Graph-structure geometric algebra deformable registration system for medical images

Abstract

The application discloses a kind of graph structure geometric algebra deformable registration systems for medical image, it is related to medical image processing technical field;Graph aggregation visual encoder carries out feature extraction and conditional position coding to input three-dimensional medical image, then dynamic graph construction and topological feature processing are carried out, GA-SPF receives the feature extracted by graph aggregation visual encoder and obtains the feature tensor after aggregation by single-pass aggregation mechanism and generates channel attention weight intelligent weighting to the geometric information of feature tensor corresponding deformation field, and then convolution is regressed out residual displacement and the deformation field corresponding to feature tensor is superimposed to generate differentiable deformation field, and space transformation module is warped or deformed according to differentiable deformation field to obtain deformation result, then smooth registration image is generated by interpolation image processing technology;The application adopts the above-mentioned graph structure geometric algebra deformable registration system for medical image, and the intelligence and accuracy of the equipment can be improved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a graph-structured geometric algebraic deformable registration system for medical images. Background Technology

[0002] Traditional registration methods, such as SyN, Elastix, and NiftiReg, typically solve for the transform field by iteratively optimizing an energy function. While these methods perform well in terms of registration accuracy, their inherent computationally intensive nature leads to slow processing speeds, severely limiting their practical application in time-sensitive clinical settings. To overcome the limitations of traditional methods, unsupervised registration methods based on deep learning have emerged in recent years. These methods learn deformation maps directly from image pairs, enabling a fast inference process, and their registration accuracy can rival or even surpass that of traditional methods. However, current mainstream deep learning registration methods mainly rely on convolutional neural networks (CNNs) or Transformer architectures. When processing 3D medical images, these architectures typically treat them as regular Euclidean grids or serialized tokens, thus lacking the ability to explicitly model the topological relationships of complex anatomical structures.

[0003] Analysis of the existing technologies reveals the following main problems: Insufficient topology awareness: Neither CNNs nor Transformers effectively and explicitly model the topological relationships between images, making it difficult to guarantee topological consistency in registration when dealing with complex anatomical structures or large deformations. Low decoding efficiency: Existing decoding strategies generally suffer from redundant feature extraction in pyramid structures, inefficient aggregation of multi-level deformation fields, and excessive attention computation overhead. These problems collectively constitute a bottleneck in the registration process. Lack of targeted model design: Many methods directly apply backbone networks designed for other tasks such as segmentation, ignoring the specific requirements of propagating topological cues and enforcing geometric consistency in the registration task, resulting in loose coupling between the encoder and decoder. Summary of the Invention

[0004] The purpose of this invention is to provide a graph-structured geometric algebraic deformable registration system for medical images, which can explicitly model and utilize the topological information of the image, while employing a deep learning registration framework with efficient decoding strategies to generate more accurate, smoother, and anatomically more reasonable deformation fields.

[0005] To achieve the above objectives, the present invention provides a graph-structured geometric algebra deformable registration system for medical images, comprising a graph aggregation visual encoder, a geometric algebraic single-pass aggregation decoder GA-SPF, and a spatial transformation module;

[0006] Given a pair of 3D medical images including floating images and fixed image Input to the graph convergence visual encoder for floating images and fixed image Stepwise feature extraction and conditional location encoding are performed separately. Then, dynamic graphs are constructed based on the features at each scale with location information, thereby generating a corresponding floating image. and fixed image Feature pyramid and ;

[0007] Features at each scale in the feature pyramid extracted by the GA-SPF receiver image aggregation visual encoder , The aggregated feature tensor is obtained through a single-pass aggregation mechanism. According to the feature tensor Based on the geometric information of the deformation field, channel attention weights are generated through geometric perception channel excitation for intelligent weighting, and then a residual displacement is regressed through a convolutional layer. This residual displacement is correlated with the feature tensor. The corresponding deformation fields are superimposed to generate a more accurate differentiable deformation field at the current scale;

[0008] Differentiable deformable field and original floating image output by GA-SPF Together they are sent to the spatial transformation module, which applies the floating image based on the differentiable deformable field. The distortion or deformation is performed to obtain the deformed result, and then a smooth registered image is generated through interpolation image processing techniques. .

