Method for evaluating angiogenesis effect under action of TSLP

Abstract

The invention discloses a method for evaluating an angiogenesis effect under the action of TSLP. The method comprises the following steps: enhancing data acquisition and preprocessing of a CT image; the preprocessed data are used for two-dimensional projection and labeling of blood vessels; the marked two-dimensional projection image and depth information are used for depth mapping and partial three-dimensional reconstruction; taking the partially reconstructed blood vessel surface model as input, and performing local feature extraction and label propagation; local features are used for global structure repairing; the globally repaired blood vessel three-dimensional model is used for evaluating the angiogenesis effect; the angiogenesis effect evaluation comprises quantitative evaluation and visual evaluation. According to the invention, artificial intelligence and medical technology are comprehensively utilized, and a comprehensive and fine angiogenesis effect evaluation method is provided.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary field of medicine and artificial intelligence science, and in particular relates to a method for evaluating the angiogenesis effect under the action of TSLP (thymic stromal lymphopoietin). Background Art

[0002] Thymic stromal lymphopoietin (TSLP) is a cytokine secreted by epithelial cells that is widely involved in various immune responses and tissue repair processes. In recent years, the potential role of TSLP in promoting angiogenesis has attracted widespread attention, especially in medical research on tumors, cardiovascular disease, and tissue regeneration. As an emerging regulatory factor, TSLP is believed to promote the formation and remodeling of blood vessels. Angiogenesis, the sprouting of new blood vessels from existing vascular networks, is an important physiological process in organisms responding to ischemia, tissue damage, and tumor growth. To further investigate the specific effects of TSLP on angiogenesis, it is particularly important to accurately evaluate its ability to induce blood vessel growth under different experimental conditions.

[0003] The mechanism of action of TSLP in angiogenesis involves the coordinated action of multiple signaling pathways and cell types. TSLP can guide the formation of new blood vessels by activating endothelial cells, promoting their proliferation, migration, and lumen formation. Furthermore, TSLP can indirectly influence angiogenesis by interacting with immune cells and regulating the inflammatory microenvironment. Studies have shown that TSLP can induce the activation of dendritic cells and helper T cells, releasing pro-angiogenic factors such as vascular endothelial growth factor, further enhancing the angiogenesis process. Therefore, TSLP plays both direct and indirect roles in regulating angiogenesis.

[0004] To further understand the regulatory effects of TSLP on angiogenesis, functional assessments are essential in addition to quantitative morphological analysis. For example, observing the perfusion capacity and hemodynamic properties of new vessels can reveal the functional quality of the new vessels. Furthermore, the effects of TSLP on the stability of new vessels are also worthy of attention, including the recruitment of vascular wall cells and the formation of the basement membrane, which are directly related to the maturity and stability of new vessels. By combining these morphological and functional assessment methods, a more comprehensive understanding of the mechanisms and effects of TSLP on angiogenesis can be achieved.

[0005] The application of relevant artificial intelligence technologies provides a comprehensive and detailed assessment framework for the specific role of TSLP in angiogenesis. This not only helps to reveal the biological mechanisms of TSLP in angiogenesis, but also provides strong data support for its potential value in clinical applications. The innovation of this approach lies in the combination of multiple quantitative evaluation indicators and multi-angle visualization techniques, ensuring accurate characterization and in-depth understanding of the growth characteristics and topological changes of angiogenesis. Summary of the Invention

[0006] In view of this, in order to more comprehensively evaluate the evaluation effect of angiogenesis, this application combines the latest deep learning technology with vascular imaging analysis technology, and proposes a combined method of quantitative evaluation and visual evaluation. In the process of quantitative evaluation, angiogenesis is three-dimensionally reconstructed and optimized to make its quantitative effect more scientific and accurate.

[0007] The present application proposes a method for evaluating the angiogenesis effect under the action of TSLP, the technical solution of which includes the following steps:

[0008] Step 1, data acquisition and preprocessing of enhanced CT images;

[0009] Step 2: Use the preprocessed data for 2D projection and annotation of blood vessels;

[0010] Step 3: Use the annotated 2D projection image and depth information for depth mapping and partial 3D reconstruction;

[0011] Step 4: Using the partially reconstructed vascular surface model as input, local feature extraction and label propagation are performed;

[0012] Step 5: Use local features to repair the global structure;

[0013] Step 6: The global repaired 3D vascular model is used to evaluate the angiogenesis effect;

[0014] The angiogenesis effect evaluation includes quantitative evaluation and visual evaluation. The quantitative evaluation includes the evaluation of changes in the number, length and diameter of new blood vessel branches. The visual evaluation refers to the use of a three-dimensional reconstructed vascular model to generate two-dimensional projection images at different angles to intuitively display the growth and topological changes of new blood vessels.

[0015] Specifically, the quantitative evaluation includes vascular skeletonization, branch number analysis, branch length analysis and branch diameter analysis; the vascular skeletonization refers to first skeletonizing the baseline and neovascular models, and using a three-dimensional morphological skeleton extraction algorithm to generate a centerline model of the blood vessel; the branch number analysis includes extracting the branch points and end points of the blood vessel from the skeleton model, the branch point refers to the location where the blood vessel bifurcates, and the end point is the end point of the blood vessel; using a topological analysis method to extract the branch points; calculating the number of branches in the baseline and neovascular models; calculating the change in the number of branches; the branch length analysis first performs branch path extraction, for each branch point, extracting all paths connected to it along the skeleton model until encountering another branch point or end point; path length calculation; calculating the total branch length in the baseline and neovascular models; calculating the change in branch length; the branch diameter analysis includes vascular diameter calculation, for each skeleton point, calculating the distance between the point and the blood vessel surface to estimate the diameter of the blood vessel; calculating the average diameter of the baseline and neovascular models; calculating the change in diameter;

[0016] The visualization evaluation includes selecting projection angles, generating maximum intensity projections, projections at non-orthogonal angles, and visualization and comparison of projection images. The selection of projection angles selects multiple typical perspectives to fully demonstrate the growth and topological characteristics of the vascular model. Typical perspectives include three orthogonal directions: front view, side view, top view, and other non-orthogonal perspectives. The maximum intensity projection generation refers to generating a maximum intensity projection for each selected projection direction to obtain a two-dimensional visualization image of the vascular model in that direction. The projection at non-orthogonal angles first rotates the three-dimensional vascular model, and performs maximum intensity projection on the rotated vascular model to obtain a maximum intensity projection image in a non-orthogonal direction. The visualization and comparison of projection images visualizes all two-dimensional projection images to form a multi-perspective display, or displays the growth of new blood vessels from different perspectives through image stitching or an interactive 3D view interface.

