A Deep Learning-Based Dynamic Segmentation Method for 3D Brain Networks

Abstract

This invention discloses a dynamic segmentation method for three-dimensional brain networks based on deep learning. The method comprises the following steps: S1. Constructing temporally and spatially consistent fused image volume data; S2. Generating a multi-scale sparse Transformer encoded feature pyramid; S3. Obtaining the same-scale brain region map structure; S4. Using the cross-scale node-aligned fused map structure as the initial output of the cross-scale node-aligned fusion mechanism; S5. Obtaining the node embedding features of the first round of fused brain region map; S6. Obtaining the updated multi-scale sparse Transformer encoded feature pyramid, and repeating steps S3 to S5 until the multi-scale sparse Transformer-GNN interactive update is completed; S7. Obtaining the dynamic segmentation result of the three-dimensional brain region. This invention achieves the collaborative extraction of local fine-grained and global coarse-grained features, which not only effectively suppresses redundant information but also enhances the modeling ability of spatial structure and functional connectivity features in cross-modal fused images.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, and in particular to a dynamic segmentation method for three-dimensional brain networks based on deep learning. Background Technology

[0002] With the rapid development of neuroimaging technology, multimodal brain imaging techniques such as functional magnetic resonance imaging, structural magnetic resonance imaging, and diffusion tensor imaging have been widely used in the study of human brain cognitive mechanisms and the early diagnosis of brain diseases. In recent years, researchers have paid more and more attention to the dynamic evolution relationship between brain functional connectivity and brain region maps. Three-dimensional brain network segmentation, as a basic link in brain imaging analysis, has gradually evolved from the traditional static brain region division to a brain network segmentation task with time dynamic characteristics.

[0003] Currently, most mainstream 3D brain network segmentation methods are based on convolutional neural networks to perform voxel-level semantic segmentation of single-modality image data. Although some progress has been made, many key problems still exist. On the one hand, traditional convolutional neural network models are difficult to fully capture the spatial relationships with strong non-Euclidean characteristics in brain image data, especially lacking global modeling capabilities when dealing with cross-scale, unstructured brain region connectivity relationships. On the other hand, there are significant differences between multimodal data in terms of spatial resolution, image registration, and signal-to-noise ratio. Traditional fusion strategies often simply use cascading or averaging operations, which cannot effectively extract deep collaborative features between modalities, thus limiting further improvement in fusion segmentation accuracy.

[0004] Furthermore, while some existing brain region segmentation methods introduce graph neural networks to model the topological structure between brain regions, they typically construct the graph structure statically, making it impossible to dynamically adjust it for different scales or time periods. This makes it difficult to support the cross-scale evolution and information feedback of brain region connectivity maps during real neural activities. In particular, when dealing with complex brain networks, the lack of an effective coupling mechanism between graph modeling and segmentation models results in poor performance in terms of boundary coherence, local details, and dynamic consistency.

[0005] Furthermore, existing brain segmentation methods generally neglect the coherence modeling across time frames, failing to maintain consistency and stability in brain region identification at consecutive time points. This is particularly crucial for capturing the evolution of brain region functional states. Existing segmentation methods typically process single-frame data during the training and inference phases, lacking the ability to model in the time dimension. This results in delays and ambiguities in the detection of dynamic brain region activities, making it difficult to meet the actual needs of neuroscience and clinical diagnosis for high spatiotemporal resolution segmentation technology.

[0006] In summary, existing three-dimensional brain network segmentation methods still have significant technical shortcomings in terms of multimodal data fusion accuracy, spatial structure representation ability, cross-scale structural consistency modeling, and dynamic segmentation continuity. There is an urgent need for a novel segmentation strategy that can fully integrate multimodal features and combine spatial topology and temporal dynamic information to achieve more accurate, stable, and time-evolutionary dynamic segmentation of three-dimensional brain regions. Summary of the Invention

[0007] One objective of this invention is to propose a dynamic segmentation method for three-dimensional brain networks based on deep learning. This invention achieves the collaborative extraction of local fine-grained and global coarse-grained features, which can not only effectively suppress redundant information, but also enhance the modeling ability of spatial structure and functional connectivity features in cross-modal fusion images.

[0008] A method for dynamic segmentation of a three-dimensional brain network based on deep learning according to an embodiment of the present invention includes the following steps:

[0009] S1. Construct temporally and spatially consistent fused image volume data;

[0010] S2. Divide the fused image volume data into cubes of different scales according to the multi-level cube partitioning strategy, perform position encoding on each scale cube and input it into the multi-scale sparse Transformer encoder to generate a multi-scale sparse Transformer encoded feature pyramid.

[0011] S3. Based on the predefined medical partition template and diffusion tensor imaging connectivity information, an initial set of brain region map nodes is established for the multi-scale sparse Transformer encoded feature pyramid at each scale to obtain the brain region map structure at the same scale.

[0012] S4. Perform cross-scale node alignment on brain region map nodes of the same name at different scales, and use the cross-scale node alignment fusion map structure as the initial output of the cross-scale node alignment fusion mechanism;

[0013] S5. Node-to-node message passing is performed within the cross-scale node-aligned fusion graph structure by graph convolution and multi-head graph attention to obtain the first round of fused brain region map node embedding features;

[0014] S6. Use the first round of fused brain region map node embedding features as residual information to feed back to the next layer input of the multi-scale sparse Transformer encoder to obtain the updated multi-scale sparse Transformer encoded feature pyramid, and repeat steps S3 to S5 until the multi-scale sparse Transformer-GNN interactive update is completed.

[0015] S7. Perform skip connection fusion and hierarchical upsampling reconstruction on the final multi-scale sparse Transformer encoded feature pyramid to generate voxel-level segmentation probability volume data;

[0016] S8. Post-process the voxel-level segmentation probability data of the training and inference phases respectively to obtain the dynamic segmentation results of the three-dimensional brain regions.

[0017] Optionally, S1 includes the following steps:

[0018] S11. Acquire structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data.

[0019] S12. Perform initial resampling processing on the structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data, respectively;

[0020] S13. Perform voxel-level rigid registration of functional magnetic resonance imaging data and diffusion tensor imaging data relative to structural magnetic resonance imaging data;

[0021] S14. Perform non-rigid registration processing on the functional magnetic resonance imaging data and diffusion tensor imaging data to establish a deformation field. The deformation field describes the small displacement of each voxel position (x,y,z) in three spatial dimensions.

