An intelligent management method for dissection teaching data

Abstract

The application relates to the technical field of data processing, in particular to an intelligent management method and system for dissection teaching data; the method comprises the following steps: acquiring dissection knowledge data, wherein the dissection knowledge data comprises three-dimensional images and text data; extracting dissection technical terms and the correlation between the dissection technical terms in the dissection knowledge data; taking the dissection technical terms as nodes and the correlation as edges to construct a multi-modal dissection knowledge graph; collecting historical interaction data of a learner; based on a learner model coupled with an Ebbinghaus forgetting curve and Bayesian inference, calculating a knowledge mastery probability vector of the learner's mastery degree of each node in the dissection knowledge graph; in response to any component in the knowledge mastery probability vector being lower than a preset threshold, using a Monte Carlo tree search algorithm in the dissection knowledge graph to plan out a personalized learning path. The application has the effects of improving learning efficiency and enhancing the firmness of knowledge mastery.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to an intelligent management method and system for dissection teaching data. BACKGROUND

[0002] Anatomy is the cornerstone of medical education, and its teaching and learning process highly depends on multi-modal data, including text description in textbooks, two-dimensional dissection atlas, three-dimensional medical images such as CT or MRI, and interactive three-dimensional models, etc. In the related art, these different modal teaching resources are usually stored and used independently, forming an "information island". For example, when a student reads the description of an organ in the text, it is difficult for him to quickly and intuitively locate the organ in the three-dimensional model and associate it with the relevant image slices and physiological function information. The fragmentation of this data organization makes it difficult to systematically express and utilize the complex spatial topological relationships (such as adjacency, inclusion) between anatomical structures, functional innervation relationships (such as innervating muscles), and tissue hierarchical relationships (such as organs, tissues, and cells).

[0003] On the other hand, traditional anatomy teaching methods, whether offline or existing online teaching platforms, mostly adopt a "one-size-fits-all" knowledge infusion mode, providing the same learning content and fixed learning path for all learners. This mode ignores the individual differences of learners, such as prior knowledge level, cognitive ability, and forgetting regularity. The system cannot assess the learners' mastery of each knowledge point in real time and accurately, nor can it dynamically adjust the teaching strategy according to their weak knowledge areas. When learners encounter difficulties in the complex anatomy knowledge system, the existing system lacks intelligent guidance mechanisms and cannot plan a personalized learning path that meets their cognitive rules and knowledge needs, resulting in low learning efficiency and insecure knowledge mastery.

[0004] In summary, the following technical problems still exist in the medical education process in the related art: 1. It is difficult to effectively integrate and manage these heterogeneous anatomy data and establish deep semantic associations between them. 2. It is difficult to achieve adaptive teaching guidance based on different learners and provide personalized learning path planning. SUMMARY

[0005] The present application provides an intelligent management method and system for dissection teaching data to solve the problems of difficulty in effectively integrating and managing these heterogeneous anatomy data and difficulty in achieving adaptive teaching guidance based on different learners in the related art.

[0006] In a first aspect, the present application provides an intelligent management method for dissection teaching data, which adopts the following technical solution:

[0007] The method comprises the following steps: acquiring anatomical knowledge data, wherein the anatomical knowledge data comprises three-dimensional image data and text data; extracting anatomical technical terms and the correlation between the anatomical technical terms in the anatomical knowledge data; taking the anatomical technical terms as nodes and the correlation as edges to construct a multi-modal anatomical knowledge graph;

[0008] The three-dimensional image data is segmented into organs and tissues to obtain spatial position codes representing the centers, directions and bounding boxes of each human tissue or organ; a bidirectional long short-term memory network model is used to obtain word embedding vectors corresponding to the nodes; a neural network model is used to fuse the spatial position codes and the word embedding vectors to obtain heterogeneous data, and the heterogeneous data is bound to the nodes;

[0009] The historical interaction data of the learner is collected, and a learner model coupled with the Ebbinghaus forgetting curve and Bayesian inference is used to calculate a knowledge mastery probability vector of the learner for each node in the anatomical knowledge graph;

[0010] When any component in the knowledge mastery probability vector is lower than a preset threshold, a Monte Carlo tree search algorithm is used in the anatomical knowledge graph to plan a personalized learning path.

