A Deformable Medical Image Registration Method Based on Dynamic Graph Convolutional Attention
By using shared-weight dual-stream feature extraction and multi-stage deformation modeling of the dynamic graph convolutional attention module, this method solves the problem that existing medical image registration methods struggle to balance global and local feature modeling, achieving efficient and accurate medical image registration applicable to scenarios such as clinical diagnosis and image-guided surgery.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-02-11
- Publication Date
- 2026-05-26
AI Technical Summary
Existing medical image registration methods struggle to balance global dependencies with local detail modeling, resulting in high computational costs and a lack of multi-stage deformation strategies, which impacts registration accuracy and stability, especially in scenarios with large deformations.
A dual-stream feature extraction encoder with shared weights is used to obtain multi-level features. Combined with a dynamic feature fusion module and a dynamic graph convolutional attention module, the deformation field is optimized through multi-stage deformation modeling. The dynamic graph convolutional attention module is used to capture global and local feature information to generate an accurate and stable deformation field.
It significantly improves the accuracy of medical image registration and the ability to align fine-grained anatomical structures, reduces computational load and memory usage, enhances stability under large deformation scenarios and the physical rationality of deformation fields, and provides a more efficient and reliable medical image registration solution.
Smart Images
Figure CN122089796A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image processing technology, specifically relating to a deformable medical image registration method based on dynamic graph convolutional attention. Background Technology
[0002] Medical image registration is a core technology in the field of medical image processing, with indispensable application value in clinical scenarios such as computer-aided diagnosis, image-guided surgery, organ atlas construction, and tumor growth monitoring. Its core objective is to achieve precise alignment of medical images acquired at different times, angles, or modalities through spatial transformation, integrating multi-source medical image information to present clinicians with more comprehensive and accurate anatomical structures and lesion information. This supports accurate disease diagnosis, treatment plan formulation, and efficacy evaluation. Deformable medical image registration, due to its ability to adapt to the nonlinear deformation of human anatomical structures, has become a research focus in this field.
[0003] In existing technologies, medical image registration methods are mainly divided into traditional iterative optimization methods and deep learning-driven registration methods. Traditional methods treat the registration process as an iterative optimization problem, continuously adjusting the deformation field to match the transformed moving image with the fixed image. However, these methods heavily rely on iterative solutions, resulting in high computational costs and long processing times, and they also have significant limitations in characterizing complex nonlinear deformations. Deep learning methods have significantly improved registration efficiency. Among them, convolutional neural network-based methods achieve fast registration through hierarchical feature extraction, but they lack the ability to model global spatial dependencies of images. Transformer-based methods effectively capture long-distance dependencies with their self-attention mechanism, resulting in better registration accuracy. However, the secondary computational complexity of the self-attention mechanism leads to extremely high computational and memory usage, making it difficult to adapt to large-scale 3D medical images. Some hybrid architectures attempt to combine the advantages of Transformer and convolutional neural networks, introducing convolutional operations in the decoder to enhance local feature extraction. However, they cannot maintain the integrity of both global and local features simultaneously. In addition, existing deep learning methods often use simple feature stitching to fuse moving and fixed image features, lacking explicit enhancement for similar anatomical structures. Furthermore, some methods directly regress the final deformation field at the decoder end, failing to fully utilize the progressive registration strategy from coarse to fine, resulting in poor stability and accuracy in large deformation scenarios.
[0004] Despite continuous technological improvements, deformable medical image registration still faces several pressing technical bottlenecks: First, it's difficult to balance global dependencies with local detail modeling; existing methods struggle to simultaneously capture effective global spatial relationships and accurately depict fine-grained anatomical structures. Second, the computational cost of global feature extraction is too high, limiting the feasibility of such methods in clinical applications of 3D medical images. Third, similar structural regions in moving and stationary images are not sufficiently enhanced; simple feature fusion methods reduce the network's focus on key anatomical regions, affecting registration accuracy. Fourth, the lack of effective multi-stage deformation modeling strategies makes it difficult to maintain registration stability and accuracy in scenarios with large deformations. These problems make it difficult for existing registration methods to achieve a balance between accuracy, efficiency, and robustness; therefore, a novel deformable medical image registration method is urgently needed to address these issues. Summary of the Invention
[0005] To address the problems existing in the background art, one aspect of the present invention provides a deformable medical image registration method based on dynamic graph convolutional attention, characterized in that it includes:
[0006] S1: Acquire fixed medical images and mobile medical images to be registered, and use a shared-weight dual-stream feature extraction encoder to extract multi-level hierarchical features from the mobile medical images and fixed medical images respectively.