[0017] Specifically, the data acquisition and preprocessing of the enhanced CT image includes the following steps:

[0018] Step 101: Enhance the CT image Normalization is performed on bones or background tissues to enhance the brightness of the vascular area, where N x ,N y ,N z Respectively represent the size of the CT image in three axial directions;

[0019] Step 102, using morphological operations to segment and remove bones or background tissue structures;

[0020] The method of using the pre-processed data for two-dimensional projection and annotation of blood vessels includes the following steps:

[0021] Step 201, projecting the three-dimensional CT image in each axis direction to generate a two-dimensional projection image;

[0022] Step 202: manually annotate the two-dimensional projection image, record the depth position where the maximum value appears, and find the depth position corresponding to the maximum value for each two-dimensional projection pixel point;

[0023] Step 203: generate depth-enhanced projections using a depth enhancement method, including forward depth projections and backward depth projections, to provide depth information about blood vessels. The forward depth refers to recording the first position where the maximum intensity is located along each projection ray. The forward depth D fw,y (x, z) represents the position where the maximum intensity at position (x, z) first appears along the y-axis, and the back depth represents the position where the maximum intensity last appears along each projection ray. The back depth D bw,y (x, z) represents the position along the y-axis where the maximum intensity at position (x, z) last appears. The depth coding projection includes a forward depth coding projection and a backward depth coding projection. The forward depth information is used to combine the maximum intensity projection with the depth value to form a forward depth coding projection: Using the backward depth information, the maximum intensity projection is combined with the backward depth value to form a backward depth encoding projection: MIP y (x,z) refers to the position (x,z) along the specified y-axis to perform maximum intensity projection to generate a two-dimensional projection image.

[0024] Furthermore, the use of the annotated two-dimensional projection image and depth information for depth mapping and partial three-dimensional reconstruction includes the following steps:

[0025] Step 301: Generate a preliminary three-dimensional depth map by combining the forward depth and the backward depth. The three-dimensional depth map has local structural information of the blood vessels, including annotation information of the two-dimensional projection and depth enhancement information.

[0026] Step 302 , using the annotation information and depth enhancement information of the two-dimensional projection, the annotation points are mapped back to the original three-dimensional space to generate a partially reconstructed blood vessel surface;

[0027] In step 303, a sparse blood vessel model is constructed based on the partially reconstructed three-dimensional structure. The sparse blood vessel model will serve as the initial model in subsequent steps.

[0028] Specifically, the method of using the annotation information and depth enhancement information of the two-dimensional projection to map the annotation points back to the original three-dimensional space to generate a partially reconstructed blood vessel surface includes the following steps:

[0029] Step 30201, assuming that the annotated two-dimensional projection image is Where A(x,z)=1 indicates that there is a marked vascular structure at this two-dimensional point;

[0030] Step 30202, create a blank depth volume, create a three-dimensional depth volume Initialize all elements to 0 to store the reconstructed partial blood vessel surface;

[0031] Step 30203, traverse the two-dimensional annotation points. For each position (x, z) in the two-dimensional annotation matrix, if A(x, z) = 1, it means that there is a vascular structure at this position. Obtain the forward and backward depth information of the annotation point: D fw (x,z) and D bw (x,z);

[0032] Step 30204: Map the two-dimensional annotation point to multiple depth layers in the three-dimensional space based on the forward depth and backward depth information. The mapping process is as follows: For a given annotation point (x, z), the forward depth D fw (x,z) and back depth D bw (x,z) determines the projection range along the y-axis in three-dimensional space; the three-dimensional depth volume is marked within the depth range: V(x,y,z)=1, For each 2D annotation point (x, z), the volume V is marked as 1 at all y values ​​within the depth range, thus generating a partial blood vessel surface in 3D space;

[0033] The method of constructing a sparse blood vessel model based on the partially reconstructed three-dimensional structure includes the following steps:

[0034] Step 30301, thinning the target, selecting a set of representative points from the existing depth map, which can describe the topological information of the entire vascular structure;

[0035] Step 30302: voxel extraction, traverse the three-dimensional depth volume V and extract all voxel points marked as 1;

[0036] Step 30303: using a uniformly spaced sampling method to thin out the blood vessel structure points;

[0037] Step 30304: sparse point set S ′ Constructed as a sparse 3D point cloud model M sparse , used to represent the overall structure of blood vessels;

[0038] Step 30305: The output sparse blood vessel model is M sparse , which contains a sparse set of points or fitted centerlines of the vascular structure.

[0039] The method of extracting local features by using the partially reconstructed blood vessel surface model as input comprises the following steps:

[0040] The sparse point cloud S ′ Projecting onto a 3D grid, generating a 3D input volume Among them I input (x, y, z) = 1 means that the point belongs to the blood vessel part;

[0041] The encoder uses a series of 3D convolutional layers and ReLU activation functions to transform the three-dimensional input volume I input Processing: F1 = ReLU (Conv3D (I input )), downsample the convolutional features to obtain higher-level feature representations: Represents the downsampled feature map of the i-th layer, Conv3D represents the 3D convolution operation, and MaxPooling3D represents the 3D maximum pooling operation;

[0042] In the decoder, the resolution of the feature map is restored by upsampling, and the upsampled features are fused with the corresponding layer features in the encoder: Represents the upsampled feature map of layer i, and UpSampling3D represents the 3D upsampling operation;

[0043] Finally, a 3D convolutional layer is used to output the reconstructed 3D volume map V output , used to represent the complete vascular area, is a probability map, where voxels with values ​​close to 1 represent blood vessels, and the obtained V output As an updated vascular model.

