Self-supervised medical image segmentation method and system based on random walk path

Abstract

The application provides a kind of self-supervised medical image segmentation method and system based on random walk path, the global feature is obtained by processing the medical image obtained;Based on the short-distance random walk strategy and the long-distance random walk strategy, the homogeneity neighborhood distribution and the heterogeneity neighborhood distribution of the center node are sampled respectively, and the random walk path is formed;Each random walk path is encoded, the weight of each random walk path is learned through attention mechanism, the path information after weight aggregation is spliced, and the global embedding representation is generated;Extract the node and edge that the topological distance of the node and the edge does not exceed the set threshold from the center node to form a subgraph, introduce random disturbance to generate a contrast view, and use the similarity between nodes to calculate embedding, generate a contrast subgraph with node features and edges;The subgraph is optimized by using structured contrast loss.The application solves the problem of long-distance node influence attenuation, and enhances the globality and locality of medical image representation.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation, specifically relating to a self-supervised medical image segmentation method and system based on random walk paths. 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 widespread application of medical image data in various fields (such as tumor diagnosis, organ segmentation, pathological image analysis, and biomedical research), graph representation learning has become a research hotspot in medical image processing. Traditional graph representation learning methods mainly rely on labeled data, using supervised learning strategies to generate embedding representations of pixels or voxels. However, due to the inherent imbalance and heterogeneity in medical image data, existing methods suffer from the following major problems when processing complex medical image structures:

[0004] (1) Topological imbalance: The connections between graph nodes are complex and diverse, manifested as asymmetric and uneven distribution of nodes in the topological space. This imbalance makes it difficult for traditional graph representation learning methods to accurately capture the relationships between nodes, affecting the quality of the representation.

[0005] (2) Homogeneous structure assumption constraint: Existing methods usually assume that graphs are homogeneous, that is, the connected nodes have similar characteristics. However, in practical applications, many graph data are heterogeneous, and traditional methods are difficult to effectively capture these complex heterogeneous relationships.

[0006] (3) Weakened information over long distances: As the topological distance between nodes increases, the influence of nodes on the central node gradually weakens, making it difficult to effectively utilize the information of distant nodes and affecting the globality of the graph representation.

[0007] (4) Label decay problem: Self-supervised learning methods need to automatically mine supervision signals from the data. However, in graph data, due to the scarcity of labeled data, the supervision signals that can be provided are naturally weak, which affects the learning effect of the model. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes a self-supervised medical image segmentation method and system based on random walk paths. By deeply exploring the complex organizational relationships and uneven distributions in medical images, this invention improves the accuracy and robustness of automatic segmentation technology when processing medical images with diverse and irregular boundaries, thereby providing more precise support for medical diagnosis and treatment planning.

[0009] According to some embodiments, the present invention adopts the following technical solution:

[0010] A self-supervised medical image segmentation method based on random walk path, comprising the following steps:

[0011] The acquired medical image is preprocessed and converted into a graph data structure, and the initial features of each node in the graph data structure are extracted, and the initial features are standardized.

[0012] The features of each node are updated based on the message passing mechanism, and the updated features of all nodes are aggregated to obtain global features.

[0013] Based on the short-range random walk strategy and the long-range random walk strategy, the homogeneity neighborhood distribution and the heterogeneity neighborhood distribution of the center node are sampled respectively to form a random walk path.

[0014] Each random walk path is encoded, the weights of each random walk path are learned through an attention mechanism, the path information after weight aggregation is spliced, and a global embedding representation is generated.

[0015] A subgraph is extracted from nodes and edges with a topological distance from the center node not exceeding a set threshold, a contrast view is generated by introducing random disturbance, and embedding calculation is performed using the similarity between nodes to generate a contrast subgraph with node features and edges.

[0016] Based on the subgraph and the contrast subgraph, the subgraph is optimized using a structured contrast loss.

[0017] As an optional implementation, the process of converting into a graph data structure includes defining each pixel or voxel in the medical image as a node in the graph, establishing edges between nodes according to the spatial proximity of pixels or voxels, and reflecting the similarity between nodes through the weight of the edge.

