A surgical operation recognition method and system based on a self-attention graph encoder

Abstract

The application provides a surgical operation recognition method and system based on a self-attention graph encoder, relates to the technical fields of medical informatization and artificial intelligence, and performs syntactic analysis on acquired surgical texts, constructs a surgical operation semantic topology graph based on word segmentation results, filters the surgical operation semantic topology graph based on the connection density of connected subgraphs to obtain a surgical operation topology graph, extracts deep semantic features of each node in the surgical operation topology graph through a self-attention graph encoder, and performs classification and recognition of surgical operations based on the deep semantic features of the nodes to obtain surgical codes; the application extracts key surgical operation sites through syntactic analysis, designs a new self-attention graph encoder structure for extracting surgical operation semantic information, effectively learns surgical operation code knowledge of large hospitals and extracts surgical codes in surgical records or diagnosis and treatment operation records, so that the application helps primary hospitals to correctly fill in surgical codes and improves medical record quality.

Description

Technical Field

[0001] This invention belongs to the field of medical information technology and artificial intelligence technology, and in particular relates to a surgical operation recognition method and system based on a self-attention map encoder. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the ongoing national medical insurance reform, surgical codes have become the primary method for recording surgical procedures on medical insurance settlement lists. Through surgical codes, hospitals can use standardized codes to more accurately record and report medical information. Simultaneously, surgical codes can effectively reduce the incidence of adverse surgical events, improve surgical quality, and enhance the efficiency and economic benefits of healthcare.

[0004] However, hospitals at different levels currently have varying levels of understanding of surgical codes. Generally speaking, large hospitals have a much higher level of understanding of surgical codes than primary hospitals. This is because large hospitals have more advanced medical equipment and experienced doctors, requiring strict surgical codes to effectively manage each surgical procedure. Primary hospitals, on the other hand, have fewer medical equipment and less experienced doctors, resulting in a lower level of understanding of surgical codes. This, in turn, affects the quality of medical records and leads to losses in medical insurance settlements. The quality of surgical code completion can be improved by learning from the surgical coding experience of large hospitals and transferring it to primary hospitals.

[0005] Currently, surgical codes are mainly filled out manually. Doctors often need to review the patient's surgical records and treatment records to extract the corresponding surgical codes. However, this method is time-consuming and labor-intensive, and manual extraction can easily lead to omissions of surgical codes, resulting in a decline in the quality of hospital medical records.

[0006] While some surgical code extraction methods based on deep learning and artificial intelligence have emerged, most existing methods input the entire surgical record text into the model, then perform feature extraction and surgical code classification, neglecting feature selection for the surgical procedures themselves. Surgical records often contain redundant text; the true information about the surgical procedures is contained within a few key descriptions. Therefore, this approach easily introduces noisy text, making it difficult to extract the correct surgical codes. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides a surgical operation recognition method and system based on a self-attention graph encoder. It uses syntactic analysis to extract key surgical operation sites and designs a new self-attention graph encoder structure to extract semantic information of surgical operations. It effectively learns the surgical operation coding knowledge of large hospitals and extracts surgical codes from surgical records or medical operation records, thereby helping primary hospitals to correctly fill in surgical codes and improve the quality of medical records.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0009] The first aspect of this invention provides a surgical operation recognition method based on a self-attention map encoder;

[0010] A surgical operation recognition method based on a self-attention map encoder includes:

[0011] Syntactic analysis was performed on the acquired surgical text, and a semantic topology graph of the surgical operation was constructed based on the obtained word segmentation results.

[0012] Based on the connection density of the connected subgraph, the semantic topology graph of the surgical operation is filtered to obtain the topology graph of the surgical operation.

[0013] Deep semantic features of each node in the surgical operation topology graph are extracted using a self-attention graph encoder.

[0014] Based on the deep semantic features of nodes, surgical operations are classified and identified to obtain surgical codes.

[0015] Furthermore, the syntactic analysis of the acquired surgical text involves using a Chinese dependency parsing tool to segment the surgical text into words, obtaining the word segments, their corresponding parts of speech, and the dependencies between the word segments; the word segments are then filtered based on their parts of speech and dependencies.

[0016] Furthermore, a directed graph approach is adopted, using word segmentation as nodes, to construct a semantic topology graph of surgical operations;

[0017] In this context, the presence of an edge between two nodes indicates a dependency relationship between the two words, with the core word pointing to the dependent word. This dependency relationship includes verb relationships, verb-modifying relationships, and noun-modifying relationships.

