A three-dimensional vascular registration method for digital subtraction angiography and magnetic resonance angiography.
By combining structural and feature-based methods with dense reversible deformation fields, DSA and MRA vascular registration is performed in three-dimensional space, solving the problems of registration accuracy and computational efficiency in existing technologies. This achieves efficient and accurate three-dimensional vascular registration and provides clinically valuable vascular anatomy structures.
Patent Information
- Application Number
- CN202610283500.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-10
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies suffer from local extrema or misregistration in the registration of 3D DSA and MRA vascular imaging across modalities. They are difficult to balance geometric errors with bifurcation correspondence consistency, and involve large computational loads and complex parameter adjustments. There is a lack of effective 3D registration methods.
By combining structure- and feature-based methods with dense reversible deformation field methods, blood vessel registration is performed in three-dimensional space. Blood vessel segmentation and three-dimensional centerline calculation are performed through a deep neural network model. Combining rigid and non-rigid registration, deformation correction is performed using a custom fractional function and a dense reversible deformation field model.
It improves the accuracy and computational efficiency of three-dimensional vascular registration, reduces the impact of noise and artifacts, ensures consistency in three-dimensional space, and provides a more intuitive vascular anatomy structure, providing an important basis for preoperative planning and postoperative evaluation.
Smart Images
Figure CN122134770A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image processing and analysis technology, specifically relating to a three-dimensional vascular registration method for digital subtraction angiography and magnetic resonance angiography. Background Technology
[0002] Cerebral vascular imaging technology plays a crucial role in the screening, prevention, diagnosis, and treatment of cerebrovascular diseases. Digital subtraction angiography (DSA) is widely considered the "gold standard" for cerebral vascular imaging due to its high resolution and precise anatomical information. However, DSA is an invasive procedure with issues such as low comfort, high risks, and radiation exposure. Magnetic resonance angiography (MRA), as a non-invasive procedure, offers high safety, but its imaging resolution and contrast are generally lower than DSA, especially in displaying the fine and distal structures of cerebral blood vessels.
[0003] High-precision registration of MRA and DSA can effectively combine the advantages of the two modalities: by aligning the MRA and DSA of the same patient in three-dimensional space, neurosurgeons can simultaneously utilize the tissue background in MRA and the high-resolution vascular information in DSA to obtain a more intuitive and comprehensive vascular anatomy, which is of great value for preoperative planning and postoperative assessment.
[0004] There have been some studies on blood vessel registration. Methods based on grayscale and similarity directly use image grayscale or similarity measures such as mutual information and correlation coefficients to perform rigid or non-rigid transformations. They usually perform well for the same modality, but in cross-modal scenes, they are prone to local extrema or misregistration [1]. Methods based on structure and features first enhance blood vessels, then extract binary blood vessels, centerlines or point clouds, and then use point cloud registration or topology-based matching. Although this type of method reduces the influence of modal differences, it is greatly affected by noise and artifacts, and is sensitive to initial pose and coverage. It is difficult to simultaneously take into account geometric errors and bifurcation correspondence consistency during the registration process [2]. The method based on dense reversible deformation field (diffeomorphic) has a large computational load and complex parameter adjustment. In the case of blood vessels, which are slender and high-frequency structures, special regularization and multi-scale strategies are often required. Otherwise, cross-branch deformation or loss of detail will occur [3]. Although some studies have improved the method by combining different types of methods, most of these studies focus on two-dimensional DSA images and lack research on three-dimensional DSA registration [4].
[0005] This invention proposes a vascular registration method for digital subtraction angiography and magnetic resonance angiography in three-dimensional space. It combines structure-based and feature-based methods with dense reversible deformation field-based methods to register blood vessels in three-dimensional space. While ensuring consistency in three-dimensional space, it directly calculates spatial distance and bifurcation topology based on the three-dimensional centerline, thereby improving the accuracy and computational efficiency of registration. Summary of the Invention
[0006] The present invention aims to propose a three-dimensional vascular registration method for digital subtraction angiography (DSA) and magnetic resonance angiography (MRA) to improve the registration accuracy of DSA and MRA vessels.
[0007] The vascular registration method proposed in this invention selects MRA as the fixed reference image and DSA as the image to be registered, and performs registration accordingly. Specifically, it consists of the following five steps:
[0008] (i) Preprocessing of digital subtraction angiography (DSA) and magnetic resonance angiography (MRA) images: using the position of the external skull in the images to perform preliminary alignment of DSA and MRA images and eliminate intensity differences between different images;
[0009] (ii) Build a deep neural network model to segment blood vessels in DSA and MRA images and remove parts of the images other than blood vessels (including removing the skull and other brain tissue from the original images).