[0009] Preferably, the conditional location encoding is set before the dynamic graph construction of the graph aggregation visual encoder. Spatial location prior information is injected into the features extracted at each scale through a depthwise separable convolution, so that the absolute and relative spatial positions of the nodes are perceived during the dynamic graph construction.

[0010] Preferably, the dynamic graph construction process in the graph aggregation visual encoder is as follows:

[0011] S1. For features at each scale in the extracted feature pyramid The image aggregation visual encoder first estimates features using an octet flip contrast method. Distance distribution statistics between nodes in an internal graph, including the mean. and standard deviation ;

[0012] S2. Assume the final output features of the dynamic graph construction. Features are processed along the three principal axes x, y, and z with different step sizes. Perform a scrolling operation to simulate finding nearby neighbors. For each scrolled feature... Calculate the features after rolling Compared with the original features The Euclidean distance between them;

[0013] S3, Set the dynamic threshold as follows If the calculated Euclidean distance is less than the dynamic threshold, the connectivity between the graph nodes at these two locations is preserved and the relative response is calculated. , The value for determining the connection relationship and at this time If the calculated Euclidean distance is not less than the dynamic threshold, then the connectivity between the graph nodes at these two locations is suppressed, and the relative response is calculated. And at this time ;

[0014] S4, obtained from S3 And S2 settings The dynamic graph constructed by performing aggregation response contains features of topological relationships. .

[0015] Preferably, after the dynamic graph of the graph aggregation visual encoder is constructed, information aggregation, refinement, and global context enhancement are performed sequentially. The specific process is as follows:

[0016] Information aggregation is the feature derived from the construction of a dynamic graph that contains topological relationships. Compared with the original features without processing The images are stitched together, taking into account both the new contextual information provided by the dynamic graph and the original features. The detailed textures are included, and the connection relationships between all graph nodes are obtained by gradually increasing the step size multiplier.

[0017] The refining process is through a Convolutional layers project the aggregated features, integrating information between channels, as shown below: ,in To refine the function, This is a concatenation function;

[0018] Global Context Augmentation (FFN) involves feeding the refined features into a feedforward network. FFN learns the nonlinear relationships between different feature channels, deeply integrating local anatomical structures with global spatial context information, and outputting a high-level feature representation that contains both fine local details and global consistency.

[0019] Preferably, the single-pass aggregation mechanism of GA-SPF is as follows:

[0020] The GA-SPF module simultaneously extracts high-level feature representations from multiple different scales in the encoder network in parallel, and adjusts all these high-level feature representations from different scales to a uniform spatial resolution at once through upsampling operations.

[0021] The high-level feature representations, normalized to all dimensions, are concatenated along the channel dimension to form a super feature tensor containing all scale information. This is then processed by a lightweight... The convolutional layer projects the super feature tensor, performing an efficient cross-scale intelligent fusion of information along the channel dimension. It learns to weight and combine information from the global context and local details, ultimately outputting a highly aggregated feature tensor. .

[0022] Preferably, the specific process of GA-eSE excitation by the geometric sensing channel in GA-SPF is as follows:

[0023] feature tensor Corresponding deformation field at the previous scale Normalization yields the local deformation direction vector. u And calculate the characteristic tensor gradient Calculate using geometric product in geometric algebra u and The inner and outer products are then combined with the geometric features obtained through global average pooling. The concatenation process generates the final channel attention weights through a fully connected layer and a sigmoid function. The weighted and refined features are obtained. .

[0024] Preferred features are those obtained after GA-eSE weighted refinement. Through a Convolutional regression produces the residual displacement at the current scale. The specific details of the current-scale deformation field update through field superposition are as follows:

[0025] ;

[0026] in, Provide the initial displacement vector at each spatial location and offer geometric prior knowledge to the attention mechanism, enabling the network to understand geometric deformations.

[0027] Preferably, the interpolation image processing procedure is as follows:

[0028] The pixel values ​​are extracted from the original floating image based on the new coordinates provided by the differentiable deformation field. Since these new coordinates are usually not in integer positions, the interpolation generates a new value by estimating the values ​​of the surrounding known pixels. Finally, all the new values ​​are stitched together to form a smooth registered image.

