Three-dimensional blood vessel structure recovery device and method fusing prior information

Abstract

The invention provides a three-dimensional vascular structure recovery device and method fusing prior information, and relates to the technical field of biomedical image processing. Comprising a data acquisition module, a data extraction module and a complex blood vessel missing structure recovery module. The method comprises the following steps: firstly, acquiring SR-CT data of a blood vessel network, and then carrying out image segmentation on the blood vessel network to obtain a binary mask; decomposing the complex blood vessel into a series of single blood vessel images according to an image segmentation result of the blood vessel network; according to a series of single blood vessel images obtained through decomposition, selecting a single blood vessel image containing a missing structure through a connected domain analysis method; a spatial transformation network based on shape prior is constructed, and a missing structure of a single blood vessel is recovered; and finally, recombining the single blood vessel image, namely the blood vessel image without the missing structure and the blood vessel image with the restored missing structure to generate a complete blood vessel network, thereby realizing missing structure restoration of the complex blood vessel network.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical image processing, and in particular to a device and method for restoring three-dimensional blood vessel structure by integrating prior information. Background Art

[0002] Breast cancer bone metastasis is a highly prevalent malignant tumor that threatens the lives of women worldwide and is often accompanied by abnormal proliferation of intraosseous blood vessels. In recent years, research into the mechanisms of breast cancer bone metastasis based on intraosseous vascular imaging and extraction has garnered significant attention. Synchrotron radiation micro-computed tomography (SR-µCT) can be combined with vascular contrast agents to provide high-quality imaging of intratibial vessels in mice, generating three-dimensional images with high contrast, high spatial resolution, and a high signal-to-noise ratio. However, the complex intraosseous vascular structure in SR-µCT images is difficult to fully extract, and recovery of missing vascular structures is also challenging.

[0003] Although existing vascular extraction technologies have advanced rapidly, they still have many shortcomings, hindering the complete extraction of complex vascular structures. For example, thresholding and region growing methods based on image grayscale distribution analysis can degrade vessel extraction quality due to image artifacts, partial volume effects, and noise, resulting in missing structures in the extracted vascular network. Methods based on shortest path algorithms and polynomial fitting are limited to discontinuous connections between vascular centerlines and ignore vessel size information, making them unable to recover missing structures in the vascular network. While methods based on level sets, random walks, and 3D-UNet neural networks can recover missing structures in vascular networks with distinct tree-like structures, these simple models, based on two-dimensional planes or single trunk vessels, are difficult to apply to complex intraosseous vascular networks with multiple trunks in three-dimensional space. Spatial transformation networks based on shape priors can recover target missing structures. However, this method requires that the shape prior and the target structure have similar structural composition. This limits its application to complex vascular networks. To date, no literature has reported a device or method for recovering missing structures in complex vascular networks based on intraosseous SR-µCT images. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a three-dimensional vascular structure recovery device and method that integrates prior information, aiming to restore the missing structure of the complex vascular network based on the SR-µCT image of the intraosseous blood vessels, thereby achieving complete extraction of the intratibial vascular network of mice.

[0005] On the one hand, a three-dimensional vascular structure restoration device integrating prior information includes a data acquisition module, a data extraction module, and a complex vascular missing structure restoration module;

[0006] The data acquisition module acquires SR-µCT data of the vascular network and sends the acquired SR-µCT data of the vascular network to the data extraction module;

[0007] The data extraction module extracts the vascular network image segmentation results required for restoring the complex vascular missing structure from the SR-µCT data of the vascular network, and sends the extracted vascular network image segmentation results to the complex vascular missing structure restoration module;

[0008] The complex vascular missing structure recovery module processes the data information sent by the data extraction module, specifically decomposing complex blood vessels, restoring the structure of a single blood vessel, reorganizing the vascular network, and outputting the recovery result of the complex vascular missing structure;

[0009] The complex vascular missing structure recovery module further includes:

[0010] The complex blood vessel decomposition module is used to decompose the complex blood vessels into a series of single blood vessels according to the blood vessel network segmentation results output by the data extraction module, and output the decomposition results to the single blood vessel structure recovery module;

[0011] The single vessel structure recovery module is used to recover the single vessel containing the missing structure based on the series of single vessels output by the complex vessel decomposition module, and output the result to the vascular network reconstruction module;

[0012] The vascular network reorganization module is used to reorganize all single blood vessels based on a series of single blood vessels with restored structures output by the single blood vessel structure recovery module, generate a vascular network, and output a complex blood vessel result with restored missing structures.

