Automatic detection method of cbct head shadow measurement marker points based on multi-geometry guidance and specific perception coding

Abstract

An automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding includes the following steps: Step S1, downsampling the 3D CBCT image and obtaining a preliminary coordinate set through a coarse localization network; Step S2, cropping image blocks centered on the coordinates and extracting local features through a visual encoder containing a shared basic encoder and a low-rank adapter; Step S3, calculating the relative position matrix of the landmarks, encoding spatial relationships using radial basis functions, and constructing a multi-anatomical heterogeneous map; Step S4, inputting visual and edge features into a multi-geometric guidance Transformer, fusing global constraints and updating features using an attention mechanism; Step S5, extracting directional geometric relationships using spherical harmonic functions to construct higher-order update terms, and dynamically updating the heterogeneous map using a gated residual mechanism; Step S6, predicting coordinate offsets through a multi-layer network and performing iterative optimization to output high-precision 3D coordinates. This method significantly improves detection accuracy and robustness.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and computer vision, specifically involving an automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding. It is used to fuse local visual features extracted from input CBCT images with dynamically constructed multi-anatomical heterogeneous map information, and accurately output high-precision three-dimensional landmark coordinates for orthodontic diagnosis and treatment planning. Background Technology

[0002] Accurate three-dimensional cephalometric landmark detection plays a crucial role in orthodontic diagnosis, treatment planning, and surgical procedures. Compared to two-dimensional lateral cephalometric mapping, which lacks depth information, cone-beam computed tomography (CBCT) provides complete craniofacial anatomical information, facilitating a more comprehensive assessment of bilateral symmetry and skeletal deformities. However, the traditional method of manually identifying landmarks by clinicians is not only time-consuming and laborious, but also prone to inconsistencies in annotation results due to observer variability and subjectivity, thus seriously affecting the reliability of clinical diagnosis and decision-making.

[0003] In recent years, deep learning-based methods have been widely applied to CBCT landmark detection, mainly divided into single-stage models and coarse-to-fine two-stage models. Single-stage models typically predict landmark locations end-to-end from downsampled images. While simple in architecture and fast inference speed, they face significant challenges in handling tiny landmarks in large-volume images and further improving accuracy. Coarse-to-fine two-stage models first perform coarse localization, followed by fine-grained prediction of the region of interest. Although this improves localization accuracy to some extent, the model structure is often more complex, the inference speed is slower, and errors introduced in the coarse localization stage can easily propagate downstream and affect the accuracy of the final result.

[0004] While the aforementioned deep learning methods have propelled the development of 3D landmark detection, existing techniques remain severely limited by the lack of explicit geometric modeling. Most attention-based methods infer structural relationships implicitly from appearance features, failing to fully utilize coordinate-driven cues available during training and inference. In fact, anatomical relationships are highly heterogeneous: constraints in hard tissues tend to be stable and almost rigid, while relationships in soft tissues are more easily deformable. Existing methods typically treat all landmark pairs uniformly, which not only obscures the inherent differences between different tissues and weakens the consistency of the global structure but also makes it difficult for models to capture complex anatomical morphologies with high accuracy. Therefore, there is an urgent need for an automatic landmark detection method based on CBCT cephalometrics that can fully perceive the local morphological differences of specific landmarks and explicitly fuse multiple geometric constraints. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides an automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding, which significantly improves the accuracy and robustness of detection. It integrates local visual features extracted from input CBCT images with dynamically constructed multi-anatomical heterogeneous map information and accurately outputs high-precision three-dimensional landmark coordinates for orthodontic diagnosis and treatment planning.

