Intelligent recognition method and system for spinal anatomical landmarks based on three-dimensional model

Abstract

The application discloses a spine anatomical landmark intelligent recognition method and system based on a three-dimensional model, relates to the field of anatomical landmark recognition, and first regards each vertebral body as a bottom layer graph composed of point clouds, and is used for finely depicting the internal geometric shape. Then, the centroids of all the vertebral bodies are regarded as nodes, and a high layer graph is constructed. Afterwards, local features are extracted on the bottom layer graph, and the local features are input into the high layer graph as initial information to perform context information propagation and enhancement, and finally, the vertebral body features fused with global structure information are obtained. When performing final landmark point judgment, the fine local geometric features of the points and the global enhanced vertebral body features are spliced to perform landmark recognition. The local decision mechanism under the global guidance effectively utilizes the anatomical prior knowledge, and solves the problems of poor robustness and anatomical fallacy caused by the lack of global context in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of anatomical landmark recognition, and more particularly, to a spinal anatomical landmark intelligent recognition method and system based on a three-dimensional model. BACKGROUND

[0002] The spine is the central axial skeleton of the human body, and its complex anatomical structure is the basis for clinical diagnosis, surgical planning, and biomechanical research. In these medical activities, it is crucial to accurately identify and locate key anatomical landmark points on the spine, such as vertebral body corner points, articular process tips, and spinous process tips. Traditionally, this work relies on manual operation by doctors on medical images such as CT and MRI, which is not only time-consuming and labor-intensive, but also inefficient, and the results are easily influenced by subjective experience, making it difficult to achieve standardization. Therefore, developing an intelligent method that can automatically and accurately identify spinal anatomical landmarks from three-dimensional models is of great significance to improve medical efficiency and accuracy, and is a technical demand that needs to be addressed in the field of smart medicine.

[0003] To improve the recognition accuracy of spinal landmark points, relevant research has been conducted in the industry. Early methods mainly rely on image processing and traditional machine learning, but when dealing with individual differences and pathological changes of the spine, they generally have poor robustness and insufficient accuracy. In recent years, deep learning, especially the point cloud processing network represented by PointNet++, has provided strong support for extracting local geometric features from three-dimensional point clouds. This type of method can effectively capture the micro-morphological features of landmark points such as spinous process tips through hierarchical local feature aggregation, significantly improving recognition accuracy. However, existing technologies still have obvious limitations: most methods treat a single vertebral body or the entire spine as an isolated point cloud, lacking a macroscopic understanding of the overall structure. When judging landmark points, the model only relies on local neighborhood information, ignoring the topological relationships and spatial position constraints between vertebral bodies. This processing approach leads to two key problems: first, the lack of global context information makes it difficult to utilize anatomical prior knowledge such as L4 being located between L3 and L5; second, in the face of severe pathological changes, scanning artifacts, or data missing, local features may be ambiguous or distorted, causing the model to make false judgments, such as misidentifying L2 as L3, which violates anatomical common sense and severely affects the robustness and anatomical consistency of the recognition results. Therefore, it is urgent to introduce overall spinal structure information based on local feature recognition to improve recognition accuracy and medical rationality.

[0004] Therefore, how to effectively fuse the local geometric details of vertebral body point clouds and the global structure context of the entire spine, so that the model can both microscopically and macroscopically, has become a key breakthrough in improving the performance of spinal anatomical landmark intelligent recognition. SUMMARY

[0005] To overcome the above-mentioned problems presented in the background art, according to an aspect of the present application, a three-dimensional model-based spinal anatomical landmark intelligent recognition method is provided, which comprises:

[0006] obtaining original CT three-dimensional image data;

[0007] preprocessing and roughly segmenting the original CT three-dimensional image data to obtain a set of vertebral clusters;

[0008] constructing a hierarchical graph for the set of vertebral clusters to obtain a set of point-level bottom graphs and an intervertebral relationship graph;

[0009] extracting bottom graph features from each point-level bottom graph in the set of point-level bottom graphs to obtain a set of vertebral shape descriptor vectors;

[0010] using each vertebral shape descriptor vector in the set of vertebral shape descriptor vectors as an initial feature of each node in the intervertebral relationship graph, inputting the initial feature into a graph convolution network-based high-level context encoder to obtain a set of context-enhanced vertebral feature vectors;

[0011] extracting a local geometric feature of a jth point of an ith point-level bottom graph from the set of point-level bottom graphs, and concatenating the local geometric feature with a context-enhanced vertebral feature vector of a vertebral body to which the jth point belongs to obtain a target point global-local concatenated feature vector;

[0012] inputting the target point global-local concatenated feature vector into a classification head to determine whether the target point belongs to a spinal anatomical landmark.

[0013] According to another aspect of the present application, a three-dimensional model-based spinal anatomical landmark intelligent recognition system is provided, which comprises:

[0014] a CT three-dimensional image acquisition module for obtaining original CT three-dimensional image data;

[0015] a CT three-dimensional image preprocessing and segmentation module for preprocessing and roughly segmenting the original CT three-dimensional image data to obtain a set of vertebral clusters;

[0016] a hierarchical graph construction module for constructing a hierarchical graph for the set of vertebral clusters to obtain a set of point-level bottom graphs and an intervertebral relationship graph;

[0017] a bottom graph feature extraction module for extracting bottom graph features from each point-level bottom graph in the set of point-level bottom graphs to obtain a set of vertebral shape descriptor vectors;

[0018] a vertebra feature context enhancement module, configured to input each vertebra shape descriptor vector in the set of vertebra shape descriptor vectors as initial features of each node in the intervertebral relationship graph, and input the initial features into a graph convolution network-based high-level context encoder to obtain a set of context-enhanced vertebra feature vectors;

[0019] a target point splicing module, configured to extract a vertebra point cloud local geometric feature of a jth point of an ith point-level bottom graph from the set of point-level bottom graphs, and splice the vertebra point cloud local geometric feature with a context-enhanced vertebra feature vector of a vertebra to which the jth point belongs to obtain a target point global-local splicing feature vector;

[0020] a landmark determination module, configured to input the target point global-local splicing feature vector into a classification head to determine whether the target point belongs to a spinal column anatomical landmark.

[0021] Compared with the prior art, the spinal column anatomical landmark intelligent identification method and system based on a three-dimensional model provided by the present application can simultaneously capture local geometric details and global structural context of a spinal column by constructing a hierarchical graph neural network framework. Specifically, each vertebra is first regarded as a bottom graph composed of point clouds, which is used to finely depict the geometric shape inside the vertebra. Then, the centroids of all the vertebrae are abstracted as nodes to construct a high-level graph, which explicitly encodes the spatial adjacency relationship between the vertebrae. Local features are extracted on the bottom graph and input as initial information into the high-level graph for context information propagation and enhancement, and finally vertebra features fused with global structural information are obtained. When making a final landmark point judgment, the fine local geometric features of a point are spliced with the context-enhanced vertebra features of the vertebra to which the point belongs. This local decision-making mechanism guided by global context effectively utilizes anatomical prior knowledge, and solves the problems of poor robustness and anatomical paradox caused by the lack of global context in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0022] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application taken in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of embodiments of the present application and are incorporated in and constitute a part of this specification, illustrate embodiments of the present application and serve to explain the present application, but do not limit the present application. In the drawings, like reference numerals refer to like parts or steps throughout.

[0023] Figure 1 A flowchart of the spinal column anatomical landmark intelligent identification method based on a three-dimensional model according to an embodiment of the present application.

[0024] Figure 2 A data flow schematic diagram of the spinal column anatomical landmark intelligent identification method based on a three-dimensional model according to an embodiment of the present application.

[0025] Figure 3 Flow chart of S2 in the method for intelligent recognition of spinal anatomical landmarks based on a three-dimensional model according to an embodiment of the present application.

[0026] Figure 4 Flow chart of S3 in the method for intelligent recognition of spinal anatomical landmarks based on a three-dimensional model according to an embodiment of the present application.

