Three-dimensional spine model reconstruction method based on neural network and 3D network architecture

Abstract

The invention provides a three-dimensional spine model reconstruction method based on a neural network and a 3D network architecture, and relates to the technical field of medical image analysis and biomedical engineering, and the method comprises the steps: carrying out the preprocessing of obtained spine medical image data, and obtaining spine image data; performing graph structure conversion on the spine image data to obtain a spine graph structure; the method comprises the following steps: based on a spine map structure, carrying out feature processing by utilizing a 3D U-Net architecture integrated with a map neural network to obtain a preliminary three-dimensional spine reconstruction model; and performing segmentation and post-processing operation on the preliminary three-dimensional spine reconstruction model to obtain a final three-dimensional spine reconstruction model. According to the method, efficient and accurate segmentation of the spine image is achieved, the quality of three-dimensional reconstruction is improved, the integrity and continuity of the spine structure are guaranteed, and the method has important value for medical applications such as auxiliary diagnosis, surgical planning and treatment monitoring.

Description

Technical Field

[0001] The present invention relates to the field of medical image analysis and biomedical engineering technology, and in particular to a three-dimensional spine model reconstruction method based on a neural network and a 3D network architecture. Background Art

[0002] The reconstruction method of the three-dimensional model of the spine usually includes the acquisition of image data (CT, MRI, X-ray), image preprocessing (denoising, standardization, alignment and registration), image segmentation, three-dimensional model construction, finite element model generation and optimization, etc. Among them, the accuracy of image segmentation directly determines the accuracy of the reconstructed model. The core idea of ​​spinal image segmentation is to accurately identify and separate the various anatomical structures of the spine from spinal medical images, including vertebral bodies, pedicles, intervertebral discs, etc. The existing segmentation methods have been continuously evolving in medical image processing and can be roughly divided into traditional image segmentation methods and deep learning methods. Deep learning methods are superior to traditional methods in terms of image segmentation accuracy, especially when dealing with complex image structures. In recent years, the application of deep learning methods in the field of spinal image segmentation has made significant progress, and important research hotspots are mainly concentrated on the model framework. Its core technology lies in the construction of network structures (such as convolutional neural networks CNN, fully convolutional networks FCN, U-Net, etc.). These networks automatically extract features layer by layer from the original medical image data, gradually transitioning from the initial low-level visual features such as edges and textures to high-level, abstract features related to anatomical structures.

[0003] Among many network structure frameworks, 3D U-Net (a deep learning model for segmentation of three-dimensional image data) has shown excellent performance in the field of spinal image segmentation because it can capture three-dimensional spatial features, especially the segmentation accuracy at the level of a single vertebra. However, the recognition ability of 3D U-Net is still insufficient for the spinal margin and internal details that are of concern in customized spinal design, as shown in the following:

[0004] (1) Optimization of segmentation of blurred boundaries and low-contrast regions. In spinal images, the boundary between the spine edge and soft tissue becomes blurred and has low contrast due to noise, tissue overlap, and signal intensity similarity. 3D U-Net requires additional optimization strategies to improve segmentation results, such as introducing an attention mechanism to highlight potential structural boundary information.

[0005] (2) Identification of subtle abnormal internal structures. For subtle abnormal structures inside the spine, 3D U-Net has limited recognition capabilities for ultra-small structures due to the limitations of convolution kernel size and step size. Therefore, it is urgent to develop a specific network structure tailored to the anatomical characteristics and image features of the spine, which can solve the problem of segmentation of blurred spine boundaries and complex internal detailed structures.

[0006] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0007] In view of this, the present invention provides a three-dimensional spine model reconstruction method based on a neural network and a 3D network architecture to solve the above-mentioned problems.

[0008] In order to solve the above problems, the specific technical solutions adopted by the present invention are as follows:

[0009] A three-dimensional spine model reconstruction method based on a neural network and a 3D network architecture, the method comprising the following steps:

[0010] S1. Preprocessing the acquired spinal medical imaging data to obtain spinal imaging data;

[0011] S2, performing graph structure conversion on the spinal image data to obtain a spinal graph structure;

[0012] S3. Based on the spinal graph structure, the 3D U-Net architecture integrated with the graph neural network is used for feature processing to obtain a preliminary three-dimensional spinal reconstruction model.