[0006] The technical solution adopted by this invention to solve its technical problem is: An automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding includes the following steps: Step S1: Obtain the high-resolution 3D CBCT image to be detected, downsample it, and obtain a preliminary set of landmark coordinates through a pre-trained coarse localization network. Step S2: Using the currently predicted coordinates of each marker point as the center, crop a local image patch from the original high-resolution CBCT image; input the image patch into the marker point-aware visual encoder to extract local visual features. The encoder includes a shared basic encoder and multiple low-rank adaptive adapters that correspond one-to-one with each marker point. Step S3: Calculate the relative position matrix between the current markers, and encode the paired spatial relationships using learnable radial basis functions; construct multiple heterogeneous graphs based on preset anatomical prior knowledge, and generate corresponding edge feature matrices; Step S4: Input the local visual features obtained in step S2 and the edge feature matrix obtained in step S3 into the multi-geometry guided Transformer; through the multi-geometry attention mechanism in the model, divide the attention head into multiple groups and make them correspond to different heterogeneous graphs respectively, combine the local visual features and geometric graph features to calculate the attention weight, output the interaction features containing global geometric constraints, and complete the update of the visual features of the marker points. Step S5: Based on the interaction features, the geometric updater in the Transformer extracts the directional geometric relationship between adjacent marker points using the spherical harmonic function and constructs a higher-order geometric update term; the gated residual mechanism is used to inject the update term into the current edge feature matrix, dynamically generating the updated multi-anatomical heterogeneous map, and passing the updated visual features and heterogeneous map to the next network layer; Step S6: After multi-layer network processing, based on the final fused features, the three-dimensional coordinate offset of each marker point is independently predicted by the decoding network and the current coordinates are updated; the updated coordinates are returned as input to step S2 for a preset number of iterative closed-loop optimizations, and finally the high-precision three-dimensional head marker point coordinates are output.

[0007] Furthermore, in step S1, after downsampling the high-resolution 3D CBCT image, a U-Net neural network is used to control memory consumption; the U-Net network extracts features through the downsampling path and predicts the Gaussian heatmap of each marker point, from which a set of coarse marker point coordinates are derived. The coarse positioning coordinates This provides initial coordinate estimation for the subsequent fine-grained localization stage, thereby establishing a two-stage detection framework from coarse to fine.

[0008] Furthermore, in step S2, the marker-aware visual encoder combines a shared base encoder and A different, low-rank adaptive (LoRA) adapter specific to the corresponding marker point; the process of extracting local visual features is as follows: Based on the current predicted coordinates Cropping a fixed-size (e.g., from the original high-resolution image) A local 3D image patch (voxel) is used to generate the first... Feature embedding of image patch ; The encoder training follows a two-stage strategy: In the initial training phase, the parameters of the base encoder and all LoRA adapters are trained end-to-end, enabling the base encoder to establish a robust general feature extraction capability; subsequently, all weights of the base encoder are frozen, and the remaining training phases are dedicated to updating and fine-tuning the landmark-specific LoRA adapters; this mechanism allows the adapters to focus on capturing the unique and subtle morphological features (such as high-frequency details) of each landmark without compromising the general anatomical knowledge acquired by the base encoder; the final output is a set of visual features containing globally robust information and precise local details. .

[0009] Furthermore, in step S3, the relative position tensor between the marker points is first calculated. Each element Indicates the marker point Relative to the marker point The displacement vector; No. A learnable radial basis function (RBF) checklist The role of the marginal vector norm in the equation is expressed as: ; in, Indicates application to The element-wise Euclidean norm of the mid-edge displacement vector; and These are learnable parameters that control the width and center of the basis functions, respectively. This is a smooth cosine truncation function used to ensure the numerical stability of the model when dealing with long-distance node interactions; Combination Using different RBF kernels, a shared intermediate feature representation is constructed: ; Based on this, four predefined heterogeneous graphs were constructed using prior anatomical knowledge. , respectively representing hard tissue connectivity ( ), soft tissue connections ( ), Inter-organizational connections ( ) and fully connected structure ( ).

[0010] The final edge feature matrix for each graph The calculation formula is: ; in, The weight projection matrix is ​​a learnable matrix. Given a binary adjacency matrix, the Hadamard product operation is performed. Perform masking to ensure it is not within a predefined edge set. The feature vectors corresponding to unconnected edges in the model are strictly transformed into zero vectors, thus explicitly injecting topological constraints.