[0018] Furthermore, the connection density of the connected subgraph is defined as the ratio of the number of edges in the connected subgraph to the total number of nodes in the surgical operation semantic topology graph, which measures the density of the subgraph connections.

[0019] Calculate the connection density of each connected subgraph, sort the connected subgraphs in descending order of connection density, and select connected subgraphs to form the surgical operation topology graph based on the ranking and the preset number of subgraphs.

[0020] Furthermore, the specific steps for extracting the deep semantic features of each node in the surgical operation topology graph are as follows:

[0021] The semantic feature vector of each character in the surgical text is obtained by using the word embedding model GloVe and the gated recurrent unit GRU.

[0022] By comprehensively considering the semantic features of each position of each character in the node through vector addition, the semantic features of each node and the adjacency matrix of the surgical operation topology graph are obtained.

[0023] Based on the semantic features of each node and the adjacency matrix of the surgical operation topology graph, a self-attention graph encoder is used to extract the deep semantic features of each node.

[0024] Furthermore, the self-attention map encoder sequentially includes a self-attention map neural network, a splicing / normalization layer, and a forward propagation normalization layer;

[0025] The self-attention graph neural network distinguishes node features based on the semantic features of nodes and their neighbors by introducing a self-attention mechanism, thereby obtaining enhanced semantic features.

[0026] Furthermore, the classification and identification of surgical procedures specifically includes:

[0027] All extracted deep semantic features are vector-added to obtain a classification feature vector for classification, and then an activation function is used for surgical classification and recognition.

[0028] A second aspect of the present invention provides a surgical operation recognition system based on a self-attention map encoder.

[0029] A surgical operation recognition system based on a self-attention graph encoder includes a syntactic analysis module, a subgraph filtering module, a semantic extraction module, and an encoding recognition module.

[0030] The syntactic analysis module is configured to perform syntactic analysis on the acquired surgical text and construct a semantic topology graph of the surgical operation based on the obtained word segmentation results.

[0031] The subgraph filtering module is configured to filter the semantic topology graph of surgical operations based on the connection density of the connected subgraphs to obtain the topology graph of surgical operations.

[0032] The semantic extraction module is configured to extract deep semantic features of each node in the surgical operation topology graph through a self-attention graph encoder.

[0033] The coding and recognition module is configured to classify and recognize surgical operations based on the deep semantic features of nodes to obtain surgical codes.

[0034] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of a surgical operation recognition method based on a self-attention map encoder as described in the first aspect of the present invention.

[0035] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a surgical operation recognition method based on a self-attention map encoder as described in the first aspect of the present invention.

[0036] The above one or more technical solutions have the following beneficial effects:

[0037] This invention provides a method for automatically identifying surgical procedures from surgical records or medical procedure records, and can learn from high-quality surgical procedure recording experience, thereby improving the quality of surgical procedure recording in primary hospitals.

[0038] This invention uses syntactic analysis to extract key surgical operation sites and designs a novel self-attention graph encoder structure to extract semantic information of surgical operations. Experimental results demonstrate the effectiveness of this invention and achieve effective identification of surgical operations.

[0039] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0040] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0041] Figure 1 This is a model structure diagram of the first embodiment.

[0042] Figure 2 This is a flowchart of the method in the first embodiment.

[0043] Figure 3 This is the semantic topology diagram of the surgical procedure in the first embodiment.

[0044] Figure 4 This is a system structure diagram of the second embodiment. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0046] Example 1

[0047] The purpose of this invention is to provide an automatic surgical code extraction method that can effectively learn the surgical operation code knowledge of large hospitals and extract the surgical codes from surgical records or treatment operation records, thereby helping primary hospitals to accurately fill in surgical codes and improve the quality of medical records.

[0048] This embodiment discloses a surgical operation recognition method based on a self-attention map encoder. The structure diagram of the entire model of this method is as follows. Figure 1 As shown, specifically, syntactic analysis technology is used to perform semantic analysis on surgical records or medical operation records, and a semantic topology graph of surgical operations is constructed based on the syntactic analysis results. Then, important semantic subgraphs that can reflect the surgical operation are selected through subgraph selection. In order to effectively extract semantic features from the semantic topology graph of surgical operations, a new self-attention graph encoder is designed based on self-attention mechanism and graph neural network to extract surgical operation features. Finally, the surgical operation is encoded. The surgical encoding extraction method described in this embodiment mainly includes the following steps: syntactic analysis, semantic subgraph selection, semantic extraction, and surgical encoding output.