[0011] By fusing three-dimensional image data and text data, extracting anatomical technical terms and their correlation, constructing a multi-modal knowledge graph, and fusing spatial position codes and word embedding vectors of anatomical entities to form heterogeneous data, and then combining a learner model and a Monte Carlo tree search to plan a personalized learning path, an anatomical knowledge graph is constructed based on anatomical technical terms and correlation, so that nodes are correlated with each other, that is, knowledge points are correlated with each other. On the one hand, it provides operation corresponding for subsequent adaptive learning path, and on the other hand, it solves the problem of information island in traditional technology and solves the problem of difficult linkage of different data in traditional technology. The neural network is used to fuse the spatial position codes and the word embedding vectors to improve the semantic integrity of the nodes. Then the fused heterogeneous data is bound to the nodes to realize the deep correlation between the three-dimensional structure and the semantic knowledge.

[0012] During the learning process of the learner, the knowledge mastery probability of the learner for the knowledge points is simulated according to the Ebbinghaus forgetting curve to judge the mastery degree of the learner for the knowledge, and for the weak links, a Monte Carlo tree search algorithm is used to plan a personalized learning path. Then, the idea of reinforcement learning is further introduced, the personalized path scheme is output by maximizing the expected knowledge gain in multiple rounds of simulation, and the defects of one-size-fits-all in traditional teaching are effectively avoided to provide different knowledge for different learners.

[0013] Optionally, the association relationship includes spatial topology, functional dominance and organizational hierarchy, wherein the functional dominance and organizational hierarchy relationship between the nodes is extracted by performing natural language processing on the text data; the three-dimensional image is segmented, and the spatial topology relationship between the nodes is determined based on the segmentation result.

[0014] The association relationship includes spatial topology, functional dominance and organizational hierarchy. Such relationship covers most of the relationships in the field of anatomy and can accurately describe the internal logic between nodes, improve the professionalism and interpretability of the anatomy knowledge graph structure, and meet the high-precision requirements of medical teaching on knowledge.

[0015] Optionally, the three-dimensional image is segmented using a cascaded 3DU-Net network.

[0016] The 3DU-Net network is mainly used for processing volume graphs, so the three-dimensional image can be segmented using the network.

[0017] Optionally, the step of segmenting the three-dimensional image at the pixel level using a cascaded 3DU-Net network includes: using a three-dimensional convolutional network in the first stage to perform rough positioning on the input complete three-dimensional image, and outputting an initial three-dimensional region containing the target organ; taking the initial three-dimensional region as the input of a three-dimensional attention convolutional network in the second stage, performing fine segmentation within the initial three-dimensional region, capturing subtle features of organ boundaries and internal textures through the attention mechanism in the three-dimensional attention convolutional network, and generating an organ segmentation mask.

[0018] In the 3DU-Net network, a two-stage segmentation method is adopted, the first node performs rough positioning on the volume image (three-dimensional image), and then high-precision segmentation is performed through a three-dimensional attention convolutional network, thereby improving the accuracy of three-dimensional image segmentation.

[0019] Optionally, the step of binding the heterogeneous data to the nodes includes: calculating the cosine similarity between the preset vectors of the heterogeneous data and each node in the knowledge graph, and binding the node with the highest similarity to the heterogeneous data.

[0020] The cosine similarity between the preset vectors of the heterogeneous data and each node in the knowledge graph is calculated, and the node with the highest similarity is bound to the heterogeneous data, realizing automatic matching and structural fusion of multi-modal information at the semantic level.