[0013] On the other hand, a method for restoring a three-dimensional vascular structure by integrating prior information is implemented based on the aforementioned device for restoring a three-dimensional vascular structure by integrating prior information, comprising the following steps:

[0014] Step 1: Acquire SR-µCT data of the vascular network;

[0015] Step 2: Segment the vascular network image to obtain a binary mask;

[0016] Step 3: Based on the image segmentation results of the vascular network, the complex blood vessels are decomposed into a series of single blood vessel images;

[0017] Step 3.1: Based on the binary mask obtained by segmentation of the vascular network image, it is converted into a three-dimensional spatial representation, i.e., a vascular point cloud, through point cloud processing to obtain the local position characteristics of the blood vessels;

[0018] Specifically, the binary mask of the blood vessels is first discretized and sampled according to a preset sampling interval. The spatial coordinate information of each sampling point is then extracted, and finally the corresponding vascular 3D point cloud data is generated. The point cloud obtained in this process is a discrete spatial representation of the vascular network in a unified spatial reference system, where each data point contains its 3D coordinates (x, y, z), that is, the local position feature.

[0019] Step 3.2: Obtain the local vascular thickness feature based on the binary mask obtained by segmenting the vascular network image;

[0020] The local thickness feature of the blood vessel is to assign a maximum sphere diameter value that can fill the interior of the blood vessel to each point on the blood vessel volume; let Ω⊂R 3 is the set of all points on the blood vessel volume, and p is any point among them. Then the local blood vessel thickness feature is defined as Th(p)=2•max({r|p∈sph(a,r)⊆Ω,a∈Q}); where R 3 Represents three-dimensional real space, sph(a,r) represents a sphere with a as the center and a radius of r;

[0021] Step 3.3: Calculate the vascular orientation based on the binary mask obtained by segmenting the vascular network image to obtain the local directional features of the vascular;

[0022] The vascular direction is the principal eigenvector of the Hessian matrix calculated at each point on the vascular volume; first, the Hessian matrix is ​​solved for each voxel point in the binary mask, and the eigenvalue and eigenvector are calculated. Then, the principal eigenvector corresponding to the maximum eigenvalue is selected as the local direction feature α representing the local principal direction of the vessel;

[0023] Step 3.4: Integrate the obtained blood vessel local position features, local thickness features, and local direction features into the existing PointNet point cloud neural network model;

[0024] Specifically, the point cloud (N, d + C) is sampled and grouped according to the local position features (x, y, z), local thickness features Th(p), and local orientation features α corresponding to each point in the vascular point cloud, and input into the PointNet point cloud neural network in the form of feature vectors to complete feature extraction. Then, feature transfer and point cloud classification (N, k) are further realized through the existing interpolation and lightweight unit point cloud neural network UnitPointNet, where N is the number of sampling points, d is the coordinate dimension, C is the number of features, and k is the label class. Finally, the category probability of the vascular branch to which each point belongs is output, and the vascular branch category to which each point belongs is determined according to the size of the probability value, thereby obtaining a series of single vascular images.

[0025] Step 4: Based on the series of single vessel images obtained by decomposition, select the single vessel image containing the missing structure through the connected domain analysis method;

[0026] Specifically, all foreground voxel groups connected in the existing 18-connectivity mode are iteratively searched in a single vessel image as connected domains; if the number of connected domains in the image exceeds 1, the image is considered to be a single vessel image containing missing structures.