[0049] like Figure 2 As shown, a surgical operation recognition method based on a self-attention map encoder includes:

[0050] Step S1: Syntactic analysis. Perform syntactic analysis on the acquired surgical text and construct a semantic topology graph of the surgical operation based on the obtained word segmentation results.

[0051] Specifically, the surgical text consists of the acquired surgical records and treatment operation records. Given a surgical record or treatment operation record text S = {s1, ..., s...} n}, where s n This represents the nth Chinese character, where n is the number of Chinese characters.

[0052] The text was segmented using the Chinese dependency parsing tool DDPaser, yielding word segments, their corresponding parts of speech, and the dependencies between them. Parts of speech included nouns, verbs, adjectives, adverbs, prepositions, etc. The dependencies between word segments are shown in Table 1.

[0053] Table 1. Set of Dependencies

[0054]

[0055] definition The word segmentation result of the text, where, This represents the m-th word segmentation result after segmenting several Chinese characters. Since the surgical operation uses verbs, nouns, and noun modifiers for description, in order to eliminate redundant text representation, this invention filters words through the dependency relationships between words, retaining only the dependency relationships of verbs, nouns, and modifiers, namely verb relationship VOB, verb-modifying relationship ADV, and noun-modifying relationship ATT.

[0056] To represent the semantic directionality, a directed graph approach is used, with word segments as nodes, to construct a semantic topology graph of surgical procedures. The resulting semantic topology graph of surgical procedures is shown below. Figure 3 As shown, there is an edge between two nodes, indicating that the two words have a dependency relationship, with the core word pointing to the dependent word. The dependency relationship includes verb relationship VOB, verb-modifying relationship ADV, and noun-modifying relationship ATT.

[0057] Step S2: Semantic subgraph filtering. Based on the connection density of the connected subgraphs, the semantic topology graph of the surgical operation is filtered to obtain the surgical operation topology graph.

[0058] Generally speaking, in semantic topological graphs formed by text fragments containing a lot of surgical procedures, the connections between nodes and edges are denser. This dense connection subgraph can help focus on key information and effectively extract the semantics of surgical procedures.

[0059] Therefore, the connection density ρ of a connected subgraph is defined. subgraph The formula used to measure the density of subgraph connections is shown below:

[0060]

[0061] in, This represents the number of edges in a connected subgraph. This represents the total number of nodes in the semantic topology graph of the surgical procedure.

[0062] Based on the formula for connection density, the connection density of each connected subgraph is calculated, and the connected subgraphs are sorted in descending order according to their connection density.

[0063] In this embodiment, the top 5 connected subgraphs are selected to form the surgical operation topology graph G = {V, R} for extracting semantic information, where V represents the set of semantic nodes and R represents the set of semantic relations.

[0064] Step S3: Semantic extraction. Using a self-attention graph encoder, deep semantic features of each node in the surgical operation topology graph are extracted. The specific steps are as follows:

[0065] Step S301: Obtain the semantic feature vector of each character in the surgical text through the word embedding model GloVe and the gated recurrent unit GRU.

[0066] To obtain the vector representation of semantic nodes, first use the word embedding model Glove to obtain the vectorized representation of characters. That is, input the text S of a surgical record or a diagnosis and treatment operation record into the word embedding model Glove, and obtain the embedded text vector representation X = {x1, …, x n}, where x n represents the vectorized representation of the nth character. To extract the semantic features of characters, use the gated recurrent unit GRU to extract the semantic features of each character in the surgical text, and its extraction formula is as follows:

[0067] h i = GRU(x i ; θ gru ), 1 ≤ i ≤ n (2)

[0068] where h i represents the semantic feature of the ith character, θ gru represents the trainable model parameters, and n is the number of characters in the surgical text.

[0069] Step S302: By vector addition, comprehensively consider the semantic features of each position of each character in the node to obtain the semantic features of each node and the adjacency matrix of the surgical operation topology graph.

[0070] In the surgical operation topology graph G, a semantic node may be composed of multiple Chinese characters, and the same Chinese character may also appear in multiple positions in the surgical record or the diagnosis and treatment operation record text.