[0029] Therefore, the graph-structured geometric algebraic deformable registration system for medical images, based on the above-described content, has the following advantages compared to the prior art:

[0030] 1. On multiple publicly available brain and cardiac MRI datasets, the key accuracy metric (Dice similarity coefficient) of this invention surpasses existing mainstream methods. Experimental data shows that it can achieve more accurate anatomical structure alignment; the deformation field has good quality and conforms to anatomical principles. While generating a high-precision deformation field, it can effectively ensure its smoothness and topological rationality, avoiding unrealistic tissue folding and tearing. The quality of the generated deformation field is close to that of the traditional gold standard method with huge computational requirements, and it is more consistent with real biomechanical characteristics.

[0031] 2. This invention, through its innovative graph aggregation encoder and single-pass aggregation decoder design, effectively reduces computational redundancy and memory consumption without sacrificing or even improving performance. Compared to other advanced models, this invention achieves competitive near-real-time processing speed while maintaining high accuracy, achieving a balance between performance and efficiency. The model is highly targeted, with each module working synergistically. It is not simply a stacking of existing modules, but is specifically designed for registration tasks. The graph aggregation encoder is used to perceive topology, and the geometric algebra decoder is used to control deformation. The two work together to improve the problem of loose design and lack of specificity in existing models. Ablation experiments fully demonstrate that each innovative module of this invention (such as dynamic graph construction and geometric perception stimulation) makes a direct and significant contribution to the final excellent results.

[0032] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0033] Figure 1 This is an overall structural diagram of a graph-structured geometric algebraic deformable registration system for medical imaging according to the present invention.

[0034] Figure 2 This is a flowchart of the graph aggregation visual encoder of a graph structure geometric algebra deformable registration system for medical imaging according to the present invention.

[0035] Figure 3 This is a flowchart of the geometric single-pass aggregation decoder of a graph-structured geometric algebraic deformable registration system for medical imaging according to the present invention. Detailed Implementation

[0036] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0037] Example

[0038] like Figures 1-3 As shown, the present invention provides a graph-structured geometric algebra deformable registration system for medical images, comprising a graph aggregation visual encoder, a geometric algebraic single-pass aggregation decoder GA-SPF, and a spatial transformation module.

[0039] Given a pair of 3D medical images including floating images and fixed image Input to the graph convergence visual encoder for floating images and fixed image Stepwise feature extraction and conditional location encoding are performed separately. Then, dynamic graphs are constructed based on the features at each scale with location information, thereby generating a corresponding floating image. and fixed image Characteristic pyramids capable of perceiving anatomical topology and , ;

[0040] Conditional location encoding is set before the dynamic graph construction of the graph aggregation visual encoder. It injects spatial location prior information into the features extracted at each scale through a depthwise separable convolution, so that the absolute and relative spatial positions of the nodes are perceived during dynamic graph construction.

[0041] The graph aggregation visual encoder includes two parallel, structurally identical, and weight-shared dynamic graph construction branches. The dual-stream independent processing design can generate aligned but unique multi-scale feature representations, laying the foundation for accurate correspondence matching in the future.

[0042] The process of constructing a dynamic graph is as follows:

[0043] S1. For features at each scale in the extracted feature pyramid The image aggregation visual encoder first estimates features using an octet flip contrast method. Distance distribution statistics between nodes in an internal graph, including the mean. and standard deviation ;

[0044] S2. Assume the final output features of the dynamic graph construction. Features are processed along the three principal axes x, y, and z with different step sizes. Perform a scrolling operation to simulate finding nearby neighbors. For each scrolled feature... Calculate the features after rolling Compared with the original features The Euclidean distance between them;

[0045] S3, Set the dynamic threshold as follows If the calculated Euclidean distance is less than the dynamic threshold, the connectivity between the graph nodes at these two locations is preserved and the relative response is calculated. , The value for determining the connection relationship and at this time If the calculated Euclidean distance is not less than the dynamic threshold, then the connectivity between the graph nodes at these two locations is suppressed, and the relative response is calculated. And at this time ;