[0022] S15. Perform intensity normalization processing on the registered structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data respectively to obtain normalized structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data.

[0023] S16. Perform artifact removal processing on the normalized structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data;

[0024] S17. The structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data, after registration, normalization, and artifact removal, are stitched and fused along the voxel dimension at the channel level to form the final fused image volume data D. fusion .

[0025] Optionally, S2 includes the following steps:

[0026] S21. Merge image volume data D fusion The system is segmented according to a multi-level cubic partitioning strategy, using different scale partitioning factors s. l The original 3D spatial voxel mesh is divided into a set of non-overlapping cubes corresponding to each scale level, with each scale level corresponding to a set of cubes at a specific resolution. Where N lThis represents the total number of cubes in the l-th layer, where l represents the scale level. Indicates the Nth level under scale level l l Each cube is a small cubic region that has been divided into three-dimensional spaces;

[0027] S22. For each cube at each scale level Perform position encoding to obtain the position encoded vector.

[0028] S23. Encode the position vector at each scale level. fused feature vector with the corresponding cube The data are concatenated to form a set of scale feature representations F. (l) ;

[0029] S24. For shallow-scale hierarchies l high , representing the scale feature set Input a local window self-attention module, and set the window size to [value]. Perform a dense self-attention mechanism within each window to output a shallow, local, fine-grained representation.

[0030] S25. For deep-scale hierarchies l low , representing the scale feature set Input a cross-window sparse self-attention module, establish connections only between a preset subset of queries and their associated keys, and output a deep global coarse-grained representation.

[0031] S26. For each shallow-scale level, the shallow local fine-grained representation is fused with the matching deep sparse global representation element-wise to form the fused encoding representation Z. (l) ;

[0032] S27. Fusion coding representation for all scale levels We will organize them in a unified manner and construct a multi-scale sparse Transformer encoded feature pyramid with cross-level structure perception capabilities.

[0033] Optionally, S3 includes the following steps:

[0034] S31. Construct a predefined medical region template to form a brain region partitioning set R. The brain region partitioning set consists of multiple brain region template regions. In each scale level l, the fused encoding representation set Z is then used. (l) The spatial position of each cube is determined to coincide with the brain region template region. The cube is mapped to the graph node in the brain region map structure through a mapping function, forming the initial node mapping relationship.

[0035] S32. At scale level l, generate an initial brain region map node set V based on node mapping relationships. (l) Each graph node in the initial set of brain region graph nodes is used to represent a set of cubes of a brain region template at the current scale. For each graph node, the average value of the encoding representation of all cubes contained in the current node is calculated to obtain the initial feature vector of the current graph node.

[0036] S33. Using the initial feature vector set of all graph nodes under scale level l, calculate the anatomical similarity between each pair of graph nodes and form an anatomical similarity matrix;

[0037] S34. Based on the normalized diffusion tensor imaging data, extract the number of white matter fiber tracts tracked between each pair of brain region templates, and calculate the normalized connectivity based on the fiber distribution density inside and outside the region. Organize the functional connectivity results between all regions into a functional similarity matrix.

[0038] S35. By linearly weighting and fusing the anatomical similarity matrix and the functional similarity matrix, a graph adjacency matrix at scale level l is constructed;

[0039] S36. Combine the set of brain region map nodes at scale level l with the graph adjacency matrix to form the same-scale brain region map structure G at the current scale level. (l) .

[0040] Optionally, S3 includes the following steps:

[0041] S41. Set a cross-scale alignment threshold, select all scale levels with resolutions lower than the cross-scale alignment threshold to form a low-resolution set, and select all scale levels with resolutions higher than or equal to the cross-scale alignment threshold to form a high-resolution set.

[0042] S42. For each brain region map structure in a low-resolution scale level, extract the spatial center coordinates of each graph node, and at the same time extract the initial feature vector of the graph node;

[0043] S43. At the high-resolution scale level, for all candidate graph nodes that have the same medical template name as the low-resolution graph nodes, calculate the spatial Euclidean distance between each candidate graph node and the low-resolution graph node, and select the candidate node with the smallest Euclidean distance as the high-resolution alignment target graph node.

[0044] S44. Form a matching pair between each pair of low-resolution graph nodes and the nearest neighbor high-resolution aligned target graph node, and construct a cross-scale aligned mapping set;

[0045] S45. For each pair of matching graph nodes in the cross-scale alignment mapping set, calculate the feature similarity using the initial feature vectors of the two graph nodes;

[0046] S46. Combine all cross-scale matching graph node pairs and their corresponding alignment weights to form a cross-scale edge set;

[0047] S47. Merge the low-resolution scale hierarchical brain region map structure with the high-resolution scale hierarchical brain region map structure and the cross-scale edge set to obtain the final cross-scale node-aligned fused brain region map structure.

[0048] Optionally, S5 includes the following steps:

[0049] S51. Based on the mapping relationship between graph nodes and cubes in the cross-scale node alignment fusion brain region map structure, the cube block encoding features at each scale level in the multi-scale sparse Transformer encoding feature pyramid are mapped to the initial attribute features of the corresponding graph nodes.

[0050] S52. The initial attribute features of all nodes in the graph at all scale levels are concatenated and integrated according to the node number to form the initial feature matrix of the cross-scale node set;

[0051] S53. Input the initial feature matrix into the graph neural network message passing module, combine it with the adjacency matrix of the brain region map structure fused by cross-scale node alignment to perform the first round of graph convolution operation, and obtain the first round of node embedding representation H. (1) ;

[0052] S54. Input the first round of graph convolutional representation into the multi-head graph attention mechanism module. The outputs of each graph node from all attention heads are integrated through a concatenation operation to generate the first round of fused brain region graph node embedding features.

[0053] Optionally, S6 includes the following steps:

[0054] S61. Based on the mapping relationship between graph nodes and original cubes, embed the first round of fused brain region graph nodes into features. Project and broadcast back to the original cubes at each scale level in the multi-scale sparse Transformer encoded feature pyramid to obtain the graph feedback residual information corresponding to each cube;

[0055] S62. For each scale level l, the graph feedback residual information R is used. (1) The corresponding residual vector The cube features of the previous Transformer encoding output Perform a joint update to form a new cube input representation.