[0011] In step S4, the multi-geometry guided Transformer includes a multi-geometry attention mechanism, which has Each attention point is divided into 4 groups to correspond to different geometric figures. Each attention head only uses the edge feature matrix of its corresponding heterogeneous graph. To calculate attention weights instead of simple feature concatenation; for the first... The iteration process The nth Transformer block, the first The calculation process for each attention head is as follows: Similarity scores are calculated based on node visual features. : ; in, and These are the learnable query and key projection matrices, respectively. Represents a batch tensor product; Then, specific geometric constraint information is fused, and the matrix is ​​normalized using a normalized exponential function (Softmax function) to calculate the final attention weight matrix. : ; in, As a learnable geometric projection matrix, this step deeply integrates the geometric spatial relationships of the markers into the allocation of visual attention weights; Calculate the first The output interaction features of each attention head : ; in, Represented by value; To dynamically fuse attention head information from different heterogeneous graphs, a small multilayer perceptron (MLP) is used to generate node-adaptive head fusion weights. : ; The outputs of each attention head are weighted and fused to form a fused interactive feature that includes multi-head structural information. : ; in, For a specific linear output projection layer Indicates the first Each attention head dynamically assigns adaptive weights to each marker point; Finally, all paired interaction features are aggregated along the source node dimension to obtain aggregated node features containing rich context and multiple geometric guidance. : ; in, Indicates all Individual source marker Aggregate information and aggregate node features. The network then sequentially passes through residual connections, layer normalization, a feedforward neural network (FFN), and further residual connections and layer normalization, ultimately forming the updated visual features of the landmark points. .

[0012] In step S5, the multi-geometry-guided Transformer further includes a geometry updater for dynamically evolving complex anatomical topological relationships between multi-layer networks. The process is as follows: First, the interaction features output based on the multi-geometric attention mechanism. A preliminary geometric descriptor is generated for each marker point using spherical harmonic functions. To capture the directional geometric relationship of the contribution of adjacent markers to it: ; in, It is a learnable linear feature projection matrix; Representation degree is And the order is The spherical harmonic function is used to represent the relative direction vector. Decomposed into basic directional components; Subsequently, an update term describing the higher-order geometric interactions between marker points is constructed based on the geometric descriptor. : ; in, There are two independent sets of learnable projection matrices; This represents the tensor product operation; This represents the summation of the components of the spherical harmonic function; Finally, using the gated residual mechanism, the update term is injected into the edge feature matrix of the corresponding heterogeneous graph. In the middle, the dynamic update of the geometric edge features is completed: ; in, It is a sigmoid linear unit activation function, used as a dynamic gate to adjust the geometric interaction information of newly computed data. The degree of integration, the updated edge feature matrix With updated visual features Together they form the input to the next layer of Transformer blocks.

[0013] In step S6, the decoding network includes Each independent linear layer maps the updated feature vector of its corresponding marker point to independently predict its 3D coordinate offset. ; The coordinate update of the marker point is accomplished by accumulating the predicted offset to the coordinates of the previous iteration: ; During the training phase, targeting the total Supervised optimization is performed on each intermediate output of the nth iteration. The coordinates output by the next iteration Using the L1 loss function Calculate its coordinates with those of actual manually labeled coordinates. The absolute error between them: ; in, For the first The true three-dimensional coordinates of each marker point, after being set. After closed-loop iterative optimization (e.g.) The model outputs a set of high-precision landmark coordinates that ultimately approximate the actual anatomical location. .

[0014] The beneficial effects of this invention are mainly reflected in the following aspects: By constructing a detection framework based on multi-geometric guidance and specific perceptual coding, local visual features and global geometric constraints are cleverly integrated; by utilizing a landmark-aware visual encoder containing a shared basic encoder and a specific low-rank adaptive adapter, the model can capture the subtle morphological features of each specific anatomical landmark in a highly targeted manner, achieving accurate extraction of high-frequency details without destroying general anatomical knowledge; at the same time, by constructing multiple heterogeneous maps based on anatomical prior knowledge and combining a multi-geometric-guided Transformer to explicitly and globally model the heterogeneous spatial geometric relationships between different tissues, the problem of weakening structural consistency caused by uniformly processing all landmark pairs in traditional methods is effectively overcome; in addition, an iterative closed-loop optimization mechanism is introduced, enabling the model to continuously integrate visual and geometric information and gradually correct coordinate deviations in multiple iterations, achieving high-precision localization of complex anatomical structures. This invention effectively solves the problems of traditional manual annotation of cephalometric landmarks being time-consuming, labor-intensive, subjective, and prone to inconsistencies. It also overcomes the shortcomings of existing deep learning methods that lack explicit geometric modeling, significantly improving the accuracy and robustness of detection. It provides a highly accurate, automated three-dimensional cephalometric landmark detection solution with extremely high clinical diagnostic consistency for orthodontic diagnosis, treatment planning, and surgical procedures. Attached Figure Description

[0015] Figure 1 This is an overall flowchart of an automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding.