[0010] (iii) Convert the blood vessels in the original image into a three-dimensional blood vessel model;
[0011] (iv) Perform rigid registration to initially align the vascular structures in physical space;
[0012] (v) Perform non-rigid registration to correct vascular deformation;
[0013] Furthermore:
[0014] The specific process for preprocessing DSA and MRA images described in step (I) is as follows:
[0015] First, the centers of the two images are aligned. Then, using the similarity of the external skull structure corresponding to the DSA and MRA as the evaluation criterion, the DSA image is subjected to an iterative affine transformation to resolve the scale difference, angle difference, and displacement between it and the MRA image, so as to obtain a DSA image with the external skull aligned with the MRA and eliminate the intensity difference between different images.
[0016] The deep neural network model described in step (II) has the following structure: Figure 2 As shown, it specifically includes Module 1, Module 2, Module 3, and the RWKV module; among which:
[0017] (1) Module 1 is an encoder module, which is used to extract multi-scale features from the input image. Its structure includes: multi-layer convolutional units, each of which consists of two convolutional layers with a kernel size of 3×3×3 and a non-linear activation function layer connected in series; at the end of each convolutional unit, a max pooling operation is set to gradually reduce the spatial resolution of the feature map and achieve downsampling ratios of 1 / 2, 1 / 4, 1 / 8 and 1 / 16; the outputs of each layer of the encoder are sequentially sent to the feature projection unit (Encoder Project) for channel mapping and unified processing of features at different scales.
[0018] (2) Module 2 is a cross-channel fusion module, which is used to realize the joint modeling of multi-scale features in the bottleneck stage of the network. Its structure includes: a feature splicing unit, which is used to concatenate the features output by each encoder layer; a channel mixing unit (CCM, Cross-Channel Mix), which is used to fuse and reweight the spliced multi-scale features in the channel dimension. The channel mixing unit includes element-wise addition, element-wise multiplication, Sigmoid activation and squared ReLU activation operations. The output fused features are distributed to each layer of the decoder through the feature splitting unit (Split).
[0019] (3) Module 3 is a decoder module, which is used to gradually restore the spatial resolution of the feature map. Its structure includes: multi-layer decoding units, each of which consists of two convolutional layers with a kernel size of 3×3×3 and a non-linear activation function layer connected in series; an upsampling operation is set at the end of each decoding unit to restore the spatial resolution; each decoding unit receives multi-scale features from the output of module 2 and maps them through the decoder projection unit to finally obtain the output result.
[0020] (4) A RWKV (Receptance Weighted Key Value) module [5] is set between module 1 and module 3. Its structure includes: the input features are first processed by 3×3×3 convolution and Q-Shift projection; global context modeling is performed through bidirectional RWKV (Bi-RWKV) units, the Bi-RWKV units containing spatial weight matrices. Numerical matrix and channel residual matrix The output is after normalization (Norm) and element-wise addition. The RWKV module is used to enhance the network's ability to model long-range dependencies.
[0021] Step (III) involves converting the blood vessels in the original image into a three-dimensional blood vessel model. The specific process is as follows:
[0022] (1) For the segmented MRA and DSA images, the centerline of the blood vessels is obtained using morphological methods. For N points on the centerline, an adjacency matrix is constructed. Indicates whether two points are adjacent;
[0023] (2) Analyze the adjacency matrix By assessing connectivity, the endpoints, bifurcation points, and geometric points of the vascular centerline are identified, thereby constructing a vascular graphical model. .in, This represents the set of vertices containing the endpoints and bifurcation points of blood vessels. This represents the set of curve edges connecting adjacent vertices of a blood vessel branch;
[0024] (3) For the vascular mapping model of DSA Vascular mapping model of MRA The vascular matching criteria for DSA and MRA in three-dimensional space can be defined as finding the optimal deformation model between the two:
[0025]
[0026] in, Used to measure the distance between two vascular maps Represents rigid or non-rigid deformation in three-dimensional space. It is the set of all possible rigid or non-rigid deformations.
[0027] Step (iv) involves rigid registration, which initially aligns the vascular structures in physical space. The specific process is as follows:
[0028] (1) Two mapping functions are introduced to establish the relationship between vessel edge matching, point matching and rigid transformation:
[0029] : Matching set of 3D blood vessel points (feasible region) Mapping to a rigid transformation ;
[0030] : Matching set of 3D blood vessel edges (feasible region) Mapped to a set of point matches .
