing visual information related to the core knowledge module "chemotherapy regimen for childhood acute lymphoblastic leukemia." These regions include tumor lesion areas (such as leukemia cell aggregation areas in bone marrow smears) and cell structure areas (such as detailed structural areas of individual leukemia cells). These regions are identified as key visual regions.

[0094] Step S1362: Extract visual features of key visual regions, including morphological features, color features, and structural features. Morphological features include outline and shape, color features include main color tone and color distribution, and structural features include internal component distribution and component connection relationships.

[0095] Visual features are extracted from the identified key visual regions, including morphological, color, and structural features. For morphological features, the outlines (e.g., regular and irregular outlines) and shapes (e.g., circles, ellipses, and irregular shapes) of cells in the key visual regions are extracted. For color features, the dominant hue (e.g., dark blue and light blue) and color distribution (e.g., uniform distribution and speckled distribution) of the key visual regions are extracted. For structural features, the distribution of internal cell components (e.g., the location of the cell nucleus and the distribution of cytoplasmic granules) and the connectivity of components (e.g., the connection between the cell membrane and the cytoplasm) are extracted.

[0096] Step S1363: Invoke the visual-text mapping rules in the field of pediatric oncology teaching to convert the extracted visual features into textual visual semantic descriptions, wherein the visual semantic descriptions include the specific content of the features and the corresponding knowledge attributes.

[0097] The system utilizes predefined visual-text mapping rules in the field of pediatric oncology education, which establish a correspondence between visual features and textual descriptions. For example, the morphological feature "irregular cell outline, polygonal shape" is mapped to the textual description "leukemia cells exhibit abnormal morphology, showing pleomorphic changes"; the structural feature "uneven density of nuclear chromatin, distributed in blocks" is mapped to the textual description "abnormal nuclear chromatin structure, indicating low cell differentiation." Through this mapping rule, the extracted visual features are converted into textualized visual semantic descriptions, which include the specific content of the features (such as the descriptions of the aforementioned morphological and structural features) and corresponding knowledge attributes (such as "abnormal morphology," "low differentiation," and other knowledge attributes related to the characteristics of leukemia cells).

[0098] Step S1364: Read the extension direction of the knowledge seed description. If it is a vertical extension, then based on the visual semantic description, deduce the formation cause, evolution process and clinical significance of the visual feature to generate vertically extended knowledge points.

[0099] Read the extended directions described by the knowledge seeds. If it's a vertical derivation, deduce the formation cause, evolution process, and clinical significance of the visual feature based on the visual semantic description. For example, if the visual semantic description is "abnormal nucleoli are visible in leukemia cells, with increased number and size," the deduction of its formation cause might be "active ribosome synthesis in tumor cells, leading to hyperactive nucleolar function"; the evolution process might be "as the disease progresses, the degree of nucleolar abnormality gradually worsens"; and the clinical significance might be "abnormal nucleoli can serve as one of the indicators for assessing the proliferative activity of leukemia cells." Based on these deductions, generate vertically extended knowledge points, such as "the formation mechanism of abnormal nucleoli in leukemia cells" and "the relationship between the degree of nucleolar abnormality and leukemia progression."

[0100] Step S1365: If it is a horizontal association, then based on the visual semantic description, find the knowledge points of other knowledge modules related to the visual feature in the pediatric oncology teaching knowledge system, and generate horizontally extended knowledge points.

[0101] If the extension direction is horizontal association, based on the visual semantic description, find knowledge points in other knowledge modules related to this visual feature in the pediatric oncology teaching knowledge system. For example, if the visual semantic description is "the proportion of leukemia cells in the bone marrow smear reaches 85%", this visual feature is related to the knowledge module "risk grouping of childhood acute lymphoblastic leukemia" (risk grouping needs to refer to the proportion of leukemia cells in the bone marrow). From this knowledge module, extract the knowledge point "the proportion of leukemia cells in the bone marrow and the criteria for risk grouping", and generate the horizontally extended knowledge point "the application of the proportion of leukemia cells in risk grouping".

[0102] Step S1366: Match the corresponding image resource fragment for each extended knowledge point. The image resource fragment comes from the key visual area of ​​the current image resource and is arranged in the logical order of the knowledge points to form an image semantic extension branch. Repeat the above steps to process all resources in the initial text resource group and the initial image resource group to generate multiple text semantic extension branches and image semantic extension branches.

[0103] For each extended knowledge point, a corresponding image resource fragment is matched. These image resource fragments come from the key visual regions of the current image resources. For example, for the extended knowledge point "correlation between leukemia cell morphological characteristics and chemotherapy efficacy," a bone marrow smear image region fragment containing typical efficacy assessment-related morphological features is matched. The extended knowledge points are arranged in their logical order, such as first describing "original morphological characteristics of leukemia cells," then describing "morphological changes after chemotherapy," and finally describing "relationship between morphological changes and efficacy." The corresponding image resource fragments are also arranged in this order, forming image semantic extension branches. Steps S1351 to S1357 are repeated to process all text resources in the initial text resource group, and steps S1361 to S1365 are repeated to process all image resources in the initial image resource group, generating multiple text semantic extension branches and image semantic extension branches.

[0104] Step S137: Input the initial video resource group into the video semantic evolution unit of the pediatric oncology teaching model, perform frame sequence analysis and audio conversion of the video resources, extract dynamic visual semantics and audio text semantics from the video, combine the extension direction described by the knowledge seed, integrate the dynamic visual semantics and audio text semantics to generate extended knowledge points, and form video semantic extension branches. The dynamic visual semantics include surgical operation steps and case demonstration process, and the audio text semantics include explanatory statements and key terms.

[0105] The initial video resource set is input into the video semantic evolution unit of the pediatric oncology teaching model. The video semantic evolution unit first performs frame sequence analysis on each video resource. For example, for the video of "Intravenous Infusion of Chemotherapy Drugs in Children with Acute Lymphoblastic Leukemia," the video frames are analyzed chronologically to extract dynamic visual semantics, such as surgical procedures like the selection of the venipuncture site, the scope and order of disinfection, the angle and depth of needle insertion, and the adjustment of the drug infusion rate. Simultaneously, the audio signals in the video are converted into text, and audio-text semantics are extracted, such as the explanation statement "When administering chemotherapy drugs intravenously, a patent venous access must be established first to ensure there is no leakage before starting the infusion," and key terms like "central venous catheter" and "infusion pump." Combining the extension direction described by the knowledge seeds, the dynamic visual semantics and audio-text semantics are integrated to generate extended knowledge points. For example, combining the vertical extension direction, the dynamic visual semantics of "selection of venipuncture angle" and the audio-text semantics of "the impact of different puncture angles on vascular damage" are integrated to generate the extended knowledge point "key points of intravenous infusion puncture technique for chemotherapy drugs," which forms a corresponding video semantic extension branch.

[0106] Step S138: Collect text semantic extension branches, image semantic extension branches, and video semantic extension branches. Using the core knowledge module in the knowledge seed description as the root node and the extended knowledge points corresponding to each extension branch as child nodes, construct the initial knowledge semantic evolution structure and delete the child nodes in the initial knowledge semantic evolution structure that have no logical connection with the core knowledge module of the root node. Logical connections include subordination, causality, and application.

[0107] Collect the text semantic extension branches, image semantic extension branches, and video semantic extension branches generated through the above steps. Using the core knowledge module "Chemotherapy Regimen for Acute Lymphoblastic Leukemia in Children" in the knowledge seed description as the root node, construct an initial knowledge semantic evolution structure by using the extended knowledge points corresponding to each extension branch as child nodes, such as "Chemotherapy Drug Combination for Low-Risk Groups" and "Chemotherapy Drug Administration Routes and Procedures." Then, examine the initial knowledge semantic evolution structure and delete child nodes that have no logical connection with the core knowledge module of the root node. Logical connections include subordinate relationships (e.g., "Chemotherapy Drug Dosage Calculation" belongs to the core knowledge module), causal relationships (e.g., "Chemotherapy Adverse Reactions Management" is caused by chemotherapy and has a causal relationship with the core knowledge module), and application relationships (e.g., "Chemotherapy Efficacy Assessment" is applied after the implementation of the chemotherapy regimen and has an application relationship with the core knowledge module). If a child node, such as "Chemotherapy Regimen for Childhood Lymphoma," has no such logical connection with the core knowledge module of the root node, it is deleted.

[0108] Step S139: Add a corresponding multimodal resource reference for each child node, add an association logic description for each branch, and delete child nodes in the initial knowledge semantic evolution structure that have no logical connection with the core knowledge module of the root node. The logical connection includes subordination, causality, and application. Generate a knowledge semantic evolution tree. The multimodal resource reference points to the initial multimodal resource that generated the child node, and the association logic description is used to explain the logical relationship between the child node and the parent node.