[0018] As an optional implementation, it further includes using an adjacency matrix to represent the topological structure of the graph, for an unweighted graph, the adjacency matrix is a binary matrix, indicating whether there is an edge connection between nodes; for a weighted graph, the adjacency matrix elements are the weight values of the edges, indicating the similarity between nodes.

[0019] As an optional implementation, the features include several of pixel values, color information, and local gradient information.

[0020] As an optional implementation, the process of aggregating the updated features of all nodes includes propagating node features through an adjacency matrix to generate new node representations, using a graph convolution network encoder to aggregate features of the graph data, and through the superposition of multiple layers of graph convolution networks, the final features of the nodes contain information of their direct neighbors and information of more distant nodes. At the last layer of the encoder, the features of all nodes are aggregated into global features.

[0021] As an alternative embodiment, in the random walk process, the non-normalized transition probability is used to maximize the topological relationship between nodes, and the transition probability is:

[0022]

[0023] where L i-1i+1 represents the shortest path length between nodes v i-1 and V i+1 , when L i-1，i+1 = 1, v i-1 is the 1-hop neighbor of v i+1 ; alpha and beta are transition parameters of random walk, which control the preference of near and far neighbors respectively; when alpha < 1 and beta > 1, the random walk strategy collects information of adjacent nodes and extracts homogeneity information; when alpha > 1 and beta < 1, the random walk strategy preferentially accesses more distant nodes and extracts heterogeneity information.

[0024] As an alternative embodiment, the Gated Recurrent Unit encoder is used to encode each random walk path, the importance weight of each path is learned through attention mechanism, the path information aggregated through weight is calculated by multi-head attention, and the result is spliced.

[0025] As an alternative embodiment, a contrast view is generated by introducing random disturbance, and the similarity between nodes is used for embedding calculation, and the process of generating a contrast view with node features and edges includes: the embedding similarity calculation between two nodes v j and v l is:

[0026] S(v j , v l ) = theta (W c [W v v j || W v v l ])

[0027] wherein, is a subgraph, W c and W v are learnable weight matrices, and theta (·) is an activation function.

[0028] Similarity-based node embedding aggregation:

[0029] W l = Softmax(sim(v j , v l ));

[0030]

[0031] where Softmax is used to normalize the weights, W l denotes the aggregated node embeddings.

[0032] As an alternative implementation, in the process of optimizing the subgraph by using the structured contrastive loss, the Wasserstein contrastive loss and the Gromov-Wasserstein contrastive loss are used to measure the similarity between nodes and edges respectively, wherein the Wasserstein contrastive loss is used to represent the transmission cost between the subgraph and its contrastive subgraph, and the optimization is achieved by minimizing the difference in node similarity between positive samples and maximizing the difference in node similarity between negative samples, and the Gromov-Wasserstein contrastive loss is used to capture the structural similarity of edges in the subgraph, and the optimization is achieved by minimizing the difference in edge similarity between positive samples and maximizing the difference in edge similarity between negative samples.

[0033] A self-supervised medical image segmentation system based on random walk paths, comprising:

[0034] A preprocessing module configured to preprocess the acquired medical image and convert it into a graph data structure, extract initial features of each node in the graph data structure, and perform standardization processing on the initial features.

[0035] A node information updating module configured to update the features of each node based on a message passing mechanism, and aggregate the updated features of all nodes to obtain global features.

[0036] A multi-hop path sampling module configured to sample homogeneity neighborhood distribution and heterogeneity neighborhood distribution of a center node based on a short-range random walk strategy and a long-range random walk strategy respectively, and form random walk paths.

[0037] A path information aggregation module configured to encode each random walk path, learn the weights of each random walk path through an attention mechanism, and concatenate the path information after the weight aggregation to generate global embedding representation.

[0038] A subgraph generation module configured to extract nodes and edges with a topological distance from a center node not exceeding a set threshold to form a subgraph, introduce random disturbance to generate a contrastive view, and use the similarity between nodes to perform embedding calculation to generate a contrastive subgraph with node features and edges.