[0007] S2: For mobile and stationary medical images, a dynamic feature fusion module and a dynamic graph convolutional attention module are constructed for the hierarchical features at each level, respectively.
[0008] S3: After applying the coarse deformation field of the previous level to the hierarchical features of the current level moving medical image, the dynamic feature fusion module corresponding to the hierarchical feature input of the current level fixed medical image obtains the initial fused feature representation of the current level.
[0009] S4: Extract the combined global and local feature information from the intermediate fusion feature representation of the previous layer and the initial fusion feature representation of the current layer through the corresponding dynamic graph convolutional attention module to obtain the intermediate fusion feature representation of the current layer;
[0010] S5: For the intermediate fusion feature representation of the current level, the residual coarse deformation field of the current level is generated by the deformable registration head. The coarse deformation field of the previous level and the residual coarse deformation field of the current level are deformed and fused to obtain the coarse deformation field of the current level.
[0011] S6: The coarse deformation field of the last level is used as the final deformation field. The moving medical image is spatially transformed using the final deformation field to obtain the registered medical image.
[0012] Another aspect of the present invention provides a deformable medical image registration system based on dynamic graph convolutional attention, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the deformable medical image registration method based on dynamic graph convolutional attention.
[0013] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements a deformable medical image registration method based on dynamic graph convolutional attention as described above.
[0014] The present invention has at least the following beneficial effects
[0015] This invention efficiently acquires multi-level features from both moving and stationary medical images using a weighted dual-stream feature extraction encoder. A dynamic feature fusion module adaptively weights and fuses key structural features from both images to enhance the representation of similar anatomical regions. A dynamic graph convolutional attention module accurately captures the combined features of global spatial dependencies and local details. A multi-stage deformation modeling strategy, from coarse to fine, iteratively optimizes the fusion of the previous-level coarse deformation field and the current-level residual coarse deformation field, ultimately generating a precise and stable deformation field. This significantly improves the accuracy of deformable medical image registration and the ability to align fine-grained anatomical structures, effectively reducing model computation and memory usage. It also enhances the stability of registration and the physical rationality of the deformation field in scenarios with large deformations, providing a more efficient and reliable medical image registration solution for clinical diagnosis, image-guided surgery, and other scenarios. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the method flow framework of the present invention;
[0017] Figure 2 This is a schematic diagram of the dual-stream feature extraction encoder of the present invention;
[0018] Figure 3 This is a schematic diagram of the dynamic feature fusion module of the present invention;
[0019] Figure 4 This is a schematic diagram of the structure of the dynamic graph convolutional attention module of the present invention. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] Please see Figure 1 One aspect of the present invention provides a deformable medical image registration method based on dynamic graph convolutional attention, characterized in that it includes:
[0022] S1: Acquire fixed medical images and mobile medical images to be registered, and use a shared-weight dual-stream feature extraction encoder to extract multi-level hierarchical features from the mobile medical images and fixed medical images respectively.
[0023] Preferably, the dual-stream feature extraction encoder includes two feature extraction branches with shared weights. Each feature extraction branch includes five convolutional blocks cascaded sequentially, with max pooling layers between adjacent convolutional blocks for downsampling. Each convolutional block contains two convolutional layers, a LeakyReLU activation function, and an instance normalization layer cascaded sequentially. The output features of all convolutional blocks in the two feature extraction branches are extracted and arranged in ascending order of scale to obtain a hierarchical feature set of the mobile medical image. and hierarchical feature set of fixed medical images ,feature and Representing mobile medical images and fixed medical images Hierarchical features at each level, .