[0016] Figure 2 This is a schematic diagram of the network structure for predicting the initial coordinates of the marker points in the coarse stage of this application.

[0017] Figure 3 This is a schematic diagram of the structure of the marker-aware visual encoder of this application.

[0018] Figure 4 This is a schematic diagram illustrating the process of extracting multiple geometric relationships and constructing multiple anatomical heterogeneous maps in this application.

[0019] Figure 5 This is a schematic diagram of the structure of the multi-geometry guided Transformer encoder module of this application. Detailed Implementation

[0020] The present invention will now be further described with reference to the accompanying drawings.

[0021] Reference Figures 1-5 An automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding includes the following steps: Step S1: Obtain the high-resolution 3D CBCT image to be detected, downsample it, and obtain a preliminary set of landmark coordinates through a pre-trained coarse localization network.

[0022] In one specific embodiment, combined with Figure 2 The diagram shows the network structure for predicting the initial coordinates of marker points in the coarse-stage prediction. Step S1 includes the following sub-steps: Step 1.1: After downsampling the high-resolution 3D CBCT image, use the U-Net network to control memory consumption; Step 1.2: The U-Net network extracts features through downsampling paths and predicts the Gaussian heatmap of each marker point, from which a set of coarse marker point coordinates are derived. ; Step 1.3, the coarse positioning result This provides initial coordinate estimation for the subsequent fine-grained localization stage, thereby establishing a two-stage detection framework from coarse to fine.

[0023] Step S2: Using the currently predicted coordinates of each marker point as the center, crop a local image block from the original high-resolution CBCT image; input the image block into the marker point perception visual encoder to extract local visual features.

[0024] In one specific embodiment, combined with Figure 3 The diagram shown illustrates the structure of a marker-aware visual encoder, which combines a shared base encoder and... A different, low-rank adaptive adapter specific to the corresponding marker point; the process of extracting local visual features includes the following sub-steps: Step 2.1: Based on the current predicted coordinates A fixed-size local 3D image patch is cropped from the original high-resolution image to generate the 3D image patch. Feature embedding of image patch ; Step 2.2: The training of the encoder follows a two-stage strategy: In the initial training stage, the parameters of the base encoder and all LoRA adapters are trained end-to-end to enable the base encoder to establish a strong general feature extraction capability. Step 2.3: Subsequently, freeze all weights of the base encoder and spend the remaining training phase updating and fine-tuning the landmark-specific LoRA adapter, allowing the adapter to focus on capturing the unique and subtle morphological features of each landmark. Step 2.4: The final output is a set of visual features containing global robust information and precise local details. .

[0025] Step S3: Calculate the relative position matrix between the current markers, and encode the paired spatial relationships using learnable radial basis functions; construct multiple heterogeneous graphs based on preset anatomical prior knowledge, and generate corresponding edge feature matrices.

[0026] In one specific embodiment, combined with Figure 4 The flowchart shown illustrates the extraction of multiple geometric relationships and the construction of multiple anatomical heterogeneous maps. Step S3 includes the following sub-steps: Step 3.1: First, calculate the relative position tensor between the marker points. Each element Indicates the marker point Relative to the marker point The displacement vector; Step 3.2, the A learnable radial basis function kernel The role of the marginal vector norm in the equation is expressed as: ; in, Indicates application to The element-wise Euclidean norm of the mid-edge displacement vector; and These are learnable parameters; It is a smooth cosine cutoff function; Step 3.3, Combining Using different RBF kernels, a shared intermediate feature representation is constructed: ; in, This represents a concatenation operation along the feature dimension; Step 3.4: Based on this, construct four predefined heteromorphic maps using prior anatomical knowledge. , respectively representing hard tissue connectivity, soft tissue connectivity, inter-tissue connectivity, and fully connected structures; Step 3.5: Final edge feature matrix for each graph The calculation formula is: ; in, Indicates the first The edge feature dimension of a heterogeneous graph; The weight projection matrix is ​​a learnable matrix. Given a binary adjacency matrix, the Hadamard product operation is performed. Perform masking to ensure that the feature vectors corresponding to unconnected edges that are not in the predefined edge set are strictly reduced to zero vectors.

