A Method and System for Locating Key Points in Orthopedic Surgery Based on Anatomical Feature Recognition

Abstract

This invention relates to the field of image feature recognition technology, specifically to a method and system for locating key surgical points in orthopedic surgery based on anatomical structure feature recognition. The method includes the following steps: constructing a prior skeletal topology map representing biomechanical connectivity based on human anatomical standards; utilizing a graph attention mechanism to perform semantic mapping and feature aggregation between the multi-scale anatomical feature map and the prior skeletal topology map to generate node-enhanced feature vectors; and performing spatial coordinate regression calculations based on these node-enhanced feature vectors to output the coordinates of the surgical key points. In this invention, by introducing prior anatomical topological knowledge as a global constraint for feature extraction, the information transmission mechanism of the graph structure effectively captures the geometric dependencies between key skeletal points. This solves the problem of local feature ambiguity caused by traditional methods in complex pathological scenarios such as fracture displacement, bone hyperplasia, or tissue occlusion, achieving high-precision and robust surgical key point localization.

Description

Technical Field

[0001] This invention relates to the field of image feature recognition technology, and in particular to a method and system for locating key points in orthopedic surgery based on anatomical structural feature recognition. Background Technology

[0002] Image feature recognition technology encompasses the application of computer vision algorithms and deep learning models to perform pixel-level analysis, semantic segmentation, and target localization of medical image data. This technology aims to automatically extract diagnostically valuable anatomical information from complex medical images through mathematical modeling and computational analysis, assisting surgeons in preoperative planning and intraoperative navigation. Traditional methods for locating key points in orthopedic surgery rely on surgeons manually marking each point on the image or using matching algorithms based on rigid geometric templates for image search and registration. These methods primarily determine key point locations by calculating the similarity between image grayscale gradients or local texture features and preset templates, typically assuming a relatively fixed geometric shape for bone structures, and the processing is often based on single-scale image information.

[0003] Traditional methods for locating key points in orthopedic surgery typically rely on manual annotation by surgeons or matching algorithms based on rigid templates. These methods utilize only grayscale or texture information from images for local searches, ignoring the inherent geometric and topological constraints between skeletal anatomical structures. When faced with complex pathological scenarios such as fracture displacement, bone hyperplasia, or camera angle deviations, they cannot establish effective global spatial context relationships. This makes the feature extraction process susceptible to noise interference or soft tissue occlusion, and makes it difficult to adapt to nonlinear deformations of bone morphology. Consequently, they suffer from large localization errors, insufficient robustness, and poor generalization ability. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a method and system for locating key points in orthopedic surgery based on anatomical structural feature recognition.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for locating key points in orthopedic surgery based on anatomical structural feature recognition, comprising the following steps:

[0006] S1: Acquire orthopedic medical imaging data, perform grayscale normalization and noise reduction preprocessing on the orthopedic medical imaging data, and input the processed orthopedic medical imaging data into a multi-level convolutional neural network model to perform feature encoding operations and extract multi-scale anatomical feature maps;

[0007] S2: Construct a priori skeletal topology graph based on standard anatomical morphology knowledge, where nodes represent anatomical landmarks and edges represent inherent biomechanical connection paths between landmarks, and perform position embedding encoding on the node states in the priori skeletal topology graph to construct a graph structure embedding matrix.

[0008] S3: Establish an attention mapping mechanism between the multi-scale anatomical feature map and the graph structure embedding matrix, calculate the semantic relevance weight matrix between image feature pixels and topological nodes, and use a graph convolutional network to aggregate, update and align the features of high-response regions to generate node-enhanced feature vectors.

[0009] S4: Input the node enhancement feature vector into the spatial coordinate regression network layer for position decoding and regression calculation, predict the probability distribution heatmap of the anatomical key points in the image coordinate system, and perform coordinate correction based on the peak response position of the probability distribution heatmap combined with geometric topological constraint rules, and output the coordinates of the surgical key points.

[0010] As a further aspect of the present invention, step S1 specifically comprises:

[0011] S11: Retrieve the original orthopedic medical image data through the medical image archiving and communication system interface, use a linear transformation algorithm to map the pixel gray values ​​of the image data to a standard range, and use an adaptive median filtering algorithm to eliminate salt-and-pepper noise interference in the image while preserving high-frequency detail information of the bone edge to generate a standardized preprocessed image sequence.

[0012] S12: Input the standardized preprocessed image sequence into the residual backbone network of the multi-level convolutional neural network model, and perform downsampling of the image through continuous convolutional and pooling layer operations to extract shallow texture detail features, mid-level shape and structure features and deep high-dimensional semantic features to form a pyramid feature set.

[0013] S13: The feature pyramid network architecture is used to perform top-down upsampling path enhancement and lateral connection fusion operations on the pyramid feature set, and the deep high-dimensional semantic features and shallow high-resolution information are pixel-level spliced ​​and channel dimensionality reduced to generate the multi-scale anatomical feature map.

[0014] As a further aspect of the present invention, step S2 specifically comprises:

[0015] S21: Based on the standard morphological structure defined by the human skeletal anatomical atlas, select anatomical landmarks with high clinical recognizability as vertices of the graph structure, and define the edges between vertices according to the physiological connection relationship of bones and the biomechanical transmission path of joint movement to construct the prior skeletal topology graph.

[0016] S22: Analyze the initial spatial coordinates of multiple vertices in the prior skeletal topology graph in the standard anatomical template, use the sinusoidal position coding function to map the two-dimensional coordinate information of the vertices into a high-dimensional dense vector, and combine the category semantic labels of the vertices to generate the initial node state vector.