[0027] Figure 5 Block diagram of the system for intelligent recognition of spinal anatomical landmarks based on a three-dimensional model according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] Embodiments of the present disclosure will be described in more detail with reference to the drawings. Although certain embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be implemented in various forms and should not be interpreted as being limited to the embodiments set forth herein, but rather these embodiments are provided so as to more thoroughly and completely understand the present disclosure. It is understood that the drawings and embodiments of the present disclosure are for exemplary purposes only and are not intended to limit the scope of protection of the present disclosure.

[0029] In view of the problems in the foregoing background art description, the present application proposes a method for intelligent recognition of spinal anatomical landmarks based on a three-dimensional model. Figure 1 Flow chart of the method for intelligent recognition of spinal anatomical landmarks based on a three-dimensional model according to an embodiment of the present application. Figure 2 Data flow schematic diagram of the method for intelligent recognition of spinal anatomical landmarks based on a three-dimensional model according to an embodiment of the present application. As shown in Figure 1 and Figure 2 The method for intelligent recognition of spinal anatomical landmarks based on a three-dimensional model according to an embodiment of the present application includes: S1, obtaining original CT three-dimensional image data; S2, pre-processing and rough segmentation of the original CT three-dimensional image data to obtain a set of vertebral clusters; S3, hierarchical graph construction of the set of vertebral clusters to obtain a set of point-level bottom graphs and an intervertebral relationship graph; S4, bottom graph feature extraction of each point-level bottom graph in the set of point-level bottom graphs to obtain a set of vertebral shape descriptor vectors; S5, taking each vertebral shape descriptor vector in the set of vertebral shape descriptor vectors as an initial feature of each node in the intervertebral relationship graph, and inputting it into a high-level context encoder based on a graph convolution network to obtain a set of context-enhanced vertebral feature vectors; S6, extracting a local geometric feature of a jth point of an ith point-level bottom graph from the set of point-level bottom graphs, and splicing it with the context-enhanced vertebral feature vector of the vertebral body to which it belongs to obtain a target point global-local splicing feature vector; and S7, inputting the target point global-local splicing feature vector into a classification head to determine whether the target point belongs to a spinal anatomical landmark.

[0030] In S1, the original CT three-dimensional image data is acquired. It should be understood that in clinical practice, computer tomography (CT) technology has become a standard imaging method for the diagnosis and evaluation of spinal diseases due to its high resolution and excellent bone tissue imaging capability. CT scanning can generate a series of continuous two-dimensional tomographic images, which are stacked to form digital image data containing complete three-dimensional spatial information. These data not only accurately record the morphology and bone density of each vertebral body, but also retain the sequence relationship and spatial position of the entire spine. Therefore, in order to realize the automatic and accurate identification of the anatomical landmarks of the spine in three-dimensional space, the most original and complete information data source, i.e. the original CT three-dimensional image data, needs to be obtained from the medical imaging system first, providing a solid foundation for subsequent three-dimensional model reconstruction, feature extraction and intelligent analysis.

[0031] In one feasible embodiment of the present application, the specific process of S1 is as follows: by specifying the DICOM file path, the entire sequence folder containing the target patient's spine scan stored in the Picture Archiving and Communication System (PACS) or local workstation is read. These data follow the Digital Imaging and Communications in Medicine (DICOM) standard format, which is represented as a series of two-dimensional slice image files, each file corresponding to a scanning layer. During the reading process, by parsing the metadata of each DICOM file, key information such as window width and window level for subsequent display, and crucial geometric information such as slice thickness, pixel spacing, and image position / direction are extracted. These geometric information defines the actual size and coordinate relationship of voxels in three-dimensional physical space.

[0032] Subsequently, all two-dimensional slice images are accurately sorted according to the slice position information in the metadata, usually along the head-foot direction (e.g. Z-axis). After sorting, all two-dimensional image matrices are stacked in order to form a complete three-dimensional volume data. For example, if a lumbar CT scan contains 200 slices, each with a resolution of 512x512 pixels, a three-dimensional array with dimensions (200, 512, 512) will be generated, where each element of the array represents a voxel, and each value is the X-ray absorption density at the corresponding spatial position point of the human body, i.e. Hounsfield Unit HU value. This three-dimensional array is the original CT three-dimensional image data, which completely and digitally reproduces the three-dimensional anatomical structure of the patient's spine.

[0033] In S2, the raw CT 3D image data is preprocessed and coarsely segmented to obtain vertebral cluster sets. Correspondingly, although the raw CT 3D image data completely records the human structure, it contains a large amount of information about soft tissues and organs unrelated to the spine, and the voxel spacing obtained from scanning with different devices varies. This brings a huge computational burden and uncertainty to subsequent feature extraction and model analysis. Directly performing detailed anatomical landmark identification on such complex and heterogeneous data is inefficient and highly susceptible to noise interference. Therefore, in order to standardize and refine the raw data, eliminate irrelevant background interference, and preprocess and coarsely segment the raw CT 3D image data to decompose the continuous spinal structure into independent, anatomically significant vertebral units, thereby providing a regular and clear input object for subsequent construction of a layered image model.

[0034] In one feasible embodiment of this application, Figure 3 This is a flowchart of step S2 in the intelligent recognition method for spinal anatomical landmarks based on a three-dimensional model according to an embodiment of this application. Figure 3 As shown, S2, preprocessing and coarse segmentation of the original CT three-dimensional image data to obtain a set of vertebral clusters, includes: S21, voxel resampling of the original CT three-dimensional image data to obtain resampled CT three-dimensional image data; S22, bone region extraction and spinal region segmentation of the resampled CT three-dimensional image data to obtain a spinal point cloud; S23, coarse segmentation of the spinal point cloud into vertebral bodies to obtain N point cloud clusters belonging to N vertebral bodies as the set of vertebral clusters.

[0035] In the above feasible embodiment, the specific procedure of S2 is as follows: in S21. It should be understood that in a clinical environment, the original three-dimensional image data generated from different medical institutions and different types of CT scanning devices has significant differences in physical resolution. Specifically, the physical size of the voxels in three dimensions (i.e., pixel spacing and slice thickness) varies, resulting in heterogeneous data. If these data with different spatial scales are directly input into the subsequent deep learning model for analysis, the model will have difficulty learning anatomical structure features with generalization ability and scale invariance, thereby seriously affecting the accuracy and stability of segmentation and recognition. Therefore, in order to standardize the space of the original data and eliminate the scale inconsistency caused by device differences, voxel resampling is required to ensure that all data are processed under a unified physical space metric, providing a homogeneous data basis for the stable operation of subsequent algorithms. Specifically, first, a target voxel spacing is determined. The target value is pre-set to balance computational efficiency and detail retention. For example, the target spacing can be set to 1.0 mm x 1.0 mm x 1.0 mm, which is isotropic. This value is selected based on experience or experiment to ensure that sufficient details required for identifying anatomical landmarks are retained, and that the subsequent computational load is not too large due to excessive resolution. Subsequently, according to the physical size of the original data and the target spacing, the dimensions of the new three-dimensional array after resampling are calculated. For example, an original image data with a size of 358.4 mm x 358.4 mm x 300 mm (corresponding to 512 x 512 x 200 voxels, with a spacing of 0.7 x 0.7 x 1.5 mm) will have a new dimension of approximately 358 x 358 x 300 after resampling to an isotropic spacing of 1.0 mm. Next, the new three-dimensional array is constructed. For each target voxel coordinate in the new array, for example, coordinate (i, j, k), the system maps it back to the physical coordinate system of the original data. Since the new and old coordinate grids cannot be perfectly aligned, the mapped coordinate may fall between the voxels of the original data grid, rather than exactly at the center of a certain original voxel. In order to determine the Hounsfield Unit (HU) value of this new voxel, a trilinear interpolation algorithm is used. The algorithm finds the nearest eight neighbor voxels around the mapping point in the original data grid. Then, according to the relative distance of the mapping point to the centers of the eight neighbor voxels, the HU values of the eight neighbor voxels are weighted and averaged, and the neighbor voxels with closer distances have larger weights. Specifically, the process first performs linear interpolation twice in the X direction to obtain two intermediate values; then performs linear interpolation on the two intermediate values in the Y direction to obtain a final intermediate value; and finally performs linear interpolation in the Z direction to calculate the final HU value of the target voxel. By repeating this interpolation process for each coordinate point in the new array, a complete and size-regularized resampled CT three-dimensional image data is generated.