[0013] S4. Segment and post-process the preliminary three-dimensional spine reconstruction model to obtain the final three-dimensional spine reconstruction model.

[0014] Preferably, the step of converting the spinal image data into a graph structure to obtain a spinal graph structure comprises the following steps:

[0015] S21, performing voxel processing on the pre-processed spinal image data to obtain voxelized data;

[0016] S22, identifying anatomical feature points in the spine using image processing technology based on the obtained voxelized data;

[0017] S23. Based on the anatomical feature points in the spine, a spinal graph structure is constructed through neighborhood analysis.

[0018] Preferably, the step of identifying anatomical feature points in the spine using image processing technology based on the obtained voxelized data comprises the following steps:

[0019] S221, performing curvature analysis on the center line of the spine by using a skeleton algorithm based on the voxelized data to obtain a curvature analysis result;

[0020] S222, analyzing the geometric shapes of the spinal vertebrae in the voxelized data to obtain shape analysis results;

[0021] S223. Identify anatomical feature points in the spine based on the curvature analysis results and the shape analysis results.

[0022] Preferably, the constructing of the spinal graph structure by neighborhood analysis based on the anatomical feature points in the spine comprises the following steps:

[0023] S231, for each anatomical feature point, calculating a neighborhood set of each anatomical feature point using spatial proximity;

[0024] S232, constructing an edge for each pair of adjacent anatomical feature points according to the proximity relationship, morphological features and gray value information, and calculating the allocation weight of each edge;

[0025] S233. Establish a spinal graph structure based on the anatomical feature points and the constructed edges, and optimize the spinal graph structure through an optimization algorithm.

[0026] Preferably, the calculation formula for calculating the neighborhood set of each anatomical feature point using spatial proximity is:

[0027] N6(v i )={v j ∈V|||v i -v j ||1=1};

[0028] In the formula, N6(v i ) represents the neighborhood set of the i-th anatomical feature point in the anatomical feature point set;

[0029] V represents the set of all anatomical feature points;

[0030] ||.||1 represents Manhattan distance;

[0031] v i and v j They respectively represent the i-th anatomical feature point and the j-th anatomical feature point in the anatomical feature point set.

[0032] Preferably, the calculation formula for calculating the allocation weight of each edge is:

[0033] w ij =α·d ij +β·m ij +γ·s ij ;

[0034] In the formula, w ij Represents the anatomical feature point v i and v j The weight of the edge between them;

[0035] d ij Represents the anatomical feature point v i and v j The spatial distance weight between them;

[0036] mij Represents the anatomical feature point v i and v j The weight of morphological features between ;

[0037] s ij Represents the anatomical feature point v i and v j The grayscale feature weight between ;

[0038] α, β, and γ all represent weight coefficients.

[0039] Preferably, the step of establishing a spinal graph structure according to the anatomical feature points and the constructed edges, and optimizing the spinal graph structure by an optimization algorithm comprises the following steps:

[0040] S2331. Sort the edges of the spine graph structure based on the weights of the edges;

[0041] S2332, taking the edge with the smallest weight as the initial point, using the Kruskal algorithm to add edges one by one to the preset spanning tree, to obtain a minimum spanning tree;

[0042] S2333, removing the edges that are not in the minimum spanning tree to obtain a primary optimized spine graph structure;

[0043] S2334, based on the primary optimized spinal graph structure, construct a similarity matrix according to the attributes of the anatomical feature points, and calculate a distance matrix between the anatomical feature points;

[0044] S2335. Calculate the graph Laplacian matrix based on the distance matrix, and cluster the anatomical feature points through K-means to obtain the final spinal graph structure.

[0045] Preferably, the method of performing feature processing based on the spinal graph structure using a 3D U-Net architecture integrated into a graph neural network to obtain a preliminary three-dimensional spinal reconstruction model comprises the following steps:

[0046] S31, using a 3D convolutional layer to extract features from the spine graph structure to obtain a multi-channel feature graph containing spatial context information;

[0047] S32. Use the transposed convolution layer to upsample and fuse the multi-channel feature map to obtain 3D volume data.