[0017] S23: The graph attention layer is used to perform weighted aggregation on the features of the initial node state vector and its neighboring nodes. The node features, including positional and semantic information, are projected into a unified embedding space through a learnable linear transformation matrix to establish the graph structure embedding matrix.

[0018] As a further aspect of the present invention, step S3 specifically comprises:

[0019] S31: The graph structure embedding matrix is ​​used as the query vector, and the pixel feature sequence after the multi-scale anatomical feature map is flattened as the key vector and value vector. The similarity measure between each topological node and the global pixel of the image is calculated by dot product operation, and the semantic relevance weight matrix is ​​generated by normalization exponential function.

[0020] S32: The semantic relevance weight matrix is ​​used to perform weighted summation and feature recombination on the multi-scale anatomical feature map. Local visual features highly related to the target anatomical landmarks are extracted from the global image features. The local visual features are then spliced ​​and fused with the corresponding graph structure node embedding features to generate fused node features.

[0021] S33: Input the fused node features into the graph convolutional network layer, propagate and aggregate feature information between adjacent nodes according to the adjacency relationship matrix of the prior skeleton topology graph, update and align the aggregated features using a nonlinear activation function, and generate the node enhancement feature vector.

[0022] As a further aspect of the present invention, step S4 specifically comprises:

[0023] S41: Input the node enhancement feature vector into the spatial coordinate regression network layer composed of multiple transposed convolutional layers, restore the spatial resolution of the feature map through progressive upsampling operations, and use a series of 1x1 convolutional layers to map the number of feature channels to the number of key point categories to generate the probability distribution heatmap.

[0024] S42: Traverse each pixel in the probability distribution heatmap, use a differentiable soft maximization operation to extract the sub-pixel position with the highest response value in the heatmap as the initial prediction coordinates, and calculate the centroid of the pixel gradient around the position to further refine the positioning of the center coordinates of the key point.

[0025] S43: Calculate the Euclidean distance and relative angle between the initial predicted coordinates, compare them with the geometric statistical parameters of the standard human skeleton model, and for abnormal coordinate points that exceed the preset physiological rationality threshold, use the penalty loss function to fine-tune the coordinate position in reverse, and output the coordinates of the key surgical points.

[0026] As a further aspect of the present invention, the process of calculating the semantic relevance weight matrix specifically includes:

[0027] Obtain the graph structure embedding matrix and the multi-scale anatomical feature map, construct an attention scoring model to measure the correlation strength between topological nodes and feature pixels, and calculate the semantic relevance weights according to the following formula:

[0028] ;

[0029] in, Representing the The topology node and the first The element values ​​in the semantic relevance weight matrix between image pixels. Represents the first in the graph structure embedding matrix The feature vector of each node The first one represented in the multi-scale anatomical feature map Feature vectors at each pixel location This represents a learnable weight matrix that maps node embeddings to query vectors. This represents a learnable weight matrix that maps image features to key vectors. The numerical value representing the feature dimension of the key vector. This represents the total number of pixels in the multi-scale anatomical feature map.

[0030] As a further aspect of the present invention, the execution process of the geometric topological constraint rules specifically includes:

[0031] Obtain the set of initial predicted coordinates, construct a geometric constraint loss function between key points based on the rigid connectivity properties of the skeleton, and iteratively correct the predicted coordinates using the gradient descent direction. The correction process is performed according to the following formula:

[0032] ;

[0033] in, This represents the corrected coordinates of the surgical key points. The initial predicted coordinates represent those obtained from heatmap regression. Represents the coordinates of adjacent keypoints that have an anatomical connection with the current keypoint. Represents the set of adjacent key points. Represents Euclidean norm operations. This represents the standard anatomical distance between two key points defined based on prior statistical knowledge. The regularization coefficient represents the strength of the geometric constraints.

[0034] As a further aspect of the present invention, the specific execution process of the position embedding encoding in S22 is as follows:

[0035] Obtain the horizontal and vertical coordinate values ​​of each anatomical landmark in the standard template coordinate system, map these coordinate values ​​to sine and cosine waveform functions of different frequencies to generate a position-sensitive high-dimensional encoding vector, and add the high-dimensional encoding vector to the corresponding anatomical type semantic embedding vector element by element. Use layer normalization to adjust the distribution of the addition result to generate the initial node state vector.

[0036] As a further aspect of the present invention, the specific execution process of S13 is as follows:

[0037] The deep high-dimensional semantic features and the mid-level shape and structure features are obtained. Bilinear interpolation upsampling is performed on the deep high-dimensional semantic features to make their spatial resolution consistent with that of the mid-level shape and structure features. Then, the mid-level shape and structure features are compressed using a 1x1 convolution kernel. The upsampled deep features and the compressed mid-level features are then fused pixel by pixel. The fused result is then smoothed again using a 3x3 convolution to eliminate aliasing effects, thereby generating the multi-scale anatomical feature map.

[0038] A system for locating key points in orthopedic surgery based on anatomical structure feature recognition, the system being used to implement the aforementioned method for locating key points in orthopedic surgery based on anatomical structure feature recognition, the system comprising:

[0039] The medical image feature encoding module is used to acquire orthopedic medical image data, perform grayscale normalization and noise reduction preprocessing on the data, and input the processed data into a multi-level convolutional neural network model to perform feature encoding operations and extract multi-scale anatomical feature maps.