[0027] Step S4: Input the local visual feature and edge feature matrix into the multi-geometry guided Transformer; calculate the attention weight through the multi-geometry attention mechanism, output the interaction features containing global geometric constraints, and complete the update of visual features.

[0028] In one specific embodiment, combined with Figure 5 The diagram shows the structure of the encoder module in a multi-geometry guided Transformer. Step S4 includes the following sub-steps: Step 4.1: Calculate similarity scores based on node visual features. : ; in, and These are the learnable query and key projection matrices, respectively. This indicates a batch tensor product operation; This represents the matrix transpose operation; Step 4.2: Subsequently, specific geometric constraint information is fused and normalized using the Softmax function to calculate the final attention weight matrix. : ; in, Represents the feature dimensions of each attention head; It is a learnable geometric projection matrix; Indicates along the feature dimension Perform a summation operation; This indicates the Hadamard product operation; Step 4.3: Calculate attention head Output interaction features Based on the calculated attention weight matrix and the corresponding value. The output of each attention point is calculated using the following formula: ; Among them, through the Hadama product operation Generate a [database name] for each pair of markers. Dimensional interactive feature vectors; Step 4.4: Generate node-adaptive head mixing weights using a small multilayer perceptron (MLP). : ; Step 4.5: Weighted fusion of the outputs of each attention head to form a fused interactive feature. : ; in, Indicates the first Each attention head dynamically assigns adaptive weights to each marker point; The output projection matrix is ​​linear; Step 4.6: Aggregate interaction features along the source node dimension to obtain aggregated node features. : ; in, Indicates all Individual source marker Aggregate information; Step 4.7, Aggregation Features The updated visual features of the marker points are then sequentially processed through residual connections, layer normalization, feedforward neural networks, and further residual connections and layer normalization to form the updated visual features. .

[0029] Step S5: Extract the directional geometric relationship between adjacent marker points using spherical harmonic functions and construct a higher-order geometric update term; dynamically generate the updated multi-anatomical heterogeneous map using a gated residual mechanism and pass it to the next network layer.

[0030] In one specific embodiment, continue to refer to Figure 5 Step S5 includes the following sub-steps: Step 5.1: Based on interaction features A preliminary geometric descriptor is generated for each marker point using spherical harmonic functions. : ; in, This indicates that for all source markers The contributions are aggregated and calculated; It is a learnable linear feature projection matrix; Representation degree is And the order is The spherical harmonic function, which acts on the unit direction vector between the marker points. superior; Step 5.2: Construct an update term based on the geometric descriptor to describe the higher-order geometric interactions between marker points. : ; in, There are two independent sets of learnable projection matrices; This represents the tensor product operation; Represents all spherical harmonic components Perform summation; Step 5.3: Using the aforementioned multi-graph construction method, employing four independent projection layers and corresponding masks, the update item... Mapped and separated into specific higher-order geometric update terms corresponding to four predefined heterogeneous graphs. ; Step 5.4: Combining the gating residual mechanism, the specific update item... Inject the edge feature matrix of the corresponding heterogeneous graph In the middle, the dynamic update of the geometry is completed: ; in, For the first The learnable weight projection matrix corresponding to each heterogeneous graph; It is an sigmoid linear unit activation function, used here as a dynamic gate.

[0031] Step S6: After multi-layer network processing, independently predict the 3D coordinate offset of each marker point based on the final fused features and update the current coordinates; return to step S2 to perform a preset number of iterative closed-loop optimizations and output high-precision coordinates.

[0032] In one specific embodiment, step S6 includes the following sub-steps: Step 6.1, the decoding network passes through Each independent linear layer predicts the 3D coordinate offset. .