[0109] After constructing the initial knowledge semantic evolution structure and deleting irrelevant child nodes, a corresponding multimodal resource reference is added to each retained child node. This reference points to the initial multimodal resource that generated the child node. For example, the multimodal resource reference for the child node "low-risk group chemotherapy drug combination" points to the text document "Clinical Guidelines for Chemotherapy Regimens for Acute Lymphoblastic Leukemia in Children V2.0" in the initial text resource group. Simultaneously, an association logic description is added to each branch. For instance, the association logic description between the child node "Chemotherapy Adverse Reaction Management" and the parent node "Chemotherapy Regimens for Acute Lymphoblastic Leukemia in Children" is "Causal relationship: Chemotherapy drug use leads to adverse reactions, which require treatment." After the above processing, a knowledge semantic evolution tree is generated.

[0110] Step S140: Traverse the knowledge semantic evolution tree, extract the multimodal resources corresponding to each branch, and construct a cross-modal dynamic association chain. The cross-modal dynamic association chain connects different modal resources in the order of knowledge evolution, and each resource node is labeled with the association logic with the knowledge seed.

[0111] After generating the knowledge semantic evolution tree, the tree is traversed, the multimodal resources corresponding to each branch are extracted, and the different modal resources are linked together according to the knowledge evolution order to construct a cross-modal dynamic association chain. The association logic between each resource node and the knowledge seed is marked.

[0112] Step S141: Starting from the root node of the knowledge semantic evolution tree, traverse all branches. Each branch is a knowledge evolution path extending from the root node to the terminal child node.

[0113] Starting from the root node "Chemotherapy regimen for childhood acute lymphoblastic leukemia" in the knowledge semantic evolution tree, all branches are traversed using either depth-first search or breadth-first search. Each branch represents a knowledge evolution path extending from the root node to the terminal child node. For example, the knowledge evolution path of one branch is: root node → low-risk group chemotherapy regimen → induction remission treatment → chemotherapy drug dosage calculation → body surface area calculation method.

[0114] Step S142: For each branch, extract the multimodal resource references corresponding to all child nodes on the knowledge evolution path, and determine the specific resource pointed to by each reference. The specific resources include text resource fragments, image resource fragments, and video resource fragments.

[0115] For each branch traversed, extract the multimodal resource references corresponding to all child nodes on the knowledge evolution path. For example, in the knowledge evolution path described above, the multimodal resource reference of the "low-risk group chemotherapy regimen" child node points to a text resource fragment, the multimodal resource reference of the "induction remission treatment" child node points to a video resource fragment, the multimodal resource reference of the "chemotherapy drug dosage calculation" child node points to a text resource fragment, and the multimodal resource reference of the "body surface area calculation method" child node points to an image resource fragment. Based on these references, determine the specific resource that each reference points to, i.e., text resource fragment, image resource fragment, video resource fragment, etc.

[0116] Step S143: Sort the extracted multimodal resources according to the knowledge evolution order. The knowledge evolution order is from the core knowledge module corresponding to the root node to the extended knowledge point corresponding to the terminal child node. The sorting is based on the logical progression relationship of the knowledge points.

[0117] The extracted multimodal resources are sorted according to the order of knowledge evolution, from the core knowledge module corresponding to the root node to the extended knowledge points corresponding to the terminal child nodes. The sorting is based on the logical progression between knowledge points. For example, the logical progression of knowledge points in the above knowledge evolution path is as follows: first, understand the overall framework of the low-risk group chemotherapy regimen; then, delve into the specific stage of induction remission treatment; next, involve the calculation of chemotherapy drug dosage in this stage; and finally, learn the body surface area calculation method required for dosage calculation. Therefore, the corresponding multimodal resource sorting is: low-risk group chemotherapy regimen text resource fragment → induction remission treatment video resource fragment → chemotherapy drug dosage calculation text resource fragment → body surface area calculation method image resource fragment.

[0118] Step S144: Analyze the modality distribution of the sorted multimodal resources. If there are multiple consecutive resources with the same modality and the knowledge content of the resources is repeated, retain the resource with the most comprehensive knowledge coverage and delete the remaining duplicate resources.

[0119] Analyze the modality distribution of the sorted multimodal resources to check for multiple consecutive resources with the same modality and duplicated knowledge content. For example, if two consecutive text resource fragments appear after sorting, both describing low-risk chemotherapy drug combinations and with largely overlapping knowledge content, retain the text resource fragment with the most comprehensive knowledge coverage (such as including drug interactions, precautions, and more), and delete the duplicate text resource fragment.

[0120] Step S145: If there is a missing modality type, based on the modality resource requirements described by the knowledge seed, the corresponding modality resources are retrieved from the multimodal resource library for pediatric oncology teaching and inserted into the corresponding knowledge evolution stage.

[0121] Check whether the sorted multimodal resources have missing modal types, i.e., whether they meet the modal resource requirements in the knowledge seed description. For example, if the modal resource requirement for the operational knowledge of "chemotherapy drug administration route operation" in the knowledge seed description is video resources, and the currently sorted resources are missing video resources at the corresponding knowledge evolution stage (such as the "induction remission treatment" stage), then based on the knowledge point description of that knowledge evolution stage (such as "intravenous infusion administration operation"), supplement the search with corresponding video resources from the pediatric oncology teaching multimodal resource library (such as "video of intravenous infusion operation of chemotherapy drugs for childhood acute lymphoblastic leukemia"), and insert the supplemented video resources into the corresponding knowledge evolution stage.

[0122] Step S1451: Extract the knowledge points corresponding to all child nodes on the current branch, arrange them in the order of knowledge evolution to obtain a knowledge point sequence. The beginning of the knowledge point sequence is the key knowledge point of the core knowledge module, and the subsequent ones are the extended knowledge points in the extension direction.

[0123] Extract the knowledge points corresponding to all child nodes on the current branch. For example, in the knowledge evolution path "root node → low-risk group chemotherapy regimen → induction remission treatment → chemotherapy drug dosage calculation → body surface area calculation method", the corresponding knowledge point sequence is: childhood acute lymphoblastic leukemia chemotherapy regimen (core knowledge module key knowledge point) → low-risk group chemotherapy regimen (extended knowledge point) → induction remission treatment (extended knowledge point) → chemotherapy drug dosage calculation (extended knowledge point) → body surface area calculation method (extended knowledge point). Arrange the above knowledge points according to the knowledge evolution order to obtain the above knowledge point sequence.

[0124] Step S1452: Analyze the knowledge attributes of each knowledge point in the knowledge point sequence. The knowledge attributes include conceptual, principle-based, operational, and case-based.

[0125] Analyze the knowledge attributes of each knowledge point in the knowledge point sequence. For example, the knowledge attribute of the knowledge point "chemotherapy regimen for childhood acute lymphoblastic leukemia" is conceptual; the knowledge attribute of the knowledge point "calculation of chemotherapy drug dosage" is principle-based; the knowledge attribute of the part involving drug administration in "induction remission treatment" is operational; and the knowledge attribute of the knowledge point "chemotherapy case of a low-risk child" is case-based.

[0126] Step S1453: Determine the logical progression relationship of knowledge points based on knowledge attributes and teaching rules. The progression order between the conceptual, principle, operational, and case types defined by the teaching rules is the conventional progression order. When there are special requirements for the extension direction of the knowledge seed description, the progression order is adjusted according to the special requirements.

[0127] The logical progression is determined based on the knowledge attributes of the knowledge points and the teaching principles. The conventional progression order defined by the teaching principles is: Conceptual → Principle-based → Operational → Case-based. For example, first learn the concept of "chemotherapy regimens for childhood acute lymphoblastic leukemia" (conceptual), then learn the "mechanism of action of chemotherapy drugs" (principle-based), then learn the "administration procedure of chemotherapy drugs" (operational), and finally learn "case applications of chemotherapy" (case-based). If the extension direction of the knowledge seeds has special requirements, such as emphasizing the priority of operational knowledge points at the clinical application level, then the progression order is adjusted to: Conceptual → Operational → Principle-based → Case-based.

[0128] Step S1454: Reorder the knowledge point sequence according to the adjusted progressive order to obtain an ordered knowledge point sequence.

[0129] Based on the adjusted progressive order, the knowledge point sequence is reordered. For example, under the conventional progressive order, the original knowledge point sequence is readjusted in the order of concept-based, principle-based, operation-based, and case-based to obtain an ordered knowledge point sequence.

[0130] Step S1455: Extract the multimodal resources corresponding to each knowledge point in the ordered knowledge point sequence, arrange them according to the knowledge point sorting order, and obtain the initial version of the resource sorting.