[0040] The prior topology graph construction module is used to construct a prior skeletal topology graph based on standard anatomical morphological knowledge, define nodes representing anatomical landmarks and edges representing biomechanical connection paths, and perform position embedding encoding on node states to construct a graph structure embedding matrix.

[0041] The graph attention feature enhancement module is used to establish an attention mapping mechanism between the multi-scale anatomical feature map and the graph structure embedding matrix, calculate the semantic relevance weight matrix, and use a graph convolutional network to aggregate, update, and align features to generate node enhancement feature vectors.

[0042] The key point coordinate regression module is used to input the node enhanced feature vector into the spatial coordinate regression network layer for position decoding and regression calculation, predict the probability distribution heatmap, and perform coordinate correction based on the heatmap peak and geometric topological constraint rules, and output the coordinates of the surgical key points.

[0043] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0044] In this invention, by constructing a prior skeletal topology map and deeply fusing it with a multi-scale anatomical feature map, and utilizing a graph attention network mechanism to capture the long-distance spatial dependencies and geometric constraints between anatomical key points, adaptive perception and correction of nonlinear deformation of skeletal morphology is achieved. This solves the technical problem of inaccurate localization in fracture displacement or occlusion scenarios caused by neglecting the topological correlation of anatomical structures in traditional methods. It significantly enhances the adaptability of feature representation to complex pathological structures and achieves high-precision and high-robust automated localization of surgical key points. Attached Figure Description

[0045] Figure 1 This is the main flowchart of the method for locating key points in orthopedic surgery according to the present invention;

[0046] Figure 2 This is a flowchart of the multi-scale anatomical feature map extraction process of the present invention;

[0047] Figure 3 This is a flowchart illustrating the construction process of the graph structure embedding matrix in this invention.

[0048] Figure 4 This is a flowchart of the node enhancement feature vector generation process of the present invention;

[0049] Figure 5 This is a flowchart of the surgical key point coordinate prediction and correction process of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0051] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0052] Please see Figure 1 and Figure 2 This invention provides a technical solution: a method for locating key points in orthopedic surgery based on anatomical structural feature recognition, comprising the following steps:

[0053] S1: Acquire orthopedic medical imaging data, perform grayscale normalization and denoising preprocessing on the orthopedic medical imaging data, and input the processed orthopedic medical imaging data into a multi-level convolutional neural network model to perform feature encoding operations and extract multi-scale anatomical feature maps.

[0054] The specific steps of S1 are as follows:

[0055] S11: Retrieve original orthopedic medical image data through the medical image archiving and communication system interface, use a linear transformation algorithm to map the pixel gray values ​​of the image data to a standard range, and use an adaptive median filtering algorithm to eliminate salt-and-pepper noise interference in the image while preserving high-frequency detail information of bone edges to generate a standardized preprocessed image sequence.

[0056] S12: Input the standardized preprocessed image sequence into the residual backbone network of the multi-level convolutional neural network model, and perform downsampling of the image through continuous convolutional and pooling layer operations to extract shallow texture detail features, mid-level shape and structure features and deep high-dimensional semantic features to form a pyramid feature set.

[0057] S13: The feature pyramid network architecture is used to perform top-down upsampling path enhancement and lateral connection fusion operations on the pyramid feature set. The deep high-dimensional semantic features and shallow high-resolution information are pixel-level spliced ​​and channel dimensionality reduced to generate multi-scale anatomical feature maps.

[0058] The specific execution process of S13 is as follows:

[0059] Deep high-dimensional semantic features and mid-level shape and structure features are obtained. Bilinear interpolation upsampling is performed on the deep high-dimensional semantic features to make their spatial resolution consistent with that of the mid-level shape and structure features. Then, the mid-level shape and structure features are compressed using a 1x1 convolution kernel. The upsampled deep features and the compressed mid-level features are then fused pixel by pixel. The fused result is then smoothed again by a 3x3 convolution to eliminate aliasing and generate a multi-scale anatomical feature map.

[0060] The original orthopedic X-ray digital image data of the patient awaiting surgery was retrieved via the interface of the hospital's internal medical image archiving and communication system, using the DICOM 3.0 communication protocol. This data is a 16-bit grayscale image matrix with a resolution range of [missing information]. to Between pixels. The photometric interpretation tag in the image metadata is read, the monochrome grayscale mode is identified, and the raw DICOM data stream is converted to a 3D tensor format, with the number of channels initialized to 1. A linear transformation algorithm is performed to map the pixel grayscale values ​​of the image data to a standard range, with a window width of 1500 HU and a window level of 450 HU, using the formula... Dynamic range compression is performed, linearly mapping 16-bit high dynamic range pixel values ​​to an 8-bit range. For the compressed image, a window size of [missing value] is used. An adaptive median filtering algorithm eliminates salt-and-pepper noise interference in images. This algorithm determines whether a pixel value belongs to a noise point by calculating the median, maximum, and minimum gray values ​​of pixels within a local window. While filtering out extreme noise, it preserves high-frequency details of the bone edges and ultimately adjusts the image size uniformly. Pixels, generating a standardized preprocessed image sequence.

[0061] in, Represents the output pixel grayscale value. Represents the original pixel grayscale value of the input. This represents the set window level value. This represents the set window width value.

[0062] The aforementioned DICOM 3.0 communication protocol refers to a medical digital imaging and communication standard used to enable the exchange, storage, and transmission of medical images and related information between medical imaging equipment and image management systems from different manufacturers.