[0027] Step 5: Construct a spatial transformation network based on shape prior to restore the missing structure of a single blood vessel;

[0028] The spatial transformation network specifically includes a feedforward neural network, a grid generator, and a sampler.

[0029] Specifically, the tubular shape prior image U and the single blood vessel structure image I are input into a feedforward neural network architecture with convolution and maximum pooling in a two-channel manner to generate the deformation parameter θ, where the tubular shape prior image U is a cylinder with a base radius of r and a height of h, and θ=f ψ (I,U),f ψ It is a neural network with parameter ψ, which is generated by the grid generator G according to the coordinates of point i in I Find the corresponding coordinates of point i in U , that is, find the pixel coordinate relationship T between the input single blood vessel structure image I and the tubular shape prior image U θ (G i ): , i is a pixel in the image; finally, according to T θ (G i ) and the sampler, which samples the pixel value of point i in U through the B-spline interpolation method to generate the output image V: ; Among them, H×W×D is the size of the single blood vessel structure image I, the tubular shape prior image U, and the output image V. H, W, and D represent the height, width, and depth of the three-dimensional image, respectively. mnl is the intensity value at position (m, n, l) on the tubular shape prior image U, n = 0, 1, 2, ..., H-1, m = 0, 1, 2, ..., W-1, l = 0, 1, 2, ..., D-1.

[0030] Step 6: Reconstructing the single vessel images, specifically including the vessel images without missing structures and the vessel images with restored missing structures, to generate a complete vascular network;

[0031] Specifically, the blood vessels are reorganized by summing the image voxel values ​​to generate a complete vascular network.

[0032] The beneficial effects of adopting the above technical solution are:

[0033] The present invention provides a three-dimensional vascular structure restoration device and method that integrates prior information. The method for restoring the missing structure of complex blood vessels proposed by the present invention performs structure restoration post-processing based on the binary mask results of blood vessel image segmentation, so that the final blood vessel extraction quality is not affected by image artifacts, partial volume effects, and noise. The present invention integrates the local thickness characteristics and tubular shape priors of the vascular network into a point cloud neural network and a spatial transformation network, respectively, to achieve restoration of blood vessel size information. The present invention inputs the local position characteristics, local thickness characteristics, and local direction characteristics of the blood vessels as additional features into the point cloud neural network model to ensure that both sides of the missing part of the blood vessel structure have the same properties, thereby marking the blood vessel branch containing the missing structure as a single blood vessel, thereby achieving simplification and accurate decomposition of the complex vascular network. The present invention first constructs a spatial transformation network based on the tubular shape prior to restore the missing structure of a single blood vessel, and then reorganizes a series of single blood vessels to achieve restoration of the missing structure of the complex vascular network. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a schematic structural diagram of a three-dimensional vascular structure restoration device integrating prior information according to a specific embodiment of the present invention;

[0035] Figure 2 This is a flowchart of a method for restoring a three-dimensional vascular structure by integrating prior information according to a specific embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of a decomposition module of a complex blood vessel according to a specific embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of a module for restoring a missing structure of a single blood vessel according to a specific embodiment of the present invention;

[0038] Figure 5 This is a schematic diagram of the restoration process and results of a complex vascular missing structure according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0040] On the one hand, a 3D vascular structure restoration device integrating prior information, such as Figure 1 As shown, it includes a data acquisition module, a data extraction module, and a complex blood vessel missing structure recovery module;

[0041] The data acquisition module acquires SR-µCT (Synchrotron Radiation Micro-computed Tomography) data of the vascular network and sends the acquired SR-µCT data of the vascular network to the data extraction module;

[0042] In this example, SR-µCT data of the vascular network within the mouse tibia were obtained;

[0043] The data extraction module extracts the vascular network image segmentation results required for restoring the complex vascular missing structure from the SR-µCT data of the vascular network, and sends the extracted vascular network image segmentation results to the complex vascular missing structure restoration module;