[0036] In S22, accordingly, after the voxel resampling, the CT three-dimensional image data is unified in the spatial scale, but its content is still a mixture containing various human tissues such as muscles, fats, organs and the like. The information of these soft tissues irrelevant to the target spine constitutes a large amount of background noise, which will seriously interfere with the subsequent segmentation and recognition of the vertebral body, a specific bone structure. In order to focus on the real analysis object and greatly reduce the computational complexity and memory consumption of the subsequent processing, the application accurately separates the region of interest from the complex whole-body image through bone region extraction and spinal region segmentation, that is, only the bone structure of the spinal part is reserved, providing a pure and clear data subset for the subsequent vertebral level fine segmentation. The specific implementation process is divided into two main stages: bone region extraction and spinal region segmentation. First, bone region extraction is performed. This stage takes advantage of the physical property that bone tissue has a high density value in CT images. The system sets a preset Hounsfield Unit (HU) threshold to distinguish between bone and soft tissue. The setting of the threshold is based on clinical experience, and a value is selected that can effectively include most of the bone while excluding most of the soft tissue, for example, the threshold can be set to 200 HU. All voxels of the resampled CT three-dimensional image data are traversed, and the voxels with a HU value greater than or equal to 200 are marked as foreground (bone), and the voxels with a value less than 200 are marked as background. This operation generates a binary three-dimensional mask, where the region with a value of 1 represents all the bone tissue preliminarily screened out, which may include the spine, part of the ribs, the pelvis and the like. Then, spinal region segmentation is performed. On the binary mask containing multiple bone parts obtained in the previous step, a three-dimensional connected component analysis algorithm is applied. This algorithm identifies all the connected foreground voxel blocks in the mask and labels each block as an independent connected component. Then, the volume of each connected component is calculated, that is, the number of voxels contained in each component is counted. In the chest and abdominal and waist regions of the human body, the spine is usually the largest and most continuous single bone structure in volume. Based on this anatomical prior knowledge, the largest connected component is automatically selected and retained, and all other smaller connected components such as disconnected rib fragments, iliac bones and the like are removed. Finally, this retained connected component representing only the complete spinal structure is converted from its voxel mask form to point cloud form. The center three-dimensional coordinates (x, y, z) of each voxel with a value of 1 in the spinal mask are extracted to form a coordinate point list. This set consisting of a large number of three-dimensional coordinate points is the final output spinal point cloud, which accurately describes the geometric shape and internal structure of the spine in the form of discrete points.

[0037] In S23, that is, after the pre-processing, the pure spine point cloud is obtained, which represents the whole spine as a continuous whole structure. However, in subsequent hierarchical graph model construction, the global structure of the spine (intervertebral relationship) needs to be separated from the local geometry (single vertebra inside) for processing. If not segmented, the subsequent algorithm will have difficulty distinguishing whether a point cloud belongs to the L3 vertebra or the L4 vertebra, and thus cannot construct a high-level graph describing the intervertebral relationship. Therefore, this step is to decompose the continuous spine point cloud into multiple independent vertebra units with clear anatomical identity, providing basic, discrete graph node elements for subsequent construction of point-level bottom graph and intervertebral relationship graph. In an embodiment of the present application, S23, the vertebra rough segmentation is performed on the spine point cloud to obtain N point cloud clusters belonging to N vertebrae as the vertebra cluster set, comprising: inputting the spine point cloud into a pre-trained three-dimensional convolutional neural network model to obtain the vertebra cluster set. It is worth mentioning that the spine and vertebra are anatomical structures with highly complex three-dimensional morphology, and their recognition depends not only on the information on a single two-dimensional slice, but more importantly, the continuity, shape and contextual relationship in three-dimensional space need to be considered. The three-dimensional convolutional neural network can directly operate on the three-dimensional voxel data through its unique three-dimensional convolution kernel, effectively capturing the feature dependency along the three spatial axes. This enables the model to learn the three-dimensional spatial configuration of complex components such as the height, curvature of the vertebra, and the pedicle, articular process, etc. Compared with traditional methods based on two-dimensional image processing or simple geometric rules, the three-dimensional convolutional neural network can automatically learn more robust and generalized vertebra shape features from large-scale data, so that it can still achieve high-precision and anatomically logical instance segmentation when facing individual differences, slight lesions or imaging artifacts, providing a reliable basis for subsequent fine analysis. Specifically, first, the input spine point cloud is converted back to its corresponding voxel grid representation in three-dimensional space. That is, each point cloud coordinate is mapped back to its corresponding voxel position, forming a binary three-dimensional array of the same size as the original resampled data, containing only the spine region, i.e. the spine voxel data, as the input of the neural network. Then, the three-dimensional array is input into a pre-trained three-dimensional convolutional neural network model. The model adopts a three-dimensional U-Net-based instance segmentation architecture. This architecture consists of an encoder path and a decoder path. The encoder path is responsible for extracting multi-level features from the input spine voxel data. It is composed of a series of repeated modules, each containing two consecutive three-dimensional convolution layers, for example using a 3x3x3 convolution kernel, followed by a rectified linear unit (ReLU) activation function, and finally a three-dimensional max pooling layer, for example using a 2x2x2 step to downsample the feature map.Through this encoding process, the model gradually reduces the spatial dimension of the feature map while increasing the number of feature channels, thereby capturing semantic information at different scales from low-level edges and textures to high-level shapes and components, forming a high-level abstract understanding of the spine structure. The decoder path is responsible for mapping the abstract features extracted by the encoder back to the original voxel space step by step to achieve accurate voxel-level classification. It restores the spatial resolution through a series of upsampling layers such as three-dimensional transpose convolution, and after each upsampling step, it splices the feature map obtained by upsampling with the feature map at the corresponding level in the encoder path through a jump connection. This allows the decoder to utilize both high-level semantic information and low-level detailed information when reconstructing the segmentation map. The last layer of the decoder path is a 1x1x1 three-dimensional convolution layer, with the number of output channels equal to the total number of classes to be identified, for example, background + 24 vertebrae = 25 classes. Finally, a Softmax activation function is used to predict the probability of each input spine voxel belonging to each vertebra class. In particular, the model's parameters such as weights and biases are obtained in advance through supervised training on a large-scale CT dataset accurately labeled by radiologists for each vertebra instance. In implementation, the model performs forward propagation on the spine voxel data and outputs a probability map with the same size as the input. According to the probability map, each voxel is assigned to the vertebra class with the highest probability. Finally, all voxel coordinate points with the same vertebra label, such as label L3, are collected to form an independent point cloud cluster. Repeat this process for all vertebra labels (L1, L2, L3...) to obtain N separate point cloud clusters, which together form the final vertebra cluster set.

[0038] In S3, the vertebra cluster set is constructed into a hierarchical graph to obtain a set of point-level bottom graphs and an inter-vertebra relationship graph. It can be understood that after the previous steps, the original CT image data has been refined into a series of independent vertebra point cloud clusters. However, these point cloud clusters are still unstructured data sets, without explicitly expressing the local geometric relationship between points within a single vertebra, nor reflecting the global topological connection between vertebrae. In order to enable the model to both microscopically identify the fine structure on the vertebra and macroscopically understand the arrangement order of the entire spine, the present application organizes the discrete vertebra point cloud data into a higher-level, more informative hierarchical graph structure through hierarchical graph construction, thereby providing an ideal data carrier that can simultaneously carry local geometric details and global context information for subsequent graph neural network processing.