[0039] A structured optimization module configured to optimize the subgraph based on the subgraph and the contrastive subgraph by using the structured contrastive loss.

[0040] An electronic device comprising a memory and a processor, and computer instructions stored on the memory and running on the processor, when the computer instructions are run by the processor, the steps in the above method are completed.

[0041] Compared with the prior art, the present application has the following beneficial effects:

[0042] The present application introduces the sub-medical image neighborhood concept of multi-hop path, comprehensively captures the relationship and local features between pixels in the medical image, overcomes the limitation of homogeneity, solves the problem of distance node influence attenuation, and enhances the globality and locality of medical image representation.

[0043] The present application uses an inter-path aggregation strategy based on an attention mechanism, aggregates information in the multi-hop path, learns the homogeneity and heterogeneity features in the path, not only retains the original attributes of the nodes, but also maintains the initial structure connection of the nodes in the random walk sequence.

[0044] The present application introduces a subgraph sampling strategy based on aggregated multi-hop path, which can capture more distant node information and reduce the negative effects caused by topological distance attenuation.

[0045] The present application combines a structured contrast loss, uses Wasserstein distance and Gromov-Wasserstein distance to measure the similarity between subgraphs, thereby maximizing the similarity of positive subgraphs and minimizing the similarity of negative subgraphs, and enhances the self-supervised signal of the model.

[0046] The present application can effectively process medical image data with complex heterogeneous structure, break through the limitation of traditional methods, and has strong adaptability; through multi-hop path sampling and structured contrast learning, the topological imbalance problem can be better solved, and the stability of medical image representation is enhanced.

[0047] The present application is suitable for graph data processing in various fields, and has wide application prospects in social network analysis, bioinformatics and knowledge graph construction.

[0048] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are used for explanation. BRIEF DESCRIPTION OF DRAWINGS

[0049] The drawings accompanying the specification of the present application are used to provide a further understanding of the present application, the schematic embodiments of the present application and the description thereof are used to explain the present application, and do not constitute an improper limitation on the present application.

[0050] Figure 1 It is a flowchart of an embodiment;

[0051] Figure 2 It is a schematic diagram of the structure of an electronic device. DETAILED DESCRIPTION

[0052] The present application will be further described below in combination with the drawings and embodiments.

[0053] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0054] It is also important to note that the use of the term "example" herein is used in the sense of serving as an example, instance, or illustration, and not meant to be limited to a specific example or instantiation. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0055] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0056] Embodiment one

[0057] A self-supervised graph topology imbalance learning method based on random walk path, as shown in Figure 1 includes the following steps:

[0058] Step S1: pre-process the obtained medical image and convert it into a graph data structure, extract the initial features of each node in the graph data structure, and standardize the initial features;

[0059] Step S2: update the features of each node based on the message passing mechanism, and aggregate the features of all nodes after updating to obtain global features;

[0060] Step S3: based on the short-distance random walk strategy and the long-distance random walk strategy, respectively sample the homogeneity neighborhood distribution and the heterogeneity neighborhood distribution of the center node to form a random walk path;

[0061] Step S4: encode each random walk path, learn the weight of each random walk path through an attention mechanism, splice the path information after weight aggregation, and generate a global embedding representation;

[0062] Step S5: extract a subgraph composed of nodes and edges whose topological distance from the center node does not exceed a set threshold, generate a contrast view by introducing random disturbance, and use the similarity between nodes to calculate the embedding to generate a contrast subgraph with node features and edges;

[0063] Step S6: based on the subgraph and the contrast subgraph, optimize the subgraph using a structured contrast loss.

[0064] The steps are described in detail as follows.

[0065] Firstly, the medical image needs to be pre-processed and converted into a graph data structure in step S1. Each pixel or voxel in the medical image is defined as a node in the graph. The edges between nodes are established according to the spatial proximity of pixels or voxels, and the weights of the edges reflect the similarity between nodes, usually calculated by pixel value difference, gradient change or texture features.