[0024] Please see Figure 2In this embodiment, three publicly available brain MRI datasets—LPBA, Mindboggle, and IXI—were used. The LPBA dataset contains 40 T1-weighted MRI volumes (30 for training and 10 for testing), all pre-aligned to MNI305 space and cropped to 160×192×160 (voxel size 1mm×1mm×1mm). The Mindboggle dataset contains 101 MRI volumes (42 for training and 20 for testing), affine-aligned to MNI152 space and cropped to 160×192×160 (voxel size 1mm×1mm×1mm). The IXI dataset contains 576 T1-weighted MRI volumes (403 for training, 58 for validation, and 115 for testing), with a cropped size of 160×192×160. The dataset is 92×224, and all datasets were preprocessed to fit the model input. The shared-weight dual-stream feature extraction encoder used contains two shared-weight feature extraction branches. Each branch consists of five concatenated convolutional blocks. A 2×2×2 max-pooling layer with a stride of 2 is used for downsampling between adjacent convolutional blocks. Each convolutional block contains two convolutional layers with a kernel size of 3×3×3, a stride of 1, and padding of 1, a LeakyReLU activation function, and an instance normalization layer. During operation, fixed and mobile medical images to be registered are first acquired. Then, features are extracted simultaneously from both types of images through the two shared-weight feature extraction branches. The output features of the five convolutional blocks in each branch are retained during extraction and arranged in ascending order of feature scale, ultimately yielding a hierarchical feature set for the mobile medical images. and hierarchical feature set of fixed medical images ;in and Corresponding to the two types of images respectively Hierarchical features at each level, The higher the level, the larger the feature scale and the richer the semantic information.
[0025] S2: For mobile and stationary medical images, a dynamic feature fusion module and a dynamic graph convolutional attention module are constructed for the hierarchical features at each level, respectively.
[0026] S3: After applying the coarse deformation field of the previous level to the hierarchical features of the current level moving medical image, the dynamic feature fusion module corresponding to the hierarchical feature input of the current level fixed medical image obtains the initial fused feature representation of the current level.
[0027] Please see Figure 3 Preferably, step S3 includes:
[0028] S31: The coarse deformation field of the previous level The displacement amplitude is magnified by 2 times, and then the deformation field is obtained by trilinear interpolation upsampling. ;
[0029] S32: Based on the deformation field Using spatial transformation networks Hierarchical features of the current level of mobile medical images Features are obtained by deformation Then, the deformed hierarchical features Hierarchical features of the current level fixed medical image The corresponding dynamic feature fusion module is input to obtain the initial fused feature representation of the current level. .
[0030] Preferably, the dynamic feature fusion module includes: a first convolutional layer, a second convolutional layer, and a weight branch, wherein the weight branch includes a third convolutional layer, a fourth convolutional layer, and a sigmoid activation function cascaded sequentially; [the module then integrates the features...] and characteristics Features are obtained by inputting them into the first convolutional layer and the second convolutional layer respectively. and characteristics ; Features and characteristics After concatenation, the input to the weight branch is processed to obtain the weighted feature map. The weight branch maps the number of channels of the features to 2 through the third and fourth convolutional layers, and generates the weighted feature map through the Sigmoid activation function. The weighted feature map is then split into two weight matrices according to the number of channels. and ;Will With features Matrix multiplication yields features ;Will With features Matrix multiplication yields features ; Features and characteristics The features are summed to obtain the initial fused feature representation of the current level. .
[0031] In this embodiment, the coarse deformation field of the previous level is... The displacement amplitude is magnified by 2 times, and upsampling is performed through trilinear interpolation (interpolation kernel size is 3×3×3, sampling step size is 1) to obtain a deformation field adapted to the current level feature scale. The Spatial Transformation Network (STN) employs a classic architecture of localization network + mesh generator + sampler. The localization network consists of two cascaded 3×3×3 convolutional layers (stride 1, padding 1), a LeakyReLU activation function, and an instance normalization layer, used to analyze the deformation field. The spatial transformation parameters are used by the mesh generator to generate a sampling mesh. The sampler employs a bilinear interpolation strategy to spatially deform the feature map, achieving precise feature alignment. Subsequently, based on the deformation field... Using this spatial transformation network (STN) to analyze the hierarchical features of the current level of mobile medical images. Perform deformation to obtain aligned features The dynamic feature fusion module includes a first convolutional layer, a second convolutional layer, and a weight branch. The first and second convolutional layers are both 3×3×3 convolutions (stride 1, padding 1), which fuse features... and characteristics Each is mapped to a feature with the same number of channels. and characteristics The weight branch consists of a cascaded third convolutional layer (3×3×3, stride 1, padding 1), a fourth convolutional layer (1×1×1, stride 1, no padding), and a sigmoid activation function. First, the features... and characteristics The data is concatenated along the channel dimension and then processed by weighted branching to generate a 2-channel weighted feature map, which is then split into a weight matrix according to the channel. and Then With features , With features Perform matrix multiplication separately, and finally obtain the features. and characteristics The features are summed to output the initial fused feature representation of the current layer. By optimizing the transfer of the coarse deformation field in the preceding layers and precisely applying the spatial transformation network, preliminary spatial alignment of the hierarchical features of mobile medical images and the features of fixed medical images was achieved. Combined with the adaptive weight allocation mechanism of the dynamic feature fusion module, the feature correlation of similar anatomical structures in the two types of images was effectively strengthened, while suppressing irrelevant noise and interference from non-critical regions. This process not only fully utilizes the registration results from the previous stage, establishing a spatial consistency foundation for feature interaction, but also improves the quality of the initial fused features through dynamic weighted fusion, providing accurate and effective input support for subsequent global and local feature extraction. This indirectly promotes the targeting and accuracy of feature characterization during the registration process, laying a solid feature foundation for the final optimization of the deformation field.