[0046] S4, obtained from S3 And S2 settings The dynamic graph constructed by performing aggregation response contains features of topological relationships. ;

[0047] After the dynamic graph is constructed, information aggregation, refinement, and global context enhancement will be performed sequentially, as follows:

[0048] Information aggregation is the feature derived from the construction of a dynamic graph that contains topological relationships. Compared with the original features without processing The images are stitched together, taking into account both the new contextual information provided by the dynamic graph and the original features. The detailed textures are included, and the connection relationships between all graph nodes are obtained by gradually increasing the step size multiplier.

[0049] The refining process is through a Convolutional layers project the aggregated features, integrating information between channels, as shown below: ,in To refine the function, This is a concatenation function;

[0050] Global Context Augmentation (FFN) involves feeding the refined features into a feedforward network. FFN learns the nonlinear relationships between different feature channels, deeply integrating local anatomical structures with global spatial context information, and outputting a high-level feature representation that contains both fine local details and global consistency.

[0051] The combination of dynamic mapping and FFN refinement ensures that the features extracted by the network during the encoding stage are anatomically meaningful and context-enhanced, providing high-quality, information-dense input for the subsequent decoder to perform accurate correspondence matching and deformation field estimation, fundamentally improving the accuracy of registration. Through this rolling-masking-aggregation mechanism, the graph aggregation visual encoder establishes connections only between regions with similar semantics or appearance, effectively preserving tissue boundaries and capturing long-distance anatomical dependencies, while avoiding the transmission of erroneous information between irrelevant regions. Compared with methods such as k-NN, it has higher computational efficiency, and the generated graph connections are variable, making it more adaptable to morphological differences among different individuals and organs.

[0052] The features extracted by the GA-SPF receiver image aggregation visual encoder are used to obtain the aggregated feature tensor through a one-pass aggregation mechanism. According to the feature tensor Based on the geometric information of the deformation field, channel attention weights are generated through geometric perception channel excitation for intelligent weighting, and then a residual displacement is regressed through a convolutional layer. This residual displacement is correlated with the feature tensor. The corresponding deformation fields are superimposed to generate a more accurate differentiable deformation field at the current scale;

[0053] The single-pass aggregation mechanism is as follows:

[0054] The GA-SPF module simultaneously extracts high-level feature representations from multiple different scales in the encoder network in parallel, and adjusts all these high-level feature representations from different scales to a uniform spatial resolution at once through upsampling operations.

[0055] The high-level feature representations, normalized to all dimensions, are concatenated along the channel dimension to form a super feature tensor containing all scale information. This is then processed by a lightweight... The convolutional layer projects this tensor, performing an efficient cross-scale intelligent fusion of information along the channel dimension. It learns to weight and combine information from the global context (deep features) and local details (shallow features), ultimately outputting a highly aggregated feature tensor. This design greatly reduces computational and memory redundancy, avoids gradient flow problems caused by dense connections, and achieves more efficient feature fusion.

[0056] The specific process of stimulating the GA-eSE via the geometric sensing channel is as follows:

[0057] feature tensor Corresponding deformation field at the previous scale Normalization yields the local deformation direction vector. u And calculate the characteristic tensor gradient Calculate using geometric product in geometric algebra u and The inner and outer products are then combined with the geometric features obtained through global average pooling. The concatenation process generates the final channel attention weights through a fully connected layer and a sigmoid function. ;

[0058] This channel attention weight The weight vector knows the relationship between each feature channel and local geometric deformation (such as rotation and shearing), so it can more intelligently enhance or suppress specific channels and guide the network to generate deformation fields that are more in line with anatomical constraints. This preserves the complete channel dimension and avoids the information loss that may be caused by dimensionality reduction in traditional SE modules.