[0044] The complex vascular missing structure recovery module processes the data information sent by the data extraction module, specifically decomposing complex blood vessels, restoring the structure of a single blood vessel, reorganizing the vascular network, and outputting the recovery result of the complex vascular missing structure;

[0045] The complex vascular missing structure recovery module further includes:

[0046] The complex blood vessel decomposition module is used to decompose the complex blood vessels into a series of single blood vessels according to the blood vessel network segmentation results output by the data extraction module, and output the decomposition results to the single blood vessel structure recovery module;

[0047] The single vessel structure recovery module is used to recover the single vessel containing the missing structure based on the series of single vessels output by the complex vessel decomposition module, and output the result to the vascular network reconstruction module;

[0048] The vascular network reorganization module is used to reorganize all single blood vessels based on a series of single blood vessels with restored structures output by the single blood vessel structure recovery module, generate a vascular network, and output a complex blood vessel result with restored missing structures.

[0049] On the other hand, a three-dimensional vascular structure restoration method integrating prior information is implemented based on the aforementioned three-dimensional vascular structure restoration device integrating prior information, such as Figure 2 As shown, the following steps are included:

[0050] Step 1: Acquire SR-µCT data of the vascular network;

[0051] Thanks to the high contrast, high spatial resolution, and high signal-to-noise ratio of SR-µCT technology, SR-µCT imaging of the mouse tibia combined with vascular contrast agents can produce strong contrast between bone tissue and blood vessels in the image, facilitating subsequent image segmentation of vascular networks and restoration of complex vascular structures. Therefore, the present invention first requires acquiring SR-µCT data of the vascular network.

[0052] Step 2: Segment the vascular network image to obtain a binary mask;

[0053] The present method for restoring complex vascular structures requires implementation based on a binary mask of the blood vessels. Therefore, it is first necessary to segment the three-dimensional vascular network in the mouse tibia SR-µCT image. Numerous literature reports have provided methods for image segmentation of intraosseous vascular networks. This embodiment employs an existing watershed image segmentation algorithm based on monogenic signal local phase asymmetry and marker-controlled watershed. For specific methods, please refer to the relevant literature (Xu H, Langer M, Peyrin F. Quantitative analysis of bone microvasculature in a mouse model using the monogenic signal phase asymmetry and marker-controlled watershed. Phys Med Biol. 2021).

[0054] Step 3: Based on the image segmentation results of the vascular network, the complex blood vessels are decomposed into a series of single blood vessel images;

[0055] Step 3.1: Based on the binary mask obtained by segmentation of the vascular network image, it is converted into a three-dimensional spatial representation, i.e., a vascular point cloud, through point cloud processing to obtain the local position characteristics of the blood vessels;

[0056] Specifically, the binary mask of the blood vessels is first discretized and sampled according to a preset sampling interval. The spatial coordinate information of each sampling point is then extracted, and finally the corresponding vascular 3D point cloud data is generated. The point cloud obtained in this process is a discrete spatial representation of the vascular network in a unified spatial reference system, where each data point contains its 3D coordinates (x, y, z), that is, the local position feature.

[0057] Step 3.2: Obtain the local vascular thickness feature based on the binary mask obtained by segmenting the vascular network image;

[0058] The local thickness feature of the blood vessel is to assign a maximum sphere diameter value that can fill the interior of the blood vessel to each point on the blood vessel volume; let Ω⊂R 3 is the set of all points on the blood vessel volume, and p is any point among them. Then the local blood vessel thickness feature is defined as Th(p)=2•max({r|p∈sph(a,r)⊆ Ω,a∈Q}); where R 3 Represents three-dimensional real space, sph(a,r) represents a sphere with a as the center and a radius of r;

[0059] Step 3.3: Calculate the vascular orientation based on the binary mask obtained by segmenting the vascular network image to obtain the local directional features of the vascular;

[0060] The vessel direction is the principal eigenvector of the Hessian matrix calculated at each point on the vessel volume. For vessels with a tubular structure, the principal eigenvector corresponding to the maximum eigenvalue of the Hessian matrix is ​​consistent with the local principal direction of the vessel. First, the Hessian matrix is ​​solved for each voxel point in the binary mask, and the eigenvalue and eigenvector are calculated. Then, the principal eigenvector corresponding to the maximum eigenvalue is selected as the local direction feature α representing the local principal direction of the vessel.