[0031] (2) Constructing a search space for edge matching based on the topological continuity of blood vessels:
[0032] Initial matching settings: Given an initial set of 3D blood vessel vertex matching pairs, the edge matching of the corresponding blood vessel graph is used as the initial state for the search;
[0033] Edge matching state expansion: Starting from the initial state, the edge matching is expanded step by step along the blood vessel topology—if the current edge matching set is Then the next round of matching set ( It is located next to the MRA vessel. (For DSA vessel edges), each extended state corresponds to a substate in the search space;
[0034] Search space formation: All possible edge matching states constitute the feasible region of the search space. Each node records the current edge matching status, the corresponding registration result, and the evaluation score.
[0035] (3) Four-step iterative search until the preset computational budget (such as the number of iterations) is reached or the theoretical highest score node is found, then stop the search:
[0036] Selection: Starting from the initial state, the Upper Confidence Bound (UCB) strategy is used to select sub-states, prioritizing sub-states with high scores and few visits to avoid getting trapped in local optima;
[0037] Expansion: If there are still unmatched blood vessel edges in the selected state, a new sub-state is added. Each sub-state corresponds to a new set of edge matching combinations, expanding the coverage of the search space.
[0038] Simulation: For a new sub-state, randomly simulate the subsequent edge matching path until all matching edges are traversed, obtain a temporary edge matching set, and calculate the corresponding registration evaluation score.
[0039] Backpropagation: The simulated evaluation score is backpropagated to all previous states of the current state, updating the cumulative score and number of visits for each node, providing a basis for the next round of selection.
[0040] (4) A custom fractional function is used to evaluate the edge matching quality, taking into account both projection overlap and scale consistency:
[0041]
[0042] The first term is the overlap term, used to calculate MRA vessel points. With DSA vessel points The minimum distance is calculated, and the distance is converted into an overlap score using an exponential function. For distance normalization parameters, The first term is the total number of 3D blood vessel points; the second term is the scale penalty term, calculated using the MRA vertex set scale. DSA vertex set scale The ratio of the penalty scales is inconsistent, when When the value is close to 1, the penalty is minimized. A higher Q value indicates better edge matching quality and a better corresponding rigid transformation.
[0043] From the optimal edge matching node, through The mapping relationship is used to derive the final rigid transformation. Rigid registration is completed.
[0044] The non-rigid registration correction of vascular deformation described in step (five) is performed as follows:
[0045] (1) Based on the obtained edge matching Further refinement yielded dense point pairs of MRA and DSA vessels. , denoted as a matrix:
[0046] MRA point set: (K is the total number of points);
[0047] DSA point set: .
[0048] (2) Modeling nonlinear deformation using Gaussian kernel function and regenerating kernel Hilbert space (RKHS):
[0049]
[0050] in, These are discrete control points randomly distributed in 3D space. For the kernel size, we can take it to be 5, then the deformation function is:
[0051]
[0052] in, To control the vertices The deformation coefficients. Coefficient matrix. C is the control point Number of kernels, kernel matrix ,in, Then the matrix form of the deformation function is:
[0053]
[0054] (3) Construct the objective function:
[0055]
[0056] in, This is a data item used to minimize the Euclidean distance between the MRA and DSA point sets to ensure registration accuracy. Let be the spatial smoothing term, where The deformation field is globally smoothed by constraining the RKHS norm (to avoid local distortion). These are weighting coefficients, typically ; The topological smoothing term is used to constrain the displacement consistency of adjacent vessel points, ensuring that the vessel topology remains unchanged. Let be the displacement matrix. For the graph Laplace matrix ( For degree matrix, ), These are weighting coefficients, typically .
[0057] (4) Calculate the gradients for each of the three terms of the objective function, and use the finite-memory quasi-Newton method (L-BFGS) to quickly iterate and solve for the optimal coefficient matrix. The final deformation field is obtained.