[0059] Features after GA-eSE weighted refinement Through a Convolutional regression produces the residual displacement at the current scale. The specific details of the current-scale deformation field update through field superposition are as follows:

[0060] ;

[0061] in, Provide the initial displacement vector at each spatial location and provide geometric prior knowledge to the attention mechanism so that the network can understand geometric deformation;

[0062] Differentiable deformable field and original floating image output by GA-SPF Together they are sent to the spatial transformation module, which applies the floating image based on the differentiable deformable field. The distortion or deformation is performed to obtain the deformed result, and then a smooth registered image is generated through interpolation image processing techniques. ;

[0063] The image interpolation process is as follows:

[0064] The pixel values ​​are extracted from the original floating image based on the new coordinates provided by the differentiable deformation field. Since these new coordinates are usually not in integer positions, the interpolation generates a new value by estimating the values ​​of the surrounding known pixels. Finally, all the new values ​​are stitched together to form a smooth registered image.

[0065] In the specific implementation process, firstly, the overall architecture of the GIENet deformable registration system is constructed, given a pair of three-dimensional medical images, i.e., floating images. and fixed image The system first processes the images in parallel using a graph aggregation visual encoder with shared weights. The graph aggregation visual encoder's role is to extract multi-scale features containing rich topological and semantic information from the two images respectively. and , These features are then fed into a geometric algebraic one-way fusion decoder; finally, the spatial transformation module transforms the final deformed field. Applied to the original floating image To obtain a fixed image Aligned and registered images .

[0066] Workflow of an image aggregation visual encoder:

[0067] It begins with a roll operation, a crucial step in generating candidate connections. The entire 3D feature is rolled or shifted along a predefined set of spatial displacement vectors (e.g., translations of different lengths along the x, y, and z axes). Each roll is equivalent to bringing a new candidate neighbor feature with a different location to each voxel in the original feature. By performing multiple rolls in different directions, the system can generate a rich set of candidate neighbors for each voxel, distributed throughout the entire image space, with extremely high efficiency.

[0068] Next, the process enters the masking stage. For each original voxel (corresponding to pixels in a 2D image; in a 3D image, it's a voxel), the algorithm calculates the similarity between its features and the features of multiple candidate neighbors generated after scrolling. Typically, distance, which has low computational cost, is used as the metric. Based on the similarity score, the system performs a screening, retaining only the top K candidate neighbors that are most similar to the current voxel's features, while masking all other neighbors with lower similarity. This step is crucial, as it dynamically constructs a personalized, sparse connection relationship for each voxel, ensuring that information exchange only occurs between anatomically relevant regions, greatly eliminating interference from irrelevant noise.

[0069] Then, the algorithm enters the aggregation stage. At this point, for each voxel, its K most important distant and nearest neighbors have been determined. The system simply sums the feature information of these K selected neighbors to obtain an aggregated feature containing rich long-distance context information. This aggregated feature is then fused with the voxel's original features.

[0070] Ultimately, through this scroll-mask-aggregate pipeline, the module outputs a brand-new, context-enhanced feature. In this new feature, the representation of each voxel is no longer isolated, but cleverly incorporates information about the anatomical structure most similar to it in the entire 3D space, providing high-quality input with a global perspective for subsequent registration tasks.

[0071] The workflow of the geometric algebra single-pass fusion decoder is as follows:

[0072] The GA-SPF module simultaneously extracts features from multiple different scales in the encoder network in parallel, and through upsampling operations, adjusts all these features from different scales to a uniform spatial resolution at once.

[0073] The features, after all dimensions are normalized, are concatenated along the channel dimension to form a super feature tensor containing all scale information. This is then processed by a lightweight... The convolutional layer projects this tensor, performing an efficient cross-scale intelligent fusion of information along the channel dimension. It learns to weight and combine information from the global context (deep features) and local details (shallow features), ultimately outputting a highly aggregated feature tensor. ;

[0074] Geometry-Aware Channel Excitation (GA-eSE) enables the channel attention mechanism to perceive geometric deformation. It not only tells the system which feature channels are important, but also how these channels are related to the current local geometric deformation (such as stretching or rotation). This is the core of the invention of this decoder. When performing channel attention weighting, it not only considers the features themselves, but also introduces the geometric information of the current deformation field.