[0021] Optionally, the historical interaction data of the learner is collected, and based on a learner model coupled with the Ebbinghaus forgetting curve and Bayesian inference, the step of calculating the knowledge mastery probability vector of the learner's mastery of each node in the anatomy knowledge graph includes:

[0022] For any node, the history interaction data of the learner, such as the correctness of the answers, the model interaction time, the data search frequency, etc., is continuously recorded;

[0023] For any moment, the knowledge mastery probability of the node at the current moment is obtained by using the Ebbinghaus forgetting curve formula; in response to the new interaction between the learner and the model, the knowledge mastery probability of the node is updated by using the Bayes formula, combining the result of the interaction and the preset guess probability and error probability of the learner; the knowledge mastery probabilities of multiple nodes constitute the mastery probability vector of the learner on the anatomy knowledge graph.

[0024] In this method, for different nodes, which can also be understood as different knowledge points, the knowledge mastery probability of the learner after learning the knowledge point is predicted by the Ebbinghaus forgetting curve, and the knowledge mastery probability of the knowledge point is updated after the learner and the model produce new interaction (such as reviewing the action of a knowledge point) according to the Bayes formula, to obtain a dynamic mastery probability vector, which is convenient for mastering the differences in the mastery of the anatomy knowledge graph by different learners, and thus can provide personalized learning paths for different learners.

[0025] Optionally, the steps of planning a personalized learning path using the Monte Carlo tree search algorithm include:

[0026] In response to detecting that the knowledge mastery probability of the learner on the phrenic nerve node is lower than a preset threshold, the Monte Carlo tree search algorithm takes the current node as the root node;

[0027] The nodes connected to the root node in the anatomy knowledge graph are taken as expansion nodes;

[0028] Starting from a new expansion node, a fast random strategy is used to simulate a learning path in the anatomy knowledge graph, and a knowledge gain value of this simulation is calculated, which is the change amplitude of the learner's knowledge mastery probability vector before and after the simulation;

[0029] The knowledge gain value is used to back-propagate all parent nodes along the learning path;

[0030] After completing a specified number of Monte Carlo tree searches, the path starting from the root node with the highest cumulative knowledge gain value is selected as the final personalized learning path.

[0031] This process is the main step of the Monte Carlo tree search algorithm, mainly used for selecting a personalized learning path for the learner.

[0032] Optionally, the number of iterations of the Monte Carlo tree search is 2000.

[0033] The spatial position encoding and the word embedding vector are fused by using the cross-modal Transformer model to obtain heterogeneous data.

[0034] The transformer model is a deep learning model architecture for processing sequence-to-sequence tasks, through which spatial position encoding and word embedding vectors can be fused.

[0035] In another aspect, the present application discloses an intelligent management system for anatomy teaching data, which adopts the following technical scheme:

[0036] An intelligent management system for anatomy teaching data comprises a processor and a memory, and the memory stores computer program instructions which, when executed by the processor, implement the intelligent management method for anatomy teaching data according to the above.

[0037] The intelligent management method for anatomy teaching data is generated into a computer program and stored in the memory to be loaded and executed by the processor, so that the system is made according to the memory and the processor, and is convenient to use.

[0038] The present application has the following technical effects:

[0039] The anatomy knowledge graph is constructed according to the association relationship between the names of anatomy techniques and the names of anatomy technical terms, different dimensional data are effectively integrated and managed, and the isolation between knowledge points is reduced. Meanwhile, the personalized learning path is specified for learners based on the difference in the mastery of the anatomy knowledge graph by different learners, so that the learners can master the anatomy knowledge more quickly, the knowledge infusion mode of “one size fits all” in the traditional teaching method is reduced, and the learning efficiency is improved. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 is a method flowchart of an anatomy teaching data intelligent management method according to an embodiment of the present application.

[0041] Figure 2 is a method flowchart of step 1 of an anatomy teaching data intelligent management method according to an embodiment of the present application.