[0033] Step 6.2, the formula for updating the coordinates of the marker point is: ; Step 6.3: During the training phase, use the L1 loss function to calculate the predicted values ​​and the true labeled values. The absolute error between them: ; in, The L1 norm of a vector; For the first The true three-dimensional coordinates of each marker point; Step 6.4, after setting After closed-loop iterative optimization, the final high-precision set of marker point coordinates is output. .

[0034] The high-precision three-dimensional cephalometric landmark coordinates output in this embodiment provide a precise digital representation of the craniofacial skeletal and soft tissue anatomy in the patient's CBCT images. This landmark coordinate set can be directly used for orthodontic diagnosis, treatment planning, and three-dimensional craniofacial morphology analysis. Through high-precision coordinate positioning, the clinical system can automatically calculate various cephalometric indices (such as bone angles and line distances), not only assisting doctors in comprehensively assessing the patient's bilateral symmetry and degree of skeletal deformity, but also providing a reliable three-dimensional geometric basis for subsequent orthognathic surgery simulation and efficacy prediction. This automated landmark detection design frees the complex CBCT cephalometric process from reliance on subjective human experience, significantly improving the standardization and consistency of clinical diagnosis.

[0035] This invention cleverly integrates local visual features with global geometric constraints by constructing a detection framework based on multi-geometric guidance and specific perceptual encoding. On one hand, it utilizes a landmark-aware visual encoder containing a specific low-rank adaptive adapter to accurately capture the subtle morphological features unique to each anatomical landmark, effectively addressing the heterogeneity of soft and hard tissues in images. On the other hand, it explicitly constructs a multi-anatomical heterogeneous map through a multi-geometric-guided Transformer, performing multi-order geometric reasoning and dynamic updates on the global topological constraints between landmarks. Coupled with an iterative closed-loop optimization mechanism from coarse to fine, this invention significantly improves the localization accuracy of minute landmarks and the consistency of the global structure, effectively solving the problems of time-consuming, labor-intensive, and inconsistent traditional manual annotation. This provides an efficient and robust underlying technical support for orthodontic diagnosis and intelligent digital oral healthcare.

[0036] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A CBCT cephalometric landmark automatic detection method based on multi-geometry guidance and specific perceptual coding, characterized in that, The method includes the following steps: Step S1: Obtain the high-resolution 3D CBCT image to be detected, downsample it, and obtain a preliminary set of landmark coordinates through a pre-trained coarse localization network. Step S2: Using the currently predicted coordinates of each marker point as the center, crop a local image patch from the original high-resolution CBCT image; input the image patch into the marker point-aware visual encoder to extract local visual features. The encoder includes a shared basic encoder and multiple low-rank adaptive adapters that correspond one-to-one with each marker point. Step S3: Calculate the relative position matrix between the current markers, and encode the paired spatial relationships using learnable radial basis functions; construct multiple heterogeneous graphs based on preset anatomical prior knowledge, and generate corresponding edge feature matrices; Step S4: Input the local visual features obtained in step S2 and the edge feature matrix obtained in step S3 into the multi-geometry guided Transformer; through the multi-geometry attention mechanism in the model, divide the attention head into multiple groups and make them correspond to different heterogeneous graphs respectively, combine the local visual features and geometric graph features to calculate the attention weight, output the interaction features containing global geometric constraints, and complete the update of the visual features of the marker points. Step S5: Based on the interaction features, the geometric updater in the Transformer extracts the directional geometric relationship between adjacent marker points using the spherical harmonic function and constructs a higher-order geometric update term; the gated residual mechanism is used to inject the update term into the current edge feature matrix, dynamically generating the updated multi-anatomical heterogeneous map, and passing the updated visual features and heterogeneous map to the next network layer; Step S6: After multi-layer network processing, based on the final fused features, the three-dimensional coordinate offset of each marker point is independently predicted by the decoding network and the current coordinates are updated; the updated coordinates are returned as input to step S2 for a preset number of iterative closed-loop optimizations, and finally the high-precision three-dimensional head marker point coordinates are output.