[0039] In one feasible embodiment of the present application, Figure 4 The flowchart of S3 in the three-dimensional model-based spinal anatomical landmark intelligent recognition method according to the embodiment of the present application. As shown in Figure 4As shown, S3, the vertebra cluster set is hierarchically graph constructed to obtain a set of point-level bottom graphs and an inter-vertebra relationship graph, comprising: S31, a point-level bottom graph is constructed for each vertebra cluster in the vertebra cluster set by farthest point sampling and K-neighbor algorithm; S32, the centroid of each vertebra cluster in the vertebra cluster set is calculated, and the centroid of each vertebra cluster is taken as a node of the inter-vertebra relationship graph; S33, if the centroid distance between the centroids of two vertebra clusters is less than a preset threshold, an edge between the corresponding two nodes in the inter-vertebra relationship graph is established.

[0040] In the above feasible embodiment, the specific process of S3 is as follows: in S31. It should be understood that although the original point cloud data contains rich geometric information, it lacks a clear structure to describe the local adjacency relationship between points. Directly processing such a large and unstructured data not only has low computational efficiency, but also makes it difficult for the model to effectively learn the fine geometric features that are crucial to identifying anatomical landmarks, such as the curvature of the pedicle surface and the sharpness of the joint process. Therefore, in order to convert each unordered vertebral point cloud cluster into a structured graph representation, the application needs to first construct a point-level bottom graph to explicitly define nodes and edges, capture and encode the local geometric topology inside the vertebral body. Specifically, taking one of the point cloud clusters representing the L3 vertebra as an example, first, the input L3 vertebra point cloud cluster, such as containing 20,000 points, is selected for node selection. In order to effectively reduce the data amount while preserving the overall shape of the vertebral body, the farthest point sampling algorithm is adopted. This algorithm first randomly selects a point from the point cloud cluster as the first node. Then, the next node is selected iteratively, each time selecting the point farthest from the set of currently selected nodes. This process is repeated until a predetermined number of nodes are selected. For example, the predetermined number of sampling points is 1024. In this way, the 1024 selected nodes can be evenly distributed on the surface of the entire L3 vertebra, forming a good skeleton representation of the shape of the vertebral body, and these 1024 points constitute the node set of the point-level bottom graph. After determining the nodes of the graph, edges are then constructed for these nodes to describe their local connection relationship. The K-neighbor algorithm is used to achieve this. For each of the 1024 nodes just selected, in the complete L3 vertebra point cloud cluster containing 20,000 points, the Euclidean distance of all points to this node is calculated, and the K nearest points are found. Here, K is a predetermined parameter determined by cross-validation experiments on the validation dataset, representing the size of the neighborhood, for example, it can be set to 16. The selection of this K value aims to capture sufficient local geometric information while avoiding the introduction of too many irrelevant connections. Subsequently, in the point-level bottom graph, edges are established between each center point and its found K neighbors. In constructing the graph, connections are usually only established between the 1024 nodes selected by the farthest point sampling. A specific approach is as follows: for each center point, only those points that also belong to the 1024 sampling nodes are selected from its K neighbors to establish edges. Another more common approach is to regard the K neighbor points as a local patch, and subsequent feature extraction will be based on the local region consisting of the center point and its K neighbors. In this scheme, edges are established between the 1024 nodes. The system will find the K nearest nodes in the remaining 1023 nodes for each node, and establish connections. By performing the above farthest point sampling and K-neighbor algorithm on the L3 vertebra cluster, a graph structure containing 1024 nodes and corresponding edges, i.e. the topology of the graph, is constructed.After the topology of the graph is defined, the initial feature of each node needs to be calculated and assigned, which is the local geometry feature of the posterior lumbar vertebra point cloud. For the jth node of the 1024 nodes, return to the original complete L3 vertebra point cloud cluster containing 20000 points, find a local neighborhood around the node (for example, its nearest 32 neighbor points). Then, the coordinates of the 32 neighbor points are normalized with respect to the coordinates of the center node j to form a coordinate matrix describing the local geometry. The matrix is input into a small, weight-shared neural network module (such as a multilayer perceptron composed of several fully connected layers), which encodes this local shape information into a fixed-dimensional high-dimensional vector (for example, 64 dimensions). The vector is the local geometry feature of the vertebra point cloud of node j, which condenses the curvature, normal direction and other fine geometric information around the point. By performing this operation on the 1024 nodes, each node is assigned a feature vector describing its local environment. Finally, by repeating this entire process for each vertebra cluster in the vertebra cluster set, a set of point-level bottom graphs is generated, each of which not only defines the connection relationship between nodes, but also carries rich local geometry features for subsequent network layer processing.

[0041] In S32, accordingly, after processing the internal geometry (bottom graph) of a single vertebra, it is necessary to turn to a higher level, i.e. the relationship between vertebrae. Directly using the complete vertebra point cloud containing tens of thousands of points to represent the relationship between vertebrae is not only extremely complex in calculation, but also full of redundant information. In order to represent and reason about the spatial arrangement between vertebrae in a concise, efficient and anatomically meaningful way, the present application abstracts and simplifies each complex vertebra point cloud cluster into a single representative entity, i.e. its geometric center as a proxy. Specifically, each vertebra point cloud cluster in the vertebra cluster set is traversed. Taking a point cloud cluster representing an L3 vertebra as an example, the point cloud cluster is composed of M points, and the three-dimensional coordinates of each point are After that, the centroid C_L3 of the point cloud cluster is calculated. The method of calculating the centroid is to calculate the arithmetic mean of the three-dimensional coordinates of all points in the cluster. The specific calculation formula is: where the summation symbol Σ indicates that the summation is performed over all points from i = 1 to M. For example, if the L3 vertebral body point cloud cluster contains 20,000 points, the x, y, and z coordinates of these 20,000 points are added separately, and then the three sums are divided by 20,000, respectively, to obtain the centroid coordinates (Cx, Cy, Cz) of the L3 vertebral body. This centroid point can be regarded as the geometric center of the vertebral body in the physical space, which succinctly summarizes the position information of the entire vertebral body in the space with a single three-dimensional coordinate point. By repeating the above centroid calculation process for all N vertebral body clusters in the vertebral body cluster set, N centroid coordinate points are obtained. After all the centroids are calculated, the nodes of the intervertebral relationship graph are constructed. Each calculated centroid is directly defined as a node in the graph. Therefore, the intervertebral relationship graph will contain N nodes, each of which not only stores its corresponding centroid three-dimensional coordinates in the data structure, but also retains its vertebral body identity such as L3.

[0042] In S33, that is, in order to let the model understand the key anatomical sequence relationship that L4 vertebrae is located below L3 vertebrae, a connection needs to be established between the nodes representing the vertebrae, so as to organize the discrete nodes into a graph that can reflect the topology of the whole spine. Therefore, by establishing edges, the connection relationship between the vertebrae (i.e. the edges of the graph) can be explicitly constructed according to the proximity of the vertebrae in the physical space, thereby completing the encoding of the global structural context of the spine. Specifically, first, a suitable preset threshold needs to be determined. This threshold is a key parameter for determining whether two vertebrae are directly adjacent. Its setting should be based on anatomical knowledge, that is, the approximate distance between the centers of adjacent vertebrae under normal circumstances. For example, by statistically analyzing a large number of lumbar CT data, it can be found that the distance between the centers of adjacent lumbar vertebrae is usually between 30 and 60 millimeters. Therefore, a relatively loose but effective threshold for excluding non-adjacent vertebrae connections can be set, for example, the preset threshold is set to 70 millimeters. After determining the threshold, all possible node pairs are traversed. For example, there are N nodes in the intervertebral relationship graph, the system will perform a combination traversal to check each pair of different nodes, for example, the relationship between node i and node j. For each pair of nodes, the coordinates of their respective centers of mass are extracted. Then, the Euclidean distance between the two centers of mass is calculated. Then, the calculated Euclidean distance is compared with the preset threshold of 70 millimeters. If the Euclidean distance is less than 70 millimeters, it is determined that the vertebrae represented by the two nodes are directly adjacent in space, and a directed edge is added between node i and node j in the data structure of the intervertebral relationship graph, indicating that there is a connection between them. If the Euclidean distance is greater than or equal to 70 millimeters, no connection is established between the two nodes. For example, when calculating the distance between the centers of mass of L3 and L4 vertebrae, the result may be 45 millimeters. Because 45 millimeters is less than 70 millimeters, a connection is established between the nodes representing L3 and L4. When calculating the distance between the centers of mass of L3 and L5, the result may be 95 millimeters, which is greater than the threshold, so no direct connection is established between L3 and L5. By repeating this process for all N(N-1) / 2 unique node pairs, the intervertebral relationship graph is constructed.