[0042] Figure 3 is a method flowchart of step 4 of an anatomy teaching data intelligent management method according to an embodiment of the present application. DETAILED DESCRIPTION

[0043] An embodiment of the present application discloses an intelligent management method for anatomy teaching data,

[0044] Referring to Figure 1 An intelligent management method for anatomy teaching data comprises steps S1-S3.

[0045] S1: Obtain anatomical knowledge data, wherein the anatomical knowledge data comprises three-dimensional images and text data, anatomical technical terms and the association relationship between the anatomical technical terms are extracted from the anatomical knowledge data, the anatomical technical terms are taken as nodes, and the association relationship is taken as edges to construct a multi-modal anatomical knowledge graph.

[0046] The nodes in the anatomical knowledge graph comprise anatomical entities, physiological functions and pathological states, other forms of nodes can also be selected in other embodiments, or some nodes are reduced for the simplification of the anatomical knowledge graph; the association relationship between the nodes is taken as edges of the anatomical knowledge graph, and the association relationship comprises spatial topology, functional domination and organizational hierarchy.

[0047] When constructing the knowledge graph, the knowledge data based on the anatomical aspect is used, wherein the knowledge data comprises text data of textbooks and three-dimensional image data. Therefore, different anatomical knowledge data is processed differently, anatomical technical terms are extracted from the text data, and the association relationship between the anatomical technical terms is analyzed from the semantics, so as to construct the anatomical knowledge graph. The spatial topology relationship between the anatomical technical terms (nodes, it should be noted that only the anatomical entity has a physical form and a spatial position relationship among the anatomical technical terms) can be obtained from the three-dimensional image data.

[0048] For the relationship such as functional domination and hierarchical organization, the natural language processing is performed on the text data in the method to extract the functional domination and hierarchical organization relationship between the nodes.

[0049] Reference Figure 2 Step S1 comprises steps S11-S12.

[0050] S11: The anatomical technical terms in the text data are recognized by using a named entity recognition technology, and the technical terms are taken as nodes of the anatomical knowledge graph.

[0051] The anatomical entities (liver, spleen, stomach), physiological functions (respiratory movement, digestive function) and pathological states (liver cirrhosis, aneurysm, appendicitis) are extracted from textbooks such as Gray's Anatomy and medical ontology libraries such as SNOMED CT as nodes by using a named entity recognition (NER) technology.

[0052] The sentences containing the anatomical technical terms are analyzed by using a relationship extraction model, and the relationship such as functional domination and hierarchical organization is extracted therefrom.

[0053] The relationship extraction model can adopt a supervised learning model based on a pre-trained language model (such as BERT, BioBERT) and combined with a conditional random field (CRF) or an attention mechanism, which is a conventional technical means in the field and will not be described here.

[0054] For example, the branch of the vagus nerve dominates the activity of the heart, so a functional dominance relationship is extracted, that is, a edge from the vagus nerve node to the heart node is connected in the anatomical knowledge graph and marked as functional dominance. Similarly, the information that liver lobule is the basic unit of liver is extracted, and a organizational hierarchical relationship from liver lobule to liver is established.

[0055] S12: segmenting the three-dimensional image, and determining the spatial topological relationship between nodes based on the segmentation result.

[0056] For the three-dimensional image, multiple organ segmentation masks are obtained by segmenting different organs (intuitively, it can also be understood that the whole model is segmented into multiple organ models, and each organ model corresponds to an organ or human tissue, and the organ segmentation mask is a storage form of the segmented organ model). The centroid coordinates, volume, surface area and circumscribed bounding box of each organ model are calculated. By comparing the positions and distances of the two bounding boxes, it can be inferred that the gallbladder is attached to the lower part of the liver, and a spatial adjacency relationship is established in the graph, that is, the liver node and the gallbladder node are connected by an edge marked as spatial adjacency.