[0066] Secondly, the initial features of each node are extracted, which can include pixel values, color information (RGB or grayscale values), local gradient information, etc. These features constitute the node feature matrix, representing the attributes of the nodes. Subsequently, the adjacency matrix is used to represent the topological structure of the graph. For an unweighted graph, the adjacency matrix is a binary matrix indicating whether there is an edge connection between nodes; for a weighted graph, the adjacency matrix elements are the weight values of the edges, representing the similarity between nodes.

[0067] Finally, the node features are standardized, which can be done using mean normalization or Z-score standardization methods to ensure that the features are within the same numerical range, which helps to improve the stability of numerical values during training.

[0068] Of course, in other embodiments, other methods can be used to standardize the data.

[0069] In step S2, the encoder based on graph convolutional network (GCN) aggregates features of the graph data through a message passing mechanism. The features of each node are determined not only by its own features but also by the information of its neighbor nodes through adjacency relationships, gradually generating global feature representations.

[0070] The feature of each node at the l-th layer is updated by the following formula:

[0071]

[0072] where H (l) is the l-th layer node feature matrix, is the adjacency matrix with self-loop added, is the degree matrix of the adjacency matrix, W (l) is the l-th layer learnable weight matrix, and σ is a nonlinear activation function (such as ReLU or LeakyReLU). This formula shows that the node features propagate through the adjacency matrix to generate new node representations.

[0073] To extract features at different levels, multiple layers of GCN are usually stacked, and each layer further integrates the node features with its neighbor nodes. Through the superposition of multiple layers, the final features of the nodes not only contain the information of their immediate neighbors but also contain the information of their more distant nodes.

[0074] In the last layer of the GCN encoder, the features of all nodes are aggregated into global features.

[0075] In step S3, the homophily neighborhood distribution and the heterophily neighborhood distribution of the center node are sampled by setting the near-distance and far-distance random walk strategies. The near-distance strategy focuses on capturing the neighbor information directly connected to the center node to obtain homophily information, while the far-distance strategy further samples the node information far away from the center node topology to obtain heterophily information.

[0076] In the random walk process, a non-normalized transition probability is used as follows to maximize the topological relationship between nodes.

[0077]

[0078] where L i-1，i+1 represents the shortest path length between nodes v i-1 and v i+1 . When L i-1，i+1 = 1, v i-1 is the 1-hop neighbor of v i+1 . When α < 1 and β > 1, the random walk is similar to a breadth-first search, collecting information from adjacent nodes and extracting homophily information; when α > 1 and β < 1, the random walk is similar to a depth-first search, preferentially accessing more distant nodes and extracting heterophily information.

[0079] In step S4, based on the multiple path information obtained by sampling, a GRU (Gated Recurrent Unit) encoder is used to encode each path, and then the importance weight of the path is learned through an attention mechanism.

[0080] The path encoding formula is as follows,

[0081] s p = GRU({g pk |k∈K})

[0082] where s p is the encoding feature of each path, g pk is the encoding feature obtained by the encoder for node v pk , and GRU is a gated recurrent unit used to obtain the encoding feature of each path.

[0083] The path weight calculation formula is as follows,

[0084]

[0085] where W path is the normalized weight of each path, W node is a learnable attention parameter, θ(·) is a LeakyReLU activation function, and the softmax function is used to normalize the weight.

[0086] The path information after weight aggregation is calculated by multi-head attention, and the results are spliced to generate a unified global embedding representation.

[0087] The multi-head attention aggregation formula is as follows,

[0088]

[0089] where W head is a learnable multi-head attention weight, e i is the global embedding representation of the center node v i , and cat(·) represents a feature splicing operation.

[0090] In step S5, based on the strategy of aggregated multi-hop path, subgraph sampling is performed to extract nodes and edges with a topological distance from the center node not exceeding a set threshold to form a subgraph.

