CN-121982426-A - CT image classification method based on graph neural network

CN121982426ACN 121982426 ACN121982426 ACN 121982426ACN-121982426-A

Abstract

The application relates to a CT image classification method based on a graph neural network, and belongs to the field of image classification. The CT image classification method comprises the steps of preprocessing CT scanning data, converting the CT scanning data into a node feature matrix, combining the node feature matrix based on distance coding and similarity coding to obtain a distance-similarity graph, updating the node feature matrix in the distance-similarity graph, calculating comprehensive scores based on a topological structure and feature dimensions, pooling the node feature matrix to obtain an updated distance-similarity graph, receiving the updated distance-similarity graph, calculating feature similarity, performing mask processing to obtain an updated similarity coding matrix, fusing the updated distance-similarity graph and the similarity coding matrix, inputting the updated distance-similarity graph and the updated similarity coding matrix into a discrimination network and a Softmax activation function, and outputting a CT image classification result. The application solves the problems of low accuracy and robustness of focus classification caused by the difficulty in accurately modeling the dynamic semantic relation between CT images in the prior art.

Inventors

JIANG WENCHAO
LI JUNHANG

Assignees

广东工业大学

Dates

Publication Date: 20260505
Application Date: 20260327

Claims (10)

1. The CT image classification method based on the graph neural network is characterized by comprising the following steps of: s1, preprocessing CT scanning data, converting the CT scanning data into a node feature matrix, and respectively obtaining a distance coding adjacent matrix and a similar coding adjacent matrix based on distance coding and similar coding; S2, updating a node feature matrix in the distance-similarity graph, calculating a comprehensive score based on a topological structure and feature dimensions, pooling the comprehensive score to obtain a pooled node matrix, and replacing the node feature matrix with the pooled node matrix to obtain an updated distance-similarity graph; S3, receiving the updated distance-similarity graph, calculating the characteristic similarity and carrying out mask processing to obtain an updated similarity coding matrix, fusing the updated distance-similarity graph and the similarity coding matrix, inputting the fused updated distance-similarity graph and similarity coding matrix into a discrimination network and a Softmax activation function, and outputting a CT image classification result.
2. The CT image classification method based on the neural network of claim 1, wherein S1 further comprises the steps of: acquiring CT scan data, performing size normalization, window width and window level adjustment and lung parenchyma cutting on the CT scan data, and simultaneously performing data enhancement to obtain A plurality of pretreated CT slices; inputting the preprocessed CT slices one by one into a backbone feature extraction network, and extracting one for each slice Deep feature vector of dimension Thereby obtaining the node characteristic matrix 。
3. The CT image classification method based on the neural network of claim 2, wherein S1 further comprises the steps of: consider each preprocessed CT slice as a graph node Collecting all graph nodes to obtain a graph node set: defining slice index based on physical location of preprocessed CT slices And establishing connectivity according to the spatial arrangement relation of the CT slices after pretreatment, and calculating any two slices And Is the physical relative distance of (2) Calculating the variable weight of the CT slice after preprocessing through distance coding according to the physical relative distance : Wherein, the Standard deviation as gaussian function; Obtaining all variable weights Obtaining a distance coding adjacency matrix 。
4. The CT image classification method based on the neural network of claim 3, wherein S1 further comprises the steps of: According to the node characteristic matrix Is used for constructing initial adjacency matrix ; Setting a reservation ratio According to the retention ratio Counting the number of edges that need to be reserved ; According to the initial adjacency matrix And the number of edges that need to be reserved Determining a first sparseness threshold Wherein, the method comprises the steps of, Is a similarity score; Based on a first sparseness threshold value For initial adjacency matrix Performing similar coding, wherein the similar coding is binarization processing, and obtaining sparse adjacent matrix elements after coding : Wherein, the Is node characteristic matrix Middle (f) Features and the first Similarity scores for the individual features; Acquiring all sparse adjacent matrix elements Obtaining similar coding adjacency matrix ; Encoding distance into adjacency matrix Similarly encoded adjacency matrix Combining to obtain a distance-similarity adjacency matrix ; Distance-similarity adjacency matrix Node characteristic matrix Combining to obtain a distance-similarity graph 。
5. The method for classifying CT images based on a neural network according to claim 4, wherein S2 further comprises the steps of: Receive distance-similarity map Node characteristic matrix of current layer through preset graph convolution network Updating and aggregating neighborhood information to obtain updated node characteristics : Wherein, the As distance-like adjacency matrix Is used for the degree matrix of the (c), In order for the weight matrix to be learnable, Is a nonlinear activation function; for updated node characteristics Node scoring based on topological structure to obtain structural score ; For updated node characteristics Node scoring is carried out based on feature dimensions to obtain feature scores ; Computing a composite score for each node , wherein, Is super parameter: According to comprehensive scoring From the slave In each node, according to a preset pooling rate Selecting Individual nodes, constitute a pool of post-point sets : Pooling the post-point set Performing pooling treatment to obtain pooling node characteristics : Pooling node features For distance-similarity diagram Replacing to obtain updated distance-similarity diagram 。
6. The method for classifying CT images based on a neural network according to claim 5, wherein S3 further comprises the steps of: Acquiring updated distance-similarity graphs According to the characteristics of the pooled nodes And transposed matrix thereof Calculating a similarity matrix : For similarity matrix The weights of all candidate edges in the list are ordered, and the candidate edges are ranked in the order Weight value of position as second sparse threshold : Filtering and screening useless edges based on a second sparse threshold For similarity matrix Performing mask processing to obtain updated similar coding matrix elements according to the following formula : Represented as pooled node features Middle (f) Features and the first The proximity of individual features at the deep semantic level; obtaining all updated similar coding matrix elements to obtain an updated similar coding matrix 。
7. The method for classifying CT images based on a neural network according to claim 6, wherein S3 further comprises the steps of: The updated similar coding matrix Distance-like adjacency matrix Performing linear weighted fusion to obtain a reconstructed adjacency matrix : Wherein, the Is a smooth weight coefficient; using GCN information propagation through reconstructed adjacency matrices Driving a new round of graph rolling network operation, and performing depth field information aggregation by the nodes according to the updated topological connection, and extracting to obtain a global feature vector : Wherein, the Is a weight matrix which can be learned; global feature vector And inputting the CT image classification result into a discrimination network and a Softmax activation function, and outputting the CT image classification result.
8. The CT image classification method based on a neural network as claimed in claim 7, characterized in that the discrimination network pairs global feature vectors Deep feature extraction is performed, and the deep feature extraction consists of cascade full-connection layers, namely The layer full join operation may be expressed as: Wherein, the Is the first The weight matrix of the layer is used to determine, Is the first The bias term of the layer is used, Is the first Layer output; The terminal output layer of the discrimination network maps the features to the category space to obtain an original scoring vector set : Wherein, the The classification total number is preset; the Softmax activation function will be the original scoring vector Conversion to mutually exclusive probability distributions First, the The probability of each category is calculated by the following formula: Wherein, the Is the first The number of original scoring vectors is the number of, Is the first A plurality of original scoring vectors; According to probability distribution Selecting the index with the maximum probability value as the final CT image classification result 。
9. The application of the CT image classification method based on the image neural network is characterized in that the method is applied to classification of pneumoconiosis, and CT scanning data are chest CT scanning data of a patient.
10. A computer readable storage medium, wherein the computer readable storage medium includes a CT image classification method program based on a graph neural network, and the CT image classification method program based on the graph neural network implements the steps of a CT image classification method based on a graph neural network according to any one of claims 1 to 8 when the CT image classification method program based on the graph neural network is executed by a processor.