2. The automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding as described in claim 1, characterized in that, In step S1, after downsampling the high-resolution 3D CBCT image, a U-Net neural network is used to control memory consumption. The U-Net network extracts features through the downsampling path and predicts the Gaussian heatmap of each marker point, from which a set of coarse marker point coordinates are derived. These coarse localization coordinates provide initial coordinate estimates for the subsequent fine-grained localization stage, thereby establishing a two-stage detection framework from coarse to fine.

3. The automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding as described in claim 1 or 2, characterized in that, In step S2, the marker-aware visual encoder combines a shared base encoder and A different, low-rank adaptive LoRA adapter specific to the corresponding marker point; the process of extracting local visual features is as follows: Based on the current predicted coordinates, a fixed-size local 3D image patch is cropped from the original high-resolution image to generate the first... Feature embedding of image patches; The encoder training follows a two-stage strategy: in the initial training stage, the parameters of the base encoder and all LoRA adapters are trained end-to-end; then, all weights of the base encoder are frozen, and the remaining training stages are dedicated to updating and fine-tuning the landmark-specific LoRA adapters; the final output is a set of visual features containing global robust information and precise local details.

4. The automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding as described in claim 1 or 2, characterized in that, In step S3, the relative position tensor between the marker points is first calculated; The effect of a learnable radial basis function (RBF) kernel on the norm of the edge vectors in the relative position tensor, combined with We construct shared intermediate features using different RBF kernels and build four predefined heterogeneous graphs based on anatomical prior knowledge, representing hard tissue connectivity, soft tissue connectivity, inter-tissue connectivity, and fully connected structures, respectively; and calculate the final edge feature matrix for each graph.

5. The automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding as described in claim 1 or 2, characterized in that, In step S4, the multi-geometry guided Transformer includes a multi-geometry attention mechanism, which has Each attention head is divided into four groups to correspond to different geometric graphs. Each attention head uses only the edge feature matrix of its corresponding heterogeneous graph to calculate the attention weight, instead of simple feature concatenation.

6. The automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding as described in claim 5, characterized in that, In step S4, for the first The iteration process The nth Transformerblock, the th The calculation process for each attention head is as follows: The similarity score is calculated based on the visual features of the nodes, then specific geometric constraint information is fused, and normalized using a normalized exponential function to calculate the final attention weight matrix. The nth... The output interaction features of each attention head; By generating adaptive head mixing weights for nodes using a small multilayer perceptron, the outputs of each attention head are weighted and fused to form a fused interactive feature containing multi-head structural information. Finally, all paired interaction features are aggregated along the source node dimension to obtain aggregated node features containing rich context and multiple geometric guidance. The aggregated node features are then passed through residual connections, layer normalization, feedforward neural networks, and residual connections and layer normalization again to finally form the updated visual features of the landmark points of the network layer.

7. The automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding as described in claim 1 or 2, characterized in that, In step S5, the multi-geometry-guided Transformer further includes a geometry updater for dynamically evolving complex anatomical topological relationships between multi-layer networks. The process is as follows: First, based on the interaction features output by the multi-geometric attention mechanism, a preliminary geometric descriptor is generated for each marker point using a spherical harmonic function to capture the directional geometric relationship of the contribution of adjacent marker points to it. Subsequently, an update term is constructed based on the geometric descriptor to describe the higher-order geometric interactions between the marker points; Finally, by combining the gated residual mechanism, the update term is injected into the edge feature matrix of the corresponding heterogeneous graph to complete the dynamic update of the geometric graph.

8. The automatic detection method for CBCT cephalometric landmarks based on multi-geometric guidance and specific perceptual coding as described in claim 1 or 2, characterized in that, In step S6, the decoding network includes Each independent linear layer maps the updated feature vector of its corresponding marker point to independently predict its 3D coordinate offset. The coordinate update of the marker point is accomplished by accumulating the predicted offset to the coordinates of the previous iteration. During the training phase, for the total... Supervised optimization is performed on each intermediate output of the nth iteration. The coordinates output in the next iteration are used to calculate the absolute error between them and the actual manually labeled coordinates using the L1 loss function, after setting... After closed-loop iterative optimization, the model outputs a set of high-precision marker coordinates that ultimately approximate the actual anatomical location.

Images