[0048] Preferably, the segmentation and post-processing of the preliminary three-dimensional spine reconstruction model to obtain the final three-dimensional spine reconstruction model comprises the following steps:

[0049] S41, constructing nodes from the obtained 3D volume data, establishing node edges based on spatial connectivity and feature similarity, and constructing an optimized graph model based on the nodes and node edges;

[0050] S42, using the maximum flow-minimum cut algorithm to segment the optimized graph model to obtain a segmentation result;

[0051] S43. Post-process and optimize the segmentation result using a boundary refinement method to obtain the final three-dimensional spine reconstruction result.

[0052] Preferably, the post-processing optimization of the segmentation result by using the boundary refinement method to obtain the final three-dimensional spine reconstruction result comprises the following steps:

[0053] S431, extracting the edge of the segmentation result using an edge detection algorithm;

[0054] S432, performing preliminary edge thinning processing on the extracted edge using a region growing algorithm;

[0055] S433, performing secondary edge thinning processing on the preliminary boundary thinning processing result by using morphological operation;

[0056] S434, smoothing the secondary edge thinning result to obtain the final three-dimensional spine reconstruction result.

[0057] The beneficial effects of the present invention are:

[0058] 1. The present invention realizes efficient and accurate segmentation of spinal images, which not only improves the quality of three-dimensional reconstruction, but also ensures the integrity and continuity of the spinal structure, and has important value for medical applications such as auxiliary diagnosis, surgical planning and treatment monitoring.

[0059] 2. The automatic reconstruction method of the three-dimensional spine model implemented by the present invention significantly improves the accuracy of image segmentation and the authenticity of the three-dimensional model, and effectively solves the problems of structural continuity and detail loss in traditional methods. The application of this technology will greatly promote the accuracy and efficiency of medical image analysis, surgical planning, disease diagnosis and treatment evaluation, and is of great significance to promoting the development of personalized medicine and precision surgery.

[0060] 3. The present invention innovatively attempts to transform the spinal structure into a graphical representation and utilizes the unique capability of GNN to learn the topological connections between nodes and the feature information of each node itself. This fusion strategy guides 3D U-Net to more accurately identify and extract features of key spinal nodes through the deep structural dependencies captured by GNN, which is expected to make up for the shortcomings of 3DU-Net in processing edge and detail features, thereby bringing higher accuracy and robustness to spinal image segmentation.

[0061] 4. The present invention improves the accuracy and reliability of spinal structure segmentation through innovative graph model representation, fusion of graph neural network (GNN) 3D U-Net model and graph cut-based post-processing strategy, and effectively improves the accuracy of the model in spinal image segmentation tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0063] Figure 1 It is a flowchart of a three-dimensional spine model reconstruction method based on a neural network and a 3D network architecture according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0065] According to an embodiment of the present invention, a three-dimensional spine model reconstruction method based on a neural network and a 3D network architecture is provided.

[0066] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to the method for reconstructing a three-dimensional spine model based on a neural network and a 3D network architecture according to an embodiment of the present invention, the method comprises the following steps:

[0067] S1. Preprocessing the acquired spinal medical imaging data to obtain spinal imaging data;

[0068] Specifically, preprocessing the acquired spinal medical imaging data to obtain spinal imaging data includes the following steps:

[0069] S11. Collect medical imaging data for three-dimensional reconstruction of the spine, such as CT or MRI images.

[0070] S12. Perform image registration processing to ensure correct alignment between different slices to facilitate subsequent structural analysis.

[0071] S13, performing preprocessing operations such as image denoising, contrast enhancement and standardization to improve image quality.

[0072] S2, performing graph structure conversion on the spinal image data to obtain a spinal graph structure;

[0073] As a preferred implementation, the graph structure conversion of the spinal image data to obtain the spinal graph structure comprises the following steps:

[0074] S21, voxelizing the preprocessed spinal image data to obtain voxelized data;

[0075] It should be noted that voxelization is performed on the original spinal image data to convert continuous spatial data into discrete data. This technology is a common technology in the field of computer-aided engineering.