[0063] The standardized preprocessed image sequence is input into the residual backbone network of a multi-level convolutional neural network model. The input tensor passes through the first layer. Convolution kernel processing, followed by The max-pooling layer expands the number of output channels to 64. The data flow sequentially passes through four main stages containing 3, 4, 23, and 3 residual blocks, respectively. Each residual block contains... , , The bottleneck layer structure is optimized, and batch normalization and ReLU activation functions are introduced. As the network depth increases, the spatial resolution of the feature maps is progressively downsampled to 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image, with the corresponding number of channels expanding to 256, 512, 1024, and 2048, respectively. In this process, the first and second stages extract shallow texture details such as bone texture and edge sharpness; the third stage extracts mid-level shape and structural features such as bone structure morphology and joint spaces; and the fourth stage extracts deep, high-dimensional semantic features including anatomical location categories and pathological patterns. These features together form a pyramid feature set.

[0064] The aforementioned multi-level convolutional neural network model refers to a deep learning architecture. In this embodiment, ResNet-101 is specifically used as the backbone network. By introducing residual skip connections, the gradient vanishing problem in deep network training is solved, thereby enabling the extraction of multi-level image features from low-level texture to high-level semantics.

[0065] A feature pyramid network architecture is employed to perform top-down upsampling path enhancement and lateral connection fusion operations on the pyramid feature set. Deep, high-dimensional semantic features are selected as the top-level features, and firstly... The convolutional layer reduces its channel count to 256, resulting in feature map P5. Bilinear interpolation upsampling is then performed on P5 to double its spatial resolution, ensuring it aligns with the spatial dimensions of the mid-layer shape and structural features. Simultaneously, utilizing... The convolutional kernel compresses the number of channels in the mid-level features to 256. The upsampled deep features are then fused with the compressed mid-level features through pixel-by-pixel addition. The fusion result is then processed again. Convolutional smoothing is applied to eliminate aliasing effects caused by upsampling, generating feature map P4. This process is repeated to generate feature maps with resolutions of [resolution values ​​to be filled in]. P3 and resolution are The final output is a multi-scale anatomical feature map containing four levels: P2, P3, P4, and P5, with each level having a uniform number of channels of 256. During model training, the mean squared error loss function is used to calculate the difference between the predicted heatmap and the true Gaussian heatmap, and the AdamW optimizer is used for parameter updates. The initial learning rate is set to [value missing]. The weight decay coefficient is The network learns the optimal feature representation by minimizing the loss function through backpropagation.

[0066] Please see Figure 1 and Figure 3S2: Construct a prior skeletal topology graph based on standard anatomical morphology knowledge, where nodes represent anatomical landmarks and edges represent the inherent biomechanical connection paths between landmarks, and perform position embedding encoding on the node states in the prior skeletal topology graph to construct a graph structure embedding matrix.

[0067] The specific steps of S2 are as follows:

[0068] S21: Based on the standard morphological structure defined by the human skeletal anatomical atlas, select anatomical landmarks with high clinical recognizability as vertices of the graph structure, and define the edges between vertices according to the physiological connection relationship of bones and the biomechanical transmission path of joint movement to construct a priori skeletal topology graph.

[0069] S22: Analyze the initial spatial coordinates of multiple vertices in the prior skeletal topology graph in the standard anatomical template, use the sinusoidal position encoding function to map the two-dimensional coordinate information of the vertices into a high-dimensional dense vector, and combine the vertices' category semantic labels to generate the initial node state vector.

[0070] The specific execution process of position embedding encoding in S22 is as follows:

[0071] Obtain the horizontal and vertical coordinate values ​​of each anatomical landmark in the standard template coordinate system, map these coordinate values ​​to sine and cosine waveform functions of different frequencies to generate a position-sensitive high-dimensional encoding vector, and add the high-dimensional encoding vector to the corresponding anatomical type semantic embedding vector element by element. Use layer normalization to adjust the distribution of the addition result to generate the initial node state vector.

[0072] S23: The graph attention layer is used to perform weighted aggregation of the initial node state vector and the features of its neighboring nodes. The node features, including positional and semantic information, are projected into a unified embedding space through a learnable linear transformation matrix to establish a graph structure embedding matrix.

[0073] Based on the standard morphological structures defined in standard anatomical atlases, highly clinically recognizable anatomical landmarks were selected as vertices of the graph structure in the pelvic and proximal femur regions. The selected key points included: the left and right anterior superior iliac spines, the left and right pubic symphysis, the left and right femoral head centers, the left and right greater trochanters, and the left and right lesser trochanters, totaling 10 anatomical landmarks. Edges between vertices were defined according to the physiological connections of the bones and the biomechanical transmission paths of joint movement. For example, connecting the ipsilateral femoral head center and the greater trochanter vertex represents the femoral neck axis, and connecting the left and right pubic symphysis represents the closed-loop structure of the pelvis. A priori skeletal topology graph was constructed, containing a set of vertices and a set of edges, and an adjacency matrix was used to describe the connections between nodes.

[0074] The initial spatial coordinates of 10 vertices in the prior skeletal topology graph within a standard anatomical template are analyzed. The coordinate values ​​are normalized to the interval 0 to 1, and a sinusoidal position encoding function is used to map the two-dimensional coordinate information of the vertices into a high-dimensional dense vector. Specifically, for each coordinate dimension, the encoded vector is calculated as follows: as well as Simultaneously, a unique category semantic label is assigned to each anatomical landmark, which is then converted into a 256-dimensional semantic embedding vector through a learnable embedding layer. The location embedding vector is then element-wise added to the anatomical type semantic embedding vector, and the distribution of the addition result is adjusted using a layer normalization operation to make its mean 0 and variance 1, thereby generating the initial node state vector.