[0091] The specific sampling process includes:

[0092] Obtain the graph G, the center node v i , the sampling path length threshold K, the parameter p and the parameter q;

[0093] Initialize the sampling path length len path =0, and the sampling subgraph contains the center node v i ;

[0094] When len_path

[0095] If len_path

[0096] Randomly select a first-order neighborhood node v j connected to v i ;

[0097] Add the edge (v I , v j ) to the sampling subgraph

[0098] Increase the value of len_path

[0099] Set v j as the new center node v i ;

[0100] Otherwise:

[0101] According to the probability parameters p and q, perform random walk to select the next node v k ;

[0102] Add the edge (v I , v k) Add the subgraph of the sample

[0103] Add 1 to the value of len_path;

[0104] Set v k as the new center node v i ;

[0105] Return the subgraph of the sample

[0106] By introducing a random perturbation of lower probability to generate a contrastive view, and using the similarity between nodes for embedding calculation, a contrastive view with the same node and edge relationship is generated.

[0107] The embedding similarity calculation formula between two nodes v j and :

[0108] S(v j , v l ) = θ(W c [W v v j || W v v l ])

[0109] Where W c and W v are learnable weight matrices, and θ(·) is a LeakyReLU activation function.

[0110] Similarity-based node embedding aggregation:

[0111] W l = softmax(sim(v j , v l ));

[0112]

[0113] Where Softmax is used to normalize the weights, and W l represents the aggregated node embedding.

[0114] Through the above steps, a contrastive subgraph with node features and edges is obtained, denoted as

[0115] In step S6, after completing subgraph sampling and contrastive view generation, a structured contrastive loss is used to optimize the model. In this embodiment, Wasserstein contrastive loss L w and Grosmov-Wasserstein contrastive loss L gw are used to measure the similarity between nodes and edges, respectively.

[0116] The loss function is as follows:

[0117] L = aL w + (1-a)L gw ;

[0118] where L is the Wasserman contrastive loss. L w is the Wasserman contrastive loss, which measures the node similarity between positive and negative samples. L gw is the Grover-Wasserman contrastive loss, which measures the edge similarity between positive and negative samples. a is a hyperparameter that controls the relative importance of L w and L gw in the total loss function.

[0119] The Wasserman contrastive loss L w :

[0120]

[0121] where is the Wasserman distance, which represents the transmission cost between subgraphs and its contrastive subgraph . n is the number of sampled subgraphs. m is the number of negative samples for each subgraph. k is a temperature parameter that controls the distribution smoothness in the contrastive loss.

[0122] The Wasserman distance

[0123]

[0124] where T is the transmission matrix, which represents the transmission plan from the node set of the source subgraph to the node set of the target subgraph. T ij is the transmission quality or probability of the source node e i to the target node , which determines how to map the node features from one distribution to another while maintaining the minimum cost. represents the transmission cost between nodes e i and . Δ(·) is the cosine similarity between node features.

[0125] The Wasserman contrastive loss L w aims to improve the robustness of the model by minimizing the difference in node similarity between positive samples, while maximizing the difference in node similarity between negative samples.

[0126] The Grover-Wasserman contrastive loss L gw :

[0127]

[0128] Gromov-Wasserstein distance The calculation formula is as follows:

[0129]

[0130] Wherein, The calculation is as follows:

[0131]

[0132] Gromov-Wasserstein contrastive loss L gw For capturing the structural similarity of edges in subgraphs, by minimizing the edge similarity difference between positive samples, while maximizing the edge similarity difference between negative samples, thereby improving the model's ability to capture the global structure of the graph.

[0133] As a verification, the method provided in the embodiment performs well in multiple eye disease medical image segmentation tasks, especially when dealing with medical data with complex topological structure, it shows strong generalization ability and stability. The experimental results on the OCT eye disease data set disclosed by Drive and Kermany show that even in the case of less labeled data, the method of the application can still achieve high segmentation accuracy.

[0134] As shown in Tables 1 and 2, by combining the graph convolution network encoder with the global embedding strategy, the local and global topological information in the eye medical image can be effectively captured, thereby significantly improving the segmentation performance. This ability is particularly important in the pixel or region classification task of eye diseases, because the annotation of ophthalmic medical data is usually expensive and limited, and the segmentation accuracy is of great significance for clinical diagnosis.

