Method and system for extracting sample multi-level structure information in pathological simulation mechanism

Abstract

The application relates to the field of artificial intelligence, and discloses a method and system for extracting multi-level structure information of a sample in a pathological simulation mechanism to assist in realizing accurate classification of the sample. The method comprises the following steps: constructing a complex adjacency matrix representing the adjacency relationship between points, edges, surfaces and bodies in a brain region-gene network; calculating an initial structure information matrix based on the complex adjacency matrix and a first learning parameter matrix, and then obtaining a structure entropy difference value matrix; performing information aggregation and updating processing on the initial structure information matrix based on the structure entropy difference value matrix and a second learning parameter matrix, and then performing reiteration on the updated structure information matrix based on the structure entropy difference value matrix and a third learning parameter matrix, introducing a new learning parameter matrix in each iteration, and determining the flow direction in the information aggregation process with the same structure entropy difference value matrix; and obtaining a target structure information matrix after at least three iterations; and outputting the target structure information matrix after flattening to a full connection layer.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for extracting multi-level structural information of samples from a pathological simulation mechanism. Background Technology

[0002] Due to the high complexity of the brain, traditional methods have significant limitations in the detection and diagnosis of diseases in their early stages. In recent years, researchers have gradually introduced the concept of complexes from topology, using this tool to capture higher-order connection patterns between different neurons in the brain, thereby advancing a more comprehensive understanding of the pathological mechanisms of AD (Alzheimer's disease). Although some progress has been made in this direction, current research still faces two key problems. First, many studies focus only on the analysis of single complexes, failing to fully explore multi-dimensional higher-order cooperative relationships. Existing cooperative relationships are limited to point-edge-face relationships, which restricts the model's expressive power in complex networks. Second, existing network modeling methods have overly complex classifications and definitions of adjacency relationships and information propagation patterns, lacking a unified and concise framework to effectively describe these higher-order connection patterns, leading to significant limitations in practical applications. To address these issues, there is an urgent need to explore more advanced mathematical tools and algorithms to more accurately capture and analyze the complex dynamic changes in the brain during AD development. Summary of the Invention

[0003] The purpose of this invention is to disclose a method and system for extracting multi-level structural information of samples from a pathological simulation mechanism, so as to assist in the accurate classification of samples.

[0004] To achieve the above objectives, the present invention discloses a method for extracting multi-level structural information of samples from a pathological simulation mechanism, comprising:

[0005] Correlation analysis between brain region features and gene features yielded brain region gene networks; Detection of all two-dimensional complexes in brain region gene networks using stringless loops; Polyhedral structures were detected from the two-dimensional complexes based on Euler's formula. Construct a complex adjacency matrix that represents the adjacency relationships between points, edges, faces, and volumes; The initial structural information matrix representing the spatial position information of the elements in the field is calculated based on the complex adjacency matrix and the first learning parameter matrix. Based on the initial structural information matrix, calculate the adjacency degree matrix representing the adjacency degree of each element, and then obtain the adjacency structure entropy matrix representing the information intensity of elements in the complex based on the adjacency degree matrix. Calculate the structural entropy difference between any two elements in the adjacency structural entropy matrix, then compare each structural entropy difference with the difference filtering threshold. Structural entropy differences that are greater than or equal to the difference filtering threshold are retained, while structural entropy differences that are less than the difference filtering threshold are set to zero. Then, determine the information flow direction between the corresponding elements based on the retained structural entropy differences to obtain the structural entropy difference matrix. The initial structural information matrix is ​​processed by information aggregation and updating based on the structural entropy difference matrix and the second learning parameter matrix. Then, the updated structural information matrix is ​​iterated again based on the structural entropy difference matrix and the third learning parameter matrix. This process is repeated, with a new learning parameter matrix introduced in each iteration and the same structural entropy difference matrix used to determine the flow of information aggregation. The target structural information matrix is ​​obtained after at least three iterations. The target structural information matrix is ​​flattened to obtain a complex feature vector of the fused point-edge-face-volume structural information output to the fully connected layer.

[0006] Preferably, the formula for calculating the initial structural information matrix, which represents the spatial location information of the elements in the field based on the complex adjacency matrix and the first learning parameter matrix, is as follows: ; in, This represents the learnable parameter matrix. This represents the activation function. Represents the complex adjacency matrix; The sum of the number of all points, edges, faces, and volumes.