[0075] in, Represents coordinate dimension or The specific value, Represents the dimension index in the encoded vector. This represents the total dimension of the embedded vector, which is set to 128 in this embodiment.

[0076] A graph attention layer is used to perform weighted aggregation of the initial node state vector and the features of its neighboring nodes. For any node... Calculate its relationship with neighboring nodes Attention coefficient between The attention coefficients are normalized using the Softmax function to obtain... And based on this, aggregate the features of neighboring nodes: This embodiment stacks three graph attention layers, capturing node dependencies in different subspaces through a multi-head attention mechanism. Finally, it projects node features, including positional and semantic information, into a unified embedding space, establishing a dimension of... The graph structure embedding matrix.

[0077] in, Representative node With nodes Attention coefficient between them Represents the weight vector. Represents a learnable linear transformation matrix. and Representing nodes respectively and nodes eigenvectors, This represents a vector concatenation operation. Represents the updated node eigenvectors, Represents the activation function. Representative node The set of neighboring nodes, This represents the normalized attention weight.

[0078] The aforementioned graph attention layer refers to an architecture based on graph neural networks that dynamically calculates the importance weights of neighboring nodes to the central node through a self-attention mechanism, thereby achieving non-uniform feature aggregation.

[0079] Please see Figure 1 and Figure 4 S3: Establish an attention mapping mechanism between multi-scale anatomical feature maps and graph structure embedding matrices, calculate the semantic relevance weight matrix between image feature pixels and topological nodes, and use graph convolutional networks to aggregate, update, and align features in high-response regions to generate node-enhanced feature vectors.

[0080] The specific steps for S3 are as follows:

[0081] S31: The graph structure embedding matrix is ​​used as the query vector, and the pixel feature sequence after the multi-scale anatomical feature map is flattened as the key vector and value vector. The similarity measure between each topological node and the global pixel of the image is calculated by dot product operation, and the semantic relevance weight matrix is ​​generated by normalization exponential function.

[0082] The process of calculating the semantic relevance weight matrix specifically includes:

[0083] Obtain the graph structure embedding matrix and multi-scale anatomical feature map, construct an attention scoring model to measure the correlation strength between topological nodes and feature pixels, and calculate the semantic relevance weights according to the following formula:

[0084] ;

[0085] in, Representing the The topology node and the first The element values ​​in the semantic relevance weight matrix between image pixels. The first element in the graph structure embedding matrix represents the graph structure embedding matrix. The feature vector of each node The first one in the multi-scale anatomical feature map Feature vectors at each pixel location This represents a learnable weight matrix that maps node embeddings to query vectors. This represents a learnable weight matrix that maps image features to key vectors. The numerical value representing the feature dimension of the key vector. Represents the total number of pixels in a multi-scale anatomical feature map;

[0086] S32: Use the semantic relevance weight matrix to perform weighted summation and feature recombination on the multi-scale anatomical feature map, extract local visual features that are highly related to the target anatomical landmarks from the global image features, and splice and fuse the local visual features with the corresponding graph structure node embedding features to generate fused node features.

[0087] S33: Input the fused node features into the graph convolutional network layer, propagate and aggregate feature information between adjacent nodes based on the adjacency matrix of the prior skeleton topology graph, and use a non-linear activation function to update and align the aggregated features to generate node enhanced feature vectors.

[0088] The dimensions to be acquired are The graph structure embedding matrix and dimension are Multi-scale anatomical feature maps. Flattening the multi-scale anatomical feature maps in space yields a multi-dimensional map. The pixel feature sequence is used to construct an attention scoring model to measure the correlation strength between topological nodes and feature pixels, and to establish an attention mapping mechanism between multi-scale anatomical feature maps and graph structure embedding matrices.

[0089] The graph embedding matrix is ​​mapped to a query vector using a linear transformation matrix, and the flattened multi-scale anatomical feature map is mapped to key and value vectors using a linear transformation matrix. The feature dimension of the key vector is set to 64. (Based on the formula...) Calculate the semantic relevance weights.

[0090] in, Representing the The topology node and the first The semantic relevance weights between image pixels The first element in the graph structure embedding matrix represents the graph structure embedding matrix. The feature vector of each node The first one in the multi-scale anatomical feature map Feature vectors at each pixel location This represents a learnable weight matrix that maps node embeddings to query vectors. This represents a learnable weight matrix that maps image features to key vectors. The numerical value representing the feature dimension of the key vector. This represents the total number of pixels in the multi-scale anatomical feature map.

[0091] Taking the first topological node, the center of the left femoral head, as an example, let's assume its query vector is... Select the 500th pixel position in the image, its key vector is Calculate the vector dot product and divide by a scaling factor of 8; assume the result is 2.0. Calculate the exponent term. The value is approximately 7.389. Assuming the sum of the exponents for all pixels is 147.78, the semantic relevance weight between this node and that pixel is 0.05. This value is significantly higher than the weight of pixels in the background region, indicating that this node has focused on the correct image anatomy region.

[0092] The value vectors of the multi-scale anatomical feature maps are weighted and summed using the generated semantic relevance weight matrix to obtain a dimension of The local visual features are then concatenated with the original graph structure node embedding features to obtain a dimension of [dimensional value missing]. The fusion feature matrix.