[0135] Table 1 Comparison of experimental results of DRIVE data set

[0136]

[0137]

[0138] Table 2 Comparison of experimental results of OCT data set

[0139]

[0140] Embodiment two

[0141] A self-supervised medical image segmentation system based on random walk path, comprising:

[0142] The preprocessing module is configured to preprocess the acquired medical image, convert it into a graph data structure, extract initial features of each node in the graph data structure, and perform standardization processing on the initial features.

[0143] The execution process of the module can refer to the process of step S1 of Embodiment 1, which will not be repeated here.

[0144] The node information updating module is configured to update the features of each node based on a message passing mechanism, and aggregate the updated features of all nodes to obtain global features.

[0145] The execution process of the module can refer to the process of step S2 of Embodiment 1, which will not be repeated here.

[0146] The multi-hop path sampling module is configured to sample homogeneity neighborhood distribution and heterogeneity neighborhood distribution of the center node based on a short-range random walk strategy and a long-range random walk strategy, respectively, to form a random walk path.

[0147] The execution process of the module can refer to the process of step S3 of Embodiment 1, which will not be repeated here.

[0148] The path information aggregation module is configured to encode each random walk path, learn the weight of each random walk path through an attention mechanism, splice the path information after the weight aggregation, and generate a global embedding representation.

[0149] The execution process of the module can refer to the process of step S4 of Embodiment 1, which will not be repeated here.

[0150] The subgraph generation module is configured to extract nodes and edges with a topological distance from the center node not exceeding a set threshold to form a subgraph, introduce random disturbance to generate a contrast view, and use the similarity between nodes for embedding calculation to generate a contrast subgraph with node features and edges.

[0151] The execution process of the module can refer to the process of step S5 of Embodiment 1, which will not be repeated here.

[0152] The structured optimization module is configured to optimize the subgraph based on the subgraph and the contrast subgraph using a structured contrast loss.

[0153] The execution process of the module can refer to the process of step S6 of Embodiment 1, which will not be repeated here.

[0154] It can be understood that the above-mentioned units / modules can be combined into one or several other units / modules respectively or entirely, or some of the units can be further split into multiple units with smaller functions to achieve the same operation, without affecting the implementation of the technical effects of the embodiments of the present application.

[0155] The above-mentioned modules of the system are divided based on logical functions. In actual applications, the functions of one module can also be implemented by multiple modules, or the functions of multiple modules can be implemented by one module. For example, the preprocessing module in the present embodiment can include the following modules in other embodiments:

[0156] The graph data structure construction module is configured to define each pixel or voxel in the medical image as a node in the graph, and the edges between the nodes are established according to the spatial proximity of the pixels or voxels, and the weight of the edge reflects the similarity between the nodes.

[0157] The adjacency matrix representation module is configured to represent the topological structure of the graph using an adjacency matrix. For an unweighted graph, the adjacency matrix is a binary matrix representing whether there is an edge connection between the nodes. For a weighted graph, the elements of the adjacency matrix are the weight values of the edges, representing the similarity between the nodes.

[0158] The initial feature processing module is configured to extract the initial features of each node in the graph data structure and perform standardization processing on the initial features.

[0159] Similarly, in other embodiments of the present application, the system can also include other units / modules. In actual applications, these functions can also be assisted by other units, and can be implemented by multiple units in cooperation.

[0160] According to another embodiment of the present application, the system described in the present embodiment can be constructed and the method of embodiment one can be implemented by running a computer program (including program code) capable of performing each step involved in the corresponding method described in embodiment one on a general computing device such as a computer including processing elements and storage elements such as a central processing unit (CPU), a random access memory (RAM), a read-only memory (ROM), etc. The computer program can be recorded on a computer-readable recording medium, for example, and loaded into the above-mentioned computing device through the computer-readable recording medium and run therein.

[0161] Embodiment three

[0162] The present implementation provides an electronic device, such as Figure 2As shown, the electronic device includes a processor 1001, a communication interface 1002, and a computer readable storage medium 1003. Among them, the processor 1001, the communication interface 1002, and the computer readable storage medium 1003 can be connected through a bus or other means.