[0131] Extract the multimodal resources corresponding to each knowledge point in the ordered knowledge point sequence. For example, text resources correspond to "chemotherapy regimens for childhood acute lymphoblastic leukemia" (conceptual type), text and video resources correspond to "mechanism of action of chemotherapy drugs" (principle type), video resources correspond to "administration procedures of chemotherapy drugs" (operation type), and text and video resources correspond to "case application of chemotherapy" (case type). Arrange the above multimodal resources according to the order of knowledge points to obtain the initial resource sorting.

[0132] Step S1456: Check whether the knowledge points of adjacent resources in the initial version of resource sorting conform to the logical progression relationship. If there is a reversal of the order, adjust the resource positions to conform to the progression order.

[0133] Check whether the knowledge points of adjacent resources in the initial version of the resource sorting conform to the logical progression relationship. For example, if the resource "chemotherapy drug administration operation" (operation type) is placed before the resource "chemotherapy drug action principle" (principle type), but the current progression order requires the principle type to come first, then adjust the positions of the two to make the resource sorting conform to the progression order.

[0134] Step S1457: After completing the resource sorting, count the modal types included in the resource sorting. Modal types include text, images, and videos. Compare the modal resource requirements described by the knowledge seeds.

[0135] After completing the resource sorting, count the number of resources for each modality type included in the sort, such as text, image, and video. Compare the statistical results with the modal resource requirements described by the knowledge seeds to check whether the corresponding modal resource requirements are met at each stage of knowledge evolution. For example, does the operational knowledge point have a corresponding video resource?

[0136] Step S1458: If any knowledge point's corresponding evolutionary stage lacks the modal type required, then use the text description of that knowledge point as the search condition and call the supplementary search interface of the Pediatric Oncology Teaching Multimodal Resource Library.

[0137] If, after comparison, it is found that any knowledge point's corresponding evolutionary stage lacks the modal type required, such as the lack of video resources for the evolutionary stage corresponding to the operational knowledge point "chemotherapy drug administration operation," then the supplementary search interface of the Pediatric Oncology Teaching Multimodal Resource Library will be invoked using the text description of the knowledge point, "chemotherapy drug administration operation," as the search condition.

[0138] Step S1459: Supplement the search interface to traverse the resource library, filter the corresponding modal resources that are semantically related to the text description of the knowledge point, and extract the core content fragments of the corresponding modal resources.

[0139] After receiving the search criteria, the supplementary search interface traverses the multimodal resource library for pediatric oncology education, filtering video modal resources semantically related to the text description of "chemotherapy drug administration procedure," such as "video of intravenous infusion procedure for chemotherapy drugs in children with acute lymphoblastic leukemia" and "video of intrathecal injection procedure for chemotherapy drugs." Core content segments are extracted from the selected video resources, such as key steps in the intravenous infusion procedure, including puncture, fixation, and adjustment of the drip rate.

[0140] Step S14510: Insert the supplementary resource fragments into the corresponding knowledge point evolution stage in the resource sorting, mark the supplementary resource with a supplementary resource identifier, and explain the reason for the supplementation.

[0141] The extracted supplementary resource fragments are inserted into the corresponding knowledge point evolution stage in the resource sorting, namely the evolution stage of the knowledge point "chemotherapy drug administration operation", and labeled with a supplementary resource identifier (such as "supplementary resource"), and the reason for supplementation is explained (such as "meeting the needs of operational knowledge video modal resources").

[0142] Step S14511: Repeat the above supplementary steps until all knowledge point evolution stages in the resource sorting meet the modal resource requirements, and finally obtain a resource sequence arranged in the order of knowledge evolution and with complete modal types, which serves as the resource basis for cross-modal dynamic association chains.

[0143] Repeat steps S1458 to S14510 to check and supplement all knowledge point evolution stages in the resource sorting until all evolution stages meet the modal resource requirements, ultimately obtaining a resource sequence arranged in the order of knowledge evolution and with complete modal types. This resource sequence serves as the resource foundation for cross-modal dynamic association chains.

[0144] Step S146: Add a logical association label to each resource node. The content of the logical association label is the logical relationship between the knowledge point corresponding to the resource and the core knowledge module in the knowledge seed. At the same time, add a transitional association label to adjacent resource nodes. The content of the transitional association label is the knowledge point association relationship between the previous resource node and the next resource node.

[0145] Add logical association annotations to each resource node. For example, if a text resource node corresponds to the knowledge point "low-risk chemotherapy drug combination," its logical relationship with the core knowledge module "chemotherapy regimen for childhood acute lymphoblastic leukemia" in the knowledge seed is subordinate. Therefore, the logical association annotation content is "Subordinate relationship: Low-risk chemotherapy drug combination is a component of the chemotherapy regimen for childhood acute lymphoblastic leukemia." Simultaneously, add transitional association annotations to adjacent resource nodes. For example, if the knowledge point corresponding to the previous resource node is "low-risk chemotherapy regimen," and the knowledge point corresponding to the next resource node is "induction remission treatment," the knowledge point relationship between the two is inclusion. The transitional association annotation content is "Inclusion relationship: Low-risk chemotherapy regimen includes the induction remission treatment phase."

[0146] Step S147: Integrate the sorted resource nodes, associated logic annotations, and associated transition annotations to form a cross-modal dynamic association chain; repeat the above steps to traverse all branches of the knowledge semantic evolution tree and generate multiple cross-modal dynamic association chains.

[0147] The sorted resource nodes, associated logical annotations, and associated transition annotations are integrated to form a complete cross-modal dynamic association chain. Steps S141 to S146 are repeated to traverse all branches of the knowledge semantic evolution tree and generate a corresponding cross-modal dynamic association chain for each branch, thereby obtaining multiple cross-modal dynamic association chains.

[0148] Step S148: Collect cross-modal dynamic association chains with complete evolution paths, no missing resource nodes, and no missing association annotations to form a set of cross-modal dynamic association chains.

[0149] The generated cross-modal dynamic association chains are filtered to collect those with complete evolution paths (no breaks from the root node to the terminal child node), no missing resource nodes (each knowledge point has a corresponding multimodal resource), and no missing association annotations (each resource node has an association logic annotation, and adjacent nodes have an association transition annotation). The association chains that meet the above conditions are integrated to form a set of cross-modal dynamic association chains.

[0150] Step S150: Determine the association chain screening criteria based on the teaching stage information, extract cross-modal dynamic association chains that meet the association chain screening criteria, integrate multimodal resources in the cross-modal dynamic association chains to generate cross-modal retrieval results, push them to the user interaction terminal, and simultaneously integrate the generated knowledge semantic evolution tree and cross-modal dynamic association chains into the knowledge association index of the pediatric oncology teaching multimodal resource library to update the basic knowledge evolution data for subsequent retrieval.

[0151] Based on the information of the teaching stage, the screening criteria for the association chain are determined. The association chains that meet the criteria are extracted from the cross-modal dynamic association chain set. The multimodal resources in the set are integrated to generate cross-modal search results and push them to the user interaction terminal. At the same time, the knowledge semantic evolution tree and the cross-modal dynamic association chain are integrated into the knowledge association index of the resource library.

[0152] Step S151: Analyze the knowledge levels in the teaching stage information. Different knowledge levels correspond to different resource emphases: the basic cognition level emphasizes the completeness of the concept of the resource, the principle analysis level emphasizes the detail of the principle of the resource, the clinical application level emphasizes the practicality of the case of the resource, and the comprehensive expansion level emphasizes the breadth of the relevance of the resource.

[0153] The knowledge hierarchy in the teaching stage information is analyzed. In this embodiment, the knowledge hierarchy is the clinical application layer. Different knowledge levels correspond to different resource focuses. The clinical application layer focuses on the practicality of the case studies, that is, it prioritizes multimodal resources that contain rich clinical cases and can guide actual clinical operations.

[0154] Step S152: Analyze the teaching objectives in the teaching stage information. Different objectives correspond to different resource type weights: text resources have high weight for knowledge understanding, video resources have high weight for skill mastery, and multimodal resources have balanced weights for comprehensive application.

[0155] The teaching objectives in the analysis of teaching stage information are guided by learning objectives. In this embodiment, the learning objectives are guided by skills mastery. Different objectives correspond to different resource type weights. Under the skills mastery orientation, video resources have a higher weight. That is, when filtering related links, related links that contain video resources and have high-quality video resources will receive higher weights.