[0032] S4: Extract the combined global and local feature information from the intermediate fusion feature representation of the previous layer and the initial fusion feature representation of the current layer through the corresponding dynamic graph convolutional attention module to obtain the intermediate fusion feature representation of the current layer;
[0033] Preferably, step S4 includes:
[0034] S41: The coarse deformation field of the previous level The displacement amplitude is magnified by 2 times, and then the deformation field is obtained by trilinear interpolation upsampling. ;
[0035] S42: Represent the intermediate fusion features of the previous level. Features are obtained by upsampling through convolution. ;
[0036] S43: Features Initial fusion feature representation with the current level After concatenation, the corresponding dynamic graph is input into the convolutional attention module to extract global and local feature information, thus obtaining the intermediate fused feature representation of the current layer. ;
[0037] S44: Represent the intermediate fusion features of the current layer. Generate the residual coarse deformation field for the current level using a deformable registration head. ;
[0038] S45: Based on the deformation field Using spatial transformation networks The initial coarse deformation field at the current level Perform deformation to obtain the initial coarse deformation field of the current level;
[0039] S46: Combine the initial coarse deformation field of the current level with the residual coarse deformation field of the current level. The coarse deformation field of the current level is obtained by summing the features. .
[0040] Please see Figure 4 Preferably, the dynamic graph convolutional attention module includes a dynamic graph convolutional block and a spatial attention module; the dynamic graph convolutional block includes a conditional location encoding layer, a first convolutional layer, a dynamic axis graph convolutional layer, and a second convolutional layer; the dynamic graph convolutional block sequentially passes the conditional location encoding layer, the first convolutional layer, the dynamic axis graph convolutional layer, and the second convolutional layer to represent the initial fused features of the current layer. Features are obtained through processing. ; Features and characteristics The feature is obtained by adding the features together. ; Features The input space attention module processes the data to obtain the intermediate fused feature representation of the current layer. .
[0041] Preferably, the dynamic axial graph convolutional layer comprises:
[0042] S401: Define the input feature representation of the dynamic axial graph convolutional layer , Representing feature dimensions; through Functions representing input features Along the depth axis Height axis and width axis Simultaneously flipping yields the flipping feature. ;
[0043] S402: Calculate the input feature representation using the L2 norm. With flip features Element-order Euclidean distance, to obtain the global distance tensor And calculate the global distance tensor. mean and standard deviation ;
[0044] S403: For input feature representation depth axis Height axis and width axis Perform the following steps S404~S408 respectively:
[0045] S404: Define the length of the current axis as T, and initialize the scroll step size. Scroll count And the connection distance parameter K;
[0046] S405: Represent the input features Scroll along the current axis in the positive direction Step to obtain scrolling features Elements that exceed the current axis boundary are filled in from the top;
[0047] S406: Calculate the input feature representation using the L2 norm. and scrolling features The element-order Euclidean distance between them yields the distance tensor. If the distance tensor The element in is less than When the mask value is 1, it is set to 1; otherwise, it is set to 0 to construct a mask matrix. ;
[0048] S407: Calculate Rolling Features With input feature representation The difference feature matrix between ; through mask matrix For difference features Element-wise multiplication is used to obtain the key feature matrix. ;
[0049] S408: Pass , Indicates the first The current axis during the next scroll The maximum response characteristic, let And repeat steps S404~S408 until... Output the current axis up to a value greater than or equal to T. Maximum response characteristics ;
[0050] S409: Depth axis Height axis and width axis Maximum response characteristics through Obtain the maximum response characteristics of all axes ;
[0051] S410: Represent the input features and the maximum response characteristics of all axes The splicing feature is obtained by splicing along the channel dimension. ; splicing features The output features of the dynamic axial graph convolutional layer are obtained after convolution mapping. .