[0076] S22, identifying anatomical feature points in the spine using image processing technology based on the obtained voxelized data;

[0077] As a preferred embodiment, the method of identifying anatomical feature points in the spine using image processing technology based on the obtained voxelized data includes the following steps:

[0078] S221, based on the voxelized data, performing curvature analysis on the center line of the spine by using a skeleton algorithm (such as Zhang-Suen algorithm extraction) to obtain a curvature analysis result;

[0079] S222, analyzing the geometric shapes of the spinal vertebrae in the voxelized data to obtain shape analysis results;

[0080] S223. Identify anatomical feature points in the spine based on the curvature analysis results and the shape analysis results.

[0081] The position where the curvature changes significantly is the corresponding anatomical feature point. The anatomical feature points are identified by analyzing the geometric shape of the vertebral body and setting the threshold of the shape parameters (such as area and circumference).

[0082] S23. Based on the anatomical feature points in the spine, a spinal graph structure is constructed through neighborhood analysis.

[0083] As a preferred embodiment, the construction of the spinal graph structure by neighborhood analysis based on the anatomical feature points in the spine includes the following steps:

[0084] S231, for each anatomical feature point, calculating a neighborhood set of each anatomical feature point using spatial proximity;

[0085] It should be noted that, based on spatial proximity, a neighbor set is defined for each identified voxel node, and this application selects six neighborhoods, that is, each voxel is adjacent to the voxels above, below, left, right, front and back.

[0086] As a preferred implementation, the calculation formula for calculating the neighborhood set of each anatomical feature point using spatial proximity is:

[0087] N6(v i )={v j ∈V|||v i -v j ||1=1};

[0088] In the formula, N6(v i ) represents the neighborhood set of the i-th anatomical feature point in the anatomical feature point set;

[0089] V represents the set of all anatomical feature points;

[0090] ||.||1 represents Manhattan distance;

[0091] v i and v j They respectively represent the i-th anatomical feature point and the j-th anatomical feature point in the anatomical feature point set.

[0092] S232, constructing an edge for each pair of adjacent anatomical feature points according to the proximity relationship, morphological features and gray value information, and calculating the allocation weight of each edge;

[0093] As a preferred implementation, the calculation formula for calculating the allocation weight of each edge is:

[0094] w ij =α·d ij +β·m ij +γ·s ij ;

[0095] In the formula, w ij Represents the anatomical feature point v i and v j The weight of the edge between them;

[0096] d ij Represents the anatomical feature point v i and v j The spatial distance weight between them;

[0097] m ij Represents the anatomical feature point v i and v j The weight of morphological features between ;

[0098] s ij Represents the anatomical feature point v i and v j The grayscale feature weight between ;

[0099] α, β, and γ all represent weight coefficients, which are used to adjust the importance of different factors. The optimal values ​​of α, β, and γ can be determined by an optimization algorithm (such as gradient descent).

[0100] Among them, for each anatomical feature point v i (x,y,z), spatial distance weight d ij Using Euclidean distance, the calculation formula is as follows:

[0101]

[0102] (2) Morphological feature weight m ij Based on the size, shape and other characteristics of the node:

[0103]

[0104] Among them, f i and f j is the anatomical feature point v i and v j The morphological characteristic value, f max represents the maximum possible difference among all eigenvalues.

[0105] (3) Gray value similarity weight s ij It can be based on the grayscale value of the voxel:

[0106]

[0107] Among them, I i and I j is the anatomical feature point v i and v j The gray value, I max is the maximum possible range of grayscale values.

[0108] S233. Establish a spinal graph structure based on the anatomical feature points and the constructed edges, and optimize the spinal graph structure through an optimization algorithm.

[0109] As a preferred implementation, the establishment of a spinal graph structure based on anatomical feature points and constructed edges, and optimization of the spinal graph structure by an optimization algorithm include the following steps:

[0110] S2331. Sort the edges of the spine graph structure based on the weights of the edges;

[0111] S2332, taking the edge with the smallest weight as the initial point, using the Kruskal algorithm to add edges one by one to the preset spanning tree, to obtain a minimum spanning tree;

[0112] S2333, removing the edges that are not in the minimum spanning tree to obtain a primary optimized spine graph structure;

[0113] S2334, based on the primary optimized spinal graph structure, construct a similarity matrix according to the attributes of the anatomical feature points, and calculate a distance matrix between the anatomical feature points;

[0114] S2335. Calculate the graph Laplacian matrix based on the distance matrix, and cluster the anatomical feature points through K-means to obtain the final spinal graph structure.