[0043] In S4, bottom-up feature extraction is performed on each of the set of point-level bottom-up maps to obtain a set of vertebral shape descriptor vectors. It should be appreciated that a point-level bottom-up map, although representing the geometric topology of a single vertebra in a structured manner, is still raw coordinates and connection relations and does not distill discriminative, high-level shape information. In order for the model to understand the complex morphology of the vertebrae, for example, to distinguish the smooth transition of the pedicle from the sharp morphology of the spinous process, meaningful features need to be learned and extracted from these bottom-up maps. More importantly, in order to perform context reasoning in the high-level map, a compact representation that summarizes the shape of the entire vertebra is needed. Therefore, in order to automatically learn and generate a condensed, fixed-dimensional feature vector, i.e., a vertebral shape descriptor, from the fine geometric structure of each vertebra, bottom-up feature extraction is performed on each point-level bottom-up map in the technical solution of the present application.

[0044] In a feasible embodiment of the present application, S4, bottom-up feature extraction is performed on each of the set of point-level bottom-up maps to obtain a set of vertebral shape descriptor vectors, including: S41, inputting the point-level bottom-up map into a bottom-up feature extractor based on a weight-sharing graph neural network to obtain a set of vertebral point cloud local geometric features; S42, performing graph pooling processing on the set of vertebral point cloud local geometric features to obtain the vertebral shape descriptor vector.

[0045] In the above feasible embodiment, the specific process of S4 is as follows: in S41. In the present application, the bottom-layer feature extractor based on weight-sharing graph neural network adopts a dynamic graph convolutional network as its core architecture. The network is stacked by multiple successive graph convolutional layers. Taking the point-level bottom-layer graph of L3 vertebra as an example, when it is input into the extractor, its 1024 nodes (each node feature is a 3-dimensional coordinate) first enter the first layer of graph convolutional layer. In this layer, for each center node, the network performs the following operations: first, in the current feature space (initially the coordinate space), dynamically find N nearest neighbor nodes (for example, the preset N = 20) for the center node. Then, calculate the difference vector between the center node feature and the feature of each neighbor node. These difference vectors and the feature vector of the center node itself are sent to a shared multi-layer perceptron (MLP). The MLP is composed of several fully connected layers and activation functions (such as ReLU), and its weight and bias parameters are shared for the neighborhood calculation of all nodes in the graph. This means that the network uses the same set of standards to learn how to extract geometric patterns from any local neighborhood. The output of the MLP is aggregated (such as max pooling) to form the updated feature vector of the center node in the first layer. This process is performed simultaneously for all 1024 nodes, resulting in 1024 new feature vectors with higher dimensions (such as 64 dimensions). These new feature vectors are then used as input for the next layer of graph convolutional layer, repeating the above dynamic neighbor search, feature transformation, and aggregation process. By stacking several such graph convolutional layers, the feature vector of each node can gradually and hierarchically aggregate neighborhood information of larger and larger range. The initial layer may learn simple edge or direction information, while deeper layers can learn more complex geometric patterns such as surfaces, concave, convex, etc. In particular, all weights and bias parameters of the bottom-layer feature extractor are learned together as part of the entire end-to-end recognition model, through backpropagation and gradient descent optimization on a large number of spine three-dimensional model datasets with joint landmark annotations. After processing by all graph convolutional layers, each node in the L3 vertebra bottom-layer graph obtains a final high-dimensional feature vector, which is the local geometric feature of the vertebra point cloud of the node, and contains rich geometric context information around the point.

[0046] In S42, a symmetric function is employed to perform the graph pooling operation to ensure that the result is independent of the order of input nodes. In this embodiment, max-pooling is employed as the symmetric function. As an example, consider the 1024 256-dimensional local geometric feature vectors of the L3 vertebral body. These vectors can be organized into a 1024 x 256 matrix. The max-pooling operation is performed along the first dimension of the matrix (i.e., the node dimension). Specifically, the system performs the operation independently on each column of the matrix (i.e., each dimension of the feature). For example, for the first dimension of the feature vectors, the system compares the values of this dimension of the 1024 vectors and selects the maximum value. Similarly, for the second dimension, the system finds the maximum value of the second dimension of the 1024 vectors. This process is repeated 256 times, covering all feature dimensions. The result of the 256 max-value operations, i.e., the maximum value of each dimension, collectively forms a new 256-dimensional vector. This vector is the vertebral body shape descriptor vector of the L3 vertebral body, which effectively condenses the most representative geometric feature information of the entire vertebral body point cloud by capturing the most significant activation value of each feature dimension. For example, if a certain feature dimension encodes sharpness, the value of this dimension in the pooled vector reflects the sharpness level of the most sharp region of the entire vertebral body. The complete process of S41 and S42 is performed independently and in parallel on each of the group of point-level base graphs. Since the base feature extractor is weight-shared, the same model is applied to all vertebral bodies, ensuring consistency in feature extraction. Finally, the system obtains N vertebral body shape descriptor vectors, which collectively form the set of vertebral body shape descriptor vectors.

[0047] In S5, each vertebral body shape descriptor vector in the set of vertebral body shape descriptor vectors is taken as the initial feature of each node in the intervertebral relationship graph, which is input into the graph convolution network-based high-level context encoder to obtain a set of context-enhanced vertebral body feature vectors. Accordingly, the vertebral body shape descriptor vectors are isolated descriptors that only contain the intrinsic information of a single vertebral body, but lack an understanding of the environment in which the vertebral body is located in the entire spine. In clinical practice, the identification of a vertebral body often needs to refer to its adjacent vertebral bodies, for example, to confirm that a vertebral body is L4, it is usually necessary to see that it is located below L3 and above L5. Without considering this context, the model is prone to confusion when facing morphologically similar vertebral bodies or abnormal morphologies caused by lesions. Therefore, in order to enable each vertebral body to see and integrate the features of its neighbor vertebral bodies, the present application enhances the isolated shape information to context-aware features that contain global anatomical sequence relationships, which are more robust and medically reasonable.

[0048] In one possible embodiment of the present application, the detailed procedure of S5 is as follows: First, each vector in the set of vertebra shape descriptor vectors is assigned to the corresponding node in the intervertebral graph as its initial feature. For example, the 256-dimensional shape descriptor vector of L3 vertebra is assigned to the node representing L3 in the intervertebral graph. After the assignment, each node in the intervertebral graph carries a high-dimensional feature describing its shape. Then, the intervertebral graph with initial node features is input into the high-level context encoder based on graph convolutional network, which adopts a standard graph convolutional network architecture. The architecture is stacked with several graph convolutional layers.