[0044] In particular, the label propagation is to generate pseudo labels for unlabeled regions in the 3D volume image, and to achieve a more comprehensive labeling of the entire vascular structure by gradually fusing existing labeled regions, including the following steps:

[0045] Enter the initial labeling matrix Among them L init (x, y, z) = 1 indicates that the location is the marked blood vessel area;

[0046] Create a weight matrix Used to measure the confidence of the pseudo label at each position; the value of the initialization weight matrix W is: This means that all labeled points have a weight of 1, while unlabeled points have an initial weight of 0;

[0047] For each unlabeled point (x, y, z), update its label using the surrounding annotations and the model prediction results;

[0048] Update the label matrix L using the adaptive label propagation formula prop , the formula is:

[0049]

[0050] Among them, N(x,y,z) represents the neighborhood set of point (x,y,z), which includes several voxel points around it, α∈[0,1] is a weight factor that controls the degree of fusion of annotation information and model prediction information, W(i,j,k) is the weight of point (i,j,k) in the neighborhood, which is used to measure the credibility of the annotation information, and L init (i, j, k) is the label value in the initial labeling matrix, 1 represents blood vessels and 0 represents non-blood vessels;

[0051] Based on the current label propagation results, the weight matrix is ​​adaptively updated to improve the accuracy of the label:

[0052] W(x,y,z)=γW(x,y,z)+(1-γ)L prop (x,y,z)

[0053] Among them, γ∈[0,1] is an adaptive update factor used to control the dynamic update of the weight matrix;

[0054] After completing several iterative updates, the pseudo label matrix L is finally obtained. final , used to describe a more complete vascular structure.

[0055] Specifically, the use of local features for global structure repair includes the following steps:

[0056] Get the pseudo label matrix L final And the locally reconstructed updated vascular model V output By fusing these two data sources, we can obtain a more comprehensive initial input: I global (x,y,z)=βL final (x,y,z)+(1-β)V output (x, y, z), β∈[0,1] is used to control the fusion weight of pseudo labels and local reconstruction information;

[0057] Use deep learning network to fusion input I globalPerform global learning, where the encoder performs multi-layer downsampling on the input to extract global features of the blood vessels, and the decoder performs step-by-step upsampling to restore resolution while ensuring the coherence of the blood vessel structure.

[0058] The loss function of the deep learning network includes reconstruction loss and structural loss. To ensure that the three-dimensional volume output by the network is as close as possible to the true label, the formula is:

[0059]

[0060] is the output of the deep learning network, indicating the probability that each voxel belongs to a blood vessel;

[0061] The structural loss maintains the structural integrity of the blood vessels through the regularization term, and the formula is:

[0062]

[0063] Where N(x,y,z) represents the neighborhood set of voxel (x,y,z);

[0064] Using the fused input I global and the pseudo-label matrix L final As training data, the optimization goal is to minimize the weighted sum of reconstruction loss and structure loss: λ is a hyperparameter used to balance the reconstruction loss and the structural loss;

[0065] After several rounds of training, the final globally optimized vascular model is obtained.

[0066] Furthermore, the continuity of the reconstruction results of the global structural repaired vascular model is improved to ensure the smoothness and anatomical rationality of the vascular model, including the following steps:

[0067] Obtain prototype features. Assume that the model contains K prototype features, each prototype feature represents an anatomical form of the blood vessel, and reconstruct the result V final The known vascular models in the training set are used to learn the prototype features of different anatomical regions and define the prototype feature matrix Where K is the number of prototypes and F is the feature dimension;

[0068] Using deep convolutional neural networks to final Perform feature extraction to obtain feature maps Each voxel point has a corresponding feature embedding, and through the feature extraction network, the feature vector representation of each voxel point in high-dimensional space is learned;

[0069] For each voxel point in the feature map Calculate the similarity with all prototype features and use cosine similarity as the measure: S(x,y,z,k) represents the similarity between the feature point and the kth prototype, P k represents the k-th prototype feature;

[0070] For each voxel point, a weighted sum is taken based on the similarity with all prototypes to obtain the optimized feature representation: w k (x,y,z) represents the weight, which is calculated as follows:

[0071] The optimized feature is represented by F opt The optimized vascular model is obtained by mapping back to the original three-dimensional spatial structure through the decoding network.

[0072] Anatomical rationality constraints are added during model training, and the constraints are implemented by introducing a regularized loss function. Represents the smoothness loss of the vascular model, with the goal of making the changes in the blood vessels more continuous and smooth;

[0073] The optimized blood vessel model V opt Perform binarization processing to obtain the final blood vessel model V binary .

[0074] The beneficial effects of this application are as follows:

[0075] The present invention applies the use of depth information for partial reconstruction, so that a relatively complete three-dimensional vascular structure can be obtained based only on the annotation of a single projection, effectively reducing information loss. Furthermore, through the label propagation and pseudo-label optimization mechanism, the accuracy of vascular reconstruction is gradually improved to ensure the coherence of local and global structures. Finally, through the close collaboration of multiple functional modules, combined with the extraction and optimization of local and global features, especially in the reconstruction process, multiple performance indicators such as reconstruction loss, structural loss and smoothness loss are integrated step by step to ensure that the complete three-dimensional model of angiogenesis is accurately reconstructed and evaluated. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 is a flow chart of the method of the present invention;

[0077] Figure 2 2 is a diagram of an evaluation method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0078] The present invention will be further described below with reference to the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention fall within the scope of protection of the present invention.

[0079] The technical solutions provided in the embodiments of the present application involve the combination of artificial intelligence machine learning technology and medical technology, which are specifically introduced and explained through the following embodiments.

[0080] like Figure 1 As shown, a method for evaluating the angiogenesis effect under the action of TSLP, the technical solution of which includes the following steps:

[0081] Step 1, data acquisition and preprocessing of enhanced CT images;

[0082] Step 2: Use the preprocessed data for 2D projection and annotation of blood vessels;

[0083] Step 3: Use the annotated 2D projection image and depth information for depth mapping and partial 3D reconstruction;

[0084] Step 4: Using the partially reconstructed vascular surface model as input, local feature extraction and label propagation are performed;

[0085] Step 5: Use local features to repair the global structure;

[0086] Step 6: The global repaired 3D vascular model is used to evaluate the angiogenesis effect;

[0087] To quantify the specific effect of thymic stromal lymphopoietin (TSLP) on angiogenesis, we focused on changes in the number, length, and diameter of vascular branches. These indicators can provide in-depth analysis of the growth and structural changes of new blood vessels, specifically assessing the complexity, density, and expansion of new blood vessels. The specific inputs include the baseline vascular model V baseline ∈{0,1} N x ×N y ×N z , namely the three-dimensional volume of blood vessels in the baseline period, and the neovascularization model V new ∈{0,1} N x ×N y ×N z , which is the three-dimensional volume of blood vessels during the neovascularization period. The output includes changes in the number of vascular branches, branch length, and diameter to assess the angiogenesis effect.