[0093] The fused node features are input into the graph convolutional network layer. Using the adjacency matrix of the prior skeletal topology graph, self-loop connections are added and normalized. The fused feature matrix is ​​then substituted into the propagation formula. By stacking two layers of graph convolutional networks, each anatomical node can aggregate the visual features of its neighboring anatomical points. The aggregated features are then updated and aligned using the ReLU nonlinear activation function to generate the final node-enhanced feature vector.

[0094] in, This represents the updated node feature matrix. Represents the ReLU activation function. This represents the normalized adjacency matrix. The fusion feature matrix representing the input. The learnable weight matrix represents the graph convolutional layer.

[0095] The aforementioned graph convolutional network layer refers to a neural network layer that can directly perform convolution operations on graph structure data. It uses the topological information of the graph to aggregate neighborhood features, thereby updating and enhancing node features.

[0096] Please see Figure 1 and Figure 5 S4: Input the node enhancement feature vector into the spatial coordinate regression network layer for position decoding and regression calculation, predict the probability distribution heatmap of anatomical key points in the image coordinate system, and perform coordinate correction based on the peak response position of the probability distribution heatmap combined with geometric topological constraint rules, and output the coordinates of surgical key points.

[0097] The specific steps for S4 are as follows:

[0098] S41: Input the node enhancement feature vector into the spatial coordinate regression network layer composed of multiple transposed convolutional layers. The spatial resolution of the feature map is restored through progressive upsampling operations. A series of 1x1 convolutional layers are used to map the number of feature channels to the number of key point categories to generate a probability distribution heatmap.

[0099] S42: Traverse each pixel in the probability distribution heatmap, use a differentiable soft maximization operation to extract the sub-pixel position with the highest response value in the heatmap as the initial prediction coordinates, and calculate the centroid of the pixel gradient around the position to further refine the positioning of the center coordinates of the key point.

[0100] S43: Calculate the Euclidean distance and relative angle between the initial predicted coordinates, compare them with the geometric and statistical parameters of the standard human skeleton model, and for abnormal coordinate points that exceed the preset physiological rationality threshold, use the penalty loss function to fine-tune the coordinate position in reverse and output the coordinates of the key surgical points.

[0101] The execution process of geometric topological constraint rules specifically includes:

[0102] Obtain the initial set of predicted coordinates, construct a geometric constraint loss function between keypoints based on the rigid connectivity properties of the skeleton, and iteratively correct the predicted coordinates using the gradient descent direction. The correction process is performed according to the following formula:

[0103] ;

[0104] in, This represents the corrected coordinates of the surgical key points. These represent the initial predicted coordinates obtained from heatmap regression. Represents the coordinates of adjacent keypoints that have an anatomical connection with the current keypoint. Represents the set of adjacent key points. Represents Euclidean norm operations. This represents the standard anatomical distance between two key points defined based on prior statistical knowledge. The regularization coefficient represents the strength of the geometric constraints.

[0105] The node-enhanced feature vectors are input into the spatial coordinate regression network layer. This network layer consists of three consecutive transposed convolutional layers, with each kernel having a size of [missing value]. The step size is 2, and the padding is 1. Node features are first reshaped and broadcast to the initial spatial feature map, and the spatial resolution of the feature map is increased from [previous value] to [current value] through progressive upsampling operations. Restore to The number of feature channels decreases layer by layer. Finally, a... The convolutional layer maps the number of feature channels to the number of keypoint categories, and applies the Sigmoid activation function to the output of each channel to generate a probability distribution heatmap describing the probability of each keypoint's existence.

[0106] The probability distribution heatmap is traversed through each channel, and sub-pixel coordinates are extracted using a differentiable soft maximization operation. Spatial normalization is then applied to the heatmap according to the formula... as well as Calculate the desired coordinates. After obtaining the initial predicted coordinates, further calculate the coordinates around these coordinates. The centroid of the pixel gradient within the neighborhood is used to finely locate the center coordinates of key points.

[0107] in, and These represent the predicted x-axis and y-axis, respectively. Representing the Key heatmap points in location The response value at that location, The temperature parameter, representing the sharpness of the control probability distribution, is set to 100 in this embodiment. and The index represents the dimension of the heatmap space being traversed.

[0108] The aforementioned soft maximization operation refers to a differentiable coordinate extraction method that replaces the traditional maximum indexing operation by calculating the expected position of the heatmap, thereby enabling the entire network to perform end-to-end training and gradient backpropagation.

[0109] Obtain the initial predicted coordinate set of all keypoints. Based on the rigid connectivity of the skeleton, construct a geometric constraint loss function between keypoints. Retrieve the geometric statistical parameters of a standard human skeleton model from a medical statistical database and set the standard anatomical distance between adjacent keypoints. Set the regularization coefficient to 0.5. For abnormal coordinate points exceeding a preset physiological reasonableness threshold, iteratively correct the predicted coordinates using the gradient descent direction. The correction process is performed according to the following formula: ;

[0110] in, This represents the corrected coordinates of the surgical key points. These represent the initial predicted coordinates obtained from heatmap regression. Represents the coordinates of adjacent keypoints that have an anatomical connection with the current keypoint. Represents the set of adjacent key points. Represents Euclidean norm operations. This represents the standard anatomical distance between two key points defined based on prior statistical knowledge. The regularization coefficient represents the strength of the geometric constraints.