[0049] When the intervertebral graph enters the first layer of graph convolutional layer, for each node in the graph, such as the L3 node, the network performs a feature aggregation and update operation. Specifically, the network looks up the neighbor nodes (L2 and L4) of the L3 node that are directly connected to it in the graph (according to the edges constructed in S33). Then, the network collects the vertebra shape descriptor vector of the L3 node itself, as well as the vertebra shape descriptor vectors of all its neighbor nodes, i.e., L2 and L4. These vectors are combined by an aggregation function such as weighted average, then multiplied by a layer-specific, learnable weight matrix, and added to a bias term, and finally processed by a nonlinear activation function such as ReLU. The essence of this process is that the new feature of the L3 node is determined by its own old feature and the old features of its neighbors. Through this layer of operation, the feature vector of the L3 node has fused the shape information of L2 and L4, thereby obtaining the first level of context awareness. This process occurs in parallel on all nodes in the graph. The updated feature vector set is output and sent to the next graph convolutional layer, repeating the same information propagation process. By stacking multiple layers of graph convolutional layers, information can be propagated further on the graph. For example, after two layers of GCN, the feature of the L3 node not only contains the information of L2 and L4, but also indirectly contains the information of L1 and L5. In particular, all weight matrices and bias vectors of the high-level context encoder are learned together as part of the entire end-to-end recognition model, through backpropagation and gradient descent optimization on the labeled dataset. After processing by all graph convolutional layers, each node in the intervertebral graph obtains a final feature vector. This vector is the context-enhanced vertebra feature vector. For example, the final vector output by the L3 node not only encodes the fine shape of L3 itself, but also incorporates the positional context information of L3 in the entire spine sequence. The set of final feature vectors of all nodes is the set of context-enhanced vertebra feature vectors.

[0050] In S6, the vertebra point cloud local geometry feature of the jth point of the ith point-level bottom-up map is extracted from the set of point-level bottom-up maps, and is spliced with the context-enhanced vertebra feature vector of the vertebra to which the jth point belongs to obtain a global-local spliced feature vector of the target point. It can be understood that the present application has distilled two key pieces of information: one is the local feature describing the fine local geometry of any point on each vertebra, and the other is the global context feature summarizing the position relationship of each vertebra in the whole spine. However, to finally determine whether a specific point is an anatomical landmark point (for example, the tip of the right superior articular process of the L3 vertebra), only the local information or only the global information is not enough. The model needs to know both what the point itself looks like (local geometry) and which vertebra the point belongs to and where it is located in the spine (global context). Therefore, in order to effectively integrate the two pieces of information from different scales and different levels, the present application needs to concatenate the two pieces of information to create a comprehensive, multi-scale feature representation for each point that needs to be determined, which contains both microscopic details and macroscopic backgrounds.

[0051] In an embodiment of the present application, S6, the vertebra point cloud local geometry feature of the jth point of the ith point-level bottom-up map is extracted from the set of point-level bottom-up maps, and is spliced with the context-enhanced vertebra feature vector of the vertebra to which the jth point belongs to obtain a global-local spliced feature vector of the target point, comprising: concatenating the vertebra point cloud local geometry feature of the jth point of the ith point-level bottom-up map with the context-enhanced vertebra feature vector to obtain the global-local spliced feature vector of the target point.

[0052] In the above embodiment, the specific process of S6 is as follows: first, the local geometry information of the target point needs to be obtained. It will accurately index to the local feature set belonging to the L3 vertebra from the set of vertebra point cloud local geometry features, and extract the vertebra point cloud local geometry feature corresponding to the jth node from the set.

[0053] At the same time, the global context information of the vertebra to which the target point belongs needs to be obtained, and the context-enhanced vertebra feature vector representing the whole L3 vertebra is indexed from the set of context-enhanced vertebra feature vectors.

[0054] Then, a splicing operation, i.e., concatenation, is performed. The two vectors extracted above, the vertebra point cloud local geometry feature representing the local geometry of point j, and the context-enhanced vertebra feature vector representing the global context of the ith vertebra to which the point belongs, are concatenated head to tail along the feature dimension, and the dimension of the spliced vector is the global-local spliced feature vector of the target point j.

[0055] In particular, when concatenating the lower node features describing the local geometry of a point with the upper node features describing the global context of the vertebra, there is a potential technical problem: the global context information can over-influence or contaminate the fine local geometry information. For example, a non-landmark point located at the edge of the vertebra with a sharp shape, whose local features should be significantly different from those of the landmark points, can be weakened in its uniqueness if directly concatenated with strong global features, leading to misclassification by the classifier. Therefore, in order to maintain or even highlight the truly discriminative parts of the local features while fusing global context information, and prevent them from being assimilated or interfered in the concatenation process, the present application optimizes the topological association of the vertebra point cloud local geometry features of the jth point of the ith point-level bottom graph, and performs a delicate purification and strengthening of the local features, so that they can maintain distinctiveness and stability in subsequent concatenation.

[0056] Based on this, in one preferred embodiment of the present application, the vertebra point cloud local geometry features of the jth point of the ith point-level bottom graph are concatenated with the context-enhanced vertebra feature vector to obtain the target point global-local concatenation feature vector, which includes:

[0057] First, the cosine similarity between the vertebra point cloud local geometry features of all points of the ith point-level bottom graph and the context-enhanced vertebra feature vector is calculated to obtain a cosine similarity component vector, wherein the cosine similarity component vector has the same length as the context-enhanced vertebra feature vector, i.e.:

[0058]

[0059] wherein, is the context-enhanced vertebra feature vector, is the vertebra point cloud local geometry feature of the jth point, is the cosine similarity calculation, is and cosine similarity between them. It should be understood that before the feature optimization, it is necessary to quantify the degree of association between each local point feature and the global context feature of the vertebra to which it belongs. Without establishing such a quantitative relationship, the subsequent optimization will lack a clear direction and basis. Direct numerical differences do not have directionality, while cosine similarity can accurately measure the consistency of two feature vectors in direction, i.e., whether the geometric patterns they describe are similar. Thus, in order to establish a basis for measuring the inherent correlation between local and global features, the present application calculates the cosine similarity between all local geometric features of the vertebra point cloud and the context-enhanced vertebra feature vector. Since the number of local points and the dimensionality of the global feature are usually inconsistent, the multiple similarity values calculated are further integrated and adjusted into a cosine similarity composition vector with the same length as the context-enhanced vertebra feature vector by linear interpolation or local pooling transformation, etc. Each element in the vector quantitatively represents the alignment degree of the local point cloud as a whole with the global context, providing a key quantitative basis for subsequent construction of the correlation matrix.

[0060] Then, the correlation matrix of the cosine similarity composition vector and the context-enhanced vertebra feature vector is calculated, i.e.,

[0061]

[0062] wherein, is the cosine similarity composition vector, is matrix multiplication, is the vector transposition operation, the vector is in column vector form, is the correlation matrix. Accordingly, after obtaining the cosine similarity composition vector, it needs to be converted into a transformation operator that can operate on local features. Thus, in order to construct an operator that can encode global context information and local-global similarity into a unified transformation rule, the present application calculates an association matrix by taking the outer product of the cosine similarity composition vector and the context-enhanced vertebra feature vector. This association matrix is a rotation projection matrix, which contains all the information on how to project local features into a feature space related to the global context according to the similarity between local and global. It is a high-dimensional, structured topological correlation calibrator.

[0063] Next, the local geometric feature of the jth point of the vertebra point cloud is mapped to the correlation matrix to obtain the local geometric mapping feature vector of the jth point of the vertebra point cloud, i.e.,

[0064]

[0065] wherein, is the jthpoint's local geometry mapped feature vector of the octree point cloud. It should be understood that the correlation matrix itself is only a transformation rule that needs to be applied to the actual local feature to produce an effect. If the original local feature is directly used for subsequent attention calculation, it lacks alignment with the global context and cannot effectively evaluate its uniqueness. Therefore, in order to obtain a local feature aligned with the global context in order to evaluate its difference with the original feature, the present application generates an octree point cloud local geometry mapped feature vector by mapping the original local geometry feature to the correlation matrix. The vector can be understood as a projection or snapshot of the original local feature from the perspective of the global context.