[0007] Preferably, ;in, It is a point-to-point adjacency matrix; It is an edge-edge adjacency matrix; It is a face-to-face adjacency matrix; It is a volume-to-volume adjacency matrix; It is a point-edge adjacency matrix; It is an edge-face adjacency matrix; Let be the face-volume adjacency matrix; where a complex is a subset of another complex, or there is an intersection between complexes of the same dimension, then an adjacency relationship exists and is assigned a value of 1; otherwise, it is assigned a value of 0. For transpose operation, This represents the multiplication operator.

[0008] Preferably, the formula for calculating the adjacency matrix representing the adjacency degree of each element based on the initial structural information matrix is ​​as follows: ;in, It is a matrix of all ones. It is the identity matrix. This is the adjacency degree matrix. Represent the Hadamard product; obtain the adjacency structure entropy matrix from the adjacency degree matrix. The calculation formula is: ; ;in, Adjacency matrix The inverse matrix, For the intermediate matrix calculated, This indicates the logarithmic operation.

[0009] Preferably, the adjacency structure entropy matrix Two elements , structural entropy difference between The calculation formula is: ;in, The threshold value is used to filter the difference.

[0010] Preferably, in any iteration process, the information aggregation and update processing of the structural information matrix based on the structural entropy difference matrix and the learning parameter matrix includes: First calculate the structural information change matrix The specific calculation formula is as follows: ;in, To learn the parameter matrix, For the first The structural information matrix before information aggregation occurs during the next iteration; Recalculate and update the first The structural information matrix after the next iteration The calculation formula is as follows: ;in, These are parameters for information control.

[0011] To achieve the above objectives, the present invention also discloses a system for extracting multi-level structural information of samples from a pathological simulation mechanism, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0012] The present invention has the following beneficial effects: 1. The brain is essentially an entropy-increasing system. High-entropy components typically influence low-entropy components, leading to a disordered and chaotic state within the complex system. This invention introduces the concept of information fields into Alzheimer's disease (AD) research. It reconstructs the dynamic distribution of brain information activity using a discretized approach and determines the aggregation direction and intensity of structural information based on the adjacency information entropy matrix. In a single iteration, a single element can aggregate all adjacent point-edge-surface-volume multi-level complex structural information. After multiple iterations, it can aggregate non-adjacent point-edge-surface-volume multi-level complex structural information. This can be used to reflect and reveal potential dynamic interaction patterns between brain regions and genes, thereby revealing potential spatiotemporal patterns in disease evolution and assisting in the accurate classification of samples.

[0013] 2. During the iteration process, the topological structure of the brain region gene network remains unchanged and the structural entropy difference matrix remains unchanged. This facilitates the output of complex feature vectors of multi-sample fusion point-edge-surface-volume structural information to the fully connected layer for classification. Then, the importance of the structural information corresponding to the point, line, surface and volume levels is back-statistically analyzed, thereby revealing potential key regulatory pathways.

[0014] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, 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 undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the method for extracting multi-level structural information of samples from the pathological simulation mechanism disclosed in the embodiments of the present invention.

[0016] Figure 2 This is a schematic diagram of the brain region-gene network construction steps disclosed in the embodiments of the present invention.

[0017] Figure 3 This is a schematic diagram of the stringless loop detection algorithm process disclosed in an embodiment of the present invention.

[0018] Figure 4 This is a schematic diagram of the complex information field model disclosed in an embodiment of the present invention. Detailed Implementation

[0019] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0020] Example 1 This embodiment discloses a method for extracting multi-level structural information of samples from a pathological simulation mechanism, such as... Figure 1 As shown, it includes the following steps: Step S1: Perform correlation analysis between brain region features and gene features to obtain the brain region gene network.

[0021] like Figure 2 As shown, this step is existing technology. First, the brain is divided into regions based on radiomics fMRI data, and then extracted and calculated... Time series corresponding to each brain region Simultaneously, based on SNP data, the corresponding... Gene expression sequences Viewing brain regions and genes as nodes in a network. , forming a node set The processed feature sequence is considered as the feature information carried by each node in the graph, denoted as... }, where, the former Each node represents a brain region time series, followed by... Each node represents a gene feature sequence; This represents the total number of nodes.

[0022] Based on this, Pearson correlation coefficients are calculated pairwise between brain region time series and gene coding sequences to measure the correlation strength between any two nodes, and then the edge weights between them are defined. This allows for the construction of brain region-gene networks. The calculation formula is as follows: ;in, Represents a node With nodes Edge weights between them Point With nodes The covariance of the corresponding sequence, This represents the standard deviation. Preferably, this step may set a threshold. Filter weights below the threshold The edge, that is, if Then place ;like Then construct a node With nodes The edges between Let the set of edges be formed by connecting edges. ,in Let be the total number of edges. The final brain region-gene network can be defined as... .