[0052] In this embodiment, the coarse deformation field of the previous level is... The displacement amplitude is magnified by 2 times, and a deformation field adapted to the current level feature scale is obtained by upsampling through trilinear interpolation with a kernel size of 3×3×3 and a step size of 1. ; Represent the intermediate fusion features of the previous level Upsampling is performed using a convolutional layer with a kernel size of 3×3×3, a stride of 1, and padding of 1 to obtain features. ;Will Initial fusion feature representation with the current level After concatenation along the channel dimension, the input is a Dynamic Graph Convolutional Attention Module (DGCAM), which consists of a Dynamic Graph Convolutional Block (DGCB) and a Spatial Attention Module (SPA). The DGCB sequentially includes a Conditional Location Encoding layer, a first convolutional layer, a Dynamic Axial Graph Convolutional layer, and a second convolutional layer. The Conditional Location Encoding layer uses a depthwise separable convolution with a kernel size of 3×3×3, a stride of 1, and padding of 1 to generate location codes, which are then added to the input features to introduce spatial information. The first and second convolutional layers both have a kernel size of 3×3×3, a stride of 1, and padding of 1, followed by a Leaky ReLU activation function and an instance normalization layer. In the Dynamic Axial Graph Convolutional layer, forward scrolling operations are performed along the depth axis (D), height axis (H), and width axis (W), respectively. For example, for the input features... When scrolling along the depth axis D, the feature can be first... Transform into Assuming the current rolling step size is 2, the feature representation after rolling is as follows: It mainly works by constructing a permutation matrix. Implementation, permutation matrix The 1 in the j-th row is located in the j-th row. The column has all other elements set to 0. The features after scrolling can be obtained by performing matrix multiplication. ; Calculate the element-wise Euclidean distance between the input feature and the rolling feature using the L2 norm, and then... The mask value corresponding to each element is set to 1, and the rest are set to 0 to construct a mask matrix. Then, the key feature matrix is obtained by element-wise multiplication of the mask matrix and the difference feature matrix. After multiple rounds of rolling, the maximum response feature of each axis is extracted and fused. Finally, the input features and the fused maximum response features are concatenated along the channel dimension and mapped back to the original number of channels through a 1×1×1 convolution with a stride of 1. The output features of DGCB and By adding features, we can obtain Then, the input and spatial attention module SPA are applied. SPA first focuses on the channel dimension. The maximum and average values are taken and concatenated. Spatial attention weights are generated using a convolutional layer with a kernel of 7×7×7, stride of 1, and padding of 3, and a sigmoid activation function. Element-wise multiplication yields the intermediate fused feature representation of the current level. ; Generating residual coarse deformation fields using deformable registration heads ,according to Using spatial transformation networks The initial coarse deformation field at the current level Perform deformation to obtain the initial coarse deformation field of the current level, and then compare the deformed initial coarse deformation field with... The coarse deformation field of the current level is obtained by summing the features. By integrating cross-level features and the synergistic effect of dynamic graph convolutional attention modules, it accurately captures global spatial dependencies and fully explores local anatomical details, effectively balancing the needs of global modeling and local characterization. At the same time, through iterative optimization of the residual coarse deformation field, it continuously corrects deformation errors, improves the accuracy and stability of the deformation field, and provides key feature support and deformation field foundation for the final high-quality medical image registration.
[0053] S5: For the intermediate fusion feature representation of the current level, the residual coarse deformation field of the current level is generated by the deformable registration head. The coarse deformation field of the previous level and the residual coarse deformation field of the current level are deformed and fused to obtain the coarse deformation field of the current level.