[0163] Among them, the communication interface 1002 is used to receive and send data, the computer readable storage medium 1003 can be stored in the memory of the electronic device, the computer readable storage medium 1003 is used to store a computer program, the computer program includes program instructions, and the processor 1001 is used to execute the program instructions stored by the computer readable storage medium 1003.

[0164] The processor 1001 (also called CPU (Central Processing Unit, Central Processing Unit)) is the computing core and control core of the electronic device, which is suitable for implementing one or more instructions, and is particularly suitable for loading and executing one or more instructions to realize the corresponding method flow or corresponding function.

[0165] The processor 1001 is configured to perform the following process:

[0166] The acquired medical image is preprocessed and converted into a graph data structure, the initial features of each node in the graph data structure are extracted, and the initial features are standardized;

[0167] The features of each node are updated based on the message passing mechanism, and the updated features of all nodes are aggregated to obtain global features;

[0168] Based on the short-range random walk strategy and the long-range random walk strategy, the homogeneity neighborhood distribution and the heterogeneity neighborhood distribution of the center node are respectively sampled to form a random walk path;

[0169] Each random walk path is encoded, the weights of each random walk path are learned through an attention mechanism, the path information after weight aggregation is spliced, and a global embedding representation is generated;

[0170] The nodes and edges with a topological distance from the center node not exceeding a set threshold are extracted to form a subgraph, a contrast view is generated by introducing random disturbance, and embedding calculation is performed using the similarity between nodes to generate a contrast subgraph with node features and edges;

[0171] Based on the subgraph and the contrast subgraph, the subgraph is optimized using a structured contrast loss.

[0172] Or the process of steps S1-S6 in Embodiment One, which will not be repeated here.

[0173] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0174] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0175] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0176] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0177] 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 by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A random walk path-based self-supervised medical image segmentation method, characterized in that, The method comprises the following steps: The acquired medical image is preprocessed and converted into a graph data structure, the initial features of each node in the graph data structure are extracted, and the initial features are standardized; The features of each node are updated based on a message passing mechanism, and the updated features of all nodes are aggregated to obtain global features; Based on the short-range random walk strategy and the long-range random walk strategy, the homogeneity neighborhood distribution and the heterogeneity neighborhood distribution of the center node are respectively sampled to form a random walk path; Each random walk path is encoded, the weights of the random walk paths are learned through an attention mechanism, the path information after weight aggregation is spliced, and a global embedding representation is generated; A subgraph composed of nodes and edges with a topological distance of the center node not exceeding a set threshold is extracted, a contrast view is generated by introducing random disturbance, and embedding calculation is performed using the similarity between nodes to generate a contrast subgraph with node features and edges; Based on the subgraph and the contrast subgraph, the subgraph is optimized using a structured contrast loss; The path weight calculation formula is: where, is the normalized weight of each path, is the learnable attention parameter, is the LeakyReLU activation function, and the Softmax function is used to normalize the weights, is the encoded feature of each path, which is obtained by using a GRU encoder; The path information after weight aggregation is calculated through multi-head attention, and the results are spliced to generate a unified global embedding representation; The multi-head attention aggregation formula is: wherein, is a learnable multi-head attention weight, is a center node global embedding representation, denotes a feature concatenation operation; The specific sampling process comprises: acquisition map , central node , sampling path length threshold , parameter and parameter ; Initialization settings are made, the sampling path length , sampling subgraph contains a central node ; When len_path < K: If len_path = 1: Randomly select a first-order neighborhood node v j Connected with v i ; The edge Join the subgraph ; The value of len_path is increased by 1; Set as new center node Set as new center node ; Otherwise: performing a random walk according to probability parameters p and q to select a next node ; The edge Join the subgraph ; The value of len_path is increased by 1; Set as new center node Set as new center node ; Return to the sampling subgraph .

2. The random walk path-based self-supervised medical image segmentation method of claim 1, wherein the converting The process of the graph data structure comprises defining each pixel or voxel in the medical image as a node in the graph, establishing edges between nodes according to the spatial proximity of the pixels or voxels, and reflecting the similarity between nodes through the weight of the edges.