[0061] Step 3.4: Integrate the obtained vascular local position features, local thickness features, and local direction features into the existing PointNet point cloud neural network model to achieve decomposition of complex vascular networks;

[0062] By inputting random order of 3D point coordinates into the PointNet point cloud neural network model, the implicit spatial relationship between 3D points can be captured and disordered data can be processed. Figure 3 As shown, it also includes the blood vessel local feature input PointNet, sampling, grouping, and interpolation process.

[0063] The local position, thickness, and orientation features of blood vessels are input as additional features of the point cloud into the existing PointNet point cloud neural network model to fully learn the key information of the blood vessel structure and ensure that the structures on both sides of the missing blood vessel have the same properties, thereby marking the missing single blood vessel as a whole.

[0064] Specifically, the point cloud (N, d + C) is sampled and grouped according to the local position features (x, y, z), local thickness features Th(p), and local orientation features α corresponding to each point in the vascular point cloud, and input into the PointNet point cloud neural network in the form of feature vectors to complete feature extraction. Then, feature transfer and point cloud classification (N, k) are further realized through the existing interpolation and lightweight unit point cloud neural network UnitPointNet, where N is the number of sampling points, d is the coordinate dimension, C is the number of features, and k is the label class. Finally, the category probability of the vascular branch to which each point belongs is output, and the vascular branch category to which each point belongs is determined according to the size of the probability value, thereby obtaining a series of single vascular images.

[0065] The present invention constructs a supervised learning model training system based on the gold standard of vascular branching: the vascular point cloud feature vector is used as the input of the point cloud neural network model, and the network parameters are optimized through multiple rounds of iterative training and back propagation algorithm to make the cross entropy loss function converge to a stable range.

[0066] Step 4: Based on the series of single vessel images obtained by decomposition, select the single vessel image containing the missing structure through the connected domain analysis method;

[0067] Specifically, all foreground voxel groups connected in the existing 18-connectivity mode are iteratively searched in a single vessel image as connected domains; if the number of connected domains in the image exceeds 1, the image is considered to be a single vessel image containing missing structures.

[0068] Step 5: Based on the decomposition of complex blood vessels in step 3 and the selection of single blood vessel images containing missing structures in step 4, a shape prior-based spatial transformation network is constructed to restore the missing structure of the single blood vessel;

[0069] Aiming at the problem that the missing blood vessel structure lacks grayscale value information and the gap geometry and size are different, this paper is inspired by the idea of ​​image registration based on graphic geometric transformation. Through the spatial transformation network, the tubular prior structure is deformed to match the target missing blood vessel structure, thereby realizing the recovery of the missing blood vessel structure. Figure 4 The spatial transformation network specifically includes a feedforward neural network, a grid generator, and a sampler.

[0070] Specifically, the tubular shape prior image U and the single blood vessel structure image I are input into a feedforward neural network architecture with convolution and maximum pooling in a two-channel manner to generate the deformation parameter θ, where the tubular shape prior image U is a cylinder with a base radius of r and a height of h, and θ=f ψ (I,U),f ψ It is a neural network with parameter ψ, which is generated by the grid generator G according to the coordinates of point i in I Find the corresponding coordinates of point i in U , that is, find the pixel coordinate relationship T between the input single blood vessel structure image I and the tubular shape prior image U θ (G i ): , i is a pixel in the image; finally, according to T θ (G i ) and the sampler, which samples the pixel value of point i in U through the B-spline interpolation method to generate the output image V: ; Among them, H×W×D is the size of the single blood vessel structure image I, the tubular shape prior image U, and the output image V. H, W, and D represent the height, width, and depth of the three-dimensional image, respectively. mnl is the intensity value at position (m, n, l) on the tubular shape prior image U, n = 0, 1, 2, ..., H-1, m = 0, 1, 2, ..., W-1, l = 0, 1, 2, ..., D-1.