[0066] Next, the jthpoint's local geometry feature of the octree point cloud is correlated and mapped to obtain the jthpoint's octree point cloud correlation and mapping mutual exclusion measure factor, i.e.

[0067]

[0068] wherein, represents the inner product of vectors and is the variance of all eigenvalues of and is the jthpoint's octree point cloud correlation and mapping mutual exclusion measure factor. Accordingly, after obtaining the original local feature and its mapped feature in the global perspective, a mechanism is needed to judge the uniqueness of the original feature. If a local feature is very different from its mapped feature in the global perspective (i.e. high mutual exclusion), it means that this local feature contains many unique details that are not included in the global information. Therefore, in order to quantitatively evaluate the uniqueness or significance of each local feature, the present application generates a mutual exclusion measure factor by correlation and mapping mutual exclusion measure. The larger the factor value, the more unique the local geometry feature of the point, the less likely it is to be summarized by the global context, and the more likely it is to be key discriminative information.

[0069] Then, the jthpoint's local geometry feature of the octree point cloud is correlated and mapped to obtain the jthpoint's octree point cloud correlation and mapping mutual exclusion measure factor, i.e.

[0070]

[0071] wherein, is the multiplication of the position point, ​​is the local geometry reinforced feature vector of the jthpoint of the point cloud of the vertebral body as the local geometry reinforced feature of the jthpoint of the point cloud of the vertebral body. It should be understood that the calculated mutual exclusivity measure factor provides a weight for feature reinforcement. Those local features with high uniqueness should be preserved and amplified, while common features highly consistent with global information can be moderately suppressed. Thus, in order to utilize the above-mentioned measure results to actively and selectively adjust the intensity of the original local features, the present application obtains the local geometry reinforced feature by using the element-by-element multiplication reinforcement of the mutual exclusivity measure factor on the original local features, in which the unique and significant geometric patterns (such as landmark point regions) are amplified, while the common and redundant patterns are suppressed, achieving the improvement of signal-to-noise ratio.

[0072] Finally, the local geometry reinforced feature vector of the jthpoint of the point cloud of the vertebral body is concatenated with the context enhanced vertebral feature vector to obtain the target point global-local spliced feature vector of the jthpoint. That is, after the above optimization, the local feature is ready for the final fusion with the global feature. At this time, the local feature has been purified, and its dominant characteristics have been highlighted. Thus, in order to finally form a comprehensive feature representation containing both macroscopic position information and optimized microscopic morphological information, the present application generates the final target point global-local spliced feature vector by concatenating the local geometry reinforced feature of the point cloud of the vertebral body with the context enhanced vertebral feature vector, which, due to the attention reinforcement of its local component, can provide a clearer and more robust discrimination signal when fed into the classifier, thereby significantly improving the accuracy of the anatomical landmark point recognition. In particular, the concatenation here and the head-to-tail splicing of the two features.

[0073] In S7, the target point global-local spliced feature vector is input into a classification head to determine whether the target point belongs to a spinal anatomical landmark. That is, the target point global-local spliced feature vector fuses microscopic geometric details with macroscopic anatomical context. This high-dimensional vector is a highly condensed and intelligent representation of the original three-dimensional data, but it is not the final discrimination result itself. In order to map this complex feature vector to a specific and meaningful anatomical category, for example, "right superior articular process tip" or "not any landmark point", a final decision module is needed. For this purpose, the present application inputs the target point global-local spliced feature vector into a classification head in order to utilize a specialized classifier to decode and judge this information-rich spliced feature, thereby making a final and deterministic classification of the target point.

[0074] In one feasible embodiment of the present application, the specific process of S7 is as follows: the global-local spliced feature vector of the target point is fed into a classification head module. The classification head in this embodiment is a multi-layer perceptron, MLP, composed of several fully connected layers.

[0075] For example, the classification head can be composed of two fully connected layers and a final output layer. The first fully connected layer receives the 512-dimensional global-local concatenated feature vector of the target point and linearly transforms it to a lower intermediate dimension, e.g., 256-dimensional. This transformation is controlled by a 512 x 256 weight matrix and a 256-dimensional bias vector. The transformed result is passed through a non-linear activation function, such as ReLU, to increase the expressive power of the model. Next, this 256-dimensional intermediate feature vector is fed into the second fully connected layer, again with linear transformation and non-linear activation, possibly to a lower dimension, e.g., 128-dimensional. Finally, this 128-dimensional vector is fed into the output layer. The output layer is a special fully connected layer whose number of neurons equals the total number of categories to be recognized. If the goal of the present invention is to recognize 13 different anatomical landmark points, such as the center of the superior and inferior endplate of the vertebral body, the tip of the left and right superior articular process, etc., plus one background category, i.e., non-landmark point, then the output layer will have 14 neurons. This layer does not use the ReLU activation function, but directly outputs raw score values. In order to get the probability of each category, these raw score values are processed by a Softmax function. The Softmax function can convert an arbitrary set of real numbers into a probability distribution, where each value is between 0 and 1, and the sum of all values is 1. For example, for the input target point, the Softmax function can output a vector containing 14 probability values, such as [0.01, 0.02, 0.9,..., 0.015].

[0076] After that, the category corresponding to the element with the largest value in the probability vector is selected as the final classification result. In the above example, the probability value of the third element, 0.9, is the highest, and if the third category is defined as the tip of the right superior articular process, then it is finally determined that the target point belongs to the anatomical landmark of the tip of the right superior articular process. In particular, all weight matrices and bias vectors of the classification head, like all learnable parameters in the previous steps, are jointly learned as part of the entire end-to-end recognition model, through backpropagation and gradient descent optimization on a large dataset of spine three-dimensional models with accurate anatomical landmark point annotations. By repeating this process for each point to be judged, the intelligent recognition of all anatomical landmark points of the entire spine is achieved, providing accurate, efficient, and standardized three-dimensional positioning basis for clinical diagnosis, surgical planning, and biomechanical analysis.

[0077] In summary, the three-dimensional model-based spinal anatomical landmark intelligent recognition method based on the embodiments of the present application is illustrated, which builds a hierarchical graph neural network framework to capture both local geometric details and global structural context of the spine. Specifically, first, each vertebral body is regarded as a bottom layer graph composed of point clouds, which is used to finely depict its internal geometry. Then, the centroids of all vertebral bodies are abstracted as nodes to build a high-level graph, which explicitly encodes the spatial adjacency relationship between the vertebral bodies. By extracting local features on the bottom layer graph and inputting them as initial information into the high-level graph for context information propagation and enhancement, the final vertebral body features fused with global structural information are obtained. In the final landmark point judgment, the fine local geometric features of the point and the context-enhanced vertebral body features are spliced. This global-guided local decision mechanism effectively utilizes anatomical prior knowledge, solving the problem of poor robustness and anatomical paradox caused by the lack of global context in the prior art.

[0078] Figure 5 The block diagram of the three-dimensional model-based spinal anatomical landmark intelligent recognition system according to the embodiments of the present application is shown. As shown in Figure 5 The three-dimensional model-based spinal anatomical landmark intelligent recognition system 100 according to the embodiments of the present application includes: a CT three-dimensional image acquisition module 110 for acquiring original CT three-dimensional image data; a CT three-dimensional image preprocessing and segmentation module 120 for preprocessing and roughly segmenting the original CT three-dimensional image data to obtain a vertebral body cluster set; a hierarchical graph construction module 130 for constructing a hierarchical graph on the vertebral body cluster set to obtain a set of point-level bottom layer graphs and an intervertebral relationship graph; a bottom layer graph feature extraction module 140 for extracting bottom layer graph features from each point-level bottom layer graph in the set of point-level bottom layer graphs to obtain a set of vertebral body shape descriptor vectors; a vertebral body feature context enhancement module 150 for inputting each vertebral body shape descriptor vector in the set of vertebral body shape descriptor vectors as the initial feature of each node in the intervertebral relationship graph into a high-level context encoder based on a graph convolution network to obtain a set of context-enhanced vertebral body feature vectors; a target point splicing module 160 for extracting the local geometric features of the jth point of the ith point-level bottom layer graph from the set of point-level bottom layer graphs, and splicing them with the context-enhanced vertebral body feature vector of the corresponding vertebral body to obtain a target point global-local splicing feature vector; and a landmark determination module 170 for inputting the target point global-local splicing feature vector into a classification head to determine whether the target point belongs to a spinal anatomical landmark.