[0156] Step S153: Based on the resource emphasis direction corresponding to the knowledge level and the resource type weight corresponding to the teaching objective orientation, determine the screening criteria for the association chain. The screening criteria for the association chain includes three dimensions: resource completeness, modal adaptability, and logical coherence. Resource completeness refers to the proportion of knowledge points covered by resources in the chain, modal adaptability refers to the degree of matching between the resource type in the chain and the objective orientation, and logical coherence refers to the logical clarity of the association annotations in the chain. The resource completeness dimension is scored according to the ratio of the number of knowledge points covered by resources to the total number of knowledge points, the modal adaptability dimension is scored according to the proportion of resource types that conform to the objective orientation, and the logical coherence dimension is scored according to the completeness and accuracy of the association annotations. In this way, the scoring rules for each dimension are determined.

[0157] Based on the resource emphasis (case practicality) corresponding to the knowledge level (clinical application level) and the resource type weight (video resources have higher weight) corresponding to the teaching objective orientation (skill mastery orientation), the screening criteria for the association chain are determined. This screening criteria includes three dimensions: resource completeness, modal adaptability, and logical coherence. The resource completeness dimension is scored according to the ratio of the number of knowledge points covered by resources in the chain to the total number of knowledge points. For example, a coverage ratio of over 90% receives a high score, 70%-90% receives a medium score, and below 70% receives a low score. The modal adaptability dimension is scored according to the proportion of resource types (video resources) that conform to the objective orientation (skill mastery orientation). A video resource ratio of over 60% receives a high score, 40%-60% receives a medium score, and below 40% receives a low score. The logical coherence dimension is scored according to the completeness (e.g., whether each resource node has a related logical annotation) and accuracy (e.g., whether the related logical description is correct). High completeness and accuracy receive a high score, partial omissions or inaccuracies receive a medium score, and severe omissions or numerous errors receive a low score. The specific scoring criteria for each dimension are determined based on the above rules.

[0158] Step S154: Traverse each association chain in the cross-modal dynamic association chain set, score the association chain in three dimensions according to the scoring rules, assign weights to each dimension based on the resource emphasis direction and resource type weight, calculate the weighted comprehensive score of the three dimensions, and filter the association chains with the weighted comprehensive score higher than the corresponding threshold as cross-modal dynamic association chains that meet the standards.

[0159] Iterate through each link in the cross-modal dynamic link set, and score each link according to the scoring rules determined in step S153, focusing on three dimensions: resource completeness, modal adaptability, and logical coherence. Based on the resource emphasis (case practicality) and resource type weight (video resources have higher weight), assign weights to each dimension. For example, modal adaptability has the highest weight (e.g., 0.4), followed by resource completeness (e.g., 0.35), and logical coherence has a weight of 0.25. Calculate the weighted comprehensive score for each link across the three dimensions (weighted comprehensive score = resource completeness score × 0.35 + modal adaptability score × 0.4 + logical coherence score × 0.25). Set a comprehensive score threshold (e.g., 75 points), and filter links with a weighted comprehensive score higher than this threshold as compliant cross-modal dynamic links.

[0160] Step S155: If the number of qualified association chains exceeds the preset number, sort them from high to low according to the weighted comprehensive score and select the first preset number of association chains; if the number of qualified association chains is lower than the preset number, lower the comprehensive score threshold by no more than the preset percentage, and re-screen until the preset number is reached or it is confirmed that there are no more qualified association chains.

[0161] The default number of matching links is 3. If the number of matching links exceeds 3 (e.g., 5), these 5 links are sorted by weighted composite score from highest to lowest, and the top 3 links are selected. If the number of matching links is less than 3 (e.g., only 2), the composite score threshold is lowered by no more than a preset percentage (e.g., 10%), and the filtering is repeated. If the number still cannot reach 3, it is confirmed that there are no more matching links, and the 2 previously filtered links are selected.

[0162] Step S156: Extract all multimodal resources from the association chains that meet the criteria, integrate them according to the sorting order of the association chains, form a resource package for each association chain, and add a search result description to each resource package. The description includes the comprehensive score of the association chain, the scores of each dimension, and the applicable teaching scenarios. The applicable teaching scenarios are derived from the information of the teaching stage. At the same time, add resource usage suggestions for each resource, which include when to use the resource in teaching.

[0163] All multimodal resources in the compliant association chains are extracted and integrated according to the sorting order of the association chains (e.g., from high to low weighted comprehensive score). Resources corresponding to each association chain form a resource package. A description of the search results is added to each resource package, including the comprehensive score of the association chain (e.g., 85 points), scores for each dimension (resource completeness 80 points, modal adaptability 90 points, logical coherence 82 points), and applicable teaching scenarios. The applicable teaching scenarios are derived from information at the teaching stage. In this embodiment, the knowledge level is the clinical application level, and the teaching objective is skills mastery; therefore, the applicable teaching scenario is "Clinical Skills Training Course on Chemotherapy Regimens for Childhood Acute Lymphoblastic Leukemia." Simultaneously, resource usage suggestions are added to each resource. For example, the suggestion for the video resource "Video on Intravenous Infusion of Chemotherapy Drugs" is to "play it when explaining the chapter on chemotherapy drug administration routes, in conjunction with practical demonstrations."

[0164] Step S157: Integrate all resource packages, search result descriptions, and resource usage suggestions to form cross-modal search results, and convert the cross-modal search results into a display format supported by the user interaction terminal. The display format includes resource preview thumbnails, resource names, logical diagrams of related chains, and search result description text.

[0165] Integrate all resource packages, search result descriptions, and resource usage suggestions to form cross-modal search results. Convert these search results into a display format supported by the user's interactive terminal, such as HTML. The display format includes resource preview thumbnails (e.g., the first frame of a video resource, the cover image of a text resource), resource names (e.g., "Clinical Guidelines for Chemotherapy Regimens for Childhood Acute Lymphoblastic Leukemia V2.0"), a logical diagram of the relational chain (showing the relationships between resource nodes in flowchart form), and search result description text (including comprehensive scores, applicable scenarios, etc.).

[0166] Step S158: Push the converted cross-modal search results to the user interaction terminal so that the terminal interface displays resource packages from high to low according to the comprehensive score of the association chain. Each resource package provides operation entry points for viewing, downloading, and collecting resources. At the same time, the terminal interface displays the search result description and resource usage suggestions.

[0167] The converted cross-modal search results are pushed to the user interaction terminal. After receiving the search results, the user interaction terminal displays resource packages on the interface, ranked from highest to lowest based on the comprehensive score of the related chains. Each resource package provides access points for viewing (clicking to open detailed resource content), downloading (saving the resource locally), and saving (marking as a frequently used resource). Simultaneously, a description of the search results (such as the comprehensive score and scores for each dimension of each resource package) and resource usage suggestions (such as when to use each resource) are displayed in a designated area of ​​the terminal interface.

[0168] Step S159: Integrate the generated knowledge semantic evolution tree and cross-modal dynamic association chain into the knowledge association index of the multimodal resource library for pediatric oncology teaching, and update the basic data of knowledge evolution for subsequent retrieval.

[0169] While pushing the cross-modal search results to the user's interactive terminal, the generated knowledge semantic evolution tree and cross-modal dynamic association chain are integrated into the knowledge association index of the pediatric oncology teaching multimodal resource library. Specifically, this process involves updating relevant entries in the knowledge association index and using the new knowledge semantic evolution tree and cross-modal dynamic association chain as the foundational data for subsequent searches, thereby improving the accuracy and efficiency of subsequent searches.

[0170] Step S1591: Extract the structural data of the generated knowledge semantic evolution tree. The structural data includes the root node, child nodes, branches, and multimodal resource references corresponding to the nodes. The root node is the core knowledge module, the child nodes are the extended knowledge points, and the branches are the related logic.

[0171] Extract the structural data of the generated knowledge semantic evolution tree. This structural data includes the root node (the core knowledge module "chemotherapy regimen for acute lymphoblastic leukemia in children"), child nodes (various extended knowledge points, such as "chemotherapy regimen for low-risk groups" and "management of adverse reactions to chemotherapy"), branches (the logical relationships between nodes, such as subordinate and causal branches), and multimodal resource references corresponding to the nodes (such as text, image, and video resources pointed to by child nodes).

[0172] Step S1592: Convert the structured data into a standardized tree-like data format. Each node contains a node identifier, knowledge attributes, associated parent node identifiers, a resource reference list, and generation time. Knowledge attributes include core and extension.

[0173] The extracted structural data is converted into a standardized tree-like data format, such as XML or JSON. In this format, each node contains a node identifier (such as a unique string ID), knowledge attributes (core or extension, with the root node being the core and the remaining child nodes being extensions), associated parent node identifier (a node identifier pointing to its parent node), a list of resource references (containing multiple multimodal resource references corresponding to this node), and generation time (a timestamp of the generation of this knowledge semantic evolution tree).

[0174] Step S1593: Extract the data of each association chain in the cross-modal dynamic association chain set generated this time. The association chain data includes chain identifier, resource node sequence, association logic annotation, association transition annotation, comprehensive score, and corresponding teaching stage information.