[0115] Specifically, the number of original voxels is huge, and the structure of the constructed spine graph is complex, which is not conducive to subsequent analysis. The graph structure is optimized by removing low-weight edges and clustering similar nodes to maintain its core topological and functional information.

[0116] The method to remove edges with low weight is as follows:

[0117] First, sort the edges: according to the edge weight w ij Sort all edges.

[0118] Then, construct the minimum spanning tree: use Kruskal's algorithm (a classic algorithm for finding a minimum spanning tree) to start with the edge with the smallest weight and add edges to the spanning tree one by one, making sure no cycles are formed.

[0119] Finally, remove edges: All edges that are not in the MST (minimum spanning tree) are considered to be lower weight edges and can be removed.

[0120] The method of clustering similar nodes is as follows:

[0121] First, construct a similarity matrix: construct a similarity matrix S based on the attributes of the nodes (such as gray value, position, etc.).

[0122] Secondly, the distance matrix D between nodes is calculated. A single linkage method is used to merge nodes step by step.

[0123] Then, calculate the graph Laplacian matrix L = DA, where A is the adjacency matrix of the graph and D is a diagonal matrix whose diagonal elements D ii is the sum of the elements in the i-th row of A. Calculate the first k eigenvectors of the Laplacian matrix. Treat these eigenvectors as points in a high-dimensional space and cluster them using the K-means algorithm.

[0124] S3. Based on the spinal graph structure, the 3D U-Net architecture integrated with the graph neural network is used for feature processing to obtain a preliminary three-dimensional spinal reconstruction model.

[0125] As a preferred implementation, the method of performing feature processing based on the spinal graph structure using a 3D U-Net architecture integrated with a graph neural network to obtain a preliminary three-dimensional spinal reconstruction model includes the following steps:

[0126] S31, using a 3D convolutional layer to extract features from the spine graph structure to obtain a multi-channel feature graph containing spatial context information;

[0127] Specifically, the 3D U-Net architecture integrated with GNN (graph neural network) combines the capabilities of the traditional 3D U-Net structure and graph neural network, including feature extraction and downsampling, upsampling and feature fusion, output and post-processing. The details are as follows:

[0128] The input spine graph structure is feature extracted using a 3D convolutional layer to obtain a multi-channel feature map containing spatial context information. Then the GCN module is introduced to use graph convolution operations beyond the voxel level to capture more complex neighborhood interactions, taking into account non-local structural dependencies and global consistency.

[0129] Here are the detailed steps and code examples to illustrate how to perform downsampling after a 3D convolutional layer:

[0130] First, 3D convolutional layers: 3D convolutional layers are applied to extract features, and each convolutional layer is usually followed by batch normalization and ReLU activation function.

[0131] Then, the maximum pooling layer: After the convolution layer, the maximum pooling layer is applied to reduce the spatial dimensions (depth, height, width) of the feature map. The maximum pooling operation selects the maximum value in each local area as the representative value of the area.

[0132] Finally, the above process is repeated: convolution and pooling operations are repeated at different levels of the network to gradually reduce the size of the feature map while increasing the number of channels of the feature map.

[0133] S32. Use the transposed convolution layer to upsample and fuse the multi-channel feature map to obtain 3D volume data.

[0134] Specifically, the transposed convolution layer is used to upsample and fuse the multi-channel feature maps to obtain 3D volume data including:

[0135] Transposed convolutional layers: Transposed convolutional layers are used for upsampling, they increase the spatial dimension of the feature map through convolution operations. Transposed convolutions are similar to the inverse process of pooling, but they do not change the number of channels of the feature map.

[0136] Skip connections: The skip connections in the 3D U-Net architecture concatenate feature maps of different resolutions in the encoder stage with feature maps of the corresponding layers in the decoder stage to preserve spatial information.

[0137] Graph Attention Mechanism: In the decoding stage, a graph attention layer is added to enhance the model’s attention to important structures. The graph attention layer can learn the dependencies between nodes and adjust the importance of nodes accordingly.