[0111] Table 1 Geometric Constraint Parameters for Key Anatomical Points

[0112] ;

[0113] Table 1 lists the standard anatomical distances and tolerance thresholds for some keypoint pairs. Taking the center of the left femoral head as an example, assuming its initial predicted coordinates are... The coordinates of its adjacent key point, the vertex of the left large rotor, are: The Euclidean distance between the two is calculated to be 100.0 pixels. Looking up the table, the standard anatomical distance is 120.5 pixels. The actual distance is less than the standard distance, and the absolute value of the deviation is 20.5 pixels, exceeding the tolerance threshold of 15.0 pixels. The unit direction vector is calculated as follows. The correction amount is calculated as follows: The corrected coordinates are then... Through the above calculations, the predicted coordinates are shifted away from the greater trochanter, making their spacing approach the anatomical standard value. Experimental data show that after introducing this geometric topological constraint rule, the average radial error of keypoint localization is reduced from 2.8 pixels to 1.5 pixels, significantly correcting the anatomical structure localization drift problem caused by image occlusion or artifacts.

[0114] A system for locating key points in orthopedic surgery based on anatomical structure feature recognition is provided. This system is used to execute the aforementioned method for locating key points in orthopedic surgery based on anatomical structure feature recognition. The system includes:

[0115] The medical image feature encoding module is used to acquire orthopedic medical image data, perform grayscale normalization and noise reduction preprocessing on the data, and input the processed data into a multi-level convolutional neural network model to perform feature encoding operations and extract multi-scale anatomical feature maps.

[0116] The prior topology graph construction module is used to construct a prior skeletal topology graph based on standard anatomical morphological knowledge, define nodes representing anatomical landmarks and edges representing biomechanical connection paths, and perform position embedding encoding on node states to construct a graph structure embedding matrix.

[0117] The graph attention feature enhancement module is used to establish an attention mapping mechanism between multi-scale anatomical feature maps and graph structure embedding matrices, calculate semantic relevance weight matrices, and use graph convolutional networks to aggregate, update, and align features to generate node-enhanced feature vectors.

[0118] The key point coordinate regression module is used to input the node enhanced feature vector into the spatial coordinate regression network layer for position decoding and regression calculation, predict the probability distribution heatmap, and perform coordinate correction based on the heatmap peak and geometric topological constraint rules, and output the coordinates of the surgical key points.

[0119] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A method for locating key points in orthopedic surgery based on anatomical structural feature recognition, characterized in that, Includes the following steps: S1: Acquire orthopedic medical imaging data, perform grayscale normalization and noise reduction preprocessing on the orthopedic medical imaging data, and input the processed orthopedic medical imaging data into a multi-level convolutional neural network model to perform feature encoding operations and extract multi-scale anatomical feature maps; S2: Construct a priori skeletal topology graph based on standard anatomical morphology knowledge, where nodes represent anatomical landmarks and edges represent inherent biomechanical connection paths between landmarks, and perform position embedding encoding on the node states in the priori skeletal topology graph to construct a graph structure embedding matrix. S3: Establish an attention mapping mechanism between the multi-scale anatomical feature map and the graph structure embedding matrix, calculate the semantic relevance weight matrix between image feature pixels and topological nodes, and use a graph convolutional network to aggregate, update and align the features of high-response regions to generate node-enhanced feature vectors. S4: Input the node enhancement feature vector into the spatial coordinate regression network layer for position decoding and regression calculation, predict the probability distribution heatmap of the anatomical key points in the image coordinate system, and perform coordinate correction based on the peak response position of the probability distribution heatmap combined with geometric topological constraint rules, and output the coordinates of the surgical key points. The specific steps of S2 are as follows: S21: Based on the standard morphological structure defined by the human skeletal anatomical atlas, select anatomical landmarks with high clinical recognizability as vertices of the graph structure, and define the edges between vertices according to the physiological connection relationship of bones and the biomechanical transmission path of joint movement to construct the prior skeletal topology graph. S22: Analyze the initial spatial coordinates of multiple vertices in the prior skeletal topology graph in the standard anatomical template, use the sinusoidal position coding function to map the two-dimensional coordinate information of the vertices into a high-dimensional dense vector, and combine the category semantic labels of the vertices to generate the initial node state vector. S23: The graph attention layer is used to perform weighted aggregation on the features of the initial node state vector and its neighboring nodes. The node features, including positional and semantic information, are projected into a unified embedding space through a learnable linear transformation matrix to establish the graph structure embedding matrix. The specific steps of S3 are as follows: S31: The graph structure embedding matrix is ​​used as the query vector, and the pixel feature sequence after the multi-scale anatomical feature map is flattened as the key vector and value vector. The similarity measure between each topological node and the global pixel of the image is calculated by dot product operation, and the semantic relevance weight matrix is ​​generated by normalization exponential function. S32: The semantic relevance weight matrix is ​​used to perform weighted summation and feature recombination on the multi-scale anatomical feature map. Local visual features highly related to the target anatomical landmarks are extracted from the global image features. The local visual features are then spliced ​​and fused with the corresponding graph structure node embedding features to generate fused node features. S33: Input the fused node features into the graph convolutional network layer, propagate and aggregate feature information between adjacent nodes according to the adjacency relationship matrix of the prior skeleton topology graph, update and align the aggregated features using a nonlinear activation function, and generate the node enhancement feature vector. The process of calculating the semantic relevance weight matrix specifically includes: Obtain the graph structure embedding matrix and the multi-scale anatomical feature map, construct an attention scoring model to measure the correlation strength between topological nodes and feature pixels, and calculate the semantic relevance weights according to the following formula: ； in, Representing the The topology node and the first The element values ​​in the semantic relevance weight matrix between image pixels. Represents the first in the graph structure embedding matrix The feature vector of each node The first one represented in the multi-scale anatomical feature map Feature vectors at each pixel location This represents a learnable weight matrix that maps node embeddings to query vectors. This represents a learnable weight matrix that maps image features to key vectors. The numerical value representing the feature dimension of the key vector. This represents the total number of pixels in the multi-scale anatomical feature map. The specific steps of S4 are as follows: S41: Input the node enhancement feature vector into the spatial coordinate regression network layer composed of multiple transposed convolutional layers, restore the spatial resolution of the feature map through progressive upsampling operations, and use a series of 1x1 convolutional layers to map the number of feature channels to the number of key point categories to generate the probability distribution heatmap. S42: Traverse each pixel in the probability distribution heatmap, use a differentiable soft maximization operation to extract the sub-pixel position with the highest response value in the heatmap as the initial prediction coordinates, and calculate the centroid of the pixel gradient around the position to further refine the positioning of the center coordinates of the key point. S43: Calculate the Euclidean distance and relative angle between the initial predicted coordinates, compare them with the geometric statistical parameters of the standard human skeleton model, and for abnormal coordinate points that exceed the preset physiological rationality threshold, use the penalty loss function to fine-tune the coordinate position in reverse, and output the coordinates of the key surgical points. The execution process of the geometric topology constraint rules specifically includes: Obtain the set of initial predicted coordinates, construct a geometric constraint loss function between key points based on the rigid connectivity properties of the skeleton, and iteratively correct the predicted coordinates using the gradient descent direction. The correction process is performed according to the following formula: ； in, This represents the corrected coordinates of the surgical key points. The initial predicted coordinates represent those obtained from heatmap regression. Represents the coordinates of adjacent keypoints that have an anatomical connection with the current keypoint. Represents the set of adjacent key points. Represents Euclidean norm operations. This represents the standard anatomical distance between two key points defined based on prior statistical knowledge. The regularization coefficient represents the strength of the geometric constraints.