[0175] Extract the data for each association chain in the generated cross-modal dynamic association chain set. Each association chain data includes a chain identifier (a unique ID that identifies the association chain), a resource node sequence (a set of resource nodes arranged in the order of knowledge evolution), an association logic annotation (a description of the association logic between each resource node and the knowledge seed), an association transition annotation (a description of the association relationship between adjacent resource nodes), a comprehensive score (a weighted comprehensive score for the association chain), and the corresponding teaching stage information (the knowledge level is the clinical application level, and the teaching objective is skills mastery orientation).

[0176] Step S1594: Convert the associated chain data into a standardized chain data format, where each resource node contains a resource identifier, modality type, knowledge point association, and insertion position.

[0177] The associated chain data is converted into a standardized chain data format. Each resource node in this format contains a resource identifier (a unique ID of the multimodal resource in the resource library), modality type (text, image, or video), knowledge point association (a description of the knowledge point corresponding to the resource), and insertion position (the sequential number in the associated chain).

[0178] Step S1595: Call the knowledge association index update interface of the multimodal resource library for pediatric oncology teaching, and input the standardized tree-structured data format and chain-structured data format into the interface.

[0179] Call the knowledge association index update interface of the multimodal resource library for pediatric oncology teaching, and input the standardized tree data format (knowledge semantic evolution tree structure data) and chain data format (cross-modal dynamic association chain data) into the interface to update the knowledge association index.

[0180] Step S1596: After receiving the data, the knowledge association index system first parses the core knowledge modules in the tree data and searches for the historical evolution tree record corresponding to the core knowledge module in the knowledge association index. If a historical evolution tree record exists, the current tree data is added to the history record as a new version, and the version generation time and corresponding teaching stage information are marked. If no historical evolution tree record exists, a new core knowledge module index entry is created, and the current tree data is stored in the entry as the initial version.

[0181] After receiving the input data, the knowledge association indexing system first parses the core knowledge module "chemotherapy regimen for childhood acute lymphoblastic leukemia" in the tree-like data. It then searches the knowledge association index for the corresponding historical evolution tree record for this core knowledge module. If a historical evolution tree record exists (e.g., a knowledge semantic evolution tree for this core knowledge module has been generated previously), the current tree-like data is added to the history as a new version, with the version generation time (current timestamp) and corresponding teaching stage information (clinical application level, skills mastery orientation). If no historical evolution tree record exists, a new core knowledge module index entry is created, and the current tree-like data is stored in that entry as the initial version.

[0182] For example, in step S15961: the knowledge association index system reads the input standardized tree data format and extracts the core knowledge module name and knowledge attributes from the root node. The knowledge attributes include the core.

[0183] The knowledge association indexing system reads the input standardized tree-structured data and extracts the core knowledge module name "chemotherapy regimen for acute lymphoblastic leukemia in children" and the knowledge attribute "core" from the root node data.

[0184] Step S15962: Call the index query module, input the core knowledge module name, and execute the index entry query. The index query module traverses the main entry list of the knowledge association index. Each main entry in the main entry list is uniquely identified by the core knowledge module name. If a main entry that matches the input core knowledge module name is found, it is determined that a historical evolution tree record exists. The evolution tree historical version list in the main entry is read, and the current tree data is added to the end of the evolution tree historical version list. A new version number is assigned to the version added this time. The version number is incremented in order of generation time.

[0185] The index query module is invoked, taking the extracted core knowledge module name "Chemotherapy Regimen for Acute Lymphoblastic Leukemia in Children" as the query parameter, and executing an index entry query. The index query module iterates through the main entry list of the knowledge association index, where each main entry is uniquely identified by its core knowledge module name. If a main entry matching the input core knowledge module name is found, a historical evolutionary tree record is confirmed to exist. The historical version list of the evolutionary tree in that main entry is read, and the current tree data is added to the end of this list. A new version number is assigned to the newly added version, incrementing in chronological order of generation (e.g., if the previous version number was V1.0, the current version number would be V2.0).

[0186] Step S15963: Add version metadata to the tree data of this version. The version metadata includes the version generation time, the corresponding teaching stage information, and the generation source. The version generation time is the current system time. The corresponding teaching stage information is the knowledge level and teaching objective orientation extracted from the retrieval request. The generation source is the identifier of this cross-modal retrieval request. Update the evolution tree historical version list in the main entry, and set this version as the latest version. When a new version is generated in the future, update the latest version identifier again.

[0187] Add version metadata to the tree data for this version, where the version generation time is the current system time (e.g., YYYY-MM-DDHH:MM:SS format). The corresponding teaching stage information is extracted from the retrieval request as "Knowledge Level: Clinical Application Level; Teaching Goal Orientation: Skill Mastery Orientation," and the generation source is a unique identifier for this cross-modal retrieval request (e.g., Request ID). Update the evolutionary tree historical version list in the main entry, add the current tree data containing version metadata to the list, and set this version as the latest version so that subsequent searches will prioritize using the latest version of the evolutionary tree data.

[0188] Step S15964: If the index query module does not find a main entry that matches the input core knowledge module name, it is determined that there is no historical evolution tree record, and a new main entry is created. The structure of the new main entry includes the core knowledge module name, creation time, evolution tree historical version list, related chain sub-entry list, and resource association list. The core knowledge module name is a unique identifier, and the creation time is the current system time.

[0189] If the index query module does not find a master entry matching the input core knowledge module name, it is determined that there is no historical evolution tree record, and a new master entry is created. The structure of the new master entry includes the core knowledge module name (as a unique identifier), creation time (current system time), a list of historical evolution tree versions (initially empty, with version data added later), a list of related chain sub-entries (used to store the related chain data corresponding to the core knowledge module), and a resource association list (a set of associated multimodal resource identifiers).

[0190] Step S15965: Use the current tree data as the initial version of the evolutionary tree historical version list. The version metadata includes the generation time, teaching stage information, and generation source.

[0191] The tree data from this instance will be used as the initial version of the evolutionary tree history version list in the newly created main entry. Its version metadata includes the generation time (current system time), teaching stage information (clinical application level, skills mastery orientation), and generation source (identifier of this search request).

[0192] Step S15966: Insert the new master entry into the master entry list of the knowledge association index. At the same time, add a search entry for the name of the new core knowledge module to the search dictionary of the knowledge association index. The search entry includes the storage location of the master entry in the master entry list, the number of associated resource identifiers, and the generation time of the latest version.

[0193] Insert the newly created master entry into the master entry list of the knowledge association index. At the same time, add a search entry for the new core knowledge module name "chemotherapy regimen for childhood acute lymphoblastic leukemia" to the search dictionary of the knowledge association index. This search entry includes the storage location of the master entry in the master entry list (such as the index address), the number of associated resource identifiers (such as 0 initially, which will be updated as resource associations are added), and the generation time of the latest version (current system time).

[0194] Step S15967: Record the result of creating or updating the main entry in the update log. The entries in the update log include the operation type, the name of the core knowledge module, the number of versions, and the operation time. The operation type includes creation and update.

[0195] The results of creating or updating the main entry are recorded in the update log. The entries in the update log include the operation type (creation or update), the name of the core knowledge module ("chemotherapy regimen for childhood acute lymphoblastic leukemia"), the number of versions (e.g., the number of versions is 1 when created and 2 when updated), and the operation time (current system time).

[0196] Step S1597: Analyze each link in the chain data. Based on the core knowledge module and teaching stage information in the link, search for the corresponding link history in the knowledge link index. If a historical link record exists, compare the current link data with the historical record, retain the link with the higher comprehensive score, and delete duplicate links with the lower comprehensive score. If no historical link record exists, add a link sub-entry in the core knowledge module index entry and store it in the current link data.

[0197] Each link in the chained data is analyzed. Based on the core knowledge module ("chemotherapy regimen for childhood acute lymphoblastic leukemia") and teaching stage information (clinical application level, skills mastery orientation) within the link, the corresponding historical link is searched in the knowledge link index. If a historical link record exists, the current link data is compared with the historical record. For example, if the overall score is compared, the link with the higher overall score is retained, and duplicate links with lower overall scores are deleted (e.g., if the historical link has an overall score of 78 points and the current link has 85 points, the current link is retained, and the historical link is deleted). If no historical link record exists, a new link sub-entry is added to the link sub-entry list of the core knowledge module index entry and stored in the current link data.

[0198] Step S1598: Extract all multimodal resource identifiers involved in the tree-like data and chain-like data, and add a knowledge evolution association tag to each resource identifier in the resource attribute table of the pediatric oncology teaching multimodal resource library. The tag content includes the associated core knowledge module, the generated evolution tree version, and the association chain identifier.