[0138] Feature fusion: The concatenated feature maps are further processed by convolutional layers to fuse information from different resolutions.

[0139] Considering that in the encoder-decoder architecture, the skip connections of 3D U-Net are still retained to fuse feature information at different resolutions. It is planned to further strengthen feature fusion in the decoding stage by adding a graph attention mechanism to enhance the model's attention to important structures and guide the network to better reconstruct the detailed spine structure.

[0140] Output and post-processing. The network output is 3D volume data, where the value of each voxel represents the probability of it belonging to the spinal structure. The network output here refers to the final output of the entire 3D U-Net network, that is, the 3D volume data obtained after the aforementioned feature extraction, downsampling, upsampling and feature fusion steps. The value of each voxel represents the probability that the voxel belongs to the spinal structure. The probability voxel data is binarized according to a pre-set threshold to generate an accurate spinal segmentation result.

[0141] S4. Segment and post-process the preliminary three-dimensional spine reconstruction model to obtain the final three-dimensional spine reconstruction model.

[0142] As a preferred embodiment, the segmentation and post-processing of the preliminary three-dimensional spine reconstruction model to obtain the final three-dimensional spine reconstruction model includes the following steps:

[0143] S41, constructing nodes from the obtained 3D volume data, establishing node edges based on spatial connectivity and feature similarity, and constructing an optimized graph model based on the nodes and node edges;

[0144] Specifically, the 3D volume data is obtained to form nodes, node edges are established based on spatial connectivity and feature similarity, and when an optimized graph model is constructed based on nodes and node edges, it is necessary to allocate inner edge energy (based on feature differences between nodes) and outer edge energy (to ensure smooth and accurate boundaries).

[0145] S42, using the maximum flow-minimum cut algorithm to segment the optimized graph model to obtain a segmentation result;

[0146] Specifically, by selecting a suitable graph cut algorithm (such as maximum flow-minimum cut), adjusting parameters (such as node weights, edge weights, etc.) to balance the accuracy and smoothness of segmentation, executing the algorithm to obtain the cutting solution with the minimum energy, and refining the spine edge, the details are as follows:

[0147] Step 1: Execute the graph cut algorithm;

[0148] Energy minimization: Use the maximum flow-minimum cut algorithm to find the cutting solution with the minimum energy.

[0149] Step 2: Result analysis;

[0150] Result Visualization: Visualize the segmentation results to intuitively evaluate the quality of the segmentation.

[0151] Step 3: Parameter adjustment;

[0152] Balance the accuracy and smoothness of segmentation: Balance the accuracy and smoothness of segmentation by adjusting node weights and edge weights.

[0153] Execute the algorithm: Re-execute the graph cut algorithm with the adjusted parameters to obtain a new segmentation result.

[0154] Step 4: Iterative optimization;

[0155] Iterative process: Repeat steps 1 to 3 until a satisfactory segmentation result is obtained.

[0156] S43. Post-process and optimize the segmentation results using a boundary refinement method to obtain the final three-dimensional spine reconstruction model.

[0157] As a preferred embodiment, the post-processing optimization of the segmentation results by using the boundary refinement method to obtain the final three-dimensional spine reconstruction model includes the following steps:

[0158] S431, extracting the edge of the segmentation result using an edge detection algorithm;

[0159] S432, performing preliminary edge thinning processing on the extracted edge using a region growing algorithm;

[0160] S433, performing secondary edge thinning processing on the preliminary boundary thinning processing result by using morphological operation;

[0161] S434, smoothing the secondary edge refinement result to obtain a final three-dimensional spine reconstruction model.

[0162] Specifically, the segmentation results are refined and smoothed, the quality is ensured through quantitative and qualitative evaluation, the parameters are adjusted according to the feedback until they are satisfactory, and finally the optimal configuration is confirmed and saved for subsequent application.

[0163] Boundary refinement aims to extract more precise edges, reduce the width of edges, and make the segmentation results clearer. Common boundary refinement methods include:

[0164] Step 1: Edge detection;

[0165] Use an edge detection algorithm (such as Canny edge detection) to extract the edges of the original segmentation result.

[0166] Step 2: Region growing;

[0167] Based on the edge detection results, the region growing algorithm is used to refine the edge and classify the voxels near the edge into the edge region.