[0058] Compared with the prior art, the present invention has the following advantages:
[0059] This invention combines structure- and feature-based methods with dense reversible deformation field methods to register DSA and MRA vessels in three-dimensional space. This effectively reduces the impact of noise and artifacts, decreases sensitivity to initial pose and coverage, better balances geometric errors and bifurcation correspondence, and significantly reduces computational load and parameter tuning. Registration in three-dimensional space ensures consistency while allowing direct calculation of spatial distances and bifurcation topology based on three-dimensional centerlines. Furthermore, it provides neurosurgeons with a more effective, intuitive, and comprehensive understanding of vascular anatomy, which is of significant value for preoperative planning and postoperative evaluation. Attached Figure Description
[0060] Figure 1 This is a flowchart of the present invention.
[0061] Figure 2 This is a diagram of the RWKV-UNet network structure.
[0062] Figure 3 This is a structural diagram of the encoder module, decoder module, and RWKV module. Detailed Implementation
[0063] Taking MRA images acquired by Philips Medical Systems Intera 3T and DSA images acquired by GE OEC 9900 C-Arm as examples, the practical application process of this invention is described in detail, including the following steps:
[0064] Step 1, Image Preprocessing: Align the image centers of MRA and DSA; using MRA as the reference image and DSA as the image to be registered, perform preliminary alignment of DSA and MRA images using the location of the external skull in the images, iteratively calculate the structural similarity between DSA and MRA, iteratively optimize the parameters, perform rigid body and affine transformations on the DSA image respectively, and eliminate intensity differences between different images.
[0065] Step 2: A deep neural network was built and trained on the publicly available MRA vessel segmentation dataset COSTA[6] and DSA images containing skull and blood vessels. Five-fold cross-validation was used, and the samples were divided into training, validation and test sets in an 8:1:1 ratio. After training the model, it was tested on the test set. The average symmetric surface distance (ASD), 95% Hausdorff distance (HD95), Dice similarity coefficient (Dice), surface Dice coefficient (SurfDice), and centerline Dice coefficient (clDice) were used as evaluation indicators. The test results on 72 MRA and 96 DSA images are shown in Table 1. It can be seen that the model can segment blood vessels in MRA and DSA images with good accuracy.
[0066] Step 3, Initialization of 3D vascular map: The segmented MRA and DSA vessels obtained in Step 2 are used to obtain the centerline of the vessels using morphological methods, the adjacency matrix of the point set on the centerline is calculated, the connectivity of the adjacency matrix is analyzed, and the endpoints, bifurcation points and geometric points of the point set are identified, thereby constructing the vascular map model.
[0067] Step 4, Rigid Registration of Blood Vessels: Two mapping functions are introduced to establish edge matching and point matching of blood vessels respectively; an edge matching search tree is constructed based on the topological continuity of blood vessels; a four-step iterative search is performed until the preset computational budget (such as the number of iterations) is reached or the node with the highest theoretical score is found, at which point the search stops; a custom score function is used to evaluate the quality of edge matching, taking into account both projection overlap and scale consistency; from the optimal edge matching node, the final rigid transformation is derived through the mapping relationship from edge matching to point matching, and from point matching to model transformation, thus completing the rigid registration.
[0068] Step 5, Non-rigid Registration of Vessels: Based on the edge matching obtained in Step 4, the dense point pairs of MRA and DSA vessels are further refined; nonlinear deformation is modeled using RKHS through a Gaussian kernel function. A data term is introduced to minimize the Euclidean distance between the MRA and DSA point sets, ensuring registration accuracy; a spatial smoothing term is introduced to constrain the global smoothness of the deformation field using the RKHS norm (avoiding local distortion); a topological smoothing term is introduced to constrain the displacement consistency of adjacent vessel points, ensuring the vessel topology remains unchanged. Gradients are calculated for each of the three terms separately, and the optimal coefficient matrix is solved iteratively using L-BFGS to obtain the final deformation field. To evaluate the effectiveness of the registration methods in Steps 4 and 5, mean square error (MSE), root mean square error (RMSE), and structural similarity (SSIM) are used as evaluation metrics. The corresponding metrics are calculated for 963 samples before registration, after rigid registration, and after both rigid and non-rigid registration. The results are shown in Table 2. It can be seen that the two-step registration method can achieve vessel registration in three-dimensional space with good accuracy.
[0069] Table 1 Results of RWKV-UNet network in vessel segmentation for MRA and DSA
[0070] .
[0071] Table 2 Comparison of registration errors and structural similarity before and after registration
[0072] .
[0073] References
[0074] [1]Guo Z., et al. Cross-modality cerebrovascular segmentation viatopology-aware domain adaptation. 2024.
[0075] [2]Ma J., et al. Point cloud registration: A comprehensive survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021.