[0023] Step S2: Detect all two-dimensional complexes in the brain region gene network using stringless loops.

[0024] Similarly, the specific execution process of this step can be referred to Figure 3 The depth-first search algorithm used mainly includes the following steps: defining a stack. To record the nodes that have been visited, whenever a new node... Pushed onto the stack In the middle, the search algorithm is based on the point-edge adjacency matrix. Iterate through all reachable adjacent nodes of the given node and check the stack. Does the data contain previously visited nodes? If the stack Visited nodes were found Then choose from the stack. arrive All nodes in the stack form a ring. During the detection process, if multiple nodes have been visited, the algorithm selects the node closest to the top of the stack. Therefore, any closed loop formed using this method must be a stringless loop. For each detected closed loop, its corresponding set of nodes and edges is defined as a surface. .

[0025] Since this algorithm can only traverse the vertices and edges of a single connected component at a time, for graphs with multiple connected components, the algorithm must be applied independently to each component. This forms a face. Then, the top node of the stack must be popped to backtrack to the previous state, continue exploring other unvisited adjacent nodes, and look for other potential closed loops.

[0026] Finally, the set of faces can be denoted as Each face A subset of a node A subset of an edge express, The edges inside need to be connected end to end to form a closed loop. The total number of faces.

[0027] Step S3: Detect polyhedral structures from each two-dimensional complex based on Euler's formula.

[0028] Euler's formula in topology can describe important topological properties of polyhedra. For a closed, non-porous polyhedron, the number of vertices... Number of sides (edges) Number of faces Satisfy the following formula: Only when a geometric structure satisfies the aforementioned formula can it form a closed, hole-free three-dimensional polyhedron. Therefore, the specific execution process for this step is as follows: 3.1 First, in order to identify the three-dimensional polyhedron in the figure, start from an initial face. Begin depth-first search and push the results onto the stack. Each search starts from the top of the stack. Start via edge-face adjacency matrix Find unvisited adjacent faces and push them onto the stack. Then, check the most recently visited 4, 5, and so on, up to the maximum number of faces visited. The geometric structure formed by the faces is examined to determine whether it satisfies Euler's formula. If it does, the geometric structure is considered to form a three-dimensional complex. It consists of a subset of nodes. A subset of edges A subset of a face constitute.

[0029] 3.2 Repeat steps 3.1 above to complete the search for all faces and obtain the set of three-dimensional complexes. , This represents the total number of three-dimensional complexes.

[0030] Step S4: Construct a complex adjacency matrix representing the adjacency relationships between points, edges, faces, and volumes.

[0031] In this step, the complex adjacency matrix is ​​constructed. Specifically: ; in, It is a point-to-point adjacency matrix; It is an edge-edge adjacency matrix; It is a face-to-face adjacency matrix; It is a volume-to-volume adjacency matrix; It is a point-edge adjacency matrix; It is an edge-face adjacency matrix; Let be the face-volume adjacency matrix; where a complex is a subset of another complex, or there is an intersection between complexes of the same dimension, then an adjacency relationship exists and is assigned a value of 1; otherwise, it is assigned a value of 0. For transpose operation, This represents the multiplication operator. In this embodiment, points and edges are considered one-dimensional complexes, faces including triangles and other polygons are considered two-dimensional complexes, and volumes conforming to Euler's formula are considered three-dimensional complexes, which will not be elaborated further.

[0032] Step S5: Calculate the initial structural information matrix representing the spatial location information of the elements in the field based on the complex adjacency matrix and the first learning parameter matrix.

[0033] This step transforms the unordered set form into an ordered vector form to describe the relative positions and information between the complexes; the elements in the field include points, edges, faces, and volumes. Preferably, the calculation formula for this step is: ;in, This represents the learnable parameter matrix. This represents the activation function; The sum of the number of all points, edges, faces, and volumes.

[0034] Step S6: Calculate the adjacency degree matrix representing the adjacency degree of each element based on the initial structural information matrix, and then obtain the adjacency structure entropy matrix representing the information intensity of elements in the complex based on the adjacency degree matrix.

[0035] In this step, the adjacency degree matrix and adjacency structure entropy matrix are terms used in the prior art. This invention innovatively introduces them to assist in the simulation of pathological mechanisms. The specific calculation process is as follows: ;in, It is a matrix of all ones. It is the identity matrix. This is the adjacency degree matrix. Represent the Hadamard product; obtain the adjacency structure entropy matrix from the adjacency degree matrix. The calculation formula is: ; ;in, Adjacency matrix The inverse matrix, For the intermediate matrix calculated, This indicates the logarithmic operation.