[0057] S2: segmenting the three-dimensional image data into organs and tissues, obtaining the spatial position code representing the centroid, direction and bounding box of each human tissue or organ; using a bidirectional long short-term memory network model to obtain a word embedding vector corresponding to a node; using a neural network model to fuse the spatial position code and the word embedding vector to obtain heterogeneous data, and binding the heterogeneous data and the node.

[0058] In the step of segmenting the three-dimensional image data into organs and tissues to obtain the spatial position code representing the centroid, direction and bounding box of each human tissue or organ: a cascaded 3D U-Net network is used to perform pixel-level segmentation on the three-dimensional image, and finally the centroid, direction and bounding box are calculated according to the organ segmentation mask to generate the spatial position code.

[0059] In the actual segmentation process: the cascaded 3D U-Net network uses the first-stage three-dimensional convolutional network to roughly locate the input complete three-dimensional image, and outputs an initial three-dimensional region containing the target organ; the initial three-dimensional region is taken as the input of the second-stage three-dimensional attention convolutional network, and fine segmentation is performed within the initial three-dimensional region. The attention mechanism in the three-dimensional attention convolutional network is used to capture the subtle features of the organ boundary and internal texture, and a pixel-level organ segmentation mask is generated; after segmentation, the centroid, direction and bounding box are calculated according to the segmentation mask to generate the spatial position code.

[0060] For example, when processing a CT image of an abdomen with a resolution of 512 by 512 by 320 pixels to segment the pancreas, the first-stage 3D convolutional network first down-samples the original image, e.g., to 128 by 128 by 80 pixels, to reduce computational complexity. The network quickly processes the down-sampled image and outputs a 3D region roughly containing the pancreas, e.g., an initial 3D region of 96 by 96 by 64 pixels. This step reduces the subsequent fine segmentation computation from over 80 million pixels to less than 600,000 pixels, greatly improving processing efficiency.

[0061] The 96 by 96 by 64 pixel initial 3D region is then fed into the second-stage 3D attention convolutional network. The network structure is embedded with an attention module, enabling it to learn and focus on feature regions more important to the segmentation task. For example, when segmenting the pancreas, the attention mechanism assigns higher weights to the ambiguous boundary regions of the pancreas and neighboring organs such as the duodenum and spleen, and also pays attention to texture changes within the pancreas caused by ducts or blood vessels running through it. The 3D attention convolutional network can more accurately distinguish the target from the background and generate a pixel-level mask of the same size as the input region, in which each pixel is accurately labeled as pancreas or non-pancreas, with a Dice coefficient of segmentation accuracy of, for example, over 0.9.

[0062] In the step of obtaining the word embedding vector corresponding to the node by using the bidirectional long short-term memory network model, the bidirectional long short-term memory network model is a BiLSTM-CRF model. The model is used to process the teaching material content to identify the anatomical technology name, i.e., the node. Then, the Word2Vec model pre-trained on a large amount of medical literature is used to convert it into a word embedding vector.

[0063] For example, from the sentence "the phrenic nerve innervates the diaphragm", the entity terms "phrenic nerve" and "diaphragm" are identified.

[0064] In the step of fusing the spatial position encoding and the word embedding vector to obtain heterogeneous data by using the neural network model, the cross-modal Transformer model is used to fuse the spatial position encoding and the word embedding vector to obtain heterogeneous data. In the actual processing process, the spatial position encoding obtained by segmenting the image entity and the word embedding vector of the corresponding term in the text are jointly input into the cross-modal Transformer model. The cross-attention mechanism of the model can learn and generate a heterogeneous data that fuses the features of the two modalities. By calculating the cosine similarity between the heterogeneous data and the preset vector of each node in the anatomical knowledge graph (which can be obtained by pre-training a language model embedding, which is a conventional technical means in the art and will not be described here), the node with the highest similarity, e.g., the liver node, is bound with the heterogeneous data (three-dimensional image data), realizing the mapping with the nodes of the anatomical knowledge graph.