[0199] Extract all multimodal resource identifiers involved in the tree-structured and chain-structured data, such as text resource IDs, image resource IDs, and video resource IDs. In the resource attribute table of the pediatric oncology teaching multimodal resource library, add a knowledge evolution association tag to each resource identifier. The tag content includes the name of the associated core knowledge module ("chemotherapy regimen for childhood acute lymphoblastic leukemia"), the generated evolution tree version (e.g., V2.0), and the association chain identifier (the chain identifier of the association chain to which this resource belongs).

[0200] Step S1599: Count the number of index entries, the number of related chain data, and the number of resource tag updates in this update. The number of index entries includes the number of newly added core knowledge module entries and the number of update history entries. Generate an update statistics report and store the update statistics report in the update log of the knowledge association index. The update log includes the update time, the update operator, and a summary of the update content.

[0201] This update includes statistics on the number of index entries (new core knowledge module entries and update history entries, e.g., 0 new entries and 1 updated entry), the number of related chain data (e.g., 3 new related chains), and the number of resource tag updates (e.g., 20 resource tags were updated). An update statistics report is generated and stored in the knowledge association index update log. The update log includes the update time (current system time), the update operator (automatic system operation, identified as "System"), and a summary of the update content (e.g., "Updated the evolutionary tree version of the core knowledge module 'Chemotherapy Regimen for Acute Lymphoblastic Leukemia in Children,' added 3 related chains, and updated 20 resource tags").

[0202] Step S15910: The knowledge association index system reconstructs the index retrieval path and generates update confirmation information, which is fed back to the pediatric oncology teaching big model to record the update status of the resource database index during this retrieval.

[0203] After completing the above update operations, the knowledge association indexing system reconstructs the index retrieval path to ensure that subsequent searches can quickly and accurately locate the updated knowledge semantic evolution tree and association chain data. Simultaneously, it generates update confirmation information, including whether the update was successful and the main content of the update, which is then fed back to the pediatric oncology teaching model. The pediatric oncology teaching model records the update status of the resource database index for this search, enabling subsequent search optimization and statistical analysis.

[0204] In this embodiment, all data involved are publicly available knowledge and teaching resource data in the field of pediatric oncology education, and do not include privacy-sensitive data. If privacy-sensitive data is collected in actual applications, privacy protection and anti-leakage technologies such as data anonymization technology (e.g., de-identifying patient personal identification information), access control technology (e.g., setting strict permission management, allowing only authorized personnel to access), and encryption technology (e.g., encrypting sensitive data during transmission and storage) will be adopted to ensure that data use complies with relevant laws and regulations.

[0205] In one exemplary embodiment, a cross-modal retrieval system based on a large-scale pediatric oncology teaching model is provided. This cross-modal retrieval system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2 As shown, this cross-modal retrieval system based on a large-scale pediatric oncology teaching model includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a cross-modal retrieval method based on a large-scale pediatric oncology teaching model. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the shell of the cross-modal retrieval system based on the large-scale teaching model of pediatric tumors. It can also be an external keyboard, touchpad, or mouse, etc.