[0075] First, the current upsampled deformation field Normalization yields the local deformation direction vector u Simultaneously, calculate aggregation features. gradient Then, using the geometric product of Geometric Algebra (GA), the following calculations are performed. u and The inner and outer products are: the inner product (scalar part) represents the projection relationship between the deformation direction and the feature gradient, reflecting the degree of stretching or compression; the outer product (double vector part) represents the directed plane spanned by the deformation direction and the feature gradient, encoding information such as local rotation and shearing.

[0076] Extracting geometric information: Instead of focusing solely on the features themselves, this approach incorporates deformation field information at the current scale, calculating the local deformation direction vectors of two key geometric elements.u and Local deformation direction vector u This represents the direction in which the image is moving at that location; feature gradient. This represents the direction in which the feature information changes the fastest.

[0077] Generating Geometric Descriptors: Using tools from geometric algebra (GA), we combine the two geometric elements mentioned above to generate a new set of descriptors that can quantitatively describe local geometric deformations;

[0078] Scalar descriptor: calculated u and The inner product is obtained, and its physical meaning is to measure the consistency between the deformation direction and the characteristic change direction. If the value is large, it means that a violent stretching or compression has occurred along the deformation direction.

[0079] Double vector descriptor: calculated u and The outer product is obtained, which mathematically defines a directed plane whose physical meaning is to encode rotation and shear information in deformation. For example, it can quantify the extent to which a tissue region is torn during registration.

[0080] Execution channel stimulus: Combine the dynamically calculated geometric descriptor (containing stretching, compression, rotation, and shearing information) with a globally averaged descriptor representing global feature content. To splice,

[0081] This highly informative combined descriptor is fed into a miniature network (fully connected layers and a sigmoid activation function) to generate the final channel attention weights. .

[0082] The spatial transformation module works as follows: GA-SPF outputs a differentiable deformable field and the original floating image. Together they are sent to the spatial transformation module, which applies the floating image based on the differentiable deformable field. The distortion or deformation is performed to obtain the deformed result, and then a smooth registered image is generated through interpolation image processing techniques. .

[0083] Therefore, this invention employs the above-mentioned graph-structured geometric algebraic deformable registration system for medical images, introducing graph neural networks (GNNs) for encoding. It utilizes graph structures to more naturally model 3D volume data and non-Euclidean topology, and adaptively establishes content-related graph structures through a dynamic graph construction (graph aggregation visual encoder) algorithm, thereby enhancing the ability to distinguish fine-grained features and capture topological relationships. It employs geometric algebra for efficient decoding, designing a GA-SPF decoder. This decoder performs multi-scale evidence aggregation only once at each stage and uses the inner-outer product descriptor between the displacement field and feature gradients to drive the geometrically sensed channel excitation, greatly reducing computational redundancy, alleviating feature entanglement, and improving the accuracy and robustness of the final deformation field estimation.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A graph-structured geometric algebraically deformable registration system for medical imaging, characterized in that: This includes a graph aggregation visual encoder, a geometric algebraic single-pass aggregation decoder (GA-SPF), and a spatial transformation module; Given a pair of 3D medical images including floating images and fixed image Input to the graph convergence visual encoder for floating images and fixed image Stepwise feature extraction and conditional location encoding are performed separately. Then, dynamic graphs are constructed based on the location-informed features at each scale, thereby generating corresponding floating images. and fixed image Feature pyramid and ; The dynamic graph construction process in the graph aggregation visual encoder is as follows: S1. For features at each scale in the extracted feature pyramid The image aggregation visual encoder first estimates features using an octet flip contrast method. Distance distribution statistics between nodes in an internal graph, including the mean. and standard deviation ; S2. Along the three principal axes x, y, z, the features are processed with different step sizes. Perform a scrolling operation to simulate finding nearby neighbors. For each scrolled feature... Calculate the features after rolling Compared with the original features The Euclidean distance between them; S3, Set the dynamic threshold as follows If the calculated Euclidean distance is less than the dynamic threshold, the connection between the graph nodes at these two locations is preserved; if the calculated Euclidean distance is not less than the dynamic threshold, the connection between the graph nodes at these two locations is suppressed. S4. Based on the determination of the connection relationship between nodes in graph S3, the features containing topological relationships are finally obtained. Features at each scale in the feature pyramid extracted by the GA-SPF receiver image aggregation visual encoder , The aggregated feature tensor is obtained through a single-pass aggregation mechanism. According to the feature tensor Based on the geometric information of the deformation field, channel attention weights are generated through geometric perception channel excitation for intelligent weighting, and then a residual displacement is regressed through a convolutional layer. This residual displacement is correlated with the feature tensor. The corresponding deformation fields are superimposed to generate a more accurate differentiable deformation field at the current scale; The single-pass aggregation mechanism of GA-SPF is as follows: The GA-SPF module simultaneously extracts high-level feature representations from multiple different scales in the encoder network in parallel, and adjusts all these high-level feature representations from different scales to a uniform spatial resolution at once through upsampling operations. The high-level feature representations, normalized to all dimensions, are concatenated along the channel dimension to form a super feature tensor containing all scale information. This is then processed by a lightweight... The convolutional layer projects the super feature tensor, performing an efficient cross-scale intelligent fusion of information along the channel dimension. It learns to weight and combine information from the global context and local details, ultimately outputting a highly aggregated feature tensor. ; Differentiable deformable field and original floating image output by GA-SPF Together they are sent to the spatial transformation module, which applies the floating image based on the differentiable deformable field. The distortion or deformation is performed to obtain the deformed result, and then a smooth registered image is generated through interpolation image processing techniques. .