[0056] S63. Input the new cube representation The input is fed into the next layer of the multi-scale sparse Transformer encoder, where local window dense self-attention and cross-window sparse self-attention mechanisms are executed. Multi-scale feature extraction is performed on the new input, and the updated multi-scale sparse Transformer encoded feature pyramid is output.

[0057]

[0058] S64. After each round of updates, repeat the same-scale brain region map structure construction process in S3 to update and obtain a new brain region map structure;

[0059] S65. Based on the new brain region map structure, repeat the cross-scale node alignment process in S4 to generate a new cross-scale node alignment fusion brain region map structure;

[0060] S66. Apply graph neural network message passing and multi-head graph attention mechanisms to the new cross-scale node-aligned fused brain region map structure to generate second-round fused brain region map node embedding features. And then feed it back to the encoder for residual update;

[0061] S67. Repeat steps S61 to S66 to perform multiple rounds of Transformer-GNN dual-path interactive updates until the set number of iterations or residual convergence threshold is reached, and finally a stable multi-scale sparse Transformer encoded feature pyramid and fused brain region map node embedding representation are obtained.

[0062] Optionally, S7 includes the following steps:

[0063] S71. Based on the stable multi-scale sparse Transformer coding feature pyramid Encoded features at all scale levels A skip connection fusion strategy is adopted to integrate information, and channel concatenation and convolution fusion are performed with the current level encoded features to generate a fused cross-scale voxel feature representation.

[0064] S72. Based on the mapping relationship between graph nodes and cubes, the embedded representation of the fused brain region graph nodes is projected back to its corresponding voxel block region through the nearest neighbor, and a graph feature field is established in the voxel space. The cross-scale voxel feature representation and the graph feature field are spliced ​​together in the channel dimension to form the final fused voxel-level segmentation feature.

[0065] S73. The final fused voxel-level segmentation features are input into the segmentation head network composed of convolutional layers. The segmentation head network performs a set of 1×1×1 convolution operations and outputs voxel-level multi-class probability predictions through the softmax function to generate voxel-level segmentation probability volume data.

[0066] Optionally, S8 specifically includes introducing neighboring time frame probability consistency constraints and deformation field smoothing regularization to the voxel-level segmentation probability data of continuous time frames during the training phase; applying recursive Bayesian filtering and B-spline-based deformation compensation to the voxel-level segmentation probability data during the inference phase; and performing conditional random field refinement, time-of-flight mapping-based hole filling, and confidence assessment on the voxel-level segmentation probability data that has undergone smooth evolution in the time domain to obtain dynamic segmentation results of three-dimensional brain regions.

[0067] The beneficial effects of this invention are:

[0068] (1) This invention constructs a multi-level cube partitioning strategy to divide the fused sMRI, fMRI and DTI image volume into cubes of different scales. It introduces a local window dense self-attention mechanism in the shallow scale and a sparse cross-window self-attention mechanism in the deep scale, thereby realizing the collaborative extraction of local fine-grained and global coarse-grained features. This not only effectively suppresses redundant information, but also enhances the ability to model the spatial structure and functional connectivity features in cross-modal fused images.

[0069] (2) Based on a predefined medical brain region template, this invention constructs initial brain region map nodes at different scales and introduces a cross-scale node alignment mechanism. Through the principle of spatial nearest neighbor and feature similarity weighting strategy, a one-to-one mapping between low-resolution and high-resolution map nodes is established and fused into a cross-scale map structure. This effectively bridges the problem of inconsistent map semantics between different resolutions, maintains spatial stability in overlapping brain regions, and maintains semantic consistency of nodes in regions with blurred anatomical structures, significantly improving the coherence and medical interpretability of the segmentation map.

[0070] (3) This invention constructs a dual-path nested interaction mechanism of Transformer and GNN. By back-projecting the node embedding feature vectors obtained by the graph neural network to the input channel of the original multi-scale Transformer, the residual update feedback of graph semantic information is realized. It also allows the graph structure and encoding features to be reconstructed in each loop, thereby achieving the co-evolution of cross-scale graph structure and encoding pyramid until the residual converges. When dealing with dynamic changes in brain regions, it can more accurately capture boundary evolution and region expansion, and improve the model's ability to perceive dynamic changes in small regions. Attached Figure Description

[0071] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0072] Figure 1 This is a flowchart of a three-dimensional brain network dynamic segmentation method based on deep learning proposed in this invention. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0074] refer to Figure 1 A dynamic segmentation method for three-dimensional brain networks based on deep learning includes the following steps:

[0075] S1. Construct temporally and spatially consistent fused image volume data;

[0076] S2. Divide the fused image volume data into cubes of different scales according to the multi-level cube partitioning strategy. Perform position encoding on cubes of each scale and input them into a multi-scale sparse Transformer encoder. Perform local window dense self-attention on cubes of the same scale in the shallow layer and perform cross-window sparse self-attention on cubes of different scales in the deep layer to generate a multi-scale sparse Transformer encoded feature pyramid.

[0077] S3. Based on predefined medical partition templates and diffusion tensor imaging connectivity information, an initial set of brain region map nodes is established for the multi-scale sparse Transformer encoded feature pyramid at each scale, and the anatomical similarity and functional similarity between nodes within the scale are calculated to obtain the brain region map structure at the same scale.

[0078] S4. Perform cross-scale node alignment on brain region map nodes with the same name at different scales. Establish alignment mapping between brain region map nodes with resolution below the threshold and brain region map nodes with resolution above the threshold through nearest neighbor matching. Under the guidance of the alignment mapping, construct a cross-scale node alignment fusion graph structure and use the cross-scale node alignment fusion graph structure as the initial output of the cross-scale node alignment fusion mechanism.

[0079] S5. Map the multi-scale sparse Transformer encoded feature pyramid to the node attributes of the cross-scale node-aligned fusion graph structure. Perform inter-node message passing within the cross-scale node-aligned fusion graph structure through graph convolution and multi-head graph attention to obtain the first round of fused brain region map node embedding features.

[0080] S6. The first round of fused brain region map node embedding features are fed back as residual information to the next layer input of the multi-scale sparse Transformer encoder. In the next layer of the multi-scale sparse Transformer encoder, they are updated together with the features of the previous layer to obtain the updated multi-scale sparse Transformer encoded feature pyramid. Steps S3 to S5 are repeated until the multi-scale sparse Transformer-GNN interactive update is completed.