[0088] In order to intuitively display the growth of new blood vessels and changes in topological structure, the 3D reconstructed blood vessel model can be visualized from multiple angles. This not only better demonstrates the expansion of new blood vessels in different directions, but also helps researchers and medical experts analyze the overall and local changes of blood vessels from different perspectives. The main method is to use the 3D reconstructed blood vessel model to generate 2D projections at different angles to intuitively display the growth of new blood vessels and topological changes. The specific input is the new blood vessel model V new ∈{0,1} N x ×N y ×N z ,The output is a two-dimensional projection image generated from different angles, which is used to compare and analyze the changes of new blood vessels in different directions.

[0089] like Figure 2 As shown, the angiogenesis effect evaluation includes quantitative evaluation and visual evaluation. The quantitative evaluation refers to the quantitative analysis of the quantitative indicators of blood vessels, and the visual evaluation refers to the use of the three-dimensional reconstructed blood vessel model to generate two-dimensional projection images at different angles to intuitively display the growth of new blood vessels and topological changes.

[0090] Specifically, the quantitative evaluation includes analysis of the number of blood vessel branches, branch length and branch diameter; the branch number analysis includes extracting branch points and end points of blood vessels from the skeleton model, where the branch point refers to the location where the blood vessel bifurcates, and the end point is the end point of the blood vessel; using a topological analysis method to extract branch points; calculating the number of branches in the baseline and neovascular models; calculating the change in the number of branches; the branch length analysis first performs branch path extraction, and for each branch point, extracts all paths connected to it along the skeleton model until another branch point or end point is encountered; path length calculation; calculating the total branch length in the baseline and neovascular models; calculating the change in branch length; the branch diameter analysis includes blood vessel diameter calculation, and for each skeleton point, calculating the distance between the point and the blood vessel surface to estimate the diameter of the blood vessel; calculating the average diameter of the baseline and neovascular models; and calculating the change in diameter.

[0091] The visualization evaluation includes selecting projection angles, generating maximum intensity projections, projections at non-orthogonal angles, and visualization and comparison of projection images. The selection of projection angles selects multiple typical perspectives to fully demonstrate the growth and topological characteristics of the vascular model. Typical perspectives include three orthogonal directions: front view, side view, top view, and other non-orthogonal perspectives. The maximum intensity projection generation refers to generating a maximum intensity projection for each selected projection direction to obtain a two-dimensional visualization image of the vascular model in that direction. The projection at non-orthogonal angles first rotates the three-dimensional vascular model, and performs maximum intensity projection on the rotated vascular model to obtain a maximum intensity projection image in a non-orthogonal direction. The visualization and comparison of projection images visualizes all two-dimensional projection images to form a multi-perspective display, or displays the growth of new blood vessels from different perspectives through image stitching or an interactive 3D view interface.

[0092] By analyzing changes in branch number, length, and diameter, the specific effects of TSLP-induced angiogenesis can be quantitatively assessed, aiding research into the mechanisms of angiogenesis. These indicators can also be used to compare changes in blood vessels before and after treatment, for example, analyzing vascular growth and expansion after drug or therapeutic intervention. Visual assessment of angiogenesis, through projections from different angles, provides a comprehensive and intuitive display of TSLP-induced angiogenesis, helping to assess the quality and growth pattern of new vessels. This visualization method has extensive applications in medical image analysis, providing physicians with intuitive diagnostic evidence and comparative views through direct two-dimensional projections.

[0093] Specifically, the data acquisition and preprocessing of the enhanced CT image includes the following steps:

[0094] Step 101: Enhance the CT image Normalization is performed on bones or background tissues to enhance the brightness of the vascular area, where N x ,N y ,N z Respectively represent the size of the CT image in three axial directions;

[0095] Step 102, using morphological operations to segment and remove bones or background tissue structures;

[0096] For angiogenesis assessment, blood vessels usually have higher intensity in CT scans. Through maximum intensity projection, these high-intensity blood vessels can be clearly presented in the two-dimensional projection image to obtain a two-dimensional projection image with highlighted vascular structure. These images are convenient for manual annotation and provide data support for subsequent three-dimensional reconstruction.

[0097] The method of using the pre-processed data for two-dimensional projection and annotation of blood vessels includes the following steps:

[0098] Step 201, projecting the three-dimensional CT image in each axis direction to generate a two-dimensional projection image;

[0099] Step 202: manually annotate the two-dimensional projection image, record the depth position where the maximum value appears, and find the depth position corresponding to the maximum value for each two-dimensional projection pixel point;

[0100] Step 203: generate depth-enhanced projections using a depth enhancement method, including forward depth projections and backward depth projections, to provide depth information about blood vessels. The forward depth refers to recording the first position where the maximum intensity is located along each projection ray. The forward depth D fw,y (x, z) represents the position where the maximum intensity at position (x, z) first appears along the y-axis, and the back depth represents the position where the maximum intensity last appears along each projection ray. The back depth D bw,y (x, z) represents the position along the y-axis where the maximum intensity at position (x, z) last appears. The depth coding projection includes a forward depth coding projection and a backward depth coding projection. The forward depth information is used to combine the maximum intensity projection with the depth value to form a forward depth coding projection: Using the backward depth information, the maximum intensity projection is combined with the backward depth value to form a backward depth encoding projection: MIP y (x,z) refers to the position (x,z) along the specified y-axis to perform maximum intensity projection to generate a two-dimensional projection image.

[0101] The generated two-dimensional projection images are convenient for clinical experts to manually label, and the labeling only needs to be completed in the two-dimensional images, thereby reducing the complexity and workload of labeling.

[0102] This technical solution generates forward and backward depth-enhanced projection images that facilitate manual vessel annotation. After annotation, the depth matrix can be used to map the annotation points from the 2D projection back to the 3D volume to generate a preliminary 3D vascular structure. The depth-enhanced projection not only contains intensity information but also provides information about the depth of the vascular structure in 3D space, helping the model better learn the spatial topology of the vessels during training.

[0103] Furthermore, the use of the annotated two-dimensional projection image and depth information for depth mapping and partial three-dimensional reconstruction includes the following steps:

[0104] Step 301: Generate a preliminary three-dimensional depth map by combining the forward depth and the backward depth. The three-dimensional depth map has local structural information of the blood vessels, including annotation information of the two-dimensional projection and depth enhancement information.