[0168] Step 3: Morphological operation;

[0169] Morphological operations (such as dilation and erosion) are applied to further refine the edges and remove noise and small holes on the edges.

[0170] Smoothing is designed to reduce noise in segmentation results and improve segmentation continuity and visual effects. Common smoothing methods include:

[0171] Step 1: Gaussian filtering;

[0172] A Gaussian filter is used to smooth the segmentation results to reduce noise and improve the continuity of segmentation.

[0173] Step 2: bilateral filtering;

[0174] Apply a bilateral filter, which considers both spatial proximity and pixel value similarity for better smoothing.

[0175] Step 3: Morphological operation;

[0176] Morphological operations (such as opening and closing) are applied to smooth the segmentation results, remove noise and improve the continuity of the segmentation.

[0177] In summary, with the help of the above technical solutions of the present invention, the present invention realizes efficient and accurate segmentation of spinal images, which not only improves the quality of three-dimensional reconstruction, but also ensures the integrity and continuity of the spinal structure, and has important value for medical applications such as auxiliary diagnosis, surgical planning and treatment monitoring. The three-dimensional spine model automatic reconstruction method implemented by the present invention significantly improves the accuracy of image segmentation and the authenticity of the three-dimensional model, and effectively solves the problems of structural continuity and detail loss in traditional methods. The application of this technology will greatly promote the accuracy and efficiency of medical image analysis, surgical planning, disease diagnosis and treatment evaluation, and is of great significance to the development of personalized medicine and precision surgery. The present invention innovatively attempts to convert the spinal structure into a graphical representation, and uses the unique ability of GNN to learn the topological connection between nodes and the feature information of each node itself. This fusion strategy guides 3D U-Net to more accurately identify and extract features of key spinal nodes through the deep structural dependencies captured by GNN, which is expected to make up for the shortcomings of 3D U-Net in processing edge and detail features, thereby bringing higher accuracy and robustness to spinal image segmentation. The present invention improves the accuracy and reliability of spinal structure segmentation through innovative graph model representation, fusion of graph neural network (GNN) 3D U-Net model and graph cut-based post-processing strategy, and effectively improves the accuracy of the model in spinal image segmentation tasks.

[0178] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) containing computer-usable program code.

[0179] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. Three-dimensional spine model reconstruction method based on neural network and 3D network architecture, It is characterized in that The method comprises the following steps: S1. Preprocessing the acquired spinal medical imaging data to obtain spinal imaging data; S2, performing graph structure conversion on the spinal image data to obtain a spinal graph structure; S3. Based on the spinal graph structure, the 3D U-Net architecture integrated with the graph neural network is used for feature processing to obtain a preliminary three-dimensional spinal reconstruction model. S4. Segment and post-process the preliminary three-dimensional spine reconstruction model to obtain the final three-dimensional spine reconstruction model.

2. The three-dimensional spine model reconstruction method based on neural network and 3D network architecture according to claim 1, It is characterized in that The graph structure conversion of the spinal image data to obtain the spinal graph structure comprises the following steps: S21, performing voxel processing on the pre-processed spinal image data to obtain voxelized data; S22, identifying anatomical feature points in the spine using image processing technology based on the obtained voxelized data; S23. Based on the anatomical feature points in the spine, a spinal graph structure is constructed through neighborhood analysis.

3. The three-dimensional spine model reconstruction method based on neural network and 3D network architecture according to claim 2, It is characterized in that The method of identifying anatomical feature points in the spine using image processing technology based on the obtained voxelized data comprises the following steps: S221, performing curvature analysis on the center line of the spine by using a skeleton algorithm based on the voxelized data to obtain a curvature analysis result; S222, analyzing the geometric shapes of the spinal vertebrae in the voxelized data to obtain shape analysis results; S223. Identify anatomical feature points in the spine based on the curvature analysis results and the shape analysis results.