[0081] S7. Perform skip connection fusion and hierarchical upsampling reconstruction on the final multi-scale sparse Transformer encoded feature pyramid, embed the final node of the cross-scale node alignment fusion graph structure into the feature projection back into the voxel space, and generate voxel-level segmentation probability volume data.

[0082] S8. Post-process the voxel-level segmentation probability data of the training and inference phases respectively to obtain the dynamic segmentation results of the three-dimensional brain regions.

[0083] In this embodiment, S1 includes the following steps:

[0084] S11. Acquire structural magnetic resonance imaging (SMRI), functional magnetic resonance imaging (fMRI), and diffusion tensor imaging (DTI) data. SMRI data is used to provide a three-dimensional anatomical reference, fMRI data characterizes fluctuations in blood oxygen levels caused by neural activity, and DTI data is used to describe the orientation of white matter fiber bundles and the connectivity information between brain regions.

[0085] S12. Perform initial resampling processing on the structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data, respectively, and standardize the spacing of the three-dimensional voxels in the x, y, and z spatial dimensions to v. x v y v z The standardized spacing is uniformly represented as the voxel space spacing Δv;

[0086] S13. Perform voxel-level rigid registration of functional magnetic resonance imaging data and diffusion tensor imaging data relative to structural magnetic resonance imaging data;

[0087] S14. After performing rigid registration, non-rigid registration processing is further performed on the functional magnetic resonance imaging data and diffusion tensor imaging data to establish a deformation field. The deformation field describes the small displacement of each voxel position (x,y,z) in three spatial dimensions, which is used to make local fine adjustments to the registration results and eliminate subtle misalignments caused by differences in brain tissue morphology or individual differences.

[0088] S15. Perform intensity normalization processing on the registered structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data respectively. Use the intramodal intensity normalization method to normalize the voxel intensity of each image using the mean and standard deviation respectively, to obtain the normalized structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data.

[0089] S16. Artifact removal processing is performed on the normalized structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data. This is done through spatial high-pass filtering and boundary enhancement operations based on morphological reconstruction to eliminate imaging artifacts, noise interference, or false structural responses caused by magnetic susceptibility effects, and to preserve anatomical boundaries and real tissue information to the greatest extent possible.

[0090] S17. The structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data, after registration, normalization, and artifact removal, are stitched and fused along the voxel dimension at the channel level to form the final fused image volume data.

[0091] In this embodiment, S2 includes the following steps:

[0092] S21. Merge image volume data D fusion The system is segmented according to a multi-level cubic partitioning strategy, using different scale partitioning factors s. l Where l represents the scale level, the original 3D spatial voxel mesh is divided into a set of non-overlapping cubes corresponding to each scale level, and each scale level corresponds to a set of cubes at a specific resolution. Where N l This represents the total number of cubes in the l-th layer. Indicates the Nth level under scale level l l Each cube is a small cubic region that has been divided into three-dimensional spaces;

[0093] The spatial structure of the fused image volume data is itself a regular three-dimensional voxel grid structure, which is regarded as the original three-dimensional spatial voxel grid of the input image.

[0094] S22. For each cube at each scale level Perform positional encoding, where i is greater than or equal to 1 and less than or equal to N. l Define the position encoding vector Where (x) i ,y i ,z i ) represents the three-dimensional spatial coordinates of the cube center in the fused image volume data, and PE(·) is a learnable position encoding function;

[0095] S23. Encode the position vector at each scale level. fused feature vector with the corresponding cube The data are concatenated to form a set of scale feature representations. in Indicates a splicing operation;

[0096] S24. For shallow-scale hierarchies l high , representing the scale feature set Input a local window self-attention module, and set the window size to [value]. Perform a dense self-attention mechanism within each window to output a shallow, local, fine-grained representation.

[0097] S25. For deep-scale hierarchies l low , representing the scale feature set Input a cross-window sparse self-attention module, establish connections only between a preset subset of queries and their associated keys, and output a deep global coarse-grained representation.

[0098] S26. For each shallow-scale level, the shallow local fine-grained representation is fused with the matching deep sparse global representation element-wise to form a fused encoded representation.

[0099] S27. Fusion coding representation for all scale levels We will organize them in a unified manner and construct a multi-scale sparse Transformer encoded feature pyramid with cross-level structure perception capabilities.

[0100] In this embodiment, S3 includes the following steps:

[0101] S31. Construct a predefined medical region template to form a brain region partitioning set R. The brain region partitioning set consists of multiple brain region template regions. In each scale level l, the fused encoding representation set Z is then used. (l) The spatial position of each cube is determined to coincide with the brain region template region. The cube is mapped to the graph node in the brain region map structure through a mapping function, forming the initial node mapping relationship.

[0102] S32. At scale level l, generate an initial brain region map node set V based on node mapping relationships. (l) Each graph node in the initial set of brain region graph nodes is used to represent the set of cubes of a brain region template at the current scale. For each graph node, the average value of the encoding representation of all cubes contained in the current node is calculated to obtain the initial feature vector of the current graph node. The initial feature vector of the graph node is used to represent the representation information of the current brain region graph node at the current scale.

[0103] S33. Using the initial feature vector set of all graph nodes under scale level l, calculate the anatomical similarity between each pair of graph nodes. The current anatomical similarity is obtained by calculating the cosine similarity between the feature vectors of the corresponding two graph nodes. The anatomical similarity is used to measure the local structural homogeneity of two brain region graph nodes in the encoding feature space, forming an anatomical similarity matrix.

[0104] S34. Based on the normalized diffusion tensor imaging data, extract the number of white matter fiber tracts tracked between each pair of brain region templates, and calculate the normalized connectivity based on the fiber distribution density inside and outside the region to represent the functional connectivity strength between the two regions. Organize the functional connectivity results between all regions into a functional similarity matrix. The functional similarity matrix is ​​used to measure the coupling relationship of neural signal pathways between brain regions.

[0105] S35. By linearly weighting and fusing the anatomical similarity matrix and the functional similarity matrix, a graph adjacency matrix at scale level l is constructed. The fusion process uses a fusion weight coefficient, which is used to control the proportion of contribution of anatomical information and functional information to the graph structure.