[0054] In this embodiment, the intermediate fusion feature representation for the current level is described. The residual coarse deformation field of the current level is generated through the deformable registration head. The registration head consists of 3×3×3 convolutional layers with a stride of 1 and padding of 1, directly outputting 3 channels of displacement information (corresponding to the x, y, and z spatial dimensions) to achieve accurate prediction of the residual deformation field; subsequently, it processes the coarse deformation field of the previous layer. Preprocessing is performed by first amplifying the displacement amplitude by a factor of 2, and then upsampling using trilinear interpolation with a kernel size of 3×3×3 and a step size of 1, so that its scale is consistent with the current hierarchical features and the coarse deformation field of the residual. To maintain consistency, a reference deformation field adapted to the current level is obtained; finally, deformation fusion is performed by feature addition, combining the preprocessed reference deformation field with the residual coarse deformation field of the current level. By adding elements one by one, we obtain the coarse deformation field optimized for the current level. This step, through a residual learning mechanism, organically integrates the fine deformation correction information of the current level with the coarse-grained deformation field of the previous level. It not only utilizes the global consistency of the previous registration results but also accurately compensates for local deformation errors through the residual field, effectively improving the fineness and continuity of the deformation field. At the same time, it continues the registration logic from coarse to fine, further enhancing the model's adaptability to complex nonlinear deformations, laying a crucial deformation field foundation for the final high-quality registration result.
[0055] S6: The coarse deformation field of the last level is used as the final deformation field. The moving medical image is spatially transformed using the final deformation field to obtain the registered medical image.
[0056] Preferably, the loss function used during training is defined as:
[0057]
[0058]
[0059]
[0060]
[0061] in, Represents the loss function; Represents similarity loss. Indicates the regularization loss. This represents the regularization weight hyperparameter; Indicates a moving image; Represents a fixed image; This represents the moving image after deformation by the final deformation field; It represents the image domain, that is, the set of all voxel locations in the image; Voxel representation Centered All voxels within the window; Represents a fixed image exist The voxel value of the location; Represents a fixed image exist Average voxel value at location; Representing an image exist The voxel value of the location; Representing an image exist Average voxel value at location; Represents any image exist Average voxel value at location; Represents any image in The voxel value of the location; Represents the final deformation field; Represents the deformation field In voxels The displacement vector of the position; Represents the spatial gradient operator; Represents the deformation field In voxels Spatial gradient of location; express The square of the L2 norm.
[0062] The composite loss function designed in this embodiment uses similarity loss. With regularization loss The weighted combination can both improve registration accuracy by calculating the correlation of voxel values within a local window to force a high degree of matching between the moving and fixed images in terms of anatomical details after registration, and ensure the smoothness and physical rationality of the deformation field by constraining the sum of squared spatial gradients of the deformation field, avoiding unnatural folds, twists, and other deformations that do not conform to the laws of anatomical structure, thus achieving a balance between registration accuracy and deformation rationality. Its parameter update method is based on the backpropagation algorithm, calculating the loss function... For the gradients of all trainable parameters in the network, optimizers such as Adam are used to iteratively update the parameters. Specifically, the update scope includes the convolutional layer weights and biases of the two-stream feature extraction encoder, the parameters of each convolutional layer and weight branch in the dynamic feature fusion module, the convolutional layer weights and learnable parameters of the dynamic graph convolutional attention module (including conditional position encoding layer, dynamic axial graph convolutional layer, and spatial attention module), and the convolutional layer weights of the deformable registration head, etc., while the regularization weight hyperparameters... As a preset hyperparameter, it is usually adjusted in the range of 0.01 to 0.1 according to the registration performance of the validation set (preferably set to 0.02 in the experiments of this application).
[0063] Another aspect of the present invention provides a deformable medical image registration system based on dynamic graph convolutional attention, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the deformable medical image registration method based on dynamic graph convolutional attention.
[0064] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements a deformable medical image registration method based on dynamic graph convolutional attention as described above.
[0065] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0066] In summary, this invention efficiently acquires multi-level features from both moving and stationary medical images through a weighted dual-stream feature extraction encoder. A dynamic feature fusion module adaptively weights and fuses key structural features from both images to enhance the representation of similar anatomical regions. A dynamic graph convolutional attention module accurately captures the combined features of global spatial dependencies and local details. Furthermore, a multi-stage deformation modeling strategy, from coarse to fine, iteratively optimizes the fusion of the previous-level coarse deformation field and the current-level residual coarse deformation field, ultimately generating a precise and stable deformation field. This significantly improves the accuracy of deformable medical image registration and the ability to align fine-grained anatomical structures, effectively reducing model computation and memory usage. It also enhances the stability of registration and the physical rationality of the deformation field in scenarios with large deformations, providing a more efficient and reliable medical image registration solution for clinical diagnosis, image-guided surgery, and other scenarios.