[0105] Step 302 , using the annotation information and depth enhancement information of the two-dimensional projection, the annotation points are mapped back to the original three-dimensional space to generate a partially reconstructed blood vessel surface;

[0106] In step 303, a sparse blood vessel model is constructed based on the partially reconstructed three-dimensional structure. The sparse blood vessel model will serve as the initial model in subsequent steps.

[0107] Specifically, the method of using the annotation information and depth enhancement information of the two-dimensional projection to map the annotation points back to the original three-dimensional space to generate a partially reconstructed blood vessel surface includes the following steps:

[0108] Assume that the annotated two-dimensional projection image is Where A(x,z)=1 indicates that there is a marked vascular structure at this two-dimensional point;

[0109] Create a blank depth volume, create a 3D depth volume Initialize all elements to 0 to store the reconstructed partial blood vessel surface;

[0110] Traverse the two-dimensional annotation points. For each position (x, z) in the two-dimensional annotation matrix, if A(x, z) = 1, it means that there is a vascular structure at that position. Get the forward and backward depth information of the annotation point: D fw (x,z) and D bw (x,z);

[0111] According to the forward depth and backward depth information, the two-dimensional annotation points are mapped to multiple depth layers in the three-dimensional space. The mapping process is as follows: For a given annotation point (x, z), the forward depth D fw (x,z) and back depth D bw (x,z) determines the projection range along the y-axis in three-dimensional space; the three-dimensional depth volume is marked within the depth range: V(x,y,z)=1, For each 2D annotation point (x, z), the volume V is marked as 1 at all y values ​​within the depth range, thus generating a partial blood vessel surface in 3D space;

[0112] The method of constructing a sparse blood vessel model based on the partially reconstructed three-dimensional structure includes the following steps:

[0113] The sparseness goal is to select a set of representative points from the existing depth map, which can describe the topological information of the entire vascular structure;

[0114] Voxel extraction, traverse the three-dimensional depth volume V and extract all voxel points marked as 1;

[0115] The vascular structure points are thinned out using the uniformly spaced sampling method;

[0116] The sparse point set S ′ Constructed as a sparse 3D point cloud model M sparse , used to represent the overall structure of blood vessels;

[0117] The output sparse blood vessel model is M sparse , which contains a sparse set of points or fitted centerlines of the vascular structure.

[0118] By combining forward and backward depth information, the position and extent of vascular structures in three-dimensional space can be precisely determined. This effectively overcomes the loss of depth information in two-dimensional projections, enabling high-precision 3D reconstructions even with weak annotations. Although this approach only generates a partial image of the vessel surface, the incorporation of depth information ensures high geometric accuracy of the generated partial structure. These partial reconstructions can then be used as input for further deep learning model training to complete the reconstruction of the entire vessel.

[0119] The method of extracting local features by using the partially reconstructed blood vessel surface model as input comprises the following steps:

[0120] The sparse point cloud S ′ Projecting onto a 3D grid, generating a 3D input volume Among them I input (x, y, z) = 1 means that the point belongs to the blood vessel part;

[0121] The encoder uses a series of 3D convolutional layers and ReLU activation functions to transform the three-dimensional input volume I input Processing: F1 = ReLU (Conv3D (I input )), downsample the convolutional features to obtain higher-level feature representations: Represents the downsampled feature map of the i-th layer, Conv3D represents the 3D convolution operation, and MaxPooling3D represents the 3D maximum pooling operation;

[0122] In the decoder, the resolution of the feature map is restored by upsampling, and the upsampled features are fused with the corresponding layer features in the encoder: Represents the upsampled feature map of layer i, and UpSampling3D represents the 3D upsampling operation;

[0123] Finally, a 3D convolutional layer is used to output the reconstructed 3D volume map V output , used to represent the complete vascular area, is a probability map, where voxels with values ​​close to 1 represent blood vessels, and the obtained V output As an updated vascular model.

[0124] The sparsification operation significantly reduces the computational complexity of vascular reconstruction, allowing subsequent deep learning models to refine the structure more efficiently. Although sparsification reduces the number of voxels, uniform or keypoint sampling can preserve key structural information of the vessels, allowing the sparse model to remain concise while still effectively expressing the topology of the vessels. The sparse model can serve as the initial input for deep learning models or other reconstruction methods, helping subsequent models refine, repair, and globally reconstruct the structure.

[0125] In this embodiment, a 3DU-Net network can be used. Through multi-level convolution and up- and down-sampling, it can effectively extract local and global features of blood vessels. Especially for sparse input data, it can achieve high-quality reconstruction with limited computing resources. At the same time, through the combination of encoder and decoder, as well as skip connections, 3DU-Net can maintain the global topology during the reconstruction process while refining local details. This solution is also applicable to sparse input. The lightweight design of 3DU-Net is particularly suitable for processing sparse input, allowing the initial reconstructed sparse model to be further refined, thereby obtaining a more accurate vascular structure.

[0126] In particular, the label propagation is to generate pseudo labels for unlabeled regions in the 3D volume image, and to achieve a more comprehensive labeling of the entire vascular structure by gradually fusing existing labeled regions, including the following steps:

[0127] Enter the initial labeling matrix Among them L init (x, y, z) = 1 indicates that the location is the marked blood vessel area;

[0128] Create a weight matrix Used to measure the confidence of the pseudo label at each position; the value of the initialization weight matrix W is: This means that all labeled points have a weight of 1, while unlabeled points have an initial weight of 0;

[0129] For each unlabeled point (x, y, z), update its label using the surrounding annotations and the model prediction results;

[0130] Update the label matrix L using the adaptive label propagation formula prop , the formula is:

[0131]

[0132] Among them, N(x,y,z) represents the neighborhood set of point (x,y,z), which includes several voxel points around it, α∈[0,1] is a weight factor that controls the degree of fusion of annotation information and model prediction information, W(i,j,k) is the weight of point (i,j,k) in the neighborhood, which is used to measure the credibility of the annotation information, and L init (i, j, k) is the label value in the initial labeling matrix, 1 represents blood vessels and 0 represents non-blood vessels;

[0133] Based on the current label propagation results, the weight matrix is ​​adaptively updated to improve the accuracy of the label:

[0134] W(x,y,z)=γW(x,y,z)+(1-γ)L prop (x,y,z)

[0135] Among them, γ∈[0,1] is an adaptive update factor used to control the dynamic update of the weight matrix;

[0136] After completing several iterative updates, the pseudo label matrix L is finally obtained. final , used to describe a more complete vascular structure.