[0106] S36. Combine the set of brain region map nodes at scale level l with the graph adjacency matrix to form the same-scale brain region map structure G at the current scale level. (l) .

[0107] In this embodiment, S3 includes the following steps:

[0108] S41. Set a cross-scale alignment threshold to determine the resolution relationship between different scale levels. Select all scale levels with resolutions lower than the cross-scale alignment threshold to form a low-resolution set, and select all scale levels with resolutions higher than or equal to the cross-scale alignment threshold to form a high-resolution set.

[0109] S42. For each brain region map structure in a low-resolution scale level, extract the spatial center coordinates of each graph node, and at the same time extract the initial feature vector of the graph node;

[0110] S43. At the high-resolution scale level, for all candidate graph nodes that have the same medical template name as the low-resolution graph nodes, calculate the spatial Euclidean distance between each candidate graph node and the low-resolution graph node, and select the candidate node with the smallest Euclidean distance as the high-resolution alignment target graph node.

[0111] S44. Each pair of low-resolution graph nodes is matched with the nearest high-resolution aligned target graph node to form a matching pair and a cross-scale alignment mapping set. Each mapping in the cross-scale alignment mapping set is used to describe the mapping path from low-resolution scale-level graph nodes to high-resolution scale-level graph nodes.

[0112] S45. For each pair of matching graph nodes in the cross-scale alignment mapping set, calculate the feature similarity using the initial feature vectors of the two graph nodes;

[0113] S46. All cross-scale matching graph node pairs and their corresponding alignment weights are combined to form a cross-scale edge set. The cross-scale edge set is used to record the connection relationship and weight strength between low-resolution graph nodes and high-resolution graph nodes.

[0114] S47. The low-resolution scale-level brain region map structure is merged with the high-resolution scale-level brain region map structure and the cross-scale edge set to obtain the final cross-scale node-aligned fused brain region map structure. The cross-scale node-aligned fused brain region map structure is composed of the graph node sets of the two scale levels, their respective graph adjacency matrices and the cross-scale edge set, and serves as the initial brain region map structure output of the cross-scale node-aligned fusion mechanism.

[0115] In this embodiment, S5 includes the following steps:

[0116] S51. Based on the mapping relationship between graph nodes and cubes in the cross-scale node alignment and fusion brain region map structure, the cube block encoding features at each scale level in the multi-scale sparse Transformer encoding feature pyramid are mapped to the initial attribute features of the corresponding graph node. The initial attribute features of the graph node are obtained by averaging all the cube block encoding features mapped to the graph node.

[0117] S52. The initial attribute features of the graph nodes in all scale levels are spliced ​​and integrated according to the node number to form an initial feature matrix of the cross-scale node set. Each row of the initial feature matrix corresponds to a graph node, and each column corresponds to a feature dimension.

[0118] S53. Input the initial feature matrix into the graph neural network message passing module, and perform the first round of graph convolution operation by combining the adjacency matrix of the brain region map structure with cross-scale node alignment. The first round of graph convolution operation uses spectral domain graph convolution to add self-loops to the adjacency matrix and then normalizes the degree matrix. The features are linearly transformed using the convolution weight matrix, and the activation function is applied to obtain the first round of node embedding representation H. (1) ;

[0119] S54. Input the first-round graph convolution representation into the multi-head graph attention mechanism module, setting the number of attention heads to K. For each graph node, at each attention head, the features of neighboring nodes are weighted and summed according to the attention weights of neighboring nodes. The attention weights are determined by the correlation between node features and the feature projection is provided by the learnable linear transformation matrix of each attention head. The outputs of each graph node obtained from all attention heads are integrated through a concatenation operation to generate the first-round fused brain region graph node embedding features.

[0120] In this embodiment, S6 includes the following steps:

[0121] S61. Based on the mapping relationship between graph nodes and original cubes, embed the first round of fused brain region graph nodes into features. Project and broadcast back to the original cubes at each scale level in the multi-scale sparse Transformer encoded feature pyramid to obtain the graph feedback residual information corresponding to each cube;

[0122] S62. For each scale level l, the graph feedback residual information R is used. (1) The corresponding residual vector The cube features of the previous Transformer encoding output Perform a joint update to form a new cube input representation.

[0123]

[0124] Where γ is the graph residual adjustment coefficient. The input represents the cube feature after fusing the residual information of the graph nodes;

[0125] S63. Input the new cube representation The input is fed into the next layer of the multi-scale sparse Transformer encoder, where local window dense self-attention and cross-window sparse self-attention mechanisms are executed. Multi-scale feature extraction is performed on the new input, and the updated multi-scale sparse Transformer encoded feature pyramid is output.

[0126] S64. After each round of updates, repeat the same-scale brain region map structure construction process in S3 to update and obtain a new brain region map structure;

[0127] S65. Based on the new brain region map structure, repeat the cross-scale node alignment process in S4 to generate a new cross-scale node alignment fusion brain region map structure;

[0128] S66. Apply graph neural network message passing and multi-head graph attention mechanisms to the new cross-scale node-aligned fused brain region map structure to generate second-round fused brain region map node embedding features. And then feed it back to the encoder for residual update;

[0129] S67. Repeat steps S61 to S66 to perform multiple rounds of Transformer-GNN dual-path interactive updates until the set number of iterations or residual convergence threshold is reached, and finally a stable multi-scale sparse Transformer encoded feature pyramid and fused brain region map node embedding representation are obtained.

[0130] In this embodiment, S7 includes the following steps:

[0131] S71. Based on the stable multi-scale sparse Transformer coding feature pyramid Encoded features at all scale levels A skip connection fusion strategy is adopted for information integration. The skip connection fusion strategy is defined as upsampling from the highest scale level l=L to the lowest scale level l=1, and then performing channel concatenation and convolution fusion with the current level encoded features to generate a fused cross-scale voxel feature representation.

[0132] S72. Based on the mapping relationship between graph nodes and cubes, the embedded representation of the fused brain region graph nodes is projected back to its corresponding voxel block region through the nearest neighbor. A graph feature field is established in the voxel space. The cross-scale voxel feature representation and the graph feature field are spliced ​​together in the channel dimension to form the final fused voxel-level segmentation feature, which represents the comprehensive expression of each voxel position in the local space and cross-node semantic dual pathways.