[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A deformable medical image registration method based on dynamic graph convolutional attention, characterized in that, include: S1: Acquire fixed medical images and mobile medical images to be registered, and use a shared-weight dual-stream feature extraction encoder to extract multi-level hierarchical features from the mobile medical images and fixed medical images respectively. S2: For mobile and stationary medical images, a dynamic feature fusion module and a dynamic graph convolutional attention module are constructed for the hierarchical features at each level, respectively. S3: After applying the coarse deformation field of the previous level to the hierarchical features of the current level moving medical image, the dynamic feature fusion module corresponding to the hierarchical feature input of the current level fixed medical image obtains the initial fused feature representation of the current level. S4: Extract the combined global and local feature information from the intermediate fusion feature representation of the previous layer and the initial fusion feature representation of the current layer through the corresponding dynamic graph convolutional attention module to obtain the intermediate fusion feature representation of the current layer; S5: For the intermediate fusion feature representation of the current level, the residual coarse deformation field of the current level is generated by the deformable registration head. The coarse deformation field of the previous level and the residual coarse deformation field of the current level are deformed and fused to obtain the coarse deformation field of the current level. S6: The coarse deformation field of the last level is used as the final deformation field. The moving medical image is spatially transformed using the final deformation field to obtain the registered medical image.
2. The deformable medical image registration method based on dynamic graph convolutional attention according to claim 1, characterized in that, The double-flow feature extraction encoder comprises two feature extraction branches sharing weights, each feature extraction branch comprises five convolution blocks cascaded in sequence, and a max pooling layer is arranged between adjacent convolution blocks for down sampling; each convolution block comprises two convolution layers, a LeakyReLU activation function and an instance normalization layer cascaded in sequence; output features of all convolution blocks in the two feature extraction branches are extracted and arranged in ascending order according to scales to obtain a hierarchical feature set of the mobile medical image and a hierarchical feature set of the fixed medical image , feature and represent hierarchical features of the first level of the mobile medical image and the fixed medical image, .
3. The deformable medical image registration method based on dynamic graph convolutional attention according to claim 1, characterized in that, Step S3 includes: S31: magnify the displacement amplitude of the coarse deformation field of the previous level by 2 times, and then up-sample the deformation field by tri-linear interpolation to obtain a deformation field S31: magnify the displacement amplitude of the coarse deformation field of the previous level by 2 times, and then up-sample the deformation field by tri-linear interpolation to obtain a deformation field ; S32: Based on the deformation field Using spatial transformation networks Hierarchical features of the current level of mobile medical images Features are obtained by deformation Then, the deformed hierarchical features Hierarchical features of the current level fixed medical image The corresponding dynamic feature fusion module is input to obtain the initial fused feature representation of the current level. .
4. The deformable medical image registration method based on dynamic graph convolutional attention according to claim 3, characterized in that, The dynamic feature fusion module includes: a first convolutional layer, a second convolutional layer, and a weight branch. The weight branch includes a third convolutional layer, a fourth convolutional layer, and a sigmoid activation function cascaded sequentially. and characteristics Features are obtained by inputting them into the first convolutional layer and the second convolutional layer respectively. and characteristics ; Features and characteristics After concatenation, the input to the weight branch is processed to obtain the weighted feature map. The weight branch maps the number of channels of the features to 2 through the third and fourth convolutional layers, and generates the weighted feature map through the Sigmoid activation function. The weighted feature map is then split into two weight matrices according to the number of channels. and ;Will With features Matrix multiplication yields features ;Will With features Matrix multiplication yields features ; Features and characteristics The features are summed to obtain the initial fused feature representation of the current level. .