2. The method for locating key points in orthopedic surgery based on anatomical structural feature recognition according to claim 1, characterized in that, The specific steps of S1 are as follows: S11: Retrieve the original orthopedic medical image data through the medical image archiving and communication system interface, use a linear transformation algorithm to map the pixel gray values ​​of the image data to a standard range, and use an adaptive median filtering algorithm to eliminate salt-and-pepper noise interference in the image while preserving high-frequency detail information of the bone edge to generate a standardized preprocessed image sequence. S12: Input the standardized preprocessed image sequence into the residual backbone network of the multi-level convolutional neural network model, and perform downsampling of the image through continuous convolutional and pooling layer operations to extract shallow texture detail features, mid-level shape and structure features and deep high-dimensional semantic features to form a pyramid feature set. S13： A feature pyramid network architecture is used to perform top-down upsampling path enhancement and lateral connection fusion operations on the pyramid feature set, and the deep high-dimensional semantic features and shallow high-resolution information are pixel-level spliced ​​and channel dimensionality reduced to generate the multi-scale anatomical feature map.

3. The method for locating key points in orthopedic surgery based on anatomical structural feature recognition according to claim 1, characterized in that, The specific execution process of the position embedding encoding in S22 is as follows: Obtain the horizontal and vertical coordinate values ​​of each anatomical landmark in the standard template coordinate system, map these coordinate values ​​to sine and cosine waveform functions of different frequencies to generate a position-sensitive high-dimensional encoding vector, and add the high-dimensional encoding vector to the corresponding anatomical type semantic embedding vector element by element. Use layer normalization to adjust the distribution of the addition result to generate the initial node state vector.

4. The method for locating key points in orthopedic surgery based on anatomical structure feature recognition according to claim 2, characterized in that, The specific execution process of S13 is as follows: The deep high-dimensional semantic features and the mid-level shape and structure features are obtained. Bilinear interpolation upsampling is performed on the deep high-dimensional semantic features to make their spatial resolution consistent with that of the mid-level shape and structure features. Then, the mid-level shape and structure features are compressed using a 1x1 convolution kernel. The upsampled deep features and the compressed mid-level features are then fused pixel by pixel. The fused result is then smoothed again using a 3x3 convolution to eliminate aliasing effects, thereby generating the multi-scale anatomical feature map.

5. A key point localization system for orthopedic surgery based on anatomical structure feature recognition, characterized in that: The system is used to implement the orthopedic surgical key point localization method based on anatomical structure feature recognition as described in any one of claims 1-4, and the system comprises: The medical image feature encoding module is used to acquire orthopedic medical image data, perform grayscale normalization and noise reduction preprocessing on the data, and input the processed data into a multi-level convolutional neural network model to perform feature encoding operations and extract multi-scale anatomical feature maps. The prior topology graph construction module is used to construct a prior skeletal topology graph based on standard anatomical morphological knowledge, define nodes representing anatomical landmarks and edges representing biomechanical connection paths, and perform position embedding encoding on node states to construct a graph structure embedding matrix. The graph attention feature enhancement module is used to establish an attention mapping mechanism between the multi-scale anatomical feature map and the graph structure embedding matrix, calculate the semantic relevance weight matrix, and use a graph convolutional network to aggregate, update, and align features to generate node enhancement feature vectors. The key point coordinate regression module is used to input the node enhanced feature vector into the spatial coordinate regression network layer for position decoding and regression calculation, predict the probability distribution heatmap, and perform coordinate correction based on the heatmap peak and geometric topological constraint rules, and output the coordinates of the surgical key points.

Images