[0206] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A cross-modal retrieval method based on a child tumor teaching large model, characterized in that, The method includes: The system receives a cross-modal retrieval request for pediatric oncology teaching initiated by a user, parses the request, and obtains the retrieval trigger content and teaching stage information. The retrieval trigger content is a single knowledge topic expression in the field of pediatric oncology teaching, and the teaching stage information includes the current knowledge level and teaching objective orientation of the teaching. The retrieval trigger content and the teaching stage information are input into the pediatric oncology teaching model. The pediatric oncology teaching model generates a knowledge seed description corresponding to the retrieval trigger content. The knowledge seed description includes core knowledge modules and knowledge extension directions. Based on the knowledge seed description, the pediatric oncology teaching multimodal resource library is invoked, and the initial multimodal resources associated with the core knowledge modules in the pediatric oncology teaching multimodal resource library are extracted. The semantic evolution processing of the initial multimodal resources is performed through the pediatric oncology teaching big model to generate a knowledge semantic evolution tree. The knowledge semantic evolution tree has the knowledge seed as the root node, the extended knowledge modules as child nodes, and semantic associations as branches. Traverse the knowledge semantic evolution tree, extract the multimodal resources corresponding to each branch, construct a cross-modal dynamic association chain, the cross-modal dynamic association chain connects different modal resources in the order of knowledge evolution, and each resource node is labeled with the association logic with the knowledge seed; Based on the teaching stage information, the association chain screening criteria are determined, cross-modal dynamic association chains that meet the association chain screening criteria are extracted, multimodal resources in the cross-modal dynamic association chains are integrated to generate cross-modal retrieval results, which are pushed to the user interaction terminal. At the same time, the knowledge semantic evolution tree generated this time and the cross-modal dynamic association chains are integrated into the knowledge association index of the pediatric oncology teaching multimodal resource library to update the basic knowledge evolution data for subsequent retrieval. 2.The method of claim 1, wherein, The step of inputting the retrieval trigger content and the teaching stage information into the pediatric oncology teaching model, and generating a knowledge seed description corresponding to the retrieval trigger content through the pediatric oncology teaching model, includes: The large-scale pediatric oncology teaching model calls upon the pediatric oncology teaching knowledge system, which is divided into knowledge levels, including a basic cognition layer, a principle analysis layer, a clinical application layer, and a comprehensive extension layer. Each knowledge level contains multiple knowledge modules, and each knowledge module contains preset core knowledge points and knowledge extension paths. The teaching stage information is analyzed to determine the knowledge level corresponding to the current teaching, and the knowledge depth requirements and knowledge extension scope under the knowledge level are extracted. The knowledge depth requirements include the level of detail of the knowledge points and the coverage of related knowledge. The knowledge extension scope includes vertical in-depth and horizontal association. Vertical in-depth refers to the in-depth expansion of the same module, and horizontal association refers to the related expansion of different modules. The search trigger content is divided into multiple search semantic units. Each search semantic unit is a text fragment that expresses a single knowledge point. Each search semantic unit contains professional terms and expression logic. The knowledge modules in the pediatric oncology teaching knowledge system are traversed, and semantic association analysis is performed between each retrieval semantic unit and the core knowledge points of the knowledge module to determine the knowledge module to which each retrieval semantic unit belongs. The knowledge modules to which all retrieval semantic units belong are statistically analyzed, and the statistical results of the belonging to the same knowledge module are merged to determine the core knowledge module corresponding to the retrieval trigger content. The core knowledge module is the knowledge module to which the most retrieval semantic units belong. Based on the aforementioned knowledge depth requirements, key knowledge points are extracted from the core knowledge modules. These key knowledge points are the content that the core knowledge module needs to cover in the current knowledge level. Based on the aforementioned knowledge extension scope, the extension direction of the core knowledge module is determined. The vertical in-depth direction corresponds to the subdivided knowledge points of the core knowledge module, and the horizontal association direction corresponds to other knowledge modules that have logical connections with the core knowledge module. The core knowledge modules, key knowledge points, and extension directions are integrated to form a knowledge seed description. The knowledge seed description also includes the modal resource requirements corresponding to each extension direction. The modal resource requirements are determined based on the knowledge presentation format, with conceptual knowledge corresponding to text resources and operational knowledge corresponding to video resources. 3.The method of claim 2, wherein, The process of traversing the knowledge modules in the pediatric oncology teaching knowledge system involves performing semantic association analysis between each retrieval semantic unit and the core knowledge points of the knowledge module to determine the knowledge module to which each retrieval semantic unit belongs, including: The first knowledge level is retrieved from the pediatric oncology teaching knowledge system, and all knowledge modules under this knowledge level are extracted to form a knowledge module list. Take the first retrieval semantic unit, perform vocabulary extraction on the retrieval semantic unit to obtain a retrieval vocabulary set. The retrieval vocabulary set includes professional terms, logical expressions, and knowledge attribute words in the field of pediatric oncology teaching. The knowledge attribute words include etiology, pathology, and treatment. Read the first knowledge module in the knowledge module list, extract the core knowledge points of the knowledge module, and form a core knowledge point list. Each core knowledge point contains standard expression terms and knowledge attribute tags. Perform association analysis between the search term set and the standard expression terms of each core knowledge point, and record the number and type of overlapping terms, including overlapping professional terms and overlapping knowledge attribute terms; Analyze the consistency between the expression logic of the retrieval semantic unit and the standard logic of the core knowledge point, extract the logical connectives between the two, including "because", "therefore", and "steps", and record whether the semantic direction of the logical connectives is consistent; Based on the number, type, and logical fit of overlapping words, determine the degree of association between the retrieval semantic unit and the current core knowledge point. If at least one core knowledge point has a degree of association with the retrieval semantic unit that is greater than the set degree, then the retrieval semantic unit is assigned to the current knowledge module. Otherwise, read the next knowledge module in the knowledge module list and repeat the above association analysis steps. If none of the knowledge modules at the current knowledge level can be associated with the search semantic unit, then the next knowledge level is read, and the above knowledge module traversal and association analysis steps are repeated until a corresponding knowledge module is found for each search semantic unit. If no matching knowledge module is found at any knowledge level, then the search semantic unit is marked as a knowledge unit to be supplemented, and the type of knowledge module to be expanded is noted in the knowledge seed description. 4.The method of claim 1, wherein, Based on the knowledge seed description, the process involves calling the pediatric oncology teaching multimodal resource library, extracting initial multimodal resources associated with the core knowledge modules from the library, and performing semantic evolution processing on the initial multimodal resources through the pediatric oncology teaching large model to generate a knowledge semantic evolution tree, including: Call the resource retrieval interface of the multimodal resource library for pediatric oncology teaching, input the core knowledge module name and key knowledge points in the knowledge seed description, and execute the resource retrieval; The resource retrieval interface is used to traverse all multimodal resources in the resource library and extract the knowledge module tags and knowledge point tags for each resource. The knowledge module tags and knowledge point tags are attribute information labeled based on the pediatric oncology teaching knowledge system. Multimodal resources whose knowledge module tags match the core knowledge module name and whose knowledge point tags contain key knowledge points are selected as initial multimodal resources. These initial multimodal resources include text resources, image resources, and video resources. The initial multimodal resources are classified according to modality type to obtain initial text resource group, initial image resource group, and initial video resource group; The initial text resource group is input into the text semantic evolution unit of the pediatric oncology teaching model. The semantic decomposition of the text resources is performed, and the core semantic components of each text resource are extracted. Based on the extension direction described by the knowledge seed, the core semantic components are extended and expanded to generate text semantic extension branches. Each text semantic extension branch corresponds to an extended knowledge point. The core semantic components include concepts, principles, and cases. The initial image resource group is input into the image semantic evolution unit of the pediatric oncology teaching model. The visual semantic transformation of the image resources is performed, and the visual features in the images are converted into textual visual semantic descriptions. Combined with the extension direction of the knowledge seed description, the extended knowledge points are deduced based on the visual semantic descriptions to generate image semantic extension branches. The visual features include tumor morphology and cell structure. The initial video resource group is input into the video semantic evolution unit of the pediatric oncology teaching model. Frame sequence analysis and audio conversion of the video resources are performed to extract dynamic visual semantics and audio text semantics from the video. Combined with the extension direction described by the knowledge seed, the dynamic visual semantics and audio text semantics are integrated to generate extended knowledge points and form video semantic extension branches. The dynamic visual semantics include surgical operation steps and case demonstration processes, and the audio text semantics include explanatory statements and key terms. Collect text semantic extension branches, image semantic extension branches, and video semantic extension branches. Take the core knowledge module in the knowledge seed description as the root node and the extended knowledge points corresponding to each extension branch as child nodes to construct an initial knowledge semantic evolution structure. Delete child nodes in the initial knowledge semantic evolution structure that have no logical connection with the core knowledge module of the root node. Logical connections include subordination, causality, and application. Add a corresponding multimodal resource reference to each child node, add an association logic description to each branch, and generate a knowledge semantic evolution tree. The multimodal resource reference points to the initial multimodal resource that generated the child node, and the association logic description is used to describe the logical relationship between the child node and the parent node. 5.The method of claim 4, wherein, The process involves inputting the initial text resource group into the text semantic evolution unit of the pediatric oncology teaching model, performing semantic decomposition of the text resources, extracting the core semantic components of each text resource, and extending and expanding the core semantic components based on the extension direction described by the knowledge seed to generate text semantic extension branches, including: Take the first text resource in the initial text resource group, divide it into multiple text paragraph units, and each text paragraph unit is a continuous semantically complete text; Each text paragraph unit is semantically annotated, with the annotation content being the semantic type corresponding to that paragraph unit. Semantic types include concept type, principle type, and case type. Extract definitions, attribute descriptions, and classification information from concept-based paragraphs to form conceptual semantic components; extract mechanism explanations, causal relationship descriptions, and process descriptions from principle-based paragraphs to form principle semantic components; extract case backgrounds, treatment processes, and result descriptions from case-based paragraphs to form case semantic components. Collect conceptual semantic components, principle semantic components, and case semantic components to form a set of core semantic components of the current text resource; Read the extension direction described by the knowledge seed. If the extension direction is vertical in-depth, then for each component in the core semantic component set, extract the subdivision content of that component. The subdivision content includes the subdivision type of the concept and the subdivision mechanism of the principle, and generate vertically extended knowledge points. If the extension direction is horizontal association, then based on the core semantic component set, find the knowledge points of other knowledge modules in the pediatric oncology teaching knowledge system that are logically related to the component, and generate horizontally extended knowledge points. The vertically or horizontally extended knowledge points are arranged in a semantically related order. Each knowledge point corresponds to a text resource fragment, which comes from the current text resource or the text resource of the related knowledge module, forming a text semantic extension branch. The process involves inputting the initial image resource set into the image semantic evolution unit of the pediatric oncology teaching model, performing visual semantic transformation of the image resources, converting visual features in the images into textual visual semantic descriptions, combining the extension directions of knowledge seed descriptions, deriving extended knowledge points based on the visual semantic descriptions, and generating image semantic extension branches, including: The first image resource in the initial image resource group is taken, converted into digital image format, and image enhancement processing is performed. The enhanced image is then divided into regions using a region segmentation algorithm to identify regions containing visual information related to the core knowledge module. These regions include tumor lesion regions and cell structure regions, which are used as key visual regions. Extract visual features of key visual regions, including morphological features, color features, and structural features. Morphological features include outline and shape, color features include dominant color and color distribution, and structural features include the distribution of internal components and the connection relationship between components. By applying the visual-text mapping rules in the field of pediatric oncology education, the extracted visual features are converted into textual visual semantic descriptions, which include the specific content of the features and their corresponding knowledge attributes. Read the extended direction of the knowledge seed description. If it is a vertical in-depth, then based on the visual semantic description, deduce the formation cause, evolution process and clinical significance of the visual feature, and generate vertically extended knowledge points. If it is a horizontal association, then based on the visual semantic description, find the knowledge points of other knowledge modules related to this visual feature in the pediatric oncology teaching knowledge system, and generate horizontally extended knowledge points; For each extended knowledge point, a corresponding image resource fragment is matched. The image resource fragment comes from the key visual area of ​​the current image resource and is arranged in the logical order of the knowledge points to form an image semantic extension branch. Repeat the above steps to process all resources in the initial text resource group and the initial image resource group to generate multiple text semantic extension branches and image semantic extension branches. 6.The method of claim 1, wherein, The process of traversing the knowledge semantic evolution tree, extracting multimodal resources corresponding to each branch, and constructing cross-modal dynamic association chains includes: Starting from the root node of the knowledge semantic evolution tree, traverse all branches. Each branch is a knowledge evolution path extending from the root node to the terminal child node. For each branch, extract the multimodal resource references corresponding to all child nodes on the knowledge evolution path, and determine the specific resource pointed to by each reference. The specific resources include text resource fragments, image resource fragments, and video resource fragments. The extracted multimodal resources are sorted according to the knowledge evolution order, which is from the core knowledge module corresponding to the root node to the extended knowledge point corresponding to the terminal child node. The sorting is based on the logical progression relationship of the knowledge points. Analyze the modality distribution of the sorted multimodal resources. If there are multiple consecutive resources with the same modality and the knowledge content of the resources is repeated, then retain the resource with the most comprehensive knowledge coverage and delete the other duplicate resources. If a modality type is missing, the corresponding modality resources will be retrieved from the multimodal resource library for pediatric oncology teaching based on the modality resource requirements described by the knowledge seed, and inserted into the corresponding knowledge evolution stage. Add a logical association label to each resource node. The content of the logical association label is the logical relationship between the knowledge point corresponding to the resource and the core knowledge module in the knowledge seed. At the same time, add a transitional association label to adjacent resource nodes. The content of the transitional association label is the knowledge point association relationship between the previous resource node and the next resource node. The sorted resource nodes, associated logic annotations, and associated transition annotations are integrated to form a cross-modal dynamic association chain; the above steps are repeated to traverse all branches of the knowledge semantic evolution tree and generate multiple cross-modal dynamic association chains. Collect cross-modal dynamic association chains with complete evolution paths, no missing resource nodes, and no omissions in association annotations to form a set of cross-modal dynamic association chains.