[0137] Specifically, the use of local features for global structure repair includes the following steps:

[0138] Get the pseudo label matrix L final And the locally reconstructed updated vascular model V output By fusing these two data sources, we can obtain a more comprehensive initial input: I global (x,y,z)=βL final (x,y,z)+(1-β)V output (x, y, z), β∈[0,1] is used to control the fusion weight of pseudo labels and local reconstruction information;

[0139] Use deep learning network to fusion input I global Perform global learning, where the encoder performs multi-layer downsampling on the input to extract global features of the blood vessels, and the decoder performs step-by-step upsampling to restore resolution while ensuring the coherence of the blood vessel structure.

[0140] The loss function of the deep learning network includes reconstruction loss and structural loss. To ensure that the three-dimensional volume output by the network is as close as possible to the true label, the formula is:

[0141]

[0142] is the output of the deep learning network, indicating the probability that each voxel belongs to a blood vessel;

[0143] The structural loss maintains the structural integrity of the blood vessels through the regularization term, and the formula is:

[0144]

[0145] Where N(x,y,z) represents the neighborhood set of voxel (x,y,z);

[0146] Using the fused input I global and the pseudo-label matrix L final As training data, the optimization goal is to minimize the weighted sum of reconstruction loss and structure loss: λ is a hyperparameter used to balance the reconstruction loss and the structural loss;

[0147] After several rounds of training, the final globally optimized vascular model is obtained.

[0148] Through the global learning module, the coherence of the vascular structure in three-dimensional space can be ensured, and any breaks or gaps that may exist in local learning can be repaired, thereby forming a complete vascular model; combined with the structural loss term, the model can maintain the overall structural characteristics of the vascular network and ensure the continuity and connectivity of the branches; by fusing pseudo-labels and local features, the global learning module can refine and expand the vascular model, ensuring that not only the global structure is preserved during the reconstruction process, but also the quality of local details is improved.

[0149] Furthermore, the continuity of the reconstruction results of the global structural repaired vascular model is improved to ensure the smoothness and anatomical rationality of the vascular model, including the following steps:

[0150] Obtain prototype features. Assume that the model contains K prototype features, each prototype feature represents an anatomical form of the blood vessel, and reconstruct the result V final The known vascular models in the training set are used to learn the prototype features of different anatomical regions and define the prototype feature matrix Where K is the number of prototypes and F is the feature dimension;

[0151] Using deep convolutional neural networks to final Perform feature extraction to obtain feature maps Each voxel point has a corresponding feature embedding, and through the feature extraction network, the feature vector representation of each voxel point in high-dimensional space is learned;

[0152] For each voxel point in the feature map Calculate the similarity with all prototype features and use cosine similarity as the measure: S(x,y,z,k) represents the similarity between the feature point and the kth prototype, P k represents the k-th prototype feature;

[0153] For each voxel point, a weighted sum is taken based on the similarity with all prototypes to obtain the optimized feature representation: w k (x,y,z) represents the weight, which is calculated as follows:

[0154] The optimized feature is represented by F opt The optimized vascular model is obtained by mapping back to the original three-dimensional spatial structure through the decoding network.

[0155] Anatomical rationality constraints are added during model training, and the constraints are implemented by introducing a regularized loss function. Represents the smoothness loss of the vascular model, the purpose is to make the changes of blood vessels more continuous and smooth, represents the ladder vector differential operator;

[0156] The optimized blood vessel model V opt Perform binarization processing to obtain the final blood vessel model V binary .

[0157] As described in the above technical solution, by fusing global and local features, it is possible to ensure that the vascular model maintains structural consistency globally while also having finer details locally; using the depth information matrix to optimize the model during the testing phase can effectively reduce breaks and discontinuous areas in the reconstruction, ensuring the smoothness of the blood vessels in three-dimensional space; through the design of structural loss, breakpoints and isolated areas in the vascular structure can be automatically repaired, ultimately forming a complete and coherent three-dimensional vascular model.

[0158] As used herein, the word "preferred" is intended to serve as an example, instance, or illustration. Any aspect or design described herein as "preferred" is not necessarily to be construed as advantageous over other aspects or designs. Rather, the use of the word "preferred" is intended to present concepts in a concrete manner. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or." That is, unless otherwise specified or clear from the context, "X employs A or B" is intended to mean any of the naturally inclusive permutations. That is, if X employs A; X employs B; or X employs both A and B, then "X employs A or B" is satisfied in any of the foregoing examples.

[0159] Moreover, although the present disclosure has been shown and described with respect to one or implementation, those skilled in the art will think of equivalent variations and modifications based on reading and understanding of this specification and the accompanying drawings. The present disclosure includes all such modifications and variations and is limited only by the scope of the appended claims. In particular, with respect to the various functions performed by the above-mentioned components (such as elements, etc.), the terms used to describe such components are intended to correspond to any component (unless otherwise indicated) that performs the specified function of the component (such as it is functionally equivalent), even if structurally different from the disclosed structure that performs the function in the exemplary implementation of the present disclosure shown herein. In addition, although the specific features of the present disclosure have been disclosed with respect to only one of several implementations, such features can be combined with one or other features of other implementations that can be desired and advantageous for a given or specific application. Moreover, insofar as the terms "including", "having", "containing" or their variations are used in specific embodiments or claims, such terms are intended to be included in a manner similar to the term "comprising".

[0160] The functional units in the embodiments of the present invention may be integrated into a single processing module, or each unit may exist physically separately, or multiple or more units may be integrated into a single module. The aforementioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium. The aforementioned storage medium may be a read-only memory, a magnetic disk, or an optical disk, etc. The aforementioned devices or systems may execute the storage method in the corresponding method embodiment.