[0071] Based on this, in this embodiment, information is aggregated by vector addition. For example, given a semantic node v 穿刺 , which is composed of two Chinese characters '穿' and '刺'. Suppose the set of position subscripts where '穿' appears in the surgical record text is I 穿 ∈{1,..., n}, and the set of position subscripts where '刺' appears is I 刺 ∈{1,..., n}. To fully consider the character semantics in the node, this example comprehensively considers the semantic feature representations of each position of each character in the node by vector addition. For example, the semantic representation of node v 穿刺 is obtained through the following formula (3) where j and k represent the position subscripts of the corresponding characters, h j represents the semantic feature representation of '穿' at position j, h k represents the semantic feature representation of '刺' at position k, |v 穿刺 | represents the number of elements in the node set, |I 穿 | represents the number of elements in the position subscript set I 穿 , and |I 刺| represents the set of position indices I 刺 The number of elements.

[0072]

[0073] Using the above methods, the semantic features of all semantic nodes in the surgical operation topology graph G are obtained.

[0074] Defining the number of nodes in the surgical operation topology graph G as t, we can obtain the feature matrix of the surgical operation topology graph G. in, Let represent the set of real numbers, and b represent the feature dimension. The adjacency matrix of the surgical operation topology graph G can be obtained by analyzing the dependencies between two points in G.

[0075] Step S303: Based on the semantic features of each node and the adjacency matrix of the surgical operation topology graph, extract the deep semantic features of each node through a self-attention graph encoder.

[0076] Most existing graph neural networks aggregate node features. This simple aggregation operation can easily lead to overly smooth node features, making it difficult to effectively extract semantic features. In order to effectively extract semantic information from the surgical operation topology graph G, this example sets up a new self-attention graph encoder. This encoder uses residual connections to ensure the extraction of deep semantic information. Secondly, this example combines self-attention with the graph Laplacian matrix to optimize the node feature aggregation method. By enhancing the feature discrimination of adjacent nodes, the feature aggregation flow is changed, effectively alleviating the oversmoothing problem.

[0077] like Figure 1 As shown in the self-attention map encoder section, the self-attention map encoder sequentially includes a self-attention map neural network, a concatenation / normalization layer, and a forward propagation normalization layer. The specific extraction steps are as follows:

[0078] (1) Self-attention graph neural network: Based on the semantic features of nodes and their neighboring nodes, a self-attention mechanism is introduced to distinguish node features and obtain enhanced semantic features.

[0079] Given a node v in the surgical operation topology graph G node Assuming its number of adjacent nodes is Based on adjacency matrix The feature matrix is ​​obtained by extracting the features of all its first-order adjacent nodes.

[0080] Using formulas (4) and (5) respectively, we obtain the matrix used to calculate attention. and in, and These are the weight matrices.

[0081] Calculate node v using formula (6). node The attention score vector.

[0082]

[0083]

[0084]

[0085] Here, softmax is an activation function. It is node v node The vector is formed by the degrees of the adjacent vertices, and T represents the matrix transpose operation.

[0086] Using the above method, this example calculates the attention score vector for all nodes. Based on the dependencies between nodes, this example assigns the values ​​of the attention score vectors to the corresponding positions in the adjacency matrix, obtaining the attention-based adjacency matrix.

[0087] The normalized attention adjacency matrix for feature extraction is obtained by performing Laplacian decomposition using formula (7).

[0088]

[0089] in, It is an attention-based adjacency matrix. Let C be the degree matrix, where C represents the total number of attention heads, c represents the c-th attention head, c is an integer and 1≤c≤C.

[0090] (2) Splicing / Normalization

[0091] Vector concatenation is performed using formula (8) to achieve semantic feature transfer on the graph:

[0092]

[0093] Where l represents the current graph neural network layer, i.e., the self-attention graph encoder layer number, || represents the vector concatenation operation, and ReLU represents an activation function. This represents the weight matrix.

[0094] To maintain feature normalization, this example uses the LayerNorm operation to normalize the concatenated node features.

[0095] (3) Forward Propagation Normalization Layer

[0096] To further extract the features of the aggregated nodes, this example uses a forward propagation normalization layer to achieve feature extraction, as shown in formula (9):

[0097]

[0098] in, It is the output feature of the forward propagation layer, U l+1 These are the spliced / normalized node features. is the weight matrix, and o is the bias vector.

[0099] The self-attention graph neural network, concatenation / normalization, and forward propagation are collectively referred to as the self-attention graph encoder E; in order to meet the need for extracting deep semantic features, this example introduces residual connections, as shown in formula (10):

[0100]

[0101] in, These are the extracted node features, and α is a weight parameter with a value between 0 and 1.