[0076] [3]Zou J., et al. A review of deep learning-based deformable medicalimage registration methods. Frontiers in Radiology, 2022.
[0077] [4]Chan H., et al. 2D-3D vascular registration between digitalsubtraction angiographic (DSA) and magnetic resonance angiographic (MRA)images. 2004 2nd IEEE International Symposium on Biomedical Imaging: Nano toMacro, Arlington, VA, USA, 2004.
[0078] [5]Jiang J., et al., RWKV-UNet: improving UNet with long-rangecooperation for effective medical image segmentation. arXiv:2501.08458v1,2025.
[0079] [6]Mou L., et al. COSTA: A multi-center TOF-MRA dataset and a styleself-consistency network for cerebrovascular segmentation. IEEE Transactionson Medical Imaging, 2024。
Claims
1. A three-dimensional vascular registration method for digital subtraction angiography and magnetic resonance angiography, characterized in that, It consists of the following five steps: (i) Preprocessing of digital subtraction angiography (DSA) and magnetic resonance angiography (MRA) images: using the position of the external skull in the images to perform preliminary alignment of DSA and MRA images and eliminate intensity differences between different images; (ii) Construct a deep neural network model to segment blood vessels in DSA and MRA images and remove the parts outside the blood vessels in the images; (iii) Convert the blood vessels in the original image into a three-dimensional blood vessel model; (iv) Perform rigid registration to initially align the vascular structures in physical space; (v) Perform non-rigid registration to correct vascular deformation.
2. The method for vascular registration of DSA and MRA in three-dimensional space according to claim 1, characterized in that, The specific process for preprocessing DSA and MRA images in step (1) is as follows: First, the centers of the two images are aligned. Then, using the similarity of the external skull structure corresponding to the DSA and MRA as the evaluation criterion, the DSA image is subjected to an iterative affine transformation to resolve the scale difference, angle difference, and displacement between it and the MRA image, so as to obtain a DSA image with the external skull aligned with the MRA and eliminate the intensity difference between different images.
3. The method for vascular registration of DSA and MRA in three-dimensional space according to claim 2, characterized in that, The deep neural network model described in step (II) includes module 1, module 2, module 3, and an RWKV module; wherein: (1) Module 1 is the encoder module, which is used to extract multi-scale features from the input image. Its structure includes: multi-layer convolutional units, each of which consists of two convolutional layers with a kernel size of 3×3×3 and a non-linear activation function layer connected in series; at the end of each convolutional unit, a max pooling operation is set to gradually reduce the spatial resolution of the feature map to achieve downsampling ratios of 1 / 2, 1 / 4, 1 / 8 and 1 / 16; the outputs of each layer of the encoder are sequentially fed into the feature projection unit for channel mapping and unified processing of features at different scales; (2) Module 2 is a cross-channel fusion module, which is used to realize joint modeling of multi-scale features in the bottleneck stage of the network. Its structure includes: a feature splicing unit, which is used to concatenate the features output by each encoder layer; a channel mixing unit, which is used to fuse and reweight the spliced multi-scale features in the channel dimension. The channel mixing unit includes element-wise addition, element-wise multiplication, Sigmoid activation and squared ReLU activation operations; the output fused features are distributed to each layer of the decoder through the feature splitting unit. (3) Module 3 is a decoder module, which is used to gradually restore the spatial resolution of the feature map. Its structure includes: a multi-layer decoding unit, each decoding unit consists of two convolutional layers with a kernel size of 3×3×3 and a non-linear activation function layer connected in series; an upsampling operation is set at the end of each decoding unit to restore the spatial resolution; each decoding unit receives multi-scale features from the output of module 2 and maps them through the decoding projection unit to finally obtain the output result; (4) The RWKV module is located between module 1 and module 3. Its structure includes: the input features are first processed by 3×3×3 convolution and Q-Shift projection; global context modeling is performed via bidirectional RWKV units, which contain spatial weight matrices. Numerical matrix and channel residual matrix Output after normalization.