[0065] S3: Collect the historical interaction data of the learner, and calculate the knowledge mastery probability vector of the learner on each node in the anatomy knowledge graph based on the learner model coupled with the Ebbinghaus forgetting curve and Bayesian inference.

[0066] For any node, the learner's historical interaction data such as correct and incorrect answers, model interaction time, and data review frequency are continuously recorded; for any time, the Ebbinghaus forgetting curve formula is used to simulate the natural decay of knowledge mastery level between two learning intervals to obtain the knowledge mastery probability of the node at the current time; in response to new interaction between the learner and the model, the model will use the Bayesian formula to update the knowledge mastery probability of the node, combining the results of this interaction and the preset learner guessing probability and error probability. It should be noted here that the knowledge mastery probabilities of multiple nodes constitute the knowledge mastery vector, and the mastery probabilities of different nodes will be updated after the learner interacts with the model, so the knowledge mastery probability vector is dynamic.

[0067] Illustratively, assume that a learner learned the knowledge point about the anatomy of the lungs three days ago, and the knowledge mastery probability at that time was evaluated as 0.9. According to the Ebbinghaus forgetting curve model, the mastery probability of the knowledge point will naturally decay over time.

[0068] The decay function is used to simulate the learner's current knowledge mastery probability for the node, and the decay function is:

[0069] , where represents the knowledge mastery probability of the learner after a time has elapsed since the learner last learned the knowledge node; represents the knowledge mastery probability after the learner last learned the knowledge node; is the time interval of 3 days, is the individual memory strength parameter; is the exponential function with the natural constant as the base.

[0070] Suppose the knowledge mastery probability of the current time of this node is attenuated to 0.7 after calculation. If a test question about the connection relationship between the pulmonary artery and the pulmonary vein is pushed to the learner at this time, the learner answers correctly. The correct answer constitutes new observation evidence. Then the knowledge mastery probability is updated by applying the Bayes inference formula. Based on the conditional probability preset by the model, that is, the probability of answering the question correctly when mastering the knowledge point is 0.95, and the probability of guessing correctly when not mastering is 0.2, combined with the prior probability of 0.7 (that is, the learner's current knowledge mastery probability of the knowledge point predicted through the attenuation function, in order to distinguish, the knowledge mastery probability before updating is called prior probability), the knowledge mastery probability of the node is recalculated, in order to distinguish, the recalculated knowledge mastery probability is called posterior mastery probability. Through calculation, the new posterior mastery probability may be significantly increased to 0.92 from 0.7, and the updated value more accurately reflects the learner's knowledge state after this successful interaction. The Bayes inference formula is a conventional technical means in the art, which will not be described here.

[0071] S4: In response to any component of the knowledge mastery probability vector being below a preset threshold, a Monte Carlo tree search algorithm is used in the anatomical knowledge graph to plan a personalized learning path.

[0072] In response to any component of the knowledge mastery probability vector being below a preset threshold, that is, the learner's knowledge mastery probability of the knowledge point corresponding to a node in the knowledge graph is low, a learning path is planned for the learner.

[0073] The Monte Carlo tree search algorithm explores different learning sequences through a large number of rapid random simulations. In each simulation, it virtually completes a learning path and evaluates the expected knowledge gain brought by the path according to a predefined utility function, for example, after mastering the phrenic nerve, it is easier to learn the diaphragm innervated by it, thereby bringing higher total gain. After thousands of simulations and backpropagation updates, the initial action with the highest average knowledge gain in the simulation is selected, that is, the learner is recommended to learn the phrenic nerve node next, and so on, iteratively constructing a personalized learning path.

[0074] Reference Figure 3 The steps of determining the personalized learning path by the Monte Carlo tree search algorithm include steps S41-S45.

[0075] S41: In response to detecting that the learner's knowledge mastery probability about the phrenic nerve node is below a preset threshold, the Monte Carlo tree search algorithm takes the current node as the root node.

[0076] The preset threshold in this embodiment is set to 0.5, which can be set according to actual conditions in other embodiments.