Description

CT image classification method based on graph neural network Technical Field The application relates to the technical field of image classification, in particular to a CT image classification method based on a graph neural network. Background Pneumoconiosis is a occupational lung disease, caused by long-term inhalation of mineral dust, and if the dust is not intervened in time, irreversible lung tissue damage is often caused, so early accurate detection is important for preventing and controlling and delaying illness. The prior art is applied to classification of pneumoconiosis CT images through a graph neural network, but has 3 problems that (1) a graph structure is constructed through full connection, chain connection, star connection and the like, the graph construction strategy of the structure can describe the relationship among CT slices to a certain extent, but is generally single and limited, the relationship among CT slices is difficult to fully characterize at the same time, (2) the graph structure expression capacity is enhanced by randomly discarding nodes and carrying out forward propagation for a plurality of times, the graph pooling strategy of the scheme is often focused on a random sampling mechanism, dynamic evaluation of the node importance is difficult to fully combine multi-level structure information, and (3) pruning is carried out on the graph structure, and although the scheme evaluates the edge weight and cuts out the connecting edge with lower weight to reduce redundant information and simplify the graph structure, the graph structure updating process usually depends on a single edge weight evaluation standard, and the complex relationship among the nodes is difficult to fully describe. Disclosure of Invention The application provides a CT image classification method based on a graph neural network, which can solve the problems of low accuracy and low robustness of focus classification caused by difficulty in accurately modeling dynamic semantic relations among CT images in the prior art. In order to achieve the above object, according to a first aspect of the present application, there is provided a CT image classification method based on a neural network, the method comprising the steps of: s1, preprocessing CT scanning data, converting the CT scanning data into a node feature matrix, and respectively obtaining a distance coding adjacent matrix and a similar coding adjacent matrix based on distance coding and similar coding; S2, updating a node feature matrix in the distance-similarity graph, calculating a comprehensive score based on a topological structure and feature dimensions, pooling the comprehensive score to obtain a pooled node matrix, and replacing the node feature matrix with the pooled node matrix to obtain an updated distance-similarity graph; S3, receiving the updated distance-similarity graph, calculating the characteristic similarity and carrying out mask processing to obtain an updated similarity coding matrix, fusing the updated distance-similarity graph and the similarity coding matrix, inputting the fused updated distance-similarity graph and similarity coding matrix into a discrimination network and a Softmax activation function, and outputting a CT image classification result. In order to achieve the above object, according to a second aspect of the present application, there is also provided an application of a CT image classification method based on a graph neural network, the method being applied to classification of pneumoconiosis, the CT scan data being chest CT scan data of a patient. In order to achieve the above object, according to a third aspect of the present application, there is also provided a computer-readable storage medium including therein a CT image classification method program based on a graph neural network, which when executed by a processor, implements the steps of a CT image classification method based on a graph neural network as described in the first aspect. According to the application, a graph structure is constructed by utilizing the physical distance relation and the deep feature similarity relation between CT sections, so that the graph model can capture the spatial continuity and the semantic relevance in the sequence image at the same time, thereby describing the distribution mode of the focus between different sections more accurately, realizing the self-adaptive screening of key nodes by comprehensively scoring the importance of the node structure and the feature discrimination capability, keeping the most valuable information for disease identification while compressing the graph scale, improving the identification capability of the model to the micro focus, and enabling the graph structure to be optimized continuously along with the feature learning process by dynamically adjusting the connection relation of the graph after the node feature update, thereby more accurately describing the semantic relation in the sequence