4. The method for vascular registration of DSA and MRA in three-dimensional space according to claim 3, characterized in that, Step (III) involves converting the blood vessels in the original image into a three-dimensional blood vessel model. The specific process is as follows: (1) For the segmented MRA and DSA vascular images, the centerline of the vascular vessel is obtained using morphological methods. For N points on the centerline, an adjacency matrix is constructed. Indicates whether two points are adjacent; (2) Analyze the adjacency matrix By assessing connectivity, the endpoints, bifurcation points, and geometric points of the vascular centerline are identified, thereby constructing a vascular graphical model. ;in, This represents the set of vertices containing the endpoints and bifurcation points of blood vessels. This represents the set of curve edges connecting adjacent vertices of a blood vessel branch; (3) For the vascular mapping model of DSA Vascular mapping model of MRA The vascular registration between DSA and MRA in three-dimensional space is defined as finding the optimal deformation model between the two: in, Used to measure the distance between two vascular maps Represents rigid or non-rigid deformation in three-dimensional space. It is the transformation domain composed of all rigid or non-rigid deformations.
5. The method for vascular registration of DSA and MRA in three-dimensional space according to claim 4, characterized in that, The specific process for step (four) is as follows: (1) Two mapping functions are introduced to establish the relationship between vessel edge matching, point matching and rigid transformation: : Matching set of 3D blood vessel points (feasible region) Mapping to a rigid transformation ; : Matching set of 3D blood vessel edges (feasible region) Mapped to a set of point matches ; (2) Constructing a search space for edge matching based on the topological continuity of blood vessels: Initial matching settings: Given an initial set of 3D blood vessel vertex matching pairs, the edge matching of the corresponding blood vessel graph is used as the initial state for the search; Edge matching state expansion: Starting from the initial state, the edge matching is expanded step by step along the blood vessel topology—if the current edge matching set is Then the next round of matching set , It is located next to the MRA vessel. For DSA vessel edges, each extended state corresponds to a substate in the search space; Search space formation: All possible edge matching states constitute the feasible region of the search space. Each node records the current edge matching status, the corresponding registration result, and the evaluation score; (3) Four-step iterative search until the preset calculation budget is reached or the theoretical highest score node is found, then stop the search: Selection: Starting from the initial state, the UCB strategy is used to select sub-states, prioritizing sub-states with high scores and few visits to avoid getting trapped in local optima; Extension: If there are still unmatched blood vessel edges in the selected state, a new sub-state is added. Each sub-state corresponds to a new set of edge matching combinations, thus expanding the coverage of the search space. Simulation: For the newly added sub-state, randomly simulate the subsequent edge matching path until all matching edges are traversed, obtain a temporary edge matching set, and calculate the corresponding registration evaluation score; Backtracking: The simulated evaluation score is backpropagated to all previous states of the current state point, updating the cumulative score and number of visits of each node, providing a basis for the next round of selection; (4) A custom fractional function is used to evaluate the edge matching quality, taking into account both projection overlap and scale consistency: The first term is the overlap term, used to calculate MRA vessel points. DSA vessel point The minimum distance is calculated, and the distance is converted into an overlap score using an exponential function. For distance normalization parameters, The first term is the total number of 3D blood vessel points; the second term is the scale penalty term, calculated using the MRA vertex set scale. DSA vertex set scale The ratio of the penalty scales is inconsistent, when When the value is close to 1, the penalty is minimized; the higher the Q value, the better the edge matching quality and the better the corresponding rigid transformation. (5) From the optimal edge matching nodes, through The mapping relationship is used to derive the final rigid transformation. Rigid registration is completed.
6. The method for vascular registration of DSA and MRA in three-dimensional space according to claim 5, characterized in that, The specific process for step (five) is as follows: (1) Based on the obtained edge matching Further refinement yielded dense point pairs of MRA and DSA vessels. , denoted as a matrix: MRA point set: K is the total number of points; DSA point set: ; (2) Modeling nonlinear deformation using Gaussian kernel function and regenerating kernel Hilbert space (RKHS): in, These are discrete control points randomly distributed in 3D space. Given a kernel size of 5, the deformation function is: in, To control the vertices Deformation coefficients; coefficient matrix C is the control point Number of kernels, kernel matrix ,in, Then the matrix form of the deformation function is: (3) Construct the objective function: in, This is a data item used to minimize the Euclidean distance between the MRA and DSA point sets to ensure registration accuracy; Let be the spatial smoothing term, where Global smoothing of the deformation field is achieved by constraining the RKHS norm. These are the weighting coefficients; The topological smoothing term is used to constrain the displacement consistency of adjacent vessel points, ensuring that the vessel topology remains unchanged. Let be the displacement matrix. For the graph Laplace matrix, For degree matrix, , These are the weighting coefficients; (4) Calculate the gradients for each of the three terms of the objective function, and use the finite-memory quasi-Newton method to quickly iterate and solve for the optimal coefficient matrix. The final deformation field is obtained.