[0071] Therefore, the shape prior-based spatial transformation network proposed in the present invention only needs to use the cross-entropy loss function to determine the parameter ψ during model training. In the test phase, the tubular shape structure in image U can be deformed with different spatial transformation parameters θ, thereby achieving its alignment with the single blood vessel structure in image I and completing the shape recovery of the missing single blood vessel structure.

[0072] Step 6: Reconstructing the single vessel images, specifically including the vessel images without missing structures and the vessel images with restored missing structures, to generate a complete vascular network;

[0073] In the series of images decomposed by step 3, the single blood vessel structure maintains the position in the image before decomposition. Therefore, the series of single blood vessels after vascular decomposition and structural restoration are specifically reconstructed by summing the image voxel values ​​to generate a complete vascular network. The recovery process and results of complex vascular missing structures are shown in Figure 2. Figure 5 shown.

[0074] The above description is merely an illustration of the preferred embodiments of the present disclosure and the technical principles employed. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present disclosure is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A three-dimensional vascular structure restoration device integrating prior information, characterized in that: It includes data acquisition module, data extraction module, and complex vascular missing structure recovery module; The data acquisition module acquires SR-µCT data of the vascular network and sends the acquired SR-µCT data of the vascular network to the data extraction module; The data extraction module extracts the vascular network image segmentation results required for restoring the complex vascular missing structure from the SR-µCT data of the vascular network, and sends the extracted vascular network image segmentation results to the complex vascular missing structure restoration module; The complex blood vessel missing structure recovery module processes the data information sent by the data extraction module, specifically including decomposing complex blood vessels, restoring single blood vessel structures, reorganizing blood vessel networks, and outputting the recovery results of the complex blood vessel missing structure.

2. The three-dimensional vascular structure restoration device integrating prior information according to claim 1, characterized in that: The complex vascular missing structure recovery module further includes: The complex blood vessel decomposition module is used to decompose the complex blood vessels into a series of single blood vessels according to the blood vessel network segmentation results output by the data extraction module, and output the decomposition results to the single blood vessel structure recovery module; The single vessel structure recovery module is used to recover the single vessel containing the missing structure based on the series of single vessels output by the complex vessel decomposition module, and output the result to the vascular network reconstruction module; The vascular network reorganization module is used to reorganize all single blood vessels based on a series of single blood vessels with restored structures output by the single blood vessel structure recovery module, generate a vascular network, and output a complex blood vessel result with restored missing structures.

3. A method for restoring a three-dimensional vascular structure by integrating prior information, based on the aforementioned device for restoring a three-dimensional vascular structure by integrating prior information, characterized in that: The following steps are involved: Step 1: Acquire SR-µCT data of the vascular network; Step 2: Segment the vascular network image to obtain a binary mask; Step 3: Based on the image segmentation results of the vascular network, the complex blood vessels are decomposed into a series of single blood vessel images; Step 4: Based on the series of single vessel images obtained by decomposition, select the single vessel image containing the missing structure through the connected domain analysis method; Step 5: Construct a spatial transformation network based on shape prior to restore the missing structure of a single blood vessel; The spatial transformation network specifically includes a feedforward neural network, a grid generator, and a sampler; Step 6: Reconstruct the single vessel images, specifically including the vessel images without missing structures and the vessel images with restored missing structures, to generate a complete vascular network.