[0133] S73. The final fused voxel-level segmentation features are input into a segmentation head network composed of convolutional layers. The segmentation head network performs a set of 1×1×1 convolution operations and outputs voxel-level multi-class probability predictions through the softmax function, generating voxel-level segmentation probability volume data P. voxel .

[0134] In this embodiment, S8 specifically includes introducing neighboring time frame probability consistency constraints and deformation field smoothing regularization into the voxel-level segmentation probability data of continuous time frames during the training phase; applying recursive Bayesian filtering and B-spline-based deformation compensation to the voxel-level segmentation probability data during the inference phase to achieve segmentation boundaries that evolve smoothly in the time domain; and performing conditional random field refinement, hole filling based on time-of-flight mapping, and confidence assessment on the voxel-level segmentation probability data that evolves smoothly in the time domain to obtain spatially refined and temporally smooth three-dimensional brain region dynamic segmentation results.

[0135] Example 1: The neurosurgery department of Hospital A admitted a patient suspected of having temporal lobe epilepsy. The surgeons needed high-resolution, dynamic, continuous, and spatially accurate brain region segmentation results preoperatively to support the development of a precise surgical plan and intraoperative navigation. The patient was a 26-year-old right-handed male with no serious comorbidities. The hospital's imaging center acquired multimodal images of his entire brain, including T1-weighted structural MRI images (spatial resolution 1 mm). 3 (Dimensions: 256×256×176) Dynamic fMRI images (TR = 0.8s, temporal length 240 frames, spatial resolution 2.5mm) 3The imaging techniques employed included 96×96×60 pixels and diffusion tensor imaging (64-axis, high resolution, spatial dimensions same as T1), along with simultaneous acquisition of medical brain region prior annotations based on the standard AAL template and Desikan-Killiany partitions. To simulate the temporal micro-deformation of brain tissue caused by intraoperative cistern infusion and traction, real-time dynamic image sequences of the brain surface and deep sulci during surgery were also acquired to verify the stability of dynamic scene segmentation.

[0136] Using the preprocessing platform of the hospital's imaging center, T1-weighted structural MRI images, fMRI, and DTI data were subjected to head motion artifact removal, intensity normalization, and voxel space resampling (unified to 1.0 mm). 3 Subsequently, based on the mutual information optimization registration algorithm, all modalities were aligned to the T1 structural space, and the cross-modal registration was further refined through non-rigid B-spline deformation compensation technology. Finally, the time-consistent and spatially consistent multi-channel image volume data were generated, with each frame having a spatial size of 256×256×176, and maintaining consistency with the original fMRI on the time axis.

[0137] In the multi-scale encoding stage, the fused image volume data is divided into multi-level patches (8×8×8, 16×16×16, 32×32×32), forming a set of high, medium, and low-level cubes. The system performs learnable positional encoding on each level of cube and inputs it into a self-developed multi-scale sparse Transformer encoder. In the shallow layer, dense self-attention captures microstructures within a local window (16×16×16), while in the deep layer, sparse global attention is established only for important key points across the window. This effectively reduces the memory and computational pressure of the original Transformer under large-scale voxels, and a single GPU card can stably support 512 voxels. 3 Real-time inference of volume data.

[0138] Using AAL template partitioning and patient-specific DTI fiber tracking data, the system automatically establishes initial brain region map structures under multi-scale partitioning, and generates node adjacency relationships based on structural feature cosine similarity and white matter fiber connectivity. For graph nodes of the same brain region under different scale partitions, cross-scale alignment mapping is automatically established using spatial nearest neighbor and feature similarity, and a cross-scale fused heterogeneous graph structure containing all scales and all brain regions is generated.

[0139] After passing the multi-scale sparse Transformer encoded features to the node attributes of the fused graph structure through the graph node mapping function, multi-layer spectral domain graph convolution and multi-head attention mechanism are used to perform message passing between nodes. The first round of node embedding features is then fed back to the next layer of Transformer, realizing the sequential global-topological dual-path cyclic interaction of segmentation features. The entire interaction process is set to three rounds of iteration, with a residual convergence threshold of 0.005, and finally stable segmentation candidate features are obtained.

[0140] In the segmentation output stage, a U-shaped skip connection is used to fuse features from each layer, and hierarchical upsampling is used to reconstruct low-resolution features to the original voxel resolution. Combined with the latest round of graph node embedding features, the voxel space is supplemented by interpolation projection, and the resulting splices serve as the input to the final segmentation head. Finally, a voxel-level segmentation probability map is output for each frame. KL consistency constraints and Kalman filter-based smoothing compensation are introduced for temporally adjacent frames, significantly improving the smoothness of segmentation over time and the realism of physiological evolution. The segmentation results are output in DICOM and NIfTI formats, which can be directly accessed by doctors through preoperative planning and intraoperative real-time navigation systems.

[0141] The research team used patient data and retrieved 25 cases of epilepsy / tumor / hemorrhagic diseases (25 subjects aged 14-62 years) from the hospital data platform between 2019 and 2024 that underwent similar segmentation requirements. They compared the method of this invention with three mainstream segmentation methods on a comparable GPU server (NVIDIA RTX4090, 128GB RAM): traditional 3D U-Net, multi-stage convolutional segmentation based on SegResNet, and a segmentation algorithm using only 2.5DSwin-Transformer. The specific comparison results are as follows:

[0142] Table 1 Comparison results of the method of the present invention with three mainstream segmentation methods

[0143]

[0144] In the test data, the number of segmentation candidate training samples was 18,000 frames, and the number of test samples was 4,100 frames. Training was performed using the AdamW optimizer with an initial learning rate of 0.0006, a batch size of 2, and 150 training epochs. All methods were performed under the same artifact, deformation, and signal drift conditions. In terms of quantitative metrics, the method of this invention outperforms traditional convolutional and pure Transformer methods in core metrics such as whole-brain segmentation Dice score (0.913), boundary continuity (Hausdorff distance 1.82mm), dynamic frame segmentation flicker rate (0.9%), and deep groove / vascular region breakage rate (2.2%). In dynamic segmentation scenarios, the fused graph neural structure can quickly adapt to changes in functional connectivity, effectively suppressing the drift and breakage problems in temporal segmentation. In terms of inference efficiency, this invention can achieve real-time inference of 1.48 seconds / frame in single-frame segmentation, which is significantly faster than most similar Transformer models, and the memory consumption is lower than the standard 2.5D Swin-Transformer scheme.