[0102] Step S4: Surgical code output. Based on the deep semantic features of the nodes, the surgical operation is classified and identified to obtain the surgical code.

[0103] When identifying surgical procedures, this example uses the surgical procedures listed in ICD-9 published by the National Medical Insurance Administration as the label category, which contains a total of 13,653 categories. Additionally, this example sets a null value class, None, to mark surgical records or treatment texts without surgical procedures. The specific steps are as follows:

[0104] First, this example performs vector addition on all extracted deep semantic features of the nodes to obtain a classification feature vector for classification.

[0105] Then, the activation function shown in formula (11) is used to classify and identify surgical procedures:

[0106]

[0107] in, Here, is the predicted surgical procedure label, 'u' is the deep semantic feature of the node, and 'sigmoid' is an activation function that can be used to achieve multi-label classification of surgical text.

[0108] In this example, the loss value Loss is calculated using the loss function shown in formula (12). After obtaining the loss, the model optimizes the entire model parameters through backpropagation.

[0109]

[0110] Where y is the true surgical operation label of the training set sample, Class represents the number of label categories, μ represents a sample in the training set Ω, and exp represents the exponent with base e. μ This represents the true surgical label of sample μ. This represents the predicted surgical label for sample μ.

[0111] The effectiveness of the method in this embodiment was verified through experiments.

[0112] During training, 500,000 high-quality surgical or clinical operation records from major hospitals were acquired and labeled. The proportions of the training set, validation set, and test set were 60%, 10%, and 30%, respectively. The self-attention graph encoder was set to 8 layers, and the label ranking accuracy was used as the evaluation metric. The experimental data are shown in Table 2.

[0113] Table 2 Experimental Results

[0114]

[0115] Compared to the traditional Gated Recurrent Unit (GRU) network, this example effectively extracts semantic information from text through syntactic analysis and graph neural networks. Compared to the traditional graph neural network extraction model (GNN), this example effectively distinguishes node features by introducing a self-attention mechanism, alleviating its oversmoothing problem. Furthermore, the introduction of forward propagation normalization layers and residual connections enables this example to extract deeper semantic information, especially in Top-1 accuracy, where it outperforms traditional graph neural networks by 12.46%. Compared to the Transformer, this example effectively utilizes the surgical operation topology graph information generated by syntactic analysis, eliminating redundant text and making the extraction of surgical operation semantics from the text more focused. In all three statistical metrics, this example outperforms the Transformer.

[0116] Example 2

[0117] This embodiment discloses a surgical operation recognition system based on a self-attention map encoder;

[0118] like Figure 4 As shown, a surgical operation recognition system based on a self-attention graph encoder includes a syntactic analysis module, a subgraph filtering module, a semantic extraction module, and an encoding recognition module.

[0119] The syntactic analysis module is configured to perform syntactic analysis on the acquired surgical text and construct a semantic topology graph of the surgical operation based on the obtained word segmentation results.

[0120] The subgraph filtering module is configured to filter the semantic topology graph of surgical operations based on the connection density of the connected subgraphs to obtain the topology graph of surgical operations.

[0121] The semantic extraction module is configured to extract deep semantic features of each node in the surgical operation topology graph through a self-attention graph encoder.

[0122] The coding and recognition module is configured to classify and recognize surgical operations based on the deep semantic features of nodes to obtain surgical codes.

[0123] Example 3

[0124] The purpose of this embodiment is to provide a computer-readable storage medium.

[0125] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a surgical operation recognition method based on a self-attention map encoder as described in Embodiment 1 of this disclosure.

[0126] Example 4

[0127] The purpose of this embodiment is to provide an electronic device.

[0128] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in a surgical operation recognition method based on a self-attention map encoder as described in Embodiment 1 of this disclosure.