[0161] In summary, the above embodiment is one implementation method of the present invention, but the implementation method of the present invention is not limited to the described embodiment. Any other changes, modifications, substitutions, combinations, and simplifications that deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A method for evaluating the angiogenesis effect under the action of TSLP, characterized in that: The following steps are involved: Step 1, data acquisition and preprocessing of enhanced CT images; Step 2: Use the preprocessed data for 2D projection and annotation of blood vessels; Step 3: Use the annotated 2D projection image and depth information for depth mapping and partial 3D reconstruction; Step 4: Using the partially reconstructed vascular surface model as input, local feature extraction and label propagation are performed; Step 5: Use local features to repair the global structure; Step 6: The global repaired 3D vascular model is used to evaluate the angiogenesis effect; The evaluation of angiogenesis effects includes quantitative evaluation and visual evaluation. The quantitative evaluation refers to quantitative analysis of vascular indicators, and the visual evaluation refers to the use of a three-dimensional reconstructed vascular model to generate two-dimensional projection images at different angles to intuitively display the growth and topological changes of new blood vessels.

2. The method for evaluating angiogenesis effect under TSLP according to claim 1, characterized in that: The method comprises the following steps: the quantitative assessment includes analysis of the number of blood vessel branches, branch length, and branch diameter; the branch number analysis includes extracting branch points and end points of blood vessels from a skeleton model, wherein a branch point refers to the location where a blood vessel bifurcates, and an end point refers to the end point of a blood vessel; extracting branch points using a topological analysis method; calculating the number of branches in baseline and neovascular models; and calculating changes in the number of branches; The branch length analysis first performs branch path extraction. For each branch point, all paths connected to it are extracted along the skeleton model until another branch point or an end point is encountered. Path length calculation; calculation of total branch length in baseline and neovascularization models; calculation of changes in branch length; The branch diameter analysis includes calculating the blood vessel diameter, for each skeleton point, calculating the distance between the point and the blood vessel surface to estimate the blood vessel diameter; calculating the average diameter of the baseline and new blood vessel models; and calculating the change value of the diameter; The visualization evaluation includes selecting projection angles, generating maximum intensity projections, projections at non-orthogonal angles, and visualization and comparison of projection images. The selection of projection angles selects multiple typical perspectives to fully demonstrate the growth and topological characteristics of the vascular model. Typical perspectives include three orthogonal directions: front view, side view, top view, and other non-orthogonal perspectives. The maximum intensity projection generation refers to generating a maximum intensity projection for each selected projection direction to obtain a two-dimensional visualization image of the vascular model in that direction. The projection at non-orthogonal angles first rotates the three-dimensional vascular model, and performs maximum intensity projection on the rotated vascular model to obtain a maximum intensity projection image in a non-orthogonal direction. The visualization and comparison of projection images visualizes all two-dimensional projection images to form a multi-perspective display, or displays the growth of new blood vessels from different perspectives through image stitching or an interactive 3D view interface.

3. The method for evaluating angiogenesis effect under TSLP according to claim 1, characterized in that: The data acquisition and preprocessing of the enhanced CT image includes the following steps: Step 101: Enhance the CT image Normalization is performed on bones or background tissues to enhance the brightness of the vascular area, where represents a real number, N x , N y , N z Respectively represent the size of the CT image in three axial directions; Step 102, using morphological operations to segment and remove bones or background tissue structures; The method of using the pre-processed data for two-dimensional projection and annotation of blood vessels includes the following steps: Step 201, projecting the three-dimensional CT image in each axis direction to generate a two-dimensional projection image; Step 202: manually annotate the two-dimensional projection image, record the depth position where the maximum value appears, and find the depth position corresponding to the maximum value for each two-dimensional projection pixel point; Step 203: generate depth-enhanced projections using a depth enhancement method, including forward depth projections and backward depth projections, to provide depth information about the blood vessels. The forward depth refers to recording the first position where the maximum intensity is located along each projection ray. The forward depth D fw,y (x, z) represents the position where the maximum intensity at position (x, z) first appears along the y-axis, and the back depth represents the position where the maximum intensity last appears along each projection ray. The back depth D bw,y (x, z) represents the position along the y-axis where the maximum intensity last appears at position (x, z); the depth coding projection includes a forward depth coding projection and a backward depth coding projection. The forward depth information is used to combine the maximum intensity projection with the depth value to form a forward depth coding projection: Using the backward depth information, the maximum intensity projection is combined with the backward depth value to form a backward depth encoding projection: MIP y (x,z) refers to the position (x,z) along the specified y-axis to perform maximum intensity projection to generate a two-dimensional projection image.

4. The method for evaluating angiogenesis effect under TSLP according to claim 3, characterized in that: The method of using the annotated two-dimensional projection image and depth information for depth mapping and partial three-dimensional reconstruction includes the following steps: Step 301: Generate a preliminary three-dimensional depth map by combining the forward depth and the backward depth. The three-dimensional depth map has local structural information of the blood vessels, including annotation information of the two-dimensional projection and depth enhancement information. Step 302 , using the annotation information and depth enhancement information of the two-dimensional projection, the annotation points are mapped back to the original three-dimensional space to generate a partially reconstructed blood vessel surface; In step 303, a sparse blood vessel model is constructed based on the partially reconstructed three-dimensional structure. The sparse blood vessel model will serve as the initial model in subsequent steps.

5. The method for evaluating angiogenesis effect under TSLP according to claim 4, characterized in that: The method of using the annotation information and depth enhancement information of the two-dimensional projection to map the annotation points back to the original three-dimensional space to generate a partially reconstructed blood vessel surface includes the following steps: Step 30201, assuming that the annotated two-dimensional projection image is Where A(x,z)=1 indicates that there is a marked vascular structure at this two-dimensional point; Step 30202, create a blank depth volume, create a three-dimensional depth volume Initialize all elements to 0 to store the reconstructed partial blood vessel surface; Step 30203, traverse the two-dimensional annotation points. For each position (x, z) in the two-dimensional annotation matrix, if A(x, z) = 1, it means that there is a blood vessel structure at that position. Obtain the forward and backward depth information of the annotation point: D fw (x,z) and D bw (x,z); Step 30204: Map the two-dimensional annotation point to multiple depth layers in the three-dimensional space based on the forward depth and backward depth information. The mapping process is as follows: For a given annotation point (x, z), the forward depth D fw (x,z) and the backward depth D bw (x,z) determines the projection range along the y-axis in three-dimensional space; the three-dimensional depth volume is marked within the depth range: V(x,y,z)=1, For each 2D annotation point (x, z), the volume V is marked as 1 at all y values ​​within the depth range, thus generating a partial blood vessel surface in 3D space; The method of constructing a sparse blood vessel model based on the partially reconstructed three-dimensional structure includes the following steps: Step 30301, thinning the target, selecting a set of representative points from the existing depth map, which can describe the topological information of the entire vascular structure; Step 30302, voxel extraction, traverse the three-dimensional depth volume V and extract all voxel points marked as 1; Step 30303: using a uniformly spaced sampling method to thin out the blood vessel structure points; Step 30304: construct the sparse point set S′ into a sparse 3D point cloud model M sparse , used to represent the overall structure of blood vessels; Step 30305: The output sparse blood vessel model is M sparse , which contains a sparse set of points or fitted centerlines of the vascular structure.