[0077] S42: Use nodes connected to the root node in the anatomical knowledge graph as extension nodes, such as the diaphragm and cervical plexus, as optional next learning actions.

[0078] S43: Starting from a new extended node, a fast random strategy is used to simulate the learning path in the anatomical knowledge graph, and the knowledge gain value of this simulation is calculated. The knowledge gain value is the magnitude of the change in the learner's knowledge mastery probability vector before and after the simulation.

[0079] Suppose a new expansion node in the Monte Carlo tree search is the heart valve. To quickly evaluate the learning value starting from this node, a fast randomization strategy is used for simulation. Based on the associations in the anatomical knowledge graph, a subsequent learning path is randomly but biasedly selected. For example, starting from the heart valve, the mitral valve is randomly selected, followed by the aortic valve, simulating a short learning process containing two new knowledge points. The change in the learner's knowledge mastery probability vector before and after the simulation is calculated. For example, before the simulation, the mastery probabilities of these two nodes are 0.2 and 0.15, respectively. After the simulation, they are assumed to have increased to 0.8 and 0.75. Therefore, the knowledge gain value of this simulation is the sum of the probability increments of the two nodes, i.e., 0.6 plus 0.6 equals 1.2.

[0080] S44: Use the knowledge gain value to backpropagate along the learning path to all parent nodes on the path.

[0081] For example, if the calculated knowledge gain is 1.2, this value will propagate back along the path of this simulation. All parent nodes along the path, including the heart valve node and its parent nodes, all the way to the root node of the search tree, will be affected by the simulation result. Each node traversed during backpropagation will have its cumulative knowledge gain increased by 1.2, and its visit count will also increase by 1. For instance, if the heart valve node had been visited 5 times before this simulation, with a cumulative gain of 4.8, its updated state will be 6 visits, with a cumulative gain of 6.0. This process makes high-yield learning paths more likely to be selected in subsequent searches.

[0082] S45: After completing a specified number of Monte Carlo tree searches, select the path with the highest cumulative knowledge gain value starting from the root node as the final personalized learning path.

[0083] Specifically, the number of iterations of the Monte Carlo tree search is set, and in this embodiment, the number of iterations is set to 2000. After completing the 2000 complete cycles containing selection, expansion, simulation and back propagation, a complex search tree is constructed, and each node in the tree records the total number of visits and the cumulative knowledge gain value. At this time, starting from the root node representing the current learning starting point, all direct child nodes are checked. For example, the child nodes of the root node are the starting point of learning path A and the starting point of learning path B. After 2000 simulations, the average knowledge gain value of the starting point of path A is 2.5, and the average knowledge gain value of the starting point of path B is 3.1. The system will select the starting point of path B with a higher average knowledge gain value as the first knowledge point of the personalized learning path.

[0084] After determining the first knowledge point, it will move to the node and repeat the above selection process. All child nodes of the current node in the search tree are examined, and the child node with the highest average knowledge gain value is again selected as the second knowledge point of the learning path. The process of greedy selection will continue until a learning path of a predetermined length, for example, containing 5 knowledge points, is generated. The sequence composed of the optimal selection at each step, for example, learning stomach first, then duodenum, and then pancreas, is recommended as the personalized learning path customized for the learner with the maximum expected knowledge gain.

[0085] The embodiment of the present application also discloses an intelligent management system for dissection teaching data, comprising a processor and a memory, and the memory stores computer program instructions which realize the intelligent management method for dissection teaching data according to the present application when executed by the processor.

[0086] The above system also includes a communication bus and a communication interface and other components well known to those skilled in the art, the settings and functions of which are known in the art, and therefore will not be described here.