[0079] Here, those skilled in the art can understand that the specific operations of each step in the above three-dimensional model-based spinal anatomical landmark intelligent recognition system have been described above with reference to Figures 1 to 4The description of the three-dimensional model-based spinal anatomic landmark intelligent recognition method of the present application is described in detail in the foregoing description, and thus, repetitive description thereof will be omitted.

Claims

1. A method for intelligent recognition of spinal anatomical landmarks based on a three-dimensional model, characterized in that, include: Acquire raw CT 3D image data; The original CT three-dimensional image data is preprocessed and coarsely segmented to obtain a set of vertebral clusters; A hierarchical graph is constructed from the vertebral cluster set to obtain a set of point-level bottom-level graphs and an inter-vertebral relationship graph; The bottom-level graph features of each point-level bottom-level graph in the set of point-level bottom-level graphs are extracted to obtain a set of cone shape descriptor vectors; Each vertebra shape descriptor vector in the set of vertebra shape descriptor vectors is used as the initial feature of each node in the intervertebral relationship graph, and then input into a high-level context encoder based on a graph convolutional network to obtain a set of context-enhanced vertebra feature vectors. Extracting the local geometric features of the cone point cloud of the j-th point in the i-th point-level bottom-level image from the set of point-level bottom-level images, and concatenating them with the context-enhanced cone feature vector of the corresponding cone to obtain the global-local concatenated feature vector of the target point, includes: calculating the cosine similarity between the local geometric features of the cone point cloud of all points in the i-th point-level bottom-level image and the context-enhanced cone feature vector to obtain a cosine similarity composition vector, wherein the cosine similarity composition vector and the context-enhanced cone feature vector have the same length; calculating the association matrix between the cosine similarity composition vector and the context-enhanced cone feature vector; and applying the local geometric features of the cone point cloud of the j-th point to the association matrix. The matrix is ​​mapped to obtain the local geometric mapping feature vector of the cone point cloud at point j; based on the local geometric mapping feature vector of the cone point cloud at point j, the local geometric features of the cone point cloud at point j are correlated and mapped to obtain the correlation mapping mutual exclusion metric factor of the cone point cloud at point j; based on the correlation mapping mutual exclusion metric factor of the cone point cloud at point j, the local geometric features of the cone point cloud at point j are enhanced by lower-level node features to obtain the local geometric enhancement feature of the cone point cloud at point j; the local geometric enhancement feature of the cone point cloud at point j is concatenated with the context-enhanced cone feature vector to obtain the global-local concatenation feature vector of the target point at point j. The global-local spliced ​​feature vector of the target point is input into the classification head to determine whether the target point belongs to the anatomical landmark of the spine.

2. The intelligent recognition method for spinal anatomical landmarks based on a three-dimensional model according to claim 1, characterized in that, The original CT 3D image data is preprocessed and coarsely segmented to obtain a set of vertebral body clusters, including: The original CT 3D image data is resampled using voxel resampling to obtain resampled CT 3D image data; The resampled CT 3D image data is used to extract the skeletal region and segment the spinal region to obtain a spinal point cloud; The spinal point cloud is coarsely segmented into vertebral bodies to obtain N point cloud clusters belonging to N vertebral bodies, which are then used as the vertebral body cluster set.

3. The intelligent recognition method for spinal anatomical landmarks based on a three-dimensional model according to claim 2, characterized in that, Performing coarse segmentation of the spinal point cloud to obtain N point cloud clusters belonging to N vertebrae as the vertebral cluster set includes: inputting the spinal point cloud into a pre-trained three-dimensional convolutional neural network model to obtain the vertebral cluster set.

4. The intelligent recognition method for spinal anatomical landmarks based on a three-dimensional model according to claim 1, characterized in that, A hierarchical graph is constructed from the vertebral cluster set to obtain a set of point-level bottom-level graphs and an inter-vertebral relationship graph, including: A point-level bottom-level graph is constructed for each vertebral cluster in the set of vertebral clusters by sampling the farthest point and the K-nearest neighbor algorithm. Calculate the centroid of each vertebral cluster in the set of vertebral clusters, and use the centroid of each vertebral cluster as a node in the inter-vertebral relationship graph; If the centroid distance between the centroids of two vertebral clusters is less than a preset threshold, an edge is established between the corresponding two nodes in the inter-vertebral relationship diagram.

5. The intelligent recognition method for spinal anatomical landmarks based on a three-dimensional model according to claim 1, characterized in that, The set of cone shape descriptor vectors is obtained by extracting features from each point-level bottom-level image in the set of point-level bottom-level images, including: The point-level low-level graph is input into the low-level feature extractor of a weight-sharing graph neural network to obtain a set of local geometric features of the cone point cloud; The set of local geometric features of the cone point cloud is subjected to graph pooling to obtain the cone shape descriptor vector.

6. A three-dimensional model-based intelligent recognition system for spinal anatomical landmarks, characterized in that, include: The CT 3D image acquisition module is used to acquire raw CT 3D image data. The CT three-dimensional image preprocessing and segmentation module is used to preprocess and coarsely segment the original CT three-dimensional image data to obtain a set of vertebral clusters. The layered graph construction module is used to construct a layered graph of the vertebral cluster set to obtain a set of point-level bottom-level graphs and an inter-vertebral relationship graph. The bottom-level graph feature extraction module is used to extract bottom-level graph features from each point-level bottom-level graph in the set of point-level bottom-level graphs to obtain a set of cone shape descriptor vectors. The vertebral feature context enhancement module is used to take each vertebral shape descriptor vector in the set of vertebral shape descriptor vectors as the initial feature of each node in the inter-vertebral relationship graph, and input it into a high-level context encoder based on graph convolutional network to obtain a set of context-enhanced vertebral feature vectors. The target point stitching module is used to extract the local geometric features of the cone point cloud of the j-th point in the i-th point-level bottom-level image from the set of point-level bottom-level images, and stitch them with the context-enhanced cone feature vector of the corresponding cone to obtain the global-local stitched feature vector of the target point. This includes: calculating the cosine similarity between the local geometric features of the cone point cloud of all points in the i-th point-level bottom-level image and the context-enhanced cone feature vector to obtain a cosine similarity composition vector, wherein the cosine similarity composition vector and the context-enhanced cone feature vector have the same length; calculating the correlation matrix between the cosine similarity composition vector and the context-enhanced cone feature vector; and stitching the local geometric features of the cone point cloud of the j-th point with... The correlation matrix is ​​mapped to obtain the local geometric mapping feature vector of the cone point cloud at point j; based on the local geometric mapping feature vector of the cone point cloud at point j, the correlation mapping mutual exclusion degree of the local geometric features of the cone point cloud at point j is measured to obtain the correlation mapping mutual exclusion degree metric factor of the cone point cloud at point j; based on the correlation mapping mutual exclusion degree metric factor of the cone point cloud at point j, the lower-level node feature enhancement of the local geometric features of the cone point cloud at point j is performed to obtain the local geometric enhancement feature of the cone point cloud at point j; the local geometric enhancement feature of the cone point cloud at point j is concatenated with the context-enhanced cone feature vector to obtain the global-local concatenation feature vector of the target point at point j. The marker determination module is used to input the global-local spliced ​​feature vector of the target point into the classification head to determine whether the target point belongs to the anatomical landmark of the spine.

Images