6. The method for evaluating angiogenesis effect under TSLP according to claim 5, characterized in that: As described above, the partially reconstructed blood vessel surface model is used as input to extract local features. The following steps are involved: Project the sparse point cloud S′ onto a 3D grid to generate a 3D input volume Among them I input (x, y, z) = 1 means that the point belongs to the blood vessel part; The encoder uses a series of 3D convolutional layers and ReLU activation functions to transform the three-dimensional input volume I input Processing: F1 = ReLU (Conv3D (I input )), downsample the convolutional features to obtain higher-level feature representations: Represents the downsampled feature map of the i-th layer, Conv3D represents the 3D convolution operation, and MaxPooling3D represents the 3D maximum pooling operation; In the decoder, the resolution of the feature map is restored by upsampling, and the upsampled features are fused with the corresponding layer features in the encoder: Represents the upsampled feature map of layer i, and UpSampling3D represents the 3D upsampling operation; Finally, a 3D convolutional layer is used to output the reconstructed 3D volume map V output , used to represent the complete vascular area, is a probability map, where voxels with values ​​close to 1 represent blood vessels, and the obtained V output As an updated vascular model, represents the upsampled feature map of layer 0, and Sigmoid represents the activation function.

7. The method for evaluating angiogenesis effect under TSLP according to claim 5, characterized in that: The label propagation described above generates pseudo labels for unlabeled regions in the 3D volume image, and achieves a more comprehensive labeling of the entire vascular structure by gradually fusing existing labeled regions. The process includes the following steps: Enter the initial labeling matrix Among them, L init (x, y, z) = 1 indicates that the location is the marked blood vessel area; Create a weight matrix Used to measure the confidence of the pseudo label at each position; the value of the initialization weight matrix W is: This means that all labeled points have a weight of 1, while unlabeled points have an initial weight of 0; For each unlabeled point (x, y, z), update its label using the surrounding annotations and the model prediction results; Update the label matrix L using the adaptive label propagation formula prop , the formula is: Among them, N(x,y,z) represents the neighborhood set of point (x,y,z), which includes several voxel points around it, α∈[0,1] is a weight factor that controls the degree of fusion of annotation information and model prediction information, W(i,j,k) is the weight of point (i,j,k) in the neighborhood, which is used to measure the credibility of the annotation information, and L init (i, j, k) is the label value in the initial labeling matrix, 1 represents blood vessels and 0 represents non-blood vessels; Based on the current label propagation results, the weight matrix is ​​adaptively updated to improve the accuracy of the label: W(x,y,z)←γW(x,y,z)+(1-γ)L prop (x,y,z) Among them, γ∈[0,1] is an adaptive update factor used to control the dynamic update of the weight matrix; After completing several iterative updates, the pseudo label matrix L is finally obtained. final , used to describe a more complete vascular structure.

8. The method for evaluating angiogenesis effect under the action of TSLP according to any one of claims 6 or 7, characterized in that: The method of using local features for global structure repair includes the following steps: Get the pseudo label matrix L final And the locally reconstructed updated vascular model V output By fusing these two data sources, we can obtain a more comprehensive initial input: I global (x,y,z)=βL final (x,y,z)+(1-β)V output (x, y, z), β∈[0,1] is used to control the fusion weight of pseudo labels and local reconstruction information; Use deep learning network to fusion input I global Perform global learning, where the encoder performs multi-layer downsampling on the input to extract global features of the blood vessels, and the decoder performs step-by-step upsampling to restore resolution while ensuring the coherence of the blood vessel structure. The loss function of the deep learning network includes reconstruction loss and structural loss. To ensure that the three-dimensional volume output by the network is as close as possible to the true label, the formula is: is the output of the deep learning network, indicating the probability that each voxel belongs to a blood vessel; The structural loss maintains the structural integrity of the blood vessels through the regularization term, and the formula is: Where N(x,y,z) represents the neighborhood set of voxel (x,y,z); Using the fused input I global and the pseudo-label matrix L final As training data, the optimization goal is to minimize the weighted sum of reconstruction loss and structure loss: λ is a hyperparameter used to balance the reconstruction loss and the structural loss; After several rounds of training, the final globally optimized vascular model is obtained.

9. A method for evaluating angiogenesis effect under TSLP action according to claim 8, characterized in that: Improve the continuity of the reconstruction results of the global structural repaired vascular model and ensure the smoothness and anatomical rationality of the vascular model, including the following steps: Obtain prototype features. Assume that the model contains K prototype features, each prototype feature represents an anatomical form of the blood vessel, and reconstruct the result V final The known vascular models in the training set are used to learn the prototype features of different anatomical regions and define the prototype feature matrix Where K is the number of prototypes and F is the feature dimension; Using deep convolutional neural networks to final Perform feature extraction to obtain feature maps Each voxel point has a corresponding feature embedding, and through the feature extraction network, the feature vector representation of each voxel point in high-dimensional space is learned; For each voxel point F in the feature map final (x,y,z), calculate the similarity with all prototype features, and use cosine similarity for measurement: S(x,y,z,k) represents the similarity between the feature point and the kth prototype, P k represents the k-th prototype feature; For each voxel point, a weighted sum is taken based on the similarity with all prototypes to obtain the optimized feature representation: w k (x,y,z) represents the weight, which is calculated as follows: The optimized feature is represented by F opt The optimized vascular model is obtained by mapping back to the original three-dimensional spatial structure through the decoding network. Anatomical rationality constraints are added during model training, and the constraints are implemented by introducing a regularized loss function. Represents the smoothness loss of the vascular model, with the goal of making the changes in the blood vessels more continuous and smooth; The optimized blood vessel model V opt Perform binarization processing to obtain the final blood vessel model V binary .

Images