[0087] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application, therefore: any equivalent changes made on the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. An intelligent management method of dissection teaching data, characterized in that, The method comprises the following steps: acquiring anatomical knowledge data, wherein the anatomical knowledge data comprises three-dimensional images and text data, extracting anatomical technical terms and the correlation between the anatomical technical terms in the anatomical knowledge data, taking the anatomical technical terms as nodes and the correlation as edges to construct a multi-modal anatomical knowledge graph; segmenting the three-dimensional image data to obtain spatial position codes representing the center, direction and bounding box of each human tissue or organ; and using a bidirectional long short-term memory network model to obtain word embedding vectors corresponding to the nodes; fusing the spatial position codes and the word embedding vectors using a neural network model to obtain heterogeneous data; binding the heterogeneous data and the nodes, including: calculating the cosine similarity between the heterogeneous data and the preset vectors of the nodes in the knowledge graph, and binding the node with the highest similarity to the heterogeneous data; collecting historical interaction data of a learner, calculating a knowledge mastery probability vector of the learner for each node in the anatomical knowledge graph based on a learner model coupled with the Ebbinghaus forgetting curve and Bayesian inference, including: for any node, continuously recording the history interaction data of the learner, such as the correctness of the answers, the model interaction time, and the data searching frequency; for any time, using the Ebbinghaus forgetting curve formula to obtain the knowledge mastery probability of the node at the current time; in response to new interaction between the learner and the model, using the Bayesian formula to update the knowledge mastery probability of the node in combination with the result of the interaction and the preset guessing probability and error probability of the learner; and the knowledge mastery probabilities of multiple nodes constitute the knowledge mastery probability vector of the learner for the anatomical knowledge graph; in response to any component in the knowledge mastery probability vector being lower than a preset threshold, using a Monte Carlo tree search algorithm in the anatomical knowledge graph to plan a personalized learning path, including: in response to detecting that the knowledge mastery probability of the learner for the phrenic nerve node is lower than the preset threshold, the Monte Carlo tree search algorithm takes the current node as the root node; the nodes connected to the root node in the anatomical knowledge graph are taken as expansion nodes; starting from a new expansion node, a fast random strategy is used to simulate a learning path in the anatomical knowledge graph and calculate the knowledge gain value of this simulation, which is the change amplitude of the knowledge mastery probability vector of the learner before and after the simulation; the knowledge gain value is used to back-propagate all parent nodes along the learning path until the root node of the search tree, and the cumulative knowledge gain value of each node passed through in the back-propagation path will increase the knowledge gain value of this simulation; after completing a specified number of Monte Carlo tree searches, the path with the highest cumulative knowledge gain value starting from the root node is selected as the final personalized learning path.

2. The method of claim 1, wherein, The correlation includes spatial topology, functional domination and organizational hierarchy, wherein the functional domination and organizational hierarchy between the nodes are extracted by natural language processing of the text data; the spatial topology between the nodes is determined based on the segmentation result of the three-dimensional image.

3. The method of claim 1, wherein, The three-dimensional image is segmented using a cascaded 3DU-Net network.

4. The method of claim 1, wherein, The steps of using the cascaded 3D U-Net network to perform pixel-level segmentation on the three-dimensional image include: using a first-stage three-dimensional convolutional network to perform rough positioning on an input complete three-dimensional image, and outputting an initial three-dimensional region containing a target organ; taking the initial three-dimensional region as an input of a second-stage three-dimensional attention convolutional network, performing fine segmentation in the initial three-dimensional region, capturing subtle features of organ boundaries and internal textures through an attention mechanism in the three-dimensional attention convolutional network, and generating an organ segmentation mask.

5. The method of claim 1, wherein, The number of iterations of the Monte Carlo tree search is 2000.

6. The method of claim 1, wherein, The spatial position encoding and the word embedding vector are fused using a cross-modal Transformer model to obtain heterogeneous data.

7. An intelligent management system for dissection teaching data, characterized by, The method comprises the following steps: A processor and a memory, the memory stores computer program instructions, when the computer program instructions are executed by the processor, a kind of anatomical teaching data intelligent management method according to any one of claims 1-6 is realized.

Images