7. The cross-modal retrieval method based on a large-scale pediatric oncology teaching model according to claim 6, characterized in that, The process of sorting the extracted multimodal resources according to the knowledge evolution order, and the process of supplementing the corresponding modal resources from the pediatric oncology teaching multimodal resource library based on the modal resource requirements described by the knowledge seed, and inserting them into the corresponding knowledge evolution stage, includes: Extract the knowledge points corresponding to all child nodes on the current branch, arrange them in the order of knowledge evolution to obtain a knowledge point sequence. The beginning of the knowledge point sequence is the key knowledge point of the core knowledge module, and the subsequent ones are the extended knowledge points in the extension direction. Analyze the knowledge attributes of each knowledge point in the knowledge point sequence. Knowledge attributes include conceptual, principle-based, operational, and case-based. The logical progression of knowledge points is determined based on knowledge attributes and teaching principles. The progression order between the conceptual, principle-based, operational, and case-based types defined by the teaching principles is the conventional progression order. When there are special requirements for the extension direction of the knowledge seed description, the progression order is adjusted according to the special requirements. Based on the adjusted progressive order, the knowledge point sequence is reordered to obtain an ordered knowledge point sequence; Extract the multimodal resources corresponding to each knowledge point in the ordered knowledge point sequence, arrange them according to the knowledge point sorting order, and obtain the initial version of the resource sorting; Check whether the knowledge points of adjacent resources in the initial resource sorting conform to a logical progression. If the order is reversed, adjust the resource positions to conform to the progression order. After completing the resource sorting, the modal types included in the resource sorting are counted. The modal types include text, images, and videos. The modal resource requirements described by the knowledge seeds are compared. If any knowledge point's corresponding evolutionary stage lacks the modal type required, then the text description of that knowledge point will be used as the search condition to call the supplementary search interface of the Pediatric Oncology Teaching Multimodal Resource Library. The supplementary retrieval interface traverses the resource library, filters out corresponding modal resources that are semantically related to the text description of the knowledge point, and extracts the core content fragments of the corresponding modal resources; Insert the supplementary resource fragments into the corresponding knowledge point evolution stage in the resource sorting, mark the supplementary resource with a supplementary resource identifier, and explain the reason for the supplementation; Repeat the above supplementary steps until all knowledge point evolution stages in the resource sorting meet the modal resource requirements, and finally obtain a resource sequence arranged in the order of knowledge evolution with complete modal types, which serves as the resource foundation for cross-modal dynamic association chains.

8. The cross-modal retrieval method based on a large-scale pediatric oncology teaching model according to claim 1, characterized in that, The process of determining the association chain screening criteria based on the teaching stage information, extracting cross-modal dynamic association chains that meet the screening criteria, integrating multimodal resources in the cross-modal dynamic association chains to generate cross-modal search results, and pushing them to the user interaction terminal includes: The knowledge levels in the teaching information are analyzed, and different knowledge levels correspond to different resource focuses: the basic cognition level focuses on the completeness of the concepts of the resources, the principle analysis level focuses on the detail of the principles of the resources, the clinical application level focuses on the practicality of the cases of the resources, and the comprehensive expansion level focuses on the breadth of the connections of the resources. The teaching objectives in the information of the teaching stage are analyzed, and different objectives correspond to different resource types with different weights: text resources are given high weight for knowledge understanding, video resources are given high weight for skills mastery, and multimodal resources are given balanced weights for comprehensive application. Based on the resource emphasis direction corresponding to the knowledge level and the resource type weight corresponding to the teaching objective orientation, the screening criteria for the association chain are determined. The screening criteria for the association chain include three dimensions: resource completeness, modal adaptability, and logical coherence. Resource completeness refers to the proportion of knowledge points covered by resources in the chain; modal adaptability refers to the degree of matching between the resource type in the chain and the objective orientation; and logical coherence refers to the logical clarity of the association annotations in the chain. The resource completeness dimension is scored according to the ratio of the number of knowledge points covered by resources to the total number of knowledge points; the modal adaptability dimension is scored according to the proportion of resource types that conform to the objective orientation; and the logical coherence dimension is scored according to the completeness and accuracy of the association annotations. In this way, the scoring rules for each dimension are determined. Traverse each link in the cross-modal dynamic link set, score the link in three dimensions according to the scoring rules, assign weights to each dimension based on the resource emphasis direction and resource type weight, calculate the weighted comprehensive score of the three dimensions, and select the links with the weighted comprehensive score higher than the corresponding threshold as cross-modal dynamic links that meet the standards. If the number of qualified related chains exceeds the preset number, they are sorted from highest to lowest by weighted comprehensive score, and the first preset number of related chains are selected. If the number of qualified related chains is lower than the preset number, the comprehensive score threshold is lowered, but the reduction does not exceed the preset percentage, and the screening is repeated until the preset number is reached or it is confirmed that there are no more qualified related chains. Extract all multimodal resources from the association chains that meet the criteria, integrate them according to the sorting order of the association chains, and form a resource package for each association chain. Add a description of the search results to each resource package, which includes the comprehensive score of the association chain, the scores of each dimension, and the applicable teaching scenarios. The applicable teaching scenarios are derived from the information of the teaching stage. At the same time, add resource usage suggestions for each resource, which include when to use the resource in teaching. Integrate all resource packages, search result descriptions, and resource usage suggestions to form cross-modal search results, and convert the cross-modal search results into a display format supported by the user interaction terminal. The display format includes resource preview thumbnails, resource names, logical diagrams of related chains, and search result description text. The converted cross-modal search results are pushed to the user's interactive terminal so that the terminal interface displays resource packages from high to low according to the comprehensive score of the association chain. Each resource package provides operation entry points for viewing, downloading, and collecting resources. At the same time, the terminal interface displays search result descriptions and resource usage suggestions.

9. The cross-modal retrieval method based on a large-scale pediatric oncology teaching model according to claim 1, characterized in that, The process of integrating the generated knowledge semantic evolution tree and cross-modal dynamic association chain into the knowledge association index of the multimodal resource library for pediatric oncology teaching, and updating the basic knowledge evolution data for subsequent searches, includes: Extract the structural data of the generated knowledge semantic evolution tree. The structural data includes the root node, child nodes, branches, and multimodal resource references corresponding to the nodes. The root node is the core knowledge module, the child nodes are the extended knowledge points, and the branches are the related logic. The structured data is converted into a standardized tree-like data format. Each node contains a node identifier, knowledge attributes, related parent node identifiers, a list of resource references, and a generation time. The knowledge attributes include core and extension. Extract the data of each association chain in the generated cross-modal dynamic association chain set. The association chain data includes the chain identifier, resource node sequence, association logic annotation, association transition annotation, comprehensive score, and corresponding teaching stage information. The associated chain data is converted into a standardized chain data format, with each resource node containing a resource identifier, modality type, knowledge point association, and insertion position; Call the knowledge association index update interface of the multimodal resource library for pediatric oncology teaching, and input the standardized tree-structured data format and chain-structured data format into the interface; After receiving the data, the knowledge association indexing system first parses the core knowledge modules in the tree data and searches for the corresponding historical evolution tree record in the knowledge association index. If a historical evolution tree record exists, the current tree data is added to the history record as a new version, and the version generation time and corresponding teaching stage information are marked. If no historical evolution tree record exists, a new core knowledge module index entry is created, and the current tree data is stored in the entry as the initial version. Each link in the chained data is analyzed. Based on the core knowledge module and teaching stage information in the link, the corresponding link history is searched in the knowledge link index. If a historical link record exists, the current link data is compared with the historical record. Links with higher comprehensive scores are retained, and duplicate links with lower comprehensive scores are deleted. If no historical link record exists, a link sub-entry is added to the core knowledge module index entry and stored in the current link data. Extract all multimodal resource identifiers involved in the tree-like data and chain-like data, and add a knowledge evolution association tag to each resource identifier in the resource attribute table of the pediatric oncology teaching multimodal resource library. The tag content includes the associated core knowledge module, the generated evolution tree version, and the association chain identifier. The number of index entries, the number of related chain data, and the number of updated resource tags are counted in this update. The number of index entries includes the number of newly added core knowledge module entries and the number of update history entries. An update statistics report is generated and stored in the update log of the knowledge association index. The update log includes the update time, the update operator, and a summary of the update content. The knowledge association indexing system reconstructs the index retrieval path and generates update confirmation information, which is then fed back to the pediatric oncology teaching model to record the update status of the resource database index during this retrieval.

10. A cross-modal retrieval system based on a large-scale pediatric oncology teaching model, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the cross-modal retrieval method based on a large-scale pediatric oncology teaching model as described in any one of claims 1 to 9 by executing the machine-executable instructions.

Images