4. The method for restoring a three-dimensional vascular structure by integrating prior information according to claim 3, characterized in that: The step 3 specifically includes the following steps: Step 3.1: Based on the binary mask obtained by segmentation of the vascular network image, it is converted into a three-dimensional spatial representation, i.e., a vascular point cloud, through point cloud processing to obtain the local position characteristics of the blood vessels; Specifically, the binary mask of the blood vessels is first discretized and sampled according to a preset sampling interval. The spatial coordinate information of each sampling point is then extracted, and finally the corresponding vascular 3D point cloud data is generated. The point cloud obtained in this process is a discrete spatial representation of the vascular network in a unified spatial reference system, where each data point contains its 3D coordinates (x, y, z), that is, the local position feature. Step 3.2: Obtain the local vascular thickness feature based on the binary mask obtained by segmenting the vascular network image; The local thickness feature of the blood vessel is to assign a maximum sphere diameter value that can fill the interior of the blood vessel to each point on the blood vessel volume; let Ω⊂R 3 is the set of all points on the blood vessel volume, and p is any point among them. Then the local blood vessel thickness feature is defined as Th(p)=2•max({r|p∈sph(a,r)⊆ Ω,a∈Q}); where R 3 Represents three-dimensional real space, sph(a,r) represents a sphere with a as the center and a radius of r; Step 3.3: Calculate the vascular orientation based on the binary mask obtained by segmenting the vascular network image to obtain the local directional features of the vascular; The vascular direction is the principal eigenvector of the Hessian matrix calculated at each point on the vascular volume; first, the Hessian matrix is ​​solved for each voxel point in the binary mask, and the eigenvalue and eigenvector are calculated. Then, the principal eigenvector corresponding to the maximum eigenvalue is selected as the local direction feature α representing the local principal direction of the vessel; Step 3.4: Integrate the obtained blood vessel local position features, local thickness features, and local direction features into the existing PointNet point cloud neural network model; Specifically, the point cloud (N, d + C) is sampled and grouped according to the local position features (x, y, z), local thickness features Th(p), and local orientation features α corresponding to each point in the vascular point cloud, and input into the PointNet point cloud neural network in the form of feature vectors to complete feature extraction. Then, feature transfer and point cloud classification (N, k) are further realized through the existing interpolation and lightweight unit point cloud neural network Unit PointNet, where N is the number of sampling points, d is the coordinate dimension, C is the number of features, and k is the label class. Finally, the category probability of the vascular branch to which each point belongs is output, and the vascular branch category to which each point belongs is determined according to the size of the probability value, thereby obtaining a series of single vascular images.

5. The method for restoring a three-dimensional vascular structure by integrating prior information according to claim 3, characterized in that: Specifically, step 4 is to iteratively search for all foreground voxel groups connected in the existing 18-connectivity mode as connected domains in the single vessel image; if the number of connected domains in the image exceeds 1, the image is determined to be a single vessel image containing missing structures.

6. The method for restoring a three-dimensional vascular structure by integrating prior information according to claim 3, characterized in that: Specifically, step 5 includes inputting the tubular shape prior image U and the single blood vessel structure image I into a feedforward neural network architecture with convolution and maximum pooling in a two-channel manner to generate the deformation parameter θ, wherein the tubular shape prior image U is a cylinder with a base radius r and a height h, and θ=f ψ (I,U),f ψ It is a neural network with parameter ψ, which is generated by the grid generator G according to the coordinates of point i in I Find the corresponding coordinates of point i in U , that is, find the pixel coordinate relationship T between the input single blood vessel structure image I and the tubular shape prior image U θ (G i ): , i is a pixel in the image; finally, according to T θ (G i ) and the sampler, which samples the pixel value of point i in U through the B-spline interpolation method to generate the output image V: ; Among them, H×W×D is the size of the single blood vessel structure image I, the tubular shape prior image U, and the output image V. H, W, and D represent the height, width, and depth of the three-dimensional image, respectively. mnl is the intensity value at position (m, n, l) on the tubular shape prior image U, n = 0, 1, 2, ..., H-1, m = 0, 1, 2, ..., W-1, l = 0, 1, 2, ..., D-1.

7. The method for restoring a three-dimensional vascular structure by integrating prior information according to claim 3, characterized in that: Specifically, step 6 is to realize the reorganization of blood vessels by summing up the image voxel values ​​to generate a complete blood vessel network.

Images