2. The graph-structured geometric algebraic deformable registration system for medical imaging according to claim 1, characterized in that: Conditional location encoding is set before the dynamic graph construction of the graph aggregation visual encoder. It injects spatial location prior information into the features extracted at each scale through a depthwise separable convolution, so that the absolute and relative spatial positions of the nodes are perceived during dynamic graph construction.

3. The graph-structured geometric algebraic deformable registration system for medical imaging according to claim 1, characterized in that: After the dynamic graph of the graph aggregation visual encoder is constructed, information aggregation, refinement, and global context enhancement are performed sequentially. The specific process is as follows: Information aggregation is the process of combining the features containing topological relationships obtained after constructing a dynamic graph with the original, unprocessed features. The images are stitched together, taking into account both the new contextual information provided by the dynamic graph and the original features. The detailed textures are included, and the connection relationships between all graph nodes are obtained by gradually increasing the step size multiplier. The refining process is through a Convolutional layers project the aggregated features, integrating information between channels; Global Context Augmentation (FFN) involves feeding the refined features into a feedforward network. FFN learns the nonlinear relationships between different feature channels, deeply integrating local anatomical structures with global spatial context information, and outputting a high-level feature representation that contains both fine local details and global consistency.

4. The graph-structured geometric algebraic deformable registration system for medical imaging according to claim 1, characterized in that: The specific process of GA-eSE excitation by the geometric sensing channel in GA-SPF is as follows: feature tensor Corresponding deformation field at the previous scale Normalization is performed to obtain the local deformation direction vector u, and the characteristic tensor is calculated. gradient Calculate u and u using the geometric product of geometric algebra. The inner and outer products are then combined with the geometric features obtained through global average pooling. The concatenation process generates the final channel attention weights through a fully connected layer and a sigmoid function. The weighted and refined features are obtained. .

5. A graph-structured geometric algebraic deformable registration system for medical imaging according to claim 4, characterized in that: Features after GA-eSE weighted refinement Through a Convolutional regression produces the residual displacement at the current scale. Deformation field at the current scale The specific details of the field overlay update are as follows: ； in, Provide the initial displacement vector at each spatial location and offer geometric prior knowledge to the attention mechanism, enabling the network to understand geometric deformations.

6. The graph-structured geometric algebraic deformable registration system for medical imaging according to claim 1, characterized in that: The image interpolation process is as follows: The pixel values ​​are extracted from the original floating image based on the new coordinates provided by the differentiable deformation field. Since these new coordinates are usually not in integer positions, the interpolation generates a new value by estimating the values ​​of the surrounding known pixels. Finally, all the new values ​​are stitched together to form a smooth registered image.

Images