[0036] Step S7: Calculate the structural entropy difference between any two elements in the adjacent structural entropy matrix, and then compare each structural entropy difference with the difference screening threshold. Structural entropy differences that are greater than or equal to the difference screening threshold are retained, while structural entropy differences that are less than the difference screening threshold are set to zero. Then, based on the retained structural entropy differences, the information flow between the corresponding elements is determined to obtain the structural entropy difference matrix.

[0037] In this step, preferably, the adjacency structure entropy matrix The two elements , structural entropy difference between The calculation formula is: ;in, Use the difference filtering threshold.

[0038] Step S8: Based on the structural entropy difference matrix and the second learning parameter matrix, perform information aggregation and update processing on the initial structural information matrix. Then, the updated structural information matrix is ​​iterated again based on the structural entropy difference matrix and the third learning parameter matrix. This process is repeated, introducing a new learning parameter matrix in each iteration and using the same structural entropy difference matrix to determine the flow of information aggregation. After at least three iterations, the target structural information matrix is ​​obtained.

[0039] Reference Figure 4This step, which involves information aggregation and updating of the structural information matrix based on the structural entropy difference matrix and the learning parameter matrix, can specifically include: 8.1 First, calculate the structural information change matrix. The specific calculation formula is as follows: ; in, To learn the parameter matrix, For the first The structural information matrix before information aggregation occurs during the next iteration.

[0040] 8.2, Recalculate and update the first... The structural information matrix after the next iteration The calculation formula is as follows: ;in, This is an information regulation parameter used to balance the contributions of historical structural information and clustering structural information. When... When the value is large, the update result relies more on existing structural information, thereby enhancing the model's preservation of global topological stability; while when When the value is small, the update result focuses more on the response to new information, so as to improve the model's sensitivity to local changes and dynamic features.

[0041] In the above steps, steps S5-S7 can be based on Figure 4 The adjacency information entropy calculation layer is implemented in step S8, which is based on... Figure 4 This is achieved through a multi-level information aggregation layer.

[0042] Step S9: Flatten the target structural information matrix to obtain the complex feature vector of the fused point-edge-face-volume structural information output to the fully connected layer.

[0043] In this step, the fully connected layer consists of several neurons, each connected to all nodes in the previous layer. Its function is to extract multi-level high-level features from the brain region-gene cavity complex and output classification results. After multiple iterations through the multi-level structural information aggregation layer, a structural information matrix representing the multi-level structural features of the brain region-gene cavity complex is obtained. Before entering the fully connected layer, the matrix needs to be flattened to obtain a one-dimensional vector representing the multi-level structural features of the brain region-gene cavity complex. This is to facilitate the output of classification results by the fully connected layer.

[0044] vector The processing related to the input of the fully connected layer is existing technology and will not be elaborated upon. However, the final step in the fully connected layer... Layer utilization The function maps the neuron's output to... Within the interval, to obtain a The vector, where, This vector represents the sample category for this disease diagnosis task. Each value represents the probability that the sample belongs to a specific category.

[0045] In the training set-based processing of this embodiment, a corresponding loss function can be constructed to determine the specific values ​​of each learning parameter matrix in the above steps. After the classification accuracy of positive and negative samples meets expectations, the model determined by the learning parameter matrix can be used for classification prediction of new samples. Such processing is existing technology and will not be elaborated further.

[0046] Example 2 This embodiment discloses a system for extracting multi-level structural information of samples from a pathological simulation mechanism, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method described in Embodiment 1 above.

[0047] In summary, the method and system for extracting multi-level structural information of samples from the pathological simulation mechanism disclosed in the embodiments of the present invention have at least the following technical effects: 1. The brain is essentially an entropy-increasing system. High-entropy components typically influence low-entropy components, leading to a disordered and chaotic state within the complex system. This invention introduces the concept of information fields into Alzheimer's disease (AD) research. It reconstructs the dynamic distribution of brain information activity using a discretized approach and determines the aggregation direction and intensity of structural information based on the adjacency information entropy matrix. In a single iteration, a single element can aggregate all adjacent point-edge-surface-volume multi-level complex structural information. After multiple iterations, it can aggregate non-adjacent point-edge-surface-volume multi-level complex structural information. This can be used to reflect and reveal potential dynamic interaction patterns between brain regions and genes, thereby revealing potential spatiotemporal patterns in disease evolution and assisting in the accurate classification of samples.