[0129] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A surgical operation recognition method based on a self-attention map encoder, characterized in that, include: Syntactic analysis was performed on the acquired surgical text, and a semantic topology graph of the surgical operation was constructed based on the obtained word segmentation results. A directed graph approach is used, with word segmentation as nodes, to construct a semantic topology graph of surgical operations; where there is an edge between two nodes, indicating that the two words have a dependency relationship, pointing from the core word to the dependent word. The dependency relationship includes verb relationship, verb modification relationship, and noun modification relationship. Based on the connection density of connected subgraphs, the semantic topology graph of surgical operations is filtered to obtain the surgical operation topology graph. The connection density of the connected subgraph is defined as the ratio of the number of edges in the connected subgraph to the total number of nodes in the semantic topology graph of surgical operations, which measures the density of subgraph connections. The connection density of each connected subgraph is calculated, and the connected subgraphs are sorted in descending order according to the connection density. Based on the ranking and the preset number of subgraphs, the connected subgraphs are selected to form the surgical operation topology graph. The deep semantic features of each node in the surgical operation topology graph are extracted by a self-attention graph encoder. The self-attention graph encoder includes a self-attention graph neural network, a splicing / normalization layer, and a forward propagation normalization layer. The self-attention graph neural network introduces a self-attention mechanism to distinguish node features based on the semantic features of the node and its neighboring nodes, thereby obtaining enhanced semantic features. Based on the deep semantic features of nodes, surgical operations are classified and identified to obtain surgical codes.

2. The surgical operation recognition method based on a self-attention map encoder as described in claim 1, characterized in that, The syntactic analysis of the acquired surgical text involves using a Chinese dependency parsing tool to segment the surgical text into words, obtaining the word segments, their corresponding parts of speech, and the dependencies between the word segments; the word segments are then filtered based on their parts of speech and dependencies.

3. The surgical operation recognition method based on a self-attention map encoder as described in claim 1, characterized in that, The specific steps for extracting the deep semantic features of each node in the surgical operation topology graph are as follows: The semantic feature vector of each character in the surgical text is obtained by using the word embedding model GloVe and the gated recurrent unit GRU. By comprehensively considering the semantic features of each position of each character in a node through vector addition, the semantic features of each node are obtained. Define the surgical operation topology graph The number of nodes in is Based on the semantic features of each node, a surgical operation topology graph is obtained. Feature matrix ,in, Represents the set of real numbers. Representing feature dimensions; through surgical operation topology graph The dependency relationship between the two points yields the surgical operation topology graph. adjacency matrix ; Based on the semantic features of each node and the adjacency matrix of the surgical operation topology graph, a self-attention graph encoder is used to extract the deep semantic features of each node.

4. The surgical operation recognition method based on a self-attention map encoder as described in claim 1, characterized in that, The classification and identification of surgical procedures specifically includes: All extracted deep semantic features are vector-added to obtain a classification feature vector for classification, and then an activation function is used for surgical classification and recognition.

5. A surgical operation recognition system based on a self-attention map encoder, characterized in that, It includes a syntactic analysis module, a subgraph filtering module, a semantic extraction module, and an encoding recognition module: The syntactic analysis module is configured to: perform syntactic analysis on the acquired surgical text, and construct a semantic topology graph of surgical operations based on the obtained word segmentation results; using a directed graph approach, with word segments as nodes, to construct the semantic topology graph of surgical operations; wherein, there is an edge between two nodes, indicating that the two words have a dependency relationship, pointing from the core word to the dependent word, and the dependency relationship includes verb relationship, verb modification relationship and noun modification relationship; The subgraph filtering module is configured to: filter the surgical operation semantic topology graph based on the connection density of the connected subgraphs to obtain the surgical operation topology graph; the connection density of the connected subgraph is defined as the ratio of the number of edges in the connected subgraph to the total number of nodes in the surgical operation semantic topology graph, which measures the density of subgraph connections; calculate the connection density of each connected subgraph, sort the connected subgraphs in descending order according to the connection density, and select connected subgraphs to form the surgical operation topology graph based on the ranking and a preset number of subgraphs; The semantic extraction module is configured to: extract deep semantic features of each node in the surgical operation topology graph through a self-attention graph encoder; the self-attention graph encoder includes a self-attention graph neural network, a splicing / normalization layer, and a forward propagation normalization layer; the self-attention graph neural network introduces a self-attention mechanism to distinguish node features based on the semantic features of the node and its neighboring nodes, thereby obtaining enhanced semantic features. The coding and recognition module is configured to classify and recognize surgical operations based on the deep semantic features of nodes to obtain surgical codes.

6. An electronic device, characterized in that it comprises: Memory is used to store computer-readable instructions in a non-transitory manner. as well as Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in any one of claims 1-4.

7. A storage medium characterized in that it non-transitory stores computer-readable instructions, wherein, When the non-transitory computer-readable instructions are executed by a computer, the instructions of the method according to any one of claims 1-4 are executed.

Images