4. The three-dimensional spine model reconstruction method based on neural network and 3D network architecture according to claim 3, It is characterized in that The method of constructing a spinal graph structure based on anatomical feature points in the spine through neighborhood analysis comprises the following steps: S231, for each anatomical feature point, calculating a neighborhood set of each anatomical feature point using spatial proximity; S232, constructing an edge for each pair of adjacent anatomical feature points according to the proximity relationship, morphological features and gray value information, and calculating the allocation weight of each edge; S233. Establish a spinal graph structure based on the anatomical feature points and the constructed edges, and optimize the spinal graph structure through an optimization algorithm.

5. The three-dimensional spine model reconstruction method based on neural network and 3D network architecture according to claim 4, It is characterized in that The calculation formula for calculating the neighborhood set of each anatomical feature point using spatial proximity is: N 6 (v i )={v j ∈V|||v i -v j || 1 =1}; Where N 6 (v i ) represents the neighborhood set of the i-th anatomical feature point in the anatomical feature point set; V represents the set of all anatomical feature points; ||.|| 1 represents the Manhattan distance; v i and v j They respectively represent the i-th anatomical feature point and the j-th anatomical feature point in the anatomical feature point set.

6. The three-dimensional spine model reconstruction method based on neural network and 3D network architecture according to claim 4, It is characterized in that The calculation formula for calculating the allocation weight of each edge is: w ij =α·d ij +b·m ij +γ·s ij ; In the formula, w ij Represents the anatomical feature point v i and v j The weight of the edge between them; d ij Represents the anatomical feature point v i and v j The spatial distance weight between them; m ij Represents the anatomical feature point v i and v j The weight of morphological features between ; s ij Represents the anatomical feature point v i and v j The grayscale feature weight between ; α, β, and γ all represent weight coefficients.

7. The three-dimensional spine model reconstruction method based on neural network and 3D network architecture according to claim 4, It is characterized in that The steps of establishing a spinal graph structure according to the anatomical feature points and the constructed edges, and optimizing the spinal graph structure by an optimization algorithm include the following steps: S2331. Sort the edges of the spine graph structure based on the weights of the edges; S2332, taking the edge with the smallest weight as the initial point, using the Kruskal algorithm to add edges one by one to the preset spanning tree, to obtain a minimum spanning tree; S2333, removing the edges that are not in the minimum spanning tree to obtain a primary optimized spine graph structure; S2334, based on the primary optimized spinal graph structure, construct a similarity matrix according to the attributes of the anatomical feature points, and calculate a distance matrix between the anatomical feature points; S2335. Calculate the graph Laplacian matrix based on the distance matrix, and cluster the anatomical feature points through K-means to obtain the final spinal graph structure.

8. The three-dimensional spine model reconstruction method based on neural network and 3D network architecture according to claim 7, It is characterized in that The method of performing feature processing based on the spinal graph structure and using the 3DU-Net architecture integrated with the graph neural network to obtain a preliminary three-dimensional spinal reconstruction model includes the following steps: S31, using a 3D convolutional layer to extract features from the spine graph structure to obtain a multi-channel feature graph containing spatial context information; S32. Use the transposed convolution layer to upsample and fuse the multi-channel feature map to obtain 3D volume data.

9. The method for reconstructing a three-dimensional spine model based on a neural network and a 3D network architecture according to claim 1, It is characterized in that The segmentation and post-processing of the preliminary three-dimensional spine reconstruction model to obtain the final three-dimensional spine reconstruction model includes the following steps: S41, constructing nodes from the obtained 3D volume data, establishing node edges based on spatial connectivity and feature similarity, and constructing an optimized graph model based on the nodes and node edges; S42, using the maximum flow-minimum cut algorithm to segment the optimized graph model to obtain a segmentation result; S43. Post-process and optimize the segmentation result using a boundary refinement method to obtain the final three-dimensional spine reconstruction result.

10. The three-dimensional spine model reconstruction method based on neural network and 3D network architecture according to claim 9, It is characterized in that The post-processing optimization of the segmentation result by using the boundary refinement method to obtain the final three-dimensional spine reconstruction result includes the following steps: S431, extracting the edge of the segmentation result using an edge detection algorithm; S432, performing preliminary edge thinning processing on the extracted edge using a region growing algorithm; S433, performing secondary edge thinning processing on the preliminary boundary thinning processing result by using morphological operation; S434, smoothing the secondary edge thinning result to obtain the final three-dimensional spine reconstruction result.

Images