[0145] This invention is based on predefined medical brain region templates, constructs initial brain region map nodes at different scales, and introduces a cross-scale node alignment mechanism. Through the principle of spatial nearest neighbor and feature similarity weighting strategy, a one-to-one mapping between low-resolution and high-resolution map nodes is established and fused into a cross-scale map structure. This effectively bridges the problem of inconsistent map semantics between different resolutions, maintains spatial stability in overlapping brain regions, and maintains semantic consistency of nodes in regions with blurred anatomical structures, significantly improving the coherence and medical interpretability of the segmentation map.

[0146] This invention constructs a nested interaction mechanism between Transformer and GNN, which achieves residual update feedback of graph semantic information by back-projecting the node embedding feature vectors obtained from the graph neural network to the input channel of the original multi-scale Transformer. It also allows the graph structure and encoded features to be reconstructed in each loop, thereby achieving the co-evolution of cross-scale graph structure and encoding pyramid until the residual converges. When dealing with dynamic changes in brain regions, it can more accurately capture boundary evolution and region expansion, and improve the model's ability to perceive dynamic changes in small regions.

[0147] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic segmentation method for three-dimensional brain networks based on deep learning, characterized in that, Includes the following steps: S1. Construct temporally and spatially consistent fused image volume data; S2. Divide the fused image volume data into cubes of different scales according to the multi-level cube partitioning strategy, perform position encoding on each scale cube and input it into the multi-scale sparse Transformer encoder to generate a multi-scale sparse Transformer encoded feature pyramid. S3. Based on the predefined medical partition template and diffusion tensor imaging connectivity information, an initial set of brain region map nodes is established for the multi-scale sparse Transformer encoded feature pyramid at each scale to obtain the brain region map structure at the same scale. S4. Perform cross-scale node alignment on brain region map nodes of the same name at different scales, and use the cross-scale node alignment fusion map structure as the initial output of the cross-scale node alignment fusion mechanism; S5. Node message passing is performed within the cross-scale node alignment fusion graph structure by graph convolution and multi-head graph attention to obtain the first round of fused brain region map node embedding features; S6. Use the first round of fused brain region map node embedding features as residual information to feed back to the next layer input of the multi-scale sparse Transformer encoder to obtain the updated multi-scale sparse Transformer encoded feature pyramid, and repeat steps S3 to S5 until the multi-scale sparse Transformer-GNN interactive update is completed. S7. Perform skip connection fusion and hierarchical upsampling reconstruction on the final multi-scale sparse Transformer encoded feature pyramid to generate voxel-level segmentation probability volume data; S8. Post-process the voxel-level segmentation probability data of the training and inference phases respectively to obtain the dynamic segmentation results of the three-dimensional brain regions.

2. The method for dynamic segmentation of a three-dimensional brain network based on deep learning according to claim 1, characterized in that, S1 includes the following steps: S11. Acquire structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data; S12. Perform initial resampling processing on the structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data, respectively; S13. Perform voxel-level rigid registration of functional magnetic resonance imaging data and diffusion tensor imaging data relative to structural magnetic resonance imaging data; S14. Perform non-rigid registration processing on the functional magnetic resonance imaging data and diffusion tensor imaging data to establish a deformation field. The deformation field describes the small displacement of each voxel position (x,y,z) in three spatial dimensions. S15. Perform intensity normalization processing on the registered structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data respectively to obtain normalized structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data. S16. Perform artifact removal processing on the normalized structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data; S17. The structural magnetic resonance imaging data, functional magnetic resonance imaging data, and diffusion tensor imaging data, after registration, normalization, and artifact removal, are stitched and fused along the voxel dimension at the channel level to form the final fused image volume data D. fusion .

3. The method for dynamic segmentation of a three-dimensional brain network based on deep learning according to claim 2, characterized in that, S2 includes the following steps: S21. Merge image volume data D fusion The system is segmented according to a multi-level cubic partitioning strategy, using different scale partitioning factors s. l The original 3D spatial voxel mesh is divided into a set of non-overlapping cubes corresponding to each scale level, with each scale level corresponding to a set of cubes at a specific resolution. Where N l This represents the total number of cubes in the l-th layer, where l represents the scale level. Indicates the Nth level under scale level l l Each cube is a small cubic region that has been divided into three-dimensional spaces; S22. For each cube at each scale level Perform position encoding to obtain the position encoded vector. S23. Encode the position vector at each scale level. fused feature vector with the corresponding cube The data are concatenated to form a set of scale feature representations F. (l) ; S24. For shallow-scale hierarchies l high , representing the scale feature set Input a local window self-attention module, and set the window size to [value]. Perform a dense self-attention mechanism within each window to output a shallow, local, fine-grained representation. S25. For deep-scale hierarchies l low , representing the scale feature set Input a cross-window sparse self-attention module, establish connections only between a preset subset of queries and their associated keys, and output a deep global coarse-grained representation. S26. For each shallow-scale level, the shallow local fine-grained representation is fused with the matching deep sparse global representation element-wise to form the fused encoding representation Z. (l) ; S27. Fusion coding representation for all scale levels We will organize them in a unified manner and construct a multi-scale sparse Transformer encoded feature pyramid with cross-level structure perception capabilities.

4. The method for dynamic segmentation of a three-dimensional brain network based on deep learning according to claim 3, characterized in that, S3 includes the following steps: S31. Construct a predefined medical partition template to form a brain region partition set R. The brain region partition set consists of multiple brain region template regions. In each scale level l, the fused encoded representation set Z is... (l) The spatial position of each cube is determined to coincide with the brain region template region. The cube is mapped to the graph node in the brain region map structure through a mapping function to form the initial node mapping relationship. S32. At scale level l, generate an initial brain region map node set V based on node mapping relationships. (l) Each graph node in the initial set of brain region graph nodes is used to represent a set of cubes of a brain region template at the current scale. For each graph node, the average value of the encoding representation of all cubes contained in the current node is calculated to obtain the initial feature vector of the current graph node. S33. Using the initial feature vector set of all graph nodes under scale level l, calculate the anatomical similarity between each pair of graph nodes and form an anatomical similarity matrix; S34. Based on the normalized diffusion tensor imaging data, extract the number of white matter fiber tracts tracked between each pair of brain region templates, and calculate the normalized connectivity based on the fiber distribution density inside and outside the region. Organize the functional connectivity results between all regions into a functional similarity matrix. S35. By linearly weighting and fusing the anatomical similarity matrix and the functional similarity matrix, a graph adjacency matrix at scale level l is constructed; S36. Combine the set of brain region map nodes at scale level l with the graph adjacency matrix to form the same-scale brain region map structure G at the current scale level. (l) .