[0048] 2. During the iteration process, the topological structure of the brain region gene network remains unchanged and the structural entropy difference matrix remains unchanged. This facilitates the output of complex feature vectors of multi-sample fusion point-edge-surface-volume structural information to the fully connected layer for classification. Then, the importance of the structural information corresponding to the point, line, surface and volume levels is statistically analyzed to reveal potential key regulatory pathways. The statistical analysis of importance and the processing of key regulatory pathways are the same as those in the "Screening Method and System for Subsets of Key Regulatory Pathways in Medical Large Models" published in CN121790010A, and will not be elaborated here.

[0049] 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 method for extracting multi-level structural information of samples from a pathological simulation mechanism, comprising: The brain region gene network is obtained by performing correlation analysis between brain region features and gene features, characterized by further comprising: Detection of all two-dimensional complexes in brain region gene networks using stringless loops; Polyhedral structures were detected from the two-dimensional complexes based on Euler's formula. Construct a complex adjacency matrix that represents the adjacency relationships between points, edges, faces, and volumes; The initial structural information matrix representing the spatial position information of the elements in the field is calculated based on the complex adjacency matrix and the first learning parameter matrix. Based on the initial structural information matrix, calculate the adjacency degree matrix representing the adjacency degree of each element, and then obtain the adjacency structure entropy matrix representing the information intensity of elements in the complex based on the adjacency degree matrix. Calculate the structural entropy difference between any two elements in the adjacency structural entropy matrix, then compare each structural entropy difference with the difference filtering threshold. Structural entropy differences that are greater than or equal to the difference filtering threshold are retained, while structural entropy differences that are less than the difference filtering threshold are set to zero. Then, determine the information flow direction between the corresponding elements based on the retained structural entropy differences to obtain the structural entropy difference matrix. The initial structural information matrix is ​​processed by information aggregation and updating based on the structural entropy difference matrix and the second learning parameter matrix. Then, the updated structural information matrix is ​​iterated again based on the structural entropy difference matrix and the third learning parameter matrix. This process is repeated, with a new learning parameter matrix introduced in each iteration and the same structural entropy difference matrix used to determine the flow of information aggregation. The target structural information matrix is ​​obtained after at least three iterations. Flatten the target structure information matrix to obtain the complex feature vector of the fused point-edge-face-volume structure information output to the fully connected layer; Wherein, the adjacency structure entropy matrix Two elements , structural entropy difference between The calculation formula is: ； in, A threshold is set for the difference.

2. The method for extracting multi-level structural information of samples according to the pathological simulation mechanism described in claim 1, characterized in that, The formula for calculating the initial structural information matrix, which represents the spatial position information of the elements in the field, based on the complex adjacency matrix and the first learning parameter matrix, is as follows: ； in, This represents the learnable parameter matrix. This represents the activation function. Represents the complex adjacency matrix; The sum of the number of all points, edges, faces, and volumes.

3. The method for extracting multi-level structural information of samples according to the pathological simulation mechanism described in claim 2, characterized in that, ； in, It is a point-to-point adjacency matrix; It is an edge-edge adjacency matrix; It is a face-to-face adjacency matrix; It is a volume-to-volume adjacency matrix; It is a point-edge adjacency matrix; It is an edge-face adjacency matrix; Let be the face-volume adjacency matrix; where a complex is a subset of another complex, or there is an intersection between complexes of the same dimension, then an adjacency relationship exists and is assigned a value of 1; otherwise, it is assigned a value of 0. For transpose operation, This represents the multiplication operator.

4. The method for extracting multi-level structural information of samples from the pathological simulation mechanism according to claim 3, characterized in that, The formula for calculating the adjacency matrix representing the adjacency degree of each element based on the initial structural information matrix is ​​as follows: ； in, It is a matrix of all ones. It is the identity matrix. This is the adjacency degree matrix. It represents the Hadamardi (or Hadama) stack; The adjacency structure entropy matrix is ​​obtained from the adjacency degree matrix. The calculation formula is: ； ； in, Adjacency matrix The inverse matrix, For the intermediate matrix calculated, This indicates the logarithmic operation.

5. The method for extracting multi-level structural information of samples from the pathological simulation mechanism according to claim 4, characterized in that, During any iteration, the information aggregation and update processing of the structural information matrix based on the structural entropy difference matrix and the learning parameter matrix includes: First calculate the structural information change matrix The specific calculation formula is as follows: ； in, To learn the parameter matrix, For the first The structural information matrix before information aggregation occurs during the next iteration; Recalculate and update the first The structural information matrix after the next iteration The calculation formula is as follows: ； in, These are parameters for information control.

6. A system for extracting multi-level structural information of samples from a pathological simulation mechanism, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1 to 5.

Images