5. The deformable medical image registration method based on dynamic graph convolutional attention according to claim 1, characterized in that, Step S4 includes: S41: The coarse deformation field of the previous level The displacement amplitude is magnified by 2 times, and then the deformation field is obtained by trilinear interpolation upsampling. ; S42: Represent the intermediate fusion features of the previous level. Features are obtained by upsampling through convolution. ; S43: Features Initial fusion feature representation with the current layer After concatenation, the corresponding dynamic graph is input into the convolutional attention module to extract global and local feature information, thus obtaining the intermediate fused feature representation of the current layer. ; S44: Represent the intermediate fusion features of the current layer. Generate the residual coarse deformation field for the current level using a deformable registration head. ; S45: Based on the deformation field Using spatial transformation networks The initial coarse deformation field at the current level Perform deformation to obtain the initial coarse deformation field of the current level; S46: Combine the initial coarse deformation field of the current level with the residual coarse deformation field of the current level. The coarse deformation field of the current level is obtained by summing the features. .
6. The deformable medical image registration method based on dynamic graph convolutional attention according to claim 5, characterized in that, The dynamic graph convolutional attention module includes a dynamic graph convolutional block and a spatial attention module; the dynamic graph convolutional block includes a conditional location encoding layer, a first convolutional layer, a dynamic axis graph convolutional layer, and a second convolutional layer; the dynamic graph convolutional block sequentially passes the conditional location encoding layer, the first convolutional layer, the dynamic axis graph convolutional layer, and the second convolutional layer to represent the initial fused features of the current layer. Features are obtained through processing ; Features and characteristics The feature is obtained by adding the features together. ; Features The input space attention module processes the data to obtain the intermediate fused feature representation of the current layer. .
7. The deformable medical image registration method based on dynamic graph convolutional attention according to claim 6, characterized in that, The dynamic axial graph convolutional layer includes: S401: Define the input feature representation of the dynamic axial graph convolutional layer , Representing feature dimensions; through Functions representing input features Along the depth axis Height axis and width axis Simultaneously flipping yields the flipping feature. ; S402: Calculate the input feature representation using the L2 norm. With flip features Element-order Euclidean distance, to obtain the global distance tensor And calculate the global distance tensor. mean and standard deviation ; S403: For input feature representation depth axis Height axis and width axis Perform the following steps S404~S408 respectively: S404: Define the length of the current axis as T, and initialize the scroll step size. Scroll count And the connection distance parameter K; S405: Represent the input features Scroll along the current axis in the positive direction Step to obtain scrolling features Elements that exceed the current axis boundary are filled in from the top; S406: Calculate the input feature representation using the L2 norm and scrolling features The element-order Euclidean distance between them yields the distance tensor. If the distance tensor The element in is less than When the mask value is 1, it is set to 1; otherwise, it is set to 0 to construct a mask matrix. ; S407: Calculate Rolling Features With input feature representation The difference feature matrix between ; through mask matrix Differential features Element-wise multiplication is used to obtain the key feature matrix. ; S408: Pass , Indicates the first The current axis during the next scroll The maximum response characteristic, let And repeat steps S404~S408 until... Output the current axis up to a value greater than or equal to T. Maximum response characteristics ; S409: Depth axis Height axis and width axis Maximum response characteristics through Obtain the maximum response characteristics of all axes ; S410: Represent the input features and the maximum response characteristics of all axes The splicing feature is obtained by splicing along the channel dimension. ; splicing features The output features of the dynamic axial graph convolutional layer are obtained after convolution mapping. .
8. The deformable medical image registration method based on dynamic graph convolutional attention according to claim 1, characterized in that, The loss function used during training is defined as: in, Represents the loss function; Represents similarity loss. Indicates the regularization loss. This represents the regularization weight hyperparameter; Indicates a moving image; Represents a fixed image; This represents the moving image after deformation by the final deformation field; It represents the image domain, that is, the set of all voxel locations in the image; Voxel representation Centered All voxels within the window; Represents a fixed image exist The voxel value of the location; Represents a fixed image exist Average voxel value at location; Representing an image exist The voxel value of the location; Representing an image exist Average voxel value at location; Represents any image exist Average voxel value at location; Represents any image in The voxel value of the location; Indicates the final deformation field; Represents the deformation field In voxels The displacement vector of the position; Represents the spatial gradient operator; Represents the deformation field In voxels Spatial gradient of location; express The square of the L2 norm.
9. A deformable medical image registration system based on dynamic graph convolutional attention, characterized in that, The system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and perform a deformable medical image registration method based on dynamic graph convolutional attention as described in any one of claims 1 to 8.
10. A computer storage medium, characterized in that, The computer storage medium stores a remote monitoring program, which, when executed by a processor, implements a deformable medical image registration method based on dynamic graph convolutional attention as described in any one of claims 1 to 8.