5. The method for dynamic segmentation of a three-dimensional brain network based on deep learning according to claim 4, characterized in that, S3 includes the following steps: S41. Set a cross-scale alignment threshold, select all scale levels with resolutions lower than the cross-scale alignment threshold to form a low-resolution set, and select all scale levels with resolutions higher than or equal to the cross-scale alignment threshold to form a high-resolution set. S42. For each brain region map structure in a low-resolution scale level, extract the spatial center coordinates of each graph node, and at the same time extract the initial feature vector of the graph node; S43. At the high-resolution scale level, for all candidate graph nodes that have the same medical template name as the low-resolution graph nodes, calculate the spatial Euclidean distance between each candidate graph node and the low-resolution graph node, and select the candidate node with the smallest Euclidean distance as the high-resolution alignment target graph node. S44. Form a matching pair between each pair of low-resolution graph nodes and the nearest neighbor high-resolution aligned target graph node, and construct a cross-scale aligned mapping set; S45. For each pair of matching graph nodes in the cross-scale alignment mapping set, calculate the feature similarity using the initial feature vectors of the two graph nodes; S46. Combine all cross-scale matching graph node pairs and their corresponding alignment weights to form a cross-scale edge set; S47. Merge the low-resolution scale hierarchical brain region map structure with the high-resolution scale hierarchical brain region map structure and the cross-scale edge set to obtain the final cross-scale node-aligned fused brain region map structure.

6. The method for dynamic segmentation of a three-dimensional brain network based on deep learning according to claim 5, characterized in that, S5 includes the following steps: S51. Based on the mapping relationship between graph nodes and cubes in the cross-scale node alignment fusion brain region map structure, the cube block encoding features at each scale level in the multi-scale sparse Transformer encoding feature pyramid are mapped to the initial attribute features of the corresponding graph nodes. S52. The initial attribute features of all nodes in the graph at all scale levels are concatenated and integrated according to the node number to form the initial feature matrix of the cross-scale node set; S53. Input the initial feature matrix into the graph neural network message passing module, combine it with the adjacency matrix of the brain region map structure fused by cross-scale node alignment to perform the first round of graph convolution operation, and obtain the first round of node embedding representation H. (1) ; S54. Input the first-round graph convolution representation into the multi-head graph attention mechanism module. The outputs of each graph node from all attention heads are integrated through a concatenation operation to generate the first-round fused brain region graph node embedding feature h. i (1) .

7. The method for dynamic segmentation of a three-dimensional brain network based on deep learning according to claim 1, characterized in that, S6 includes the following steps: S61. Based on the mapping relationship between graph nodes and original cubes, embed the first round of fused brain region graph nodes into features. Project and broadcast back to the original cubes at each scale level in the multi-scale sparse Transformer encoded feature pyramid to obtain the graph feedback residual information corresponding to each cube; S62. For each scale level l, the graph feedback residual information R is used. (1) The corresponding residual vector The cube features of the previous Transformer encoding output Perform a joint update to form a new cube input representation. S63. Input the new cube representation The input is fed into the next layer of the multi-scale sparse Transformer encoder, where local window dense self-attention and cross-window sparse self-attention mechanisms are executed. Multi-scale feature extraction is performed on the new input, and the updated multi-scale sparse Transformer encoded feature pyramid is output. S64. After each round of updates, repeat the same-scale brain region map structure construction process in S3 to update and obtain a new brain region map structure; S65. Based on the new brain region map structure, repeat the cross-scale node alignment process in S4 to generate a new cross-scale node alignment fusion brain region map structure; S66. Apply graph neural network message passing and multi-head graph attention mechanisms to the new cross-scale node-aligned fused brain region map structure to generate second-round fused brain region map node embedding features. And then feed it back to the encoder for residual update; S67. Repeat steps S61 to S66 to perform multiple rounds of Transformer-GNN dual-path interactive updates until the set number of iterations or residual convergence threshold is reached, and finally a stable multi-scale sparse Transformer encoded feature pyramid and fused brain region map node embedding representation are obtained.

8. The method for dynamic segmentation of a three-dimensional brain network based on deep learning according to claim 7, characterized in that, S7 includes the following steps: S71. Based on the stable multi-scale sparse Transformer encoding feature pyramid Encoded features at all scale levels A skip connection fusion strategy is adopted to integrate information, and channel splicing and convolution fusion are performed with the current level encoded features to generate a fused cross-scale voxel feature representation. S72. Based on the mapping relationship between graph nodes and cubes, the embedded representation of the fused brain region graph nodes is projected back to its corresponding voxel block region through the nearest neighbor, and a graph feature field is established in the voxel space. The cross-scale voxel feature representation and the graph feature field are spliced ​​together in the channel dimension to form the final fused voxel-level segmentation feature. S73. The final fused voxel-level segmentation features are input into the segmentation head network composed of convolutional layers. The segmentation head network performs a set of 1×1×1 convolution operations and outputs voxel-level multi-class probability predictions through the softmax function to generate voxel-level segmentation probability volume data.

9. The method for dynamic segmentation of a three-dimensional brain network based on deep learning according to claim 8, characterized in that, Specifically, S8 includes introducing neighboring time frame probability consistency constraints and deformation field smoothing regularization into the voxel-level segmentation probability data of continuous time frames during the training phase; applying recursive Bayesian filtering and B-spline-based deformation compensation to the voxel-level segmentation probability data during the inference phase; and performing conditional random field refinement, time-of-flight mapping-based hole filling, and confidence assessment on the voxel-level segmentation probability data that has undergone smooth evolution in the time domain to obtain dynamic segmentation results of three-dimensional brain regions.

Images