3. The random walk path-based self-supervised medical image segmentation method according to claim 1 or 2, characterized in that, It also includes using an adjacency matrix to represent the topological structure of the graph, for an unweighted graph, the adjacency matrix is a binary matrix representing whether there is an edge connection between nodes; for a weighted graph, the adjacency matrix elements are the weight values of the edges, representing the similarity between nodes.

4. The random walk path-based self-supervised medical image segmentation method of claim 1, wherein, The process of aggregating the updated features of all nodes comprises propagating node features through an adjacency matrix to generate new node representations, using an encoder of a graph convolutional network to aggregate features of the graph data, and through the superposition of multiple layers of graph convolutional networks, the final features of the nodes contain information of their direct neighbors and information of nodes at a further distance. In the last layer of the encoder, the features of all nodes are aggregated into global features.

5. The random walk path-based self-supervised medical image segmentation method of claim 1, wherein the step of generating the random walk path comprises: generating a random walk path by using a random walk algorithm based on the medical image data. In the random walk process, the non-normalized transition probability is used to maximize the topological relationship between nodes, and the transition probability is: wherein, denotes the shortest path length between nodes and , when , is 1-hop neighbor of , is the transition parameter of random walk, respectively controlling the preference of near-neighbor and far-neighbor nodes; when and , the random walk strategy collects information of neighboring nodes, extracting homogeneity information; when and , the random walk strategy preferentially accesses more distant nodes, extracting heterogeneity information.

6. The random walk path-based self-supervised medical image segmentation method of claim 1, wherein the introduction of The process of generating a contrast view using random perturbation and using similarity between nodes for embedding computation, generating a contrast view with node features and edges includes: and between two nodes wherein, , is a subgraph, and is a learnable weight matrix, is an activation function; Based on the similarity, the node embedding aggregation comprises: 。 7. The random walk path-based self-supervised medical image segmentation method of claim 1, wherein, In the process of optimizing the subgraph using a structured contrast loss, the Wasserstein contrast loss and the Gromov-Wasserstein contrast loss are used to measure the similarity between nodes and edges respectively, wherein the Wasserstein contrast loss is used to represent the transmission cost between the subgraph and its contrast subgraph, and the optimization is realized by minimizing the difference in node similarity between positive samples and maximizing the difference in node similarity between negative samples. The Gromov-Wasserstein contrast loss is used to capture the structural similarity of edges in the subgraph, and the optimization is realized by minimizing the difference in edge similarity between positive samples and maximizing the difference in edge similarity between negative samples.

8. A random walk path based self-supervised medical image segmentation system, applying the method of claim 1, characterized in that, The method comprises the following steps: The preprocessing module is configured to preprocess the acquired medical image, convert it into a graph data structure, extract initial features of each node in the graph data structure, and perform standardization processing on the initial features; The node information updating module is configured to update the features of each node based on a message passing mechanism, and aggregate the updated features of all nodes to obtain global features; The multi-hop path sampling module is configured to sample homogeneity neighborhood distribution and heterogeneity neighborhood distribution of the center node based on a short-range random walk strategy and a long-range random walk strategy, respectively, to form random walk paths; The inter-path information aggregation module is configured to encode each random walk path, learn the weights of the random walk paths through an attention mechanism, splice the path information after the weight aggregation, and generate global embedding representation; The subgraph generation module is configured to extract nodes and edges with a topological distance from the center node not exceeding a set threshold to form a subgraph, introduce random disturbance to generate a contrast view, and use the similarity between nodes to perform embedding calculation to generate a contrast subgraph with node features and edges; The structured optimization module is configured to optimize the subgraph based on the subgraph and the contrast subgraph using a structured contrast loss.

9. An electronic device, characterized by comprising: The computer program product comprises a memory and a processor, and computer instructions stored on the memory and running on the processor, and when the computer instructions are run by the processor, the steps in the method of any one